Student Success Intelligent Assistant: An Academic Risk Prediction System Integrating Rule-Based Reasoning and Machine Learning

This article introduces an academic risk prediction system for students that integrates rule-based reasoning and machine learning. It provides personalized advice via an AI-driven chat interface to help students identify risk factors and improve their academic performance. The system's core goals are early warning, interpretability, action orientation, and easy accessibility. It aims to transform students from passive managers to active self-managers and serve as a powerful assistant for teachers.

Today's students face challenges such as increasing course difficulty, scattered resources, and rising psychological pressure. Traditional academic early warning systems are lagging and one-dimensional, making it difficult to help at-risk students in a timely manner. Artificial intelligence can analyze multi-dimensional academic data to provide early warnings before problems worsen and offer personalized interventions. The Student Success Intelligent Assistant is a practical application of this concept.

System Design Philosophy

The core philosophy is "prevention is better than cure", with goals including early warning, interpretability, action orientation, and easy accessibility.

Dual-Engine Prediction Model

• Rule-Based Reasoning Engine: Establishes clear rules for attendance, assignment submission, grade trends, etc., based on educational experts' knowledge. It has high transparency and is easy to adjust.
• Machine Learning Model: Trained using historical data, with features covering academic, behavioral, demographic, and temporal aspects. Models like gradient boosting trees and logistic regression are selected. SMOTE is used to handle imbalanced data, and time-series cross-validation is applied to avoid data leakage.

Fusion Decision Mechanism

Risk score = α × Rule Engine Score + (1-α) × ML Model Score. α is flexibly configured based on the semester phase (higher weight for rules at the beginning of the semester, higher weight for ML in the middle of the semester).

Dialogue Capability Design

• Risk Explanation: Lists specific triggering factors, explains the impact mechanism, and provides improvement paths.
• Personalized Advice: Generates customized suggestions for issues like time management, learning methods, and psychological pressure.

Technical Implementation

• Retrieval-Augmented Generation (RAG): Builds a knowledge base of school policies and course information to ensure accurate responses.
• Few-Shot Prompting: Generates warm and encouraging responses based on student profiles, risk factors, and available resources.
• Safety Guardrails: Filters inappropriate content, identifies crisis signals and transfers to humans, and limits professional advice.

Data Collection

• Progressive Collection: Only asks for basic information in the first conversation; collects more data naturally in ongoing dialogues and automatically syncs with school system data.
• Natural Language Input: Students describe their situation in natural language, and the AI automatically parses and stores it in a structured format.

Privacy and Ethics

• Data Minimization: Only collects necessary data, cleans it regularly, and students can view or delete their data.
• Transparent Usage: Clearly informs users of data purposes, provides a privacy policy, and obtains explicit consent.
• Fairness Assurance: Regularly audits model biases to avoid algorithmic discrimination.

Case 1: Freshman Adaptation Difficulties

Student Zhang (Computer Science major) was at high risk after midterms. The system analyzed his declining attendance, unsubmitted assignments, and failing calculus. After an AI conversation, sleep issues were identified. Suggestions included communicating with roommates/switching dorms, contacting teaching assistants, and attending tutoring classes. By the end of the semester, his risk level dropped to low.

Case 2: Junior Student with Overload Course Load

Student Li (junior) took 6 courses (24 credits). The system warned of heavy credit load, late assignments, and reduced study time. The AI suggested dropping one course, prioritizing core courses, and creating a schedule. He achieved good grades in the end.

Key Metrics

• Prediction Accuracy: Precision, recall, F1 score.
• Intervention Effect: Grade improvement rate, risk conversion ratio, student satisfaction.
• System Usage: Number of dialogues, advice adoption rate, user retention rate.

Continuous Optimization

• Model Iteration: Retrain every semester, conduct A/B testing on algorithm features.
• Knowledge Base Update: Regularly update policy resources and optimize advice templates.
• Rule Optimization: Adjust rules based on teacher feedback and identify new risk patterns.

Cross-School Adaptation

Supports field mapping for different schools, rule configuration, and multi-school knowledge base versions.

Functional Expansion Directions

• Career Planning Integration: Recommend career paths, internship opportunities, and skill gap analysis.
• Mental Health Support: Identify risk signals and link with academic support.
• Social Learning: Recommend study partners and form interest groups.

Conclusion

The system does not replace teachers; instead, it assists teachers in focusing on students and empowers students with self-awareness. In the future, it will be implemented in more schools, and educators need to embrace such tools to improve educational quality.