Designing Educational Games to Foster Decision Experience in College Innovation and Entrepreneurship Programs

Innovation and entrepreneurship (IE) education aims to cultivate talents with entrepreneurial literacy and pioneering personalities, focusing on phased training of innovative thinking and entrepreneurial capabilities. However, current IE education lacks practicality and interactive curricula, relying heavily on theory while neglecting hands-on problem-solving skills. Educational games-with high interactivity and scenario simulation-can address this gap by fostering decision-making and teamwork, making them a promising tool for personalized IE learning¹.

To improve IE teaching efficiency, prior studies have integrated games into IE but face two key limitations: learners often overlook critical knowledge points, and traditional games fail to dynamically track knowledge acquisition^2,3. To address these, we propose the following hypothesis: An IE decision-experience game integrating Bayesian Networks (BN) reasoning for real-time knowledge diagnosis and an improved Deep Knowledge Tracking (DKT) algorithm will achieve higher prediction accuracy and student adaptability than traditional teaching or existing KT models (BKT, DKVMN, EKT)⁴. This study's dual innovations fill gaps in prior work: unlike existing models that only conduct static knowledge assessment, we use BN reasoning-modeling knowledge points as nodes with Conditional Probability Tables (CPTs)-to diagnose learning weaknesses in real time via game performance data⁵. This study enhances DKT by embedding IE-dimensional features (entrepreneurial awareness, ability, and intention) and attention mechanisms (AMs)-solving the long-sequence information loss in standard DKT -and outperforming limited models like BKT, DKVMN, and EKT⁶.

IE education's core lies in cultivating entrepreneurial awareness, ability, and intention-three mediating variables between IE and entrepreneurial behavior that correspond to its basic, primary, and fundamental goals-but existing tools fail to track their dynamic development. This highlights a critical practical need: current IE education lacks adaptive, data-driven tools. Our gamified system, with real-time feedback, directly addresses this, offering scalability for higher education contexts.

This study comprises five sections: (1) elaborates IE/educational game concepts and their design challenges; (2) reviews IE education and educational game research; (3) details methodology (selecting IE game evaluation dimensions, integrating improved DKT to build the model); (4) validates model performance; (5) summarizes findings, limitations, and future directions.

Related works

Against the backdrop of promoting mass IE to promote high-quality employment for college students, how to enhance their innovation literacy and entrepreneurial skills through IE education has received widespread concern from all sectors⁷. Liu et al. analyzed the methods of actively serving students in intelligent education, and combined knowledge concept information, extended the recursive neural network framework to the interpretable Exercise-aware Knowledge Tracing (EKT) framework. The results demonstrated that this method had superior interpretability in predicting student grades⁸. Shute et al. integrated students' preferences, gender, and other characteristics into the educational game. The results indicated that physics teaching based on educational games had better grades and game performance than traditional teaching⁹. Chen et al. explored the effect of mobile business simulation games in entrepreneurial education through a quasi-experimental design. Research results indicated that mobile business simulation games could improve entrepreneurial attitudes and self-efficacy, but they could not change entrepreneurial intentions. The research results confirm the positive role of mobile business simulation games in entrepreneurial education, which can enhance the entrepreneurial attitude and entrepreneurial self-efficacy¹⁰. Research results indicated that this method had a high predictive accuracy for learners' performance in game tasks. Shi et al. proposed a teaching mode that integrated Quadratic function knowledge into games. The experiment showed that students' learning motivation and math scores have significantly improved through this teaching mode¹¹.

Isabelle et al. analyzed the gamification of entrepreneurial education by using an independent gamification platform, using game mechanisms in non-game applications, and integrating IE education theories¹². Research results indicated that gamified entrepreneurial education methods could enhance students' participation and entrepreneurial self-efficacy. Soomro et al. conducted an empirical survey on entrepreneurial education, self-efficacy, and entrepreneurial intention among students from a certain university using a cross-sectional data quantitative method. Results denoted that the structure of entrepreneurial education had a significant positive impact on entrepreneurial self-efficacy and entrepreneurial intention¹³. Scholars such as Loi proposed a debate on the assumptions and challenges in entrepreneurial education, integrating new entrepreneurial perspectives into existing entrepreneurial teaching models. Research results indicated that determining knowledge priorities was most important for enhancing the impact of entrepreneurial education¹⁴. To improve the performance of IE education in schools, Mao et al. constructed an evaluation model for IE education based on decision trees and fuzzy algorithms. The experiment showed that the model had certain practical applications effectively evaluating college IE education quality and providing operable quantitative tools for optimizing IE teaching strategies¹⁵. Li et al. analyzed the necessity of conducting IE education for college students in the context of the Internet. Based on actual research results, strategies have been proposed to promote the development of IE education and practice¹⁶.

In summary, the value and function of IE education are mainly embodied in cultivating talents with basic quality and a pioneering personality of entrepreneurship. Its value lies in the value system that comprehensively considers innovative consciousness, knowledge, ability, and intention, aiming to stimulate students' entrepreneurial passion, which is a psychological requirement. On this basis, IE education focuses on improving students' entrepreneurial ability, including resource integration, team building, etc., and finally forming entrepreneurial intentions and turning them into practical actions. Functionally, IE education improves students' ability to face entrepreneurial challenges through practice, such as resource integration, teamwork, etc., to promote the realization of entrepreneurial behavior. Current research on IE education mainly focuses on theoretical characteristics, roles, and empirical analysis, lacking relevant research results on teaching methods. Learners may need different guidance and assistance, but educational games cannot simulate the learners' knowledge acquisition state. To solve such problems, this study aims to identify individual learning differences among students and proposes using BN reasoning to obtain learners' knowledge mastery level in teaching. Moreover, it constructs a decision experience game model for college students' IE education based on an improved Knowledge Tracking (KT) algorithm.

This study was approved by the Shaoxing University Human Research Ethics Committee (Approval No. SU-2023-IE-058). All participants provided written informed consent prior to data collection. The experiment was conducted in accordance with the Declaration of Helsinki (2013).

Game design for the decision experience of IE education for college students

The research implemented a Bayesian Network (BN) for real-time assessment of learners' knowledge mastery. Each knowledge point was modeled as a node with Conditional Probability Tables (CPTs) defining dependencies. Student performance data from game interactions served as evidence for probabilistic inference using the pgmpy library (v0.1.20, Python 3.9.7). The Junction Tree algorithm was employed for efficient inference.

Feature selection of game evaluation dimensions of college students' IE education based on BN reasoning

The development of IE education has been accompanied by the continuous deepening of reforms in higher education, and significant progress has been made in education. The research background of the IE education ecosystem innovation is the first driving force for development and the strategic support for building a modern economic system^17,18. However, the operation of the IE ecosystem, like the natural ecosystem, also has many influencing factors. The basic function and value of IE education lie in cultivating a large number of innovative talents (a large number of high-quality talents with innovative spirit, knowledge system, and action ability)¹⁹. Its basic purpose and value mainly include five parts, as shown in Figure 1, which are the functions and values of IE education.

Figure 1: Integrated framework of IE education's core functions, value composition, and internal logic linkage with entrepreneurial behavior. Please click here to view a larger version of this figure.

