LeetCode Rating Predictor: A Contest Score Prediction System Based on Deep Neural Networks

LeetCode Rating Predictor is a full-stack application for algorithm contest enthusiasts. It uses a dense neural network (TensorFlow) with 15-dimensional features to predict contest rating changes, including a FastAPI backend and React frontend, supporting real-time prediction. Target users include contest participants, algorithm learners, and tech enthusiasts, providing an intuitive prediction tool and learning reference.

Background

To provide a real-time rating prediction tool for contest participants, combining machine learning and web development practices to meet users' needs such as pre-contest assessment and progress tracking.

Target Users

• Contest Participants: Assess pre-contest status and set goals
• Algorithm Learners: Track progress and analyze weak points
• Tech Enthusiasts: Learn full-stack ML application architecture and feature engineering

Backend: FastAPI + TensorFlow

• Asynchronous high-performance framework with auto-generated API documentation
• Data acquisition layer crawls user historical data with caching strategy
• 15-dimensional features: basic statistics, difficulty distribution, contest history, time series
• Model structure: Input layer → 64-unit ReLU → Dropout(0.3) → 32-unit ReLU → Dropout(0.2) → Linear output layer

Frontend: React + Data Visualization

• React18 + TypeScript + Tailwind CSS + Recharts
• Core components: Search, Dashboard (trend chart/radar chart), Prediction results
• Responsive design supporting mobile devices and dark mode

Data Processing

• Sources: LeetCode API, anonymized user data, historical contest archives
• Process: Cleaning → Feature calculation → Standardization → Time window splitting

Training Configuration

• Optimizer Adam(0.001), loss MSE, evaluation metrics MAE/RMSE/R²
• Early stopping mechanism to prevent overfitting

Performance

• Offline test set RMSE ~45 points, R² ~0.72
• Online evaluation collects real feedback and conducts A/B testing

LeetCode Rating Predictor is an excellent full-stack ML project that demonstrates the practical value of combining deep learning and web technologies. Core contributions:

1. Provides a reference for the complete ML application development process
2. Reflects the importance of feature engineering in real-world problems
3. Balances model performance and engineering complexity It provides a progress tracking tool for contest enthusiasts and a practical case for tech learners.

Current Limitations

• Data relies on LeetCode API, insufficient data for new users
• Large prediction deviation in extreme cases, insufficient function interpretability

Future Improvements

• Model upgrade: LSTM/Transformer, graph neural networks
• Function expansion: Recommendation system, learning path planning
• Engineering optimization: Feature platform, model serving (MLflow/Seldon)