AutoJudge: An Intelligent Programming Problem Difficulty Prediction System Based on Random Forest Algorithm

AutoJudge is an AI tool that uses the random forest algorithm to analyze the text features of programming problems, aiming to solve the pain point of judging problem difficulty in programming learning and competitions. It automatically evaluates the difficulty of programming problems using machine learning technology, helping users quickly determine the complexity of challenges. Its application scenarios cover competition participants, education platforms, and corporate interviews. Although it has limitations, it can be optimized in the future through directions like multi-modal fusion.

In programming learning and competition preparation, how to judge the real difficulty of a problem is a common pain point. The difficulty labels of existing online judge platforms (OJ) are mostly based on subjective experience or simple pass rate statistics, lacking in-depth analysis of the text features of problems. Choosing overly difficult problems can easily undermine confidence, while overly easy ones waste time—thus AutoJudge was born.

AutoJudge chooses the random forest algorithm because it has advantages such as interpretable feature importance, strong anti-overfitting ability, and efficient handling of high-dimensional data. Text feature engineering includes the lexical level (keyword frequency, text length, etc.), semantic level (word embedding to capture semantics), and structural level (paragraph structure, input-output complexity, etc.).

The core features of AutoJudge include: 1. Web data crawling: Built-in crawler to obtain problem data from platforms like LeetCode and Codeforces; 2. Real-time difficulty prediction: Users can get Easy/Medium/Hard evaluation results in seconds by pasting the problem description; 3. Multi-language support: Adapted for programming learners from different language backgrounds.

The application scenarios of AutoJudge include: 1. Programming competition participants: Quickly filter problems suitable for their level and develop a scientific problem-solving plan; 2. Online education platforms: Integrate the API to automatically label problem difficulty, reducing manual burden; 3. Corporate technical interviews: Evaluate the rationality of self-designed problem difficulty to ensure the quality of interview questions.

Current limitations: It relies on problem text features, so its accuracy is limited for problems that require code analysis; there are differences in difficulty definitions across different platforms, requiring domain-specific fine-tuning. Future directions: Multi-modal fusion (combining text, test cases, and historical data); personalized recommendation (combining user problem-solving records); deep learning upgrade (exploring pre-trained models like BERT).