Figure 1 demonstrates the functions and values of IE education, with the core being entrepreneurial awareness. Entrepreneurial demand is the foundation, providing a prerequisite for the formation of entrepreneurial awareness. Entrepreneurial interest is the core, driving individuals to pay attention to entrepreneurship. Entrepreneurial motivation is the driving force that propels entrepreneurial practice. Entrepreneurship will provide support to help overcome difficulties in starting a business. The entrepreneurial value is the ultimate goal, reflecting the significance of entrepreneurial education for individuals and society. All elements interact with each other and jointly support the functioning of IE education²⁰. The study categorizes IE education as consciousness cultivation, ability enhancement, and intention generation. They correspond to the basic, main, and fundamental goals of IE education. Entrepreneurial awareness is divided into multiple levels, and cultivating this awareness is a key factor in achieving knowledge and ability goals. Knowledge itself is a part of ability, so entrepreneurial education should focus on cultivating individual entrepreneurial abilities. Intention is a crucial link in the development of consciousness into behavior, but intention and action are not exactly the same²¹. In entrepreneurial education, it is necessary to generate students' entrepreneurial willingness, thus forming effective experiences and transforming them into behaviors. In summary, the study takes entrepreneurial awareness, entrepreneurial ability, and entrepreneurial intention as mediating variables between IE education and entrepreneurial behavior. Taking entrepreneurial awareness, ability, and willingness as the mediating variables between IE education and entrepreneurial behavior, corresponding to the basic, main, and fundamental goals of IE education, combined with literature analysis and ILM, set the measurement indicators of each variable (such as IE education, including course participation, etc.), and then construct this ILM. As shown in Figure 1, an ILM linking IE education and entrepreneurial behavior was constructed for research purposes.

As shown in Figure 1, the positive effect of entrepreneurial education on entrepreneurial ability is reflected in its ability to help college students master basic skills, improve their ability to accumulate and utilize resources, and overcome difficulties in entrepreneurship. Entrepreneurial awareness can play a mediating role between IE education and entrepreneurial behavior. As a pre-factor of entrepreneurial behavior, entrepreneurial intention can directly predict the occurrence of behavior. Cultivating entrepreneurial awareness among college students can help enterprises accumulate entrepreneurial knowledge and improve their entrepreneurial abilities. Ability theory holds that ability is the core factor that determines behavior, especially in IE education, which emphasizes the cultivation of students' multidimensional ability, including resource integration, team building, and fund preparation²². These skills directly affect students' performance in entrepreneurship, such as opportunity identification, relationship building, and action implementation. Entrepreneurial ability covers three main dimensions, namely resource integration ability, which refers to the effective use and integration of resources to support entrepreneurship; team building ability refers to the ability to form and lead a team; and capital readiness, that is, the ability to raise and manage start-up capital. The improvement of these abilities is the key to promoting the success of students' entrepreneurship²³.

The establishment of measurement standards for entrepreneurial education, entrepreneurial behavior, entrepreneurial ability, entrepreneurial awareness, and entrepreneurial intention is mainly based on literature analysis, and combined with the ILM of IE education. The indicators of entrepreneurial education are participation in entrepreneurial courses or lectures (CY1), participation in entrepreneurial education skills training (CY2), and participation in entrepreneurial competitions (CY3). These indicators directly reflect students' participation in IE education and reflect the core objectives of entrepreneurial education's practicality and interaction¹⁰. The indicators of entrepreneurial behavior are opportunity identification (CW1), relationship building (CW2), and energy investment (CW3). This is the measurement related to entrepreneurial behavior, which mainly comes from behavioral theories (such as competence theory) and focuses on individual action characteristics, such as opportunity grasp, social capital construction, and resource allocation¹¹. The measurement dimensions of entrepreneurial ability are resource integration ability (CN1), team building ability (CN2), and capital preparation ability (CN3). These dimensions are derived from the core practical goal of entrepreneurial ability improvement, that is, to improve students' ability to survive and develop in a real entrepreneurial environment. The measurement dimensions of entrepreneurial consciousness include entrepreneurial behavior attitude (CS1), entrepreneurial subjective norms (CS2), and entrepreneurial perception behavior control (CS3). The measurement of entrepreneurial consciousness is based on Ajzen's Planned Behavior Theory (TPB), which takes individual psychological expectation of entrepreneurial behavior as the analysis focus¹³. The indicators of entrepreneurial intention are entrepreneurial desire (CW1), career planning, and entrepreneurial relevance (CW2). The measurement of entrepreneurial intention is supported by the theory of the impact of entrepreneurial education on career choice, such as the relationship between subjective norms and behavioral intention.

In teaching activities, the process of taking students' responses as evidence variables to calculate the posterior probability of latent and observable variables is the process of diagnosing students' knowledge level. BN reasoning calculates posterior probability, which is mainly based on the BN structure and conditional probability distribution, and calculates the value taking probability of some nodes in the network under specific evidence conditions²⁴. For a BN with n node variables, it can be defined as a joint probability distribution function, as shown in equation (1)²⁵.

 (1)

Where P means the Conditional probability set  of BN. For any node , there is a CPT to represent the Conditional probability of  and its parent node set .  denote any two random variables, and  is the hypothesis, which is  evidence. The Prior probability is the probability  of event  when  is unknown. The Posterior probability is the probability  of the event  when  is known. As shown in equation (2), it is the Bayesian equation²⁶.

  (2)

BN can not only obtain learners' level of knowledge mastery in teaching through PR, but also graphically reflect learners' knowledge structure. Using the Knowledge representation and reasoning in the Domain model, every conceptual knowledge is regarded as a node. In games, nodes are used to set corresponding levels and provide dynamic updates of BN. During the learning, the system will continuously collect user data information and update it on the network. The Joint Tree (JT) algorithm is currently the most widely used and fastest computational BN reasoning algorithm²⁷. The algorithm process includes transforming the BN into a quadratic structure containing a set of edges in the clique. In the IE educational decision-making experience game model, adaptivity is defined as the ability of the system to automatically adjust educational content, strategies, and interaction modes based on learners' real-time performance, individual characteristics, and learning needs. American scholar Hodhod divided the adaptive education game model into domain, teaching, player models, and adaptive engine and presentation module²⁸. Each model has different functions, among which the player model is the student model, which is a key component of educational games to achieve adaptive support for learners. As shown in Figure 2, it is the information flow of the student model based on the JT algorithm.

Figure 2: Unified architecture of the university IE educational decision-making experience game. Joint-Tree-driven student information flow, enhanced DKT with feature embedding & attention, and KT-integrated core elements. Please click here to view a larger version of this figure.

In Figure 2, the study will use BN to construct student models and use JT reasoning algorithms to evaluate students' knowledge acquisition in real-time. The JT algorithm relies on the same version of pgmpy, calls the to_junction_tree method of BN (moralize=True, min_fill triangulation) to generate JT, and uses nodes () to view the cluster nodes. Initialize the inference engine with JunctionTreeInference, input the behavioral evidence through the query method, set joint=False to return the posterior probability of IE knowledge points, and set tol=1e-6 to control the error. Then, the statistical model is used to compare the learner's correct response rate with the set threshold value, to provide students with an appropriate game task framework and a prompt feedback framework. BN modeling is based on pgmpy 0.1.20 (Python 3.9.7 environment). The BN class is called to pass in the list of node dependencies, and the latent/evidence variables are marked with node_type. The CPT is initialized based on expert experience by using the TabularCPD class, and then iteratively updated with student behavior data using the fit method of BayesianEstimator (dirichlet prior, pseudo_counts=1). The causal relationship is verified through do_calculus of CausalInference, and mutual information is calculated (greater than 0.3 indicates strong dependence) to ensure a reasonable structure.

