Building an Intelligent Learning Recommendation System from Scratch: A Python-Based Heuristic Search Practice

This article introduces the open-source project Study_Recommender, developed by Sayan Mondal, a first-year student at VIT Bhopal University, as a practical project for an AI/ML course. The system uses only Python standard libraries and Pandas to implement full learning material recommendation functionality. Its core technologies include intelligent agents, constraint search, and heuristic sorting, without relying on deep learning frameworks like TensorFlow, demonstrating the feasibility of using classic AI techniques to solve personalized learning recommendation problems.

The design philosophy of Study_Recommender is "return to basics": in an era dominated by deep learning, it chooses a classic AI tech stack to build the recommendation system. This choice stems from an understanding of the learning material recommendation scenario—clear constraints (subject, difficulty, material type) and interpretable rules are more suitable than black-box neural networks. The system simulates the logic of human teachers to recommend videos, notes, and practice problems, and optimizes recommendation quality based on user feedback.

The core of the system is an intelligent agent that follows the perception-action-learning cycle, consisting of three modules:

1. Perception module: Collects user input (subject preferences, difficulty requirements, material type preferences) as constraints;
2. Decision module: Filters candidate materials through constraints, then sorts them using heuristic scoring;
3. Learning module: Updates the adaptive score of materials from users' 1-5 star ratings to optimize future recommendations.

The recommendation algorithm is divided into two parts: constraint filtering and heuristic scoring.

• Constraint filtering: Precisely matches candidate materials based on subject, difficulty, and type;
• Heuristic scoring formula: Final Score = adaptive_score × difficulty_weight × type_weight
  • Adaptive score: Uses Bayesian average, formula: (base_score + Σ ratings)/(1 + N), where base_score is initially 3.5 and N is the number of ratings, solving the cold start problem;
  • Difficulty weight: 1.2x for difficult content, 1.0x for easy content;
  • Type weight: 1.1x for practice problems, 1.0x for videos.

The adaptive capability comes from the user feedback loop: after a user rates a material, the system updates its adaptive score—high ratings increase its ranking, while low ratings decrease it. This mechanism simulates the recommendation pattern of human teachers and works without a large amount of user data, making it suitable for cold start scenarios. The system uses CSV files to store materials (materials.csv) and feedback records (feedback.csv), which is lightweight and easy to deploy without requiring a database server.

The project code is well-organized, with core modules including:

• agent.py: Implements the perception-action-learning cycle;
• recommender.py: Encapsulates constraint search and heuristic sorting;
• feedback.py: Handles rating updates;
• main.py: Command-line interface;
• utils.py: Auxiliary functions. Dependencies only include Python 3.x and Pandas, with a low deployment threshold, making it suitable for demonstration in teaching environments.

Application scenarios:

• Teaching cases: Demonstrates classic AI concepts such as intelligent agents, constraint satisfaction, and heuristic search;
• Technical prototype: Suitable for scenarios with limited resources or requiring interpretable recommendation results. Expansion directions:
• Add features like learning duration and history to optimize scoring;
• Add a collaborative filtering layer to enable user knowledge sharing;
• Migrate data storage to a database to support large-scale applications.

Study_Recommender proves that classic AI techniques can build effective recommendation systems, with its value lying in simplicity, interpretability, and ease of deployment. It is an excellent starting point for beginners to understand the principles of recommendation systems, and provides a lightweight reference framework for developers. The core competitiveness of AI technology lies in understanding the essence of problems and mastering classic methods, rather than complexity.