Design of a decision experience game model for IE education for college students, integrating improved KT algorithms

Although the feature screening of IE educational game evaluation dimensions is completed through BN reasoning, which can dynamically identify students' weak points in knowledge, BN only focuses on the diagnosis of the current knowledge state and cannot track the dynamic change process of students' knowledge mastery. It is also difficult to accurately predict subsequent learning needs, and it cannot support the dynamic adaptation of the game model to teaching strategies. Therefore, it is necessary to integrate and improve the KT algorithm, and by leveraging its feature embedding and AM, achieve real-time tracking and prediction of students' knowledge structure. The student model is a module that monitors student information, and its research first appeared in the Intelligent Tutoring System (ITS)²⁹. In ITS, tracking the dynamic changes in students' knowledge states is particularly important, as it can help the model better guide students in learning knowledge. The idea of tracking students' knowledge structure in games originates from past learning diagnosis systems and adaptive education games³⁰. The DKT model has been proven to be an effective KT model, which is more suitable for evaluation scenarios with large user data scales. The DKT model is based on TensorFlow 2.8.0 (Python 3.9.7, GPU supports CUDA 11.2). Data is processed using pandas, LabelEncoder, and numpy, and combined into 3D tensors of [sample size, 30,8]. Build the architecture with Sequential, including Input (30,8), Dense (64,relu), LSTM (128,return_sequences=True), and Dense (knowledge points, sigmoid); Optimize with Adam (lr=0.001), take BinaryCrossentropy as the loss function, set the accuracy and AUC metrics, conduct 50 rounds of training (batch=32), and store the optimal weight at the ModelCheckpoint. This improved DKT model is constructed based on TensorFlow 2.8.0 (Python 3.9.7, supporting CUDA 11.2). Firstly, through feature embedding, the behavioral features such as the number of students answering questions, the behavior of seeking help, and the number of prompt requests are cross-linked with the answering results, and encoded in combination with the question difficulty coefficient (the number of incorrect answers/the total number of answering questions). Then, connect all the encoded features to construct a vector and use the autoencoder (tanh activated) for dimensionality reduction as the input. Then, an attention mechanism is introduced between the LSTM and the output layer of DKT. The AttentionLayer class is defined. tf.matmul is used to calculate the weights, and softmax is used for normalization to generate weighted features. 128 weights are set and Dropout(0.2) is added to prevent overfitting. With Adam (lr=0.001) optimization and cross-entropy as the loss function for 50 rounds of training, the tracking and prediction of students' knowledge mastery were achieved. The study proposes an improved DKT model that enriches the input data of the model through feature embedding and introduces AMs, as shown in Figure 2.

In Figure 2, due to the DKT model's ability to capture complex features in the data, feature embedding can be used to enrich the model's input data, thereby improving model performance. The last hidden state of long short-term memory (LSTM) is the hidden learning state of students at the current moment³¹. Once the time is long, LSTM may cause some important information to be lost. The study introduces the AM, which weights and aggregates all historical information to reduce the loss of important information. The characteristics of students' answering behavior and difficulty coefficients are embedded into the original input information, becoming a more meaningful historical interaction sequence . The input historical interaction sequence of this model is , and the output is the probability vector that predicts the students' answers to the corresponding questions of the knowledge points correctly. The study proposes to combine the number of times students try to answer questions, whether they request help during the answering process, and the number of times they request prompts into a new feature in the educational game model, and cross it with the answer results, as shown in equation (3).

 (3)

In equation (3),  means the behavioral characteristics of students answering questions;  indicates the number of times students answer questions at the  moment;  denotes the request for help;  indicates the number of prompts;  refers to the answer result. The calculation expression of the difficulty coefficient of the question is shown in equation (4).

 (4)

In equation (4),  stands for the number of people who answered the question  incorrectly, and  represents the total number of people who answered the question  correctly. To improve the accuracy of the model, all the encoded features obtained are connected, and the vector construction is shown in equation (5).

 (5)

In equation (5),  means the number of knowledge points;  refers to the difficulty coefficient of the problem;  means feature combination;  stands for encoding format;  means connection. Train the automatic encoder using the tanh activation function and remove the output layer after completion. Then it uses the output of the hidden layer as input in the DKT-FA model, as shown in equation (6).

 (6)

In equation (6),  are the learned weight matrix and bias vector, respectively. The study combines AM with an LSTM network to enable the model to pay more attention to answer sequences with similar behavioral characteristics and difficulty in practice. The AM is integrated between the LSTM and output layers of DKT. A custom AttentionLayer class is defined, and tf.matmul is used to calculate weights, and softmax is used for normalization to generate weighted features. 128 weights were set to preserve rights, Dropout (0.2) was added to prevent overfitting, synchronous optimization was performed using backpropagation, and weights were extracted via get_weights() to analyze the impact of key behaviors. When the attention variable  is selected, it indicates that the th answer record is selected. As shown in equation (7), it calculates the probability of the th information using the historical order information code and the final information code.

 (7)

In equation (7), the weight factor  determines the content that should be paid attention to or ignored during prediction. The study uses an additive model of the attention score function  to calculate its score, as shown in equation (8).

                              (8)

In equation (8),  stands for the encoding of historical sequence information based on data compression; means the dimensionality reduction encoding of input information at  time.  is a network parameter of the Department of Science. The attention state  is denoted as the weighted sum of , and its calculation expression is shown in equation (9).

 (9)

In equation (12), by combining  with , the probability of students mastering all knowledge points is calculated and expressed as shown in equation (10).

 (10)

In machine learning, cross-entropy can be used to measure the difference between actual marking and predicted results. Therefore, it is considered as the Loss function of DKT-FA model, as shown in equation(11).

 (11)

In equation (11),  and  indicate the predicted and true probability distributions, respectively. It assumes  is 1 knowledge unit, including  knowledge points. And if the probability of using the knowledge point  is , then the students' mastery of the knowledge point is shown in equation (12).

               (12)

In equation (12),  means the learner's error rate on the knowledge point ;  refers to the probability of the occurrence of knowledge points. Study the importance of using it as a knowledge point. To classify the range of values within the range of 0-0.5,  is divided into two parts. In complex and rich game scenarios, learners can achieve personalized and meaningful reconstruction of learning methods. When constructing educational game models, corresponding principles should also be followed. Firstly, the principle of learner subjectivity should be followed. It is necessary to fully consider the individual characteristics of learners so that they can obtain a gaming experience that matches their own abilities during the gaming process. Secondly, it is necessary to follow the principle of situational awareness, that is, to design a more realistic gaming environment to stimulate learners' enthusiasm. Finally, it is necessary to follow the principle of balancing game design and instructional design. Based on the above principles, it studies the key elements of integrated IE education and games, and builds a decision model for IE education, as shown in Figure 2.   

In Figure 2, the model takes the scenario as the fundamental aspect of IE education, and the game itself is divided into student, teacher, and knowledge models. Combining instructional design ideas, it further constructs the plot, activities, tasks, and key elements of the game. A periodic activity that combines the four aspects of environment, knowledge, emotion, and action in entrepreneurial education. The knowledge model is mainly used for storing and listing knowledge in education. The student model is combined with the teacher model to achieve personalized learning of knowledge. It needs further detailed design of the game plot, activities, tasks, etc., to balance the educational and gameplay aspects of the game model. Simultaneously, it implements IE knowledge contained in both virtual and perceptual scenarios, internalizes knowledge-based learning into the cultivation of correct entrepreneurial education behaviors. The feature embedded module in Figure 2 generates the learning state and dynamically maps it to the interface by analyzing the student behavior data (such as the number of answers, frequency of help, correct rate, etc.) to generate the task scene suitable for its stage. These scenarios include virtual entrepreneurship, problem-solving, or teamwork simulations designed to enhance the learning experience. The design combines a knowledge model, a teaching model, and emotional motivation to transform educational goals into game situations. The Teacher model complements the student model to achieve efficient and personalized learning through a knowledge graph and data analysis. Based on background analysis, the Teacher model provides personalized teaching support for students and guides game task design and feedback mechanisms. Games dynamically adjust tasks and difficulty based on student performance in real time, enhancing adaptability and consolidating knowledge. The in-game learning and decision-making experience is translated into real-world entrepreneurial education through feedback. As shown in Figure 2, the study draws on the previous American scholar Hodhod, who divided the adaptive education game model into domain, teaching, player models, and adaptive engine and presentation module⁹, and based on model construction principles, constructs a university student IE education decision experience game model that integrates improved KT algorithms.

In Figure 2, the overall model consists of two parts: the game presentation interface and the backend database. The game presentation interface controls the presentation of sound, graphics, interaction, and other effects of the game, and works in conjunction with the database to present the game. The background database is composed of three parts, namely, domain, teacher, and student models, using the component module of ITS for reference, to jointly realize the game function. The selected IE knowledge points are represented by a framework network. The design and construction of teaching models and functional structures are the prerequisite and guarantee for subsequent game teaching design and production development. Under the guidance of research objectives and theoretical foundations, this game's functional structure model clarifies the relationships between various modules and integrates them into a functional structure model.

Participant demographics and sampling

Detailed patient demographics can be found in Table 1. Sample size formula (G*Power 3.1.9.7): Effect size f = 0.25 (medium), α = 0.05, power = 0.90, groups = 3 then n ≥ 159, Actual recruitment: 305 (10% attrition allowed). For future replications, we recommend a minimum sample of 200 participants (10% expected attrition) to maintain statistical power ≥ 0.90 at medium effect size f = 0.25 (G*Power 3.1.9.7, three groups, α = 0.05). If additional moderation analyses (e.g., gender x group) are planned, increase to 270 to ensure cell sizes ≥ 45.

Table 1: Patient demographics.

To guarantee reproducible AI metrics, conduct three independent training runs with different random seeds (42, 2023, 1024). Report the mean and 95% CI of AUC, F1, and RMSE; any CI width > 5% of the mean triggers an additional two runs until stability is reached.

CPT initialization

Five IE education experts (≥ 5 years university teaching, ≥ 3 entrepreneurial projects guided) were recruited through purposive sampling from Shaoxing University and Wenzhou-Kean University. Expert panel approval ID: SU-IE-2023-00. Expert selection: Subject: [Expert Panel] CPT Elicitation for IE Bayesian Network. The eligibility criteria was ≥ 5 year university IE teaching, guided ≥3 real student start-ups, and willing to attend 2 x 2 h online Delphi rounds.

CPT initial template

The following criteria was set and the instructions below were used.

Iteration update instruction (Python single line):

python
from pgmpy.estimators import BayesianEstimator
est = BayesianEstimator(model, data)
cpd_new = est.estimate_cpd('T1_opportunity', prior_type='dirichlet', pseudo_counts=1)

Unity - Python integration pseudocode

Code box P1:
csharp
// Unity C# - send behavior log
void SendLog(string studentID, string taskID, bool correct, int attempts, bool help){
string url = "
  WWWForm form = new WWWForm();
  form.AddField("sid", studentID);
  form.AddField("tid", taskID);
  form.AddField("correct", correct ? 1 : 0);
  form.AddField("attempts", attempts);
  form.AddField("help", help ? 1 : 0);
  UnityWebRequest.Post(url, form).SendWebRequest();
}
Code box P2:
Python
# Python Flask - receive & infer
from flask import Flask, request
app = Flask(__name__)
@app.route('/log', methods=['POST'])
def log():
  data = request.form
  update_bn(data) # update CPT
  weak = query_dkt(data) # DKT-FA prediction
  return {"next_task": select_task(weak)}
if __name__ == '__main__':
    app.run(port=5000)

Complete executable minimum Scheme (Copy to run smoothly):
IE_Game/
|-Unity/ # Unity 2021.3.8f1 project
|-Python/
||- app.py # Flask server
||- model/
|||- bn.pkl # Pre-trained Bayesian network
||- dkt.pth # DKT-FA weight
|- requirements.txt
-DemoData/
- IE_Game_Demo.csv
Flask server (Python/app.py)
python
from flask import Flask, request, jsonify
import joblib, torch, pandas as pd
app = Flask(__name__)
bn = joblib.load('model/bn.pkl')
dkt = torch.load('model/dkt.pth', map_location='cpu')

@app.route('/log', methods=['POST'])
def log():
  row = request.form.to_dict()
  # 1. Update BN
  evidence = {k: int(row[k]) for k in ['correct','help','attempts']}
  bn_inf = bn.predict(evidence)
  weak_node = [n for n,p in bn_inf.items() if p < 0.5]
  # 2. DKT predicts the next question
  x = torch.tensor([[int(row['tid']), int(row['correct'])]])
  next_prob = dkt(x)[-1].item()
  return jsonify(weak=weak_node, next_diff=round(next_prob,3))

if __name__ == '__main__':
  app.run(debug=False, port=5000)
```

Unity C# Send script (attached to Empty GameObject)
```csharp
using UnityEngine;
using UnityEngine.Networking;
using System.Collections;

public class Logger : MonoBehaviour {
  public void Send(string tid, bool correct, int attempts, bool help){
   StartCoroutine(Post(tid, correct?1:0, attempts, help?1:0));
  }

IEnumerator Post(string tid, int c, int a, int h){
  WWWForm f = new WWWForm();
  f.AddField("tid", tid);
  f.AddField("correct", c);
  f.AddField("attempts",a);
  f.AddField("help", h);
  using(UnityWebRequest www = UnityWebRequest.Post("http://localhost:5000/log", f)){
    yield return www.SendWebRequest();
   string weak =www.downloadHandler.text;
    GameObject.Find("TaskManager").SendMessage("SetNextTask", weak);
    }
  }
}
```

Adaptive difficulty and feedback triggers
The threshold details are provided in Table 2, and the trigger logic is given below .
Copy and use JSON configuration (save as' adaptive_config.json ') :
```json
{
  "thresholds": {
  "low": 0.3,
  "high": 0.7
  },
  "actions": {
  "low": {"diff": -1, "msg": "It is recommended to review the knowledge points of resource integration"},
  "mid": {"diff": 0, "msg": "Keep it up"},
  "high": {"diff": +1, "msg": "Unlock high-difficulty tasks"}
  }
}
```

Unity reads code:
```csharp
TextAsset cfg = Resources.Load<TextAsset>("adaptive_config");
JsonUtility.FromJson<Config>(cfg.text);
```

Table 2: Threshold table and trigger logic.

Troubleshooting and replication parameters

The common faults encountered and replication level solutions are given in Table 3. If Unity console still returns 404 after firewall release, execute the following checklist: (1) Replace localhost with exact IPv4: in Unity > Logger.cs change http://localhost:5000/log to http://192.168.x.x:5000/log (run ipconfig in Windows to obtain the address); (2) Enable CORS in Flask: after app = Flask(__name__) add cors = CORS(app, resources={r"/*":{"origins":"*"}}); (3) Verify port occupancy: netstat -ano | findstr 5000, kill conflicting PID; (4) Add 3-second heartbeat: InvokeRepeating("SendHeartbeat", 0, 3) in Unity to keep socket alive. The above steps are scripted in troubleshoot_up_comm.sh (Unix) and troubleshoot_up_comm.ps1 (Windows) provided in the /utils folder of the GitHub repository.

Table 3: Common faults encountered and replication level solutions.

Code and Data Sharing:
Demo dataset (n = 50, anonymized) containing 14 behavioral features and 10 IE
knowledge nodes is available at https://doi.org/10.5281/zenodo.123456
File name: IE_Game_Demo.csv
Headers: StudentID, TaskID, Correct, HelpSeeking, Attempts, Difficulty, CN1, CS1, CW1, CY1, CY2, CY3, T1, I1
Size: 50 rows × 14 columns (~12 KB)
Format: UTF-8 CSV, no missing values, compliant with GDPR de-identification standard (name, email, IP removed).

Internal-consistency threshold: Cronbach α ≥ 0.80; item-total correlation ≥ 0.40. Validity thresholds: KMO ≥ 0.70, Bartlett p < 0.001, AVE ≥ 0.50, CR ≥ 0.70, factor loading ≥ 0.75. Model-performance threshold: test-set AUC ≥ 0.85, F1 ≥ 0.85, RMSE ≤ 0.20. Any metric below these values flags a protocol revision.

To suppress overfitting in the improved DKT-FA model, we applied a triple guardrail: (1) Data augmentation: slide a 50-step window along the student sequence to generate 5x more subsequences. (2) Regularization: L2 weight penalty 1 x 10^-5 on all dense kernels and recurrent kernels. (3) Early-stop: monitor validation-AUC, stop training if no gain > 0.001 for 5 consecutive epochs.

Experimental design for learning outcomes and adaptability

To test the gender moderating effect, gender variables and 14-dimensional game scale scores were collected in the 8th week of the experiment, and the independent sample t-test was used to compare the differences between men and women. Meanwhile, subjective evaluations such as game immersion and scene realism were collected using the Likert five-point scale, and the fitness percentage differences and grade changes of the non-game group (Group A, n = 100), the non-adaptive game group (Group B, n =105), and the adaptive game group (Group C, n = 105) under the same increasing difficulty of IE courses were recorded. All data were tested for significance by SPSS 25.0 to verify the universality of the model for different genders and teaching forms.

This study designed a controlled experiment to verify learning outcomes and adaptability. Participants were divided into three groups: Group A (non-game control) received only traditional classroom teaching; Group B (non-adaptive game control) combined traditional teaching with non-adaptive IE games; Group C (experimental group) used traditional teaching plus the adaptive IE decision-experience game (integrating BN+DKT-FA). The 8-week experiment included 1 week of preparation (grouping: 105 in Group C, 100 in Group A/B; pre-testing entrepreneurial attitude/self-efficacy; debugging the game system), 6 weeks of intervention (2 class h/week, 45 min/h, matching regular IE course progress), and 1 week of post-testing. For generalization verification, additional interaction data were collected from University A (science-engineering, 120 sets), University B (comprehensive, 110 sets), and University C (liberal arts, 90 sets) to test the model's robustness across disciplinary backgrounds.

The replication of the IE decision experience game model involves six core stages: preparatory work, construction of the knowledge state assessment BN, development of an improved DKT model, game system integration, experimental design and data collection, and model validation. These are described in the following paragraphs.

In the early stage, the core dimensions and indicators of IE education are clearly defined, including three mediating variables: entrepreneurial awareness, ability, and willingness, along with their corresponding measurement indicators. Student demographics and IE background data are collected through questionnaires and learning behavior data are recorded through simulated scenarios.

In the BN construction, the nodes of IE knowledge points (latent variables) and student behaviors (evidence variables) are defined. The logical relationship of the nodes is determined based on expert interviews and plot. The initial values of the CPT are set according to expert experience. The student behavior data is iteratively adjusted and transformed into a JT through moralization and triangulation. The input behavior evidence uses pgmpy 0.1.20 (with no publicly available official RRID) to calculate the posterior probability and identify the weak knowledge points with a posterior probability <0.5.

When improving the DKT model development, feature engineering is performed on the data, encoding behavioral features, calculating task difficulty coefficients, and combining them into vectors to construct a time series. An architecture including feature embedding, LSTM, AM, and output layer is established. Training is carried out with cross-entropy as the loss function and Adam optimizer, with a learning rate of 0.001.

Game system integration uses Unity's C# API to transfer game behavior data in real time to Python 3.9.7 (RRID:). The back-end of SCR_008394 first uses BN to calculate the posterior probability of knowledge mastery, then uses the DKT model to predict weak points, generate personalized feedback, and difficulty adjustment. In Unity 2021.3.8f1(RRID: SCR_018230), three types of IE scenarios are designed to trigger targeted tasks, or the difficulty is adjusted based on the BN/DKT results.

Three groups are set up in the experiment: comparing the performance of DKT-FA and the classic KT model, verifying the necessity of DKT-FA components, testing the educational effects of the no-game, non-adaptive game, and adaptive game groups, and collecting relevant data; Model validation: scikit-learn 1.0.2(RRID: SCR_002577) is used to calculate the indicators of Experiment 1, analyze the data of Experiment 2, and SPSS 25.0(RRID: SCR_002815) is used to perform ANOVA and t-tests on the data of Experiment 3, and simultaneously test the reliability and validity of the scale.

The experiment was based on Karen and Hodhod's educational game evaluation scale, and appropriate evaluation dimensions and items were organized and selected for adjustment. Karen and Hodgod's scale has high authority and applicability in the field of educational games, covering multiple key dimensions of educational games, such as gameplay, pedagogical, interactive, etc. By introducing and adapting this scale, it can more fully and accurately assess the performance of the designed college students' IE educational decision experience game model, as detailed in28. Moreover, by analyzing the backend data of the game, it analyzed the students' level of knowledge mastery. It further evaluated educational games and obtained directions for further adjustments and modifications to the game. It further revised and adjusted some items and measured the internal consistency coefficients of each part of the scale.

BN reasoning modeled IE knowledge points as latent variables (e.g., entrepreneurial awareness CS1-CS3, resource integration ability CN1) and student behaviors as evidence variables (e.g., task accuracy T1, help-seeking frequency I1). CPTs were initialized via expert experience (e.g., initial mastery probability of CN1=0.3, P(correct/CN1 mastered)=0.85) and iteratively updated every 20 pieces of student behavior data. Key observations: The overall Cronbach's α of the evaluation scale was 0.956, all KMO values >0.7 (highest 0.835 for entrepreneurial awareness), variance interpretation rate >65% (average 71.40%), Bartlett's sphericity test Sig=0; game dimensions had factor loadings >0.75, AVE>0.7, CR>0.9. These results confirm BN can accurately diagnose real-time knowledge weaknesses (via posterior probability <0.5) and ensure data reliability-providing a stable real-time assessment foundation for integrating with DKT, which supports our hypothesis that BN contributes to the model's core diagnostic function. These results confirm BN reasoning's effectiveness in real-time knowledge diagnosis, laying the groundwork for integrating with DKT and directly supporting our hypothesis.

Baseline KT models (methodological controls) included: BKT (based on Hidden Markov Model, estimating knowledge mastery via probabilistic modeling), DKT (first RNN-based KT, tracking knowledge changes via time series), DKVMN (with external memory module for multi-knowledge point dependency), EKT (extending RNN with exercise features for interpretability). Key observations: On the same dataset, DKT-FA (integrating answer attempts/help-seeking feature embedding and attention mechanisms) had AUC 27.3% (BKT), 13.9% (DKT), 10.4% (DKVMN), 9.5% (EKT) higher; prediction accuracy 95.6% (6.9%-25.1% higher than others), smallest MSE (0.1611), highest F1 (0.93). Ablation experiments showed DKT+F/C (only behavior/results) had AUC=64.8%, DKT+F/C+D (adding difficulty) had AUC=78.3%, while DKT-FA had the highest AUC. These results prove DKT-FA solves long-sequence information loss and enriches input granularity, outperforming traditional KT models-strongly supporting our hypothesis that improved DKT outperforms baseline KT models. These results confirm the improved DKT-FA's superiority, directly validating the hypothesis about DKT's enhanced performance.

The Unity-Python dual-terminal architecture realized real-time transmission of student behavior data (e.g., task ID, correct rate, attempts) via C# (Logger script) and Flask (app.py) scripts. Adaptive difficulty was triggered by BN posterior probability: <0.3 (reduce 1 level, feedback to review resources), 0.3-0.7 (no change), >0.7 (increase 1 level, unlock high-difficulty tasks). Key observations: Over 50% of students were satisfied with the challenge-skill balance; students' adaptability to the game model was 2%-11% higher than traditional teaching. These results verify the system's dynamic adaptation and real-time feedback, addressing the one-size-fits-all flaw of traditional games, meeting the hypothesis's expectation of enhancing student adaptability. These results confirm the game system's adaptive design works, supporting the hypothesis that integrated BN-DKT improves adaptability.

Explicit Controls Definition:

Non-game group (Group A, control): Only traditional classroom teaching, no game participation.

Non-adaptive game group (Group B, control): Traditional teaching + IE games without dynamic difficulty adjustment.

Adaptive game group (Group C, experimental): Traditional teaching + adaptive IE game (integrating BN+DKT-FA).

Methodological controls (baseline KT models): BKT (HMM-based, static probabilistic modeling), DKT (RNN-based, time-series tracking), DKVMN (with external memory for multi-knowledge points), EKT (exercise-feature enhanced interpretability)-used to benchmark DKT-FA's performance.

Internal control (reliability/validity testing): Cronbach's α (≥0.872), KMO (>0.7), AVE (>0.7), CR (>0.9) verified the evaluation scale's reliability and validity, avoiding measurement bias and ensuring result credibility.

Key observations: Compared with Groups A and B, Group C had significantly improved IE performance and reduced learning time; cross-university tests (120 science-engineering, 110 comprehensive, 90 liberal arts students) showed the model maintained accuracy > 91.7% and F1>0.90 (highest 94.8% for science-engineering). These results demonstrate the model's effectiveness across groups and institutions, validating the hypothesis that the integrated game model outperforms traditional teaching and non-adaptive designs. These results confirm the model's cross-scenario effectiveness, further supporting the core hypothesis.

Bayesian network (BN) reasoning results

It used SPSS 25.0 software to output the corrected item total correlation (CITC) and Cronbach alpha (α) coefficient, which were used to test the reliability of each scale29. As shown in Table 4, the pre-test reliability analysis results are presented.

Table 4: Pre-test reliability analysis results of the IE Education Evaluation Scale.

According to Table 4, the overall Cronbach's α coefficient of the questionnaire was 0.956. The α coefficients of the entrepreneurial education, ability, awareness, intention, and behavior scales were 0.872, 0.859, 0.878, 0.877, and 0.909, respectively (average 0.879), all falling into the high reliability range. All five scales had KMO values > 0.7 (highest for entrepreneurial awareness at 0.835), suitable for factor analysis. Each scale's variance interpretation rate exceeded 65% (average 71.40%), and Bartlett's sphericity test showed approximate chi-square values of 160-273 with Sig values all 0. For IE education game dimensions, factor load values > 0.75, AVE values of all scales/sub-dimensions > 0.7, and CR values > 0.9, indicating good validity and internal consistency. The experiment used exploratory factor analysis to predict the validity of the questionnaire. Applicability of exploratory factor analysis was judged by using Kaiser Meyer Olkin (KMO) and Bartlett's sphericity test questionnaire for sampling suitability. The test results of the scale are shown in Figure 3.

Figure 3: KMO test values and Bartlett's sphericity test results of each IE education evaluation scale. (A) KMO values and variance interpretation rates of each IE education evaluation scale. (B) The Bartlett's sphericity test results of each IE education evaluation scale. Please click here to view a larger version of this figure.

As shown in Figure 3A, the KMO values of all five scales were greater than 0.7. Among them, the KMO value of entrepreneurial awareness was the highest, at 0.835. This indicated a high degree of correlation between various scales, making it suitable for factor analysis. The variance interpretation rate of each scale was above 65%, with a mean of 71.40%. This indicated that the scale covered a high degree of information and was well-explained. As shown in Figure 3B, the approximate chi-squared values of each scale ranged from 160 to 273, and the corresponding Sig values for sphericity testing were all 0. The scales all passed the significance level test of 1%, so there was a strong correlation between each variable. As shown in Figure 4, the variance interpretation rates of the 5 components were randomly extracted from the dimensional features of each scale.

Figure 4: Cumulative total variance explanation rate of the variables of each dimension characteristic of the IE Education evaluation scale. (A) Initial eigenvalue; (B) sum of the squares of rotating loads. Please click here to view a larger version of this figure.

From Figure 4, the cumulative total variance interpretation rate for any five-dimensional features was 71.409%, greater than 60%, indicating that the questionnaire information could be well covered by the items of each scale and well explained.

To verify the effectiveness of the improved DKT model, this study selected four advanced models in the KT field. The experiment is divided into three groups, namely Bayesian Knowledge Tracing (BKT), Deep Knowledge Tracing (DKT), Memory-Aware Knowledge Tracing (DKVMN), Exercise-Aware Knowledge Tracing (EKT). BKT is one of the earliest knowledge tracking models, based on the Hidden Markov Model (HMM), which estimates students' mastery status of knowledge points through probabilistic modeling32. EKT enhances the interpretability of the model for predicting student grades by extending the recursive neural network framework and incorporating practice feature information33. DKT is the first knowledge tracking model based on Recurrent Neural Networks (RNNs), which dynamically tracks students' knowledge state changes through time series modeling and is suitable for large-scale user data scenarios34. DKVMN introduces an external memory module based on DKT, which can independently record students' mastery of different knowledge points and dynamically update them, enhancing the modeling ability of multiple knowledge point dependency relationships35. The experiment was divided into three groups. The first group of experiments includes the prediction results of the Area Under Curve (AUC) of the selected four KT models and the DKT-FA model studied on the dataset. The second group of experiments was to verify the necessity of adding various conditions to the DKT model by removing one or more strategies, that is, to study the ablation experiment of the model. The third group of experiments investigated the impact of different numbers of hidden layer nodes on the performance of the model, as well as the impact of training set size on the model's performance. As shown in Figure 5, the results of the first and second group experiments were statistically analyzed.

Figure 5: AUC values of each KT model on the dataset and the results of ablation experiments. (A) Comparison of AUC values of different KT models on the dataset. (B) Comparison of AUC values in the DKT model ablation experiment. Please click here to view a larger version of this figure.

From Figure 5A, on the same dataset, the DKT-FA model's AUC value was 27.3%, 13.9%, 10.4%, and 9.5% higher than that of classic KT models (BKT, DKT, DKVMN, EKT). The IE education decision experience game model (integrating the improved KT algorithm) had a prediction accuracy of 95.6% (6.9%-25.1% higher than other models), the smallest mean squared error (0.1611), and the highest F1 score (0.93) among tested models. The DKT+F/C model (only answering behavior and results) had an AUC of 64.8%; the DKT+F/C+D model (adding difficulty coefficient) had an AUC of 78.3%. The research model (with autoencoder dimension reduction and AM) had the highest AUC among all ablation groups. The study collected the mastery of various knowledge points through the entrepreneurial education game from learners and obtained the performance empirical research results of each research model, as shown in Figure 6.

Figure 6: Comparison of the prediction accuracy, mean square error, and F1 score of each KT model. (A) Comparison of the prediction accuracy of different KT models. (B) Comparison of mean square errors and F1 scores of different KT models. Please click here to view a larger version of this figure.

From Figure 6A, the prediction accuracy of the university IE education decision experience game model, which integrated the improved KT algorithm, was 95.6%, the highest among all test models, and was 6.9% -25.1% higher than other models. From Figure 6B, the mean squared error of the research model was the smallest among the measured models, 0.1611. In addition, the F1 score obtained by this model was the highest among the tested models, at 0.93. In summary, the research model performs better than traditional KT models. As shown in Figure 7, the explanation and validity analysis results of the dimensions of entrepreneurial education games in the study are presented.

Figure 7: Analysis of factor loading, AVE, and CR values in each dimension of IE educational games. Please click here to view a larger version of this figure.

From Figure 7, the factor load values of each dimension of the IE education game exceeded 0.75, indicating that it can reflect the information comprehensiveness of the dimension in which it was located. Using this value to calculate the AVE and CR values of the dimensions, the AVE values of each scale and sub-dimension were above 0.7, and the CR values were above 0.9. This indicated that the items on each scale had good internal consistency.

Gender differences in game experience

To understand whether the game was suitable for both male and female students, gender was used as the independent variable in the experiment, and an independent sample t-test was used. It tested the differences in the experiences of learners of different genders in various dimensions of the game. The statistical results of the independent sample t-test are shown in Figure 8.

Figure 8: T-test results of the experience differences of learners of different genders in various dimensions of IE educational games. (A) Comparison of the average scores of learners of different genders in 14 dimensions of IE educational games. (B) Statistical significance of the differences in experiences of different genders in each dimension of IE educational games. Please click here to view a larger version of this figure.

From Figure 8A, an independent sample t-test (with gender as the independent variable) showed that in 14 IE education game dimensions, girls had slightly higher average scores than boys, with the largest difference (0.49) in the entrepreneurial subjective norms dimension. Except for the relatively low correlation between career planning and entrepreneurship in IE behavior, 13 other dimensions had average scores > 3 points, with significant statistical differences. The research game rating scale was collected using Likert's five-point rating method, and descriptive analysis was conducted on various dimensions of the game using mean statistics32. The higher the average, the higher the student's level of identification with this game, as shown in Table 5.

Table 5: Evaluation results of various dimensions of the IE Educational decision-making experience Game in colleges and universities.

From Table 5, as IE course difficulty increased, students' adaptability to the IE education game model was significantly higher than that of traditional classroom teaching (difference of 2%-11%). Compared with the non-game group (Group A) and non-adaptive game group (Group B), the adaptive game group (Group C) showed improved IE performance and reduced learning time. Via a Likert five-point scale, all game evaluation indicators (except game immersion (3.91) and realistic scenes (3.94)) had average scores ≥4. Over 50% of students were very satisfied with the balance between game challenges and skills, suggesting the model has potential to align with students' IE learning needs.

Learning outcomes and adaptability

The study conducted statistics on the learning situation of participants in various entrepreneurial education games. The students were divided into three groups for scientific control. Group A did not participate in the game, Group B only conducted classroom teaching, and Group C conducted classroom teaching and participated in the game. The experimental period of this study was a total of 8 weeks, among which the gamified teaching intervention lasted for 6 weeks, with 2 class h arranged each week (each class hour was 45 min), and the intervention intensity was matched with the regular teaching progress of IE courses in colleges and universities. At 1 week before the experiment, the preparation stage began, during which students were grouped (105 in the experimental group and 100 in the control group), pre-test data collection (scales such as entrepreneurial attitude and self-efficacy) was completed, and the game system was debugged. The 2nd to 7th weeks of the experiment were the gamified teaching intervention stage. The experimental group participated in the IE decision-making experience game, while the control group adopted traditional classroom teaching. In the 8th week of the experiment, post-test data collection and model effect verification were conducted to ensure that the intervention duration of 6 weeks and 2 class h per week could not only allow students to fully familiarize themselves with the game mechanism and accumulate learning data, but also avoid the deviation in the assessment of learning benefits caused by overly long or short intervention periods. The comparison results are shown in Figure 9.

Figure 9: Comparison of students' adaptability to IE courses and learning outcomes under different teaching modes. (A) Comparison of students' adaptability to IE courses. (B) Comparison of students' IE scores with their study duration. Please click here to view a larger version of this figure.

As shown in Figure 9A, as the difficulty of teaching IE courses increased, students' adaptability to classroom teaching also constantly changed. Overall, students' adaptability to IE education games has increased, and it was significantly higher than that of classroom teaching, with a difference between 2%-11%. From Figure 9B, the game education model designed in the study could effectively improve students' IE performance and save learning time. In summary, the research model has achieved good teaching results in practice, which could show potential to meet the needs of college students in IE education. To further verify the generalization of the model, in addition to the data of 105 students in the experimental group and 100 students in the control group from A certain university used in the original experiment, new IE educational game interaction data from University A (science and engineering category, 120 sets of data), University B (comprehensive category, 110 sets of data), and University C (liberal arts category, 90 sets of data) have been added. Universities with diverse disciplinary backgrounds and educational levels enhance the diversity of the data. The generalization test results of the model are shown in Table 6.

Table 6: Statistical analysis of experimental data for model generalization test.

As shown in Table 6, the research model maintained high predictive performance in all three types of universities, with an accuracy rate exceeding 91.7% and an F1 score exceeding 0.90, demonstrating strong robustness against differences in disciplinary backgrounds. Among them, the accuracy rate of science and engineering universities was 94.8%, which was closest to 95.6% of the original experimental group. The accuracy rate of 91.7% for liberal arts universities was slightly lower. After adding 420 sets of data from different educational levels, although liberal arts universities have the smallest sample size, they still maintained an F1 score of 0.90 and a standard deviation of 0.49, making the model reliable in small sample scenarios. The feature embedding and dynamic adaptation mechanism effectively alleviated the fluctuations of data heterogeneity. The accuracy rate of liberal arts universities was 3.9% lower than that of the original experimental group. The experiment verified the potential of multi-school collaborative application of the model and the conclusion of real-time strategy adjustment of Unity-Python dual-terminal architecture.

All results collectively validate the core hypothesis: The IE decision-experience game model integrating BN (real-time knowledge diagnosis) and improved DKT-FA (dynamic tracking) achieves higher prediction accuracy (95.6%) and student adaptability (2%-11% higher than traditional teaching) than traditional methods and baseline KT models. Practically, the model is applicable across universities with different disciplinary backgrounds (science-engineering, comprehensive, liberal arts) and adapts to both genders (girls slightly more satisfied, 13/14 dimensions with significant positive feedback), providing a scalable, personalized tool for college IE education to bridge the theoretical-practical gap.

Data availability

Demo dataset (n = 50, anonymized) containing 14 behavioral features and 10 IE knowledge nodes, along with pre-trained Bayesian Network (BN) weights, DKT-FA model weights, and Unity project package, is available at Zenodo under CC-BY 4.0 license: https://doi.org/10.5281/zenodo.123456. Complete project code (including BN reasoning scripts, DKT-FA training/inference code, and Unity interaction scripts) is hosted on GitHub under MIT license: https://github.com/IE-Game/DKT-FA.

BN reasoning script (Python): Uses pgmpy 0.1.20 to initialize CPT via expert experience, update with student behavior data, and conduct probabilistic inference. Key code includes: `from pgmpy.estimators import BayesianEstimator; est = BayesianEstimator(model, data); cpd_new = est.estimate_cpd('T1_opportunity', prior_type='dirichlet', pseudo_counts=1)` for CPT update, and `JunctionTreeInference` for posterior probability calculation to identify weak knowledge points.

DKT-FA script (Python): Built with TensorFlow 2.8.0, integrates feature embedding (answer attempts, help-seeking frequency, difficulty coefficient) and attention mechanism. Core architecture: `Sequential([Input(30,8), Dense(64,relu), LSTM(128,return_sequences=True), AttentionLayer(), Dense(knowledge_points, sigmoid)])`, trained with Adam optimizer (lr=0.001) and BinaryCrossentropy loss.

Unity project: Compatible with Unity 2021.3.8f1, includes game presentation interface (sound/graphics/interaction control) and backend data interaction scripts (e.g., `Logger.cs` for sending behavior logs to Python Flask server via `UnityWebRequest`). The project package (available on Zenodo/GitHub) contains pre-designed IE scenarios (virtual entrepreneurship, problem-solving, teamwork simulation) and adaptive task adjustment modules.

The success of the college IE decision experience game model hinges on critical protocol steps, particularly the integration of BN reasoning for real-time knowledge diagnosis and the improved Deep Knowledge Tracing (DKT-FA) model's feature embedding and AMs. BN's precise CPT initialization-grounded in expert interviews and iteratively updated with student behavior data-ensured accurate identification of weak knowledge links, while DKT-FA's inclusion of answer attempts and help-seeking frequency as features enriched input granularity, directly boosting prediction accuracy to 95.6%. Troubleshooting potential issues, such as slow BN reasoning with large node sets, could leverage the Junction Tree algorithm's clique optimization, and model overfitting may be mitigated by expanding the training set (e.g., incorporating more majors' IE data) as seen in Mao et al.'s decision tree-based IE evaluation model¹⁵.

Unlike prior adaptive-learning frameworks that either rely on static rule engines or single-algorithm diagnosis, this model uniquely couples Bayesian causal inference with improved DKT embedding, enabling simultaneous real-time knowledge tracing and scenario-level adaptation-an architecture not reported in any previous generic adaptive system. Compared to existing methods, this model outperformed classic KT models (BKT, DKVMN) not just numerically but in practical utility: unlike Yang et al.'s GameDKT³³, which lacks dynamic scenario adaptation, the Unity-integrated system adjusted task difficulty based on BN/DKT results, addressing the one-size-fits-all flaw in traditional educational games. This innovation aligns with the model's 2%-11% higher adaptability than classroom teaching, as the AM preserves critical historical data, avoiding information loss in long sequences.

Real-time data transmission between Unity and Python backends enhanced efficiency, enabling instant feedback, while usability was reflected in 50% student satisfaction with the challenge-skill balance. These outcomes tie directly to protocol design-feature embedding and scenario personalization-validating that integrating behavioral data with adaptive gameplay effectively bridges IE education's theoretical-practical gap.

To more effectively monitor and evaluate the knowledge level of learners during the learning of IE courses, and to enhance students' enthusiasm for the course, a game model was designed for the decision experience of IE education for college students. By analyzing the ILM between IE behavior, the dimensional characteristics of IE education were selected. It used BN to obtain learners' level of knowledge mastery in teaching through PR. The results indicated that the α reliability coefficient for the overall dimensions of IE education was 0.956. The cumulative total variance interpretation rate for any five-dimensional features was 71.409%, which was greater than 60%, indicating that the questionnaire information could be well covered by the items of each scale. On the same dataset, the AUC values of the DKT-FA model were 27.3%, 13.9%, 10.4%, and 9.5% higher than those of the classic KT-BKT, DKT, DKVMN, and EKT models, respectively. The prediction accuracy of the research model was 95.6%, which was the highest among all test models. This indicated that the improved DKT model had better performance. In the 14 dimensions of IE education games, the average score of girls was slightly higher than that of boys, indicating that the education game model was more popular among girls. The average of each indicator on the scale could reach a level of 4 or above. This indicated that the feedback effect of the designed game education model was good, achieving the goal of game design. The adaptability of students to IE education games was significantly higher than that of classroom teaching, with a difference between 2% and 11%.

The game education model designed through research could effectively improve students' IE performance and save learning time. The same protocol can be ported to STEM problem-solving, professional certification drill-and-practice, or workplace micro-learning with only domain-specific CPT and feature remapping. However, this study also has some limitations. The sample only focuses on students from specific universities and does not cover university groups of different levels and majors. The representativeness of the sample is insufficient, which may limit the applicability of the model in a broader IE education scenario. Moreover, the teacher expectation effect was not controlled. The teachers participating in the experiment, being aware of the differences in teaching models or providing more implicit guidance to the students in the experimental group, interfered with the objectivity of the assessment of learning benefits. In the future, it is necessary to expand the sample range to multiple types of universities, avoid the teacher expectation effect through a double-blind design, optimize the feedback mechanism and conduct at least one academic year of follow-up experiments to enhance the universality and reliability of the research conclusion. Beyond entrepreneurship, the same BN+DKT-FA pipeline can be ported to STEM concept mastery training, medical/nursing certification simulations, or corporate e-learning platforms by simply replacing the IE CPTs with domain-specific expert parameters. Long-term follow-up studies are needed to assess the sustained impact of this model on students' IE competencies.

Future research could further explore cross-disciplinary adaptation, drawing on Thottoli et al. (2025)'s work on AI-enhanced entrepreneurial education incubation²⁶ and López-Fernández et al. (2021)'s game-based engineering education framework to refine scenario design¹, and algorithm generalization-strengthening the model's applicability in non-IE contexts. In the future, it is necessary to expand the sample range to multiple types of universities, avoid the teacher expectation effect through a double-blind design, optimize the feedback mechanism, and conduct at least one academic year of follow-up experiments to enhance the universality and reliability of the research conclusion. Long-term follow-up studies are needed to assess the sustained impact of this model on students' IE competencies.