Machine Learning Analysis of LeetCode Problems

Overview

This project applies machine learning techniques to analyze and predict various attributes of LeetCode problems. The dataset used is the Leetcode Problem Dataset, containing 1,825 problem entries with 17 features.

Objectives

• Explore the dataset to uncover relationships between different features.
• Build predictive models for:
  • Problem difficulty classification
  • Related topics prediction
  • Number of accepted submissions prediction
  • Likes and dislikes prediction
• Develop an interactive UI for user-friendly access to these predictions.

Data Description

The dataset consists of the following key features:

• Id: Problem ID
• Title & Description: Name and textual description of the problem
• Is_Premium: Whether a premium subscription is required (1 = Yes, 0 = No)
• Difficulty: Easy, Medium, or Hard
• Acceptance_rate: Frequency of correct submissions
• Companies & Related Topics: Companies that ask the problem and relevant topics
• Likes & Dislikes: Community engagement metrics
• Similar Questions: Related problems and difficulty levels

Data Processing

Several preprocessing steps were applied:

• Multi-hot encoding for categorical variables (e.g., related_topics, companies)
• Label encoding for difficulty and title
• Conversion of numerical shorthand (e.g., K for thousands, M for millions)
• Text preprocessing (lemmatization, stopword removal, TF-IDF vectorization)

Machine Learning Models

Different ML models were used for various tasks:

1. Difficulty Classification

• Models Used: Logistic Regression, Random Forest, SVM, KNN, Gradient Boosting, XGBoost, and a simple Neural Network.
• Best Performing Model: Random Forest.
• Feature Importance: is_premium, acceptance_rate, rating, discuss_count, and top words from TF-IDF.

2. Related Topics Prediction (Multi-label Classification)

• Models Used: Logistic Regression (One-vs-Rest), SVM, Random Forest, KNN, Gradient Boosting, XGBoost.
• Threshold Optimization: Used Stratified K-Fold Cross-Validation to optimize prediction thresholds per label.
• Best Performing Model: Random Forest.

3. Predicting Accepted Submissions

• Model Used: Linear Regression.
• Key Features: submissions, difficulty, discuss_count, is_premium.
• Evaluation Metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE).

4. Predicting Likes & Dislikes

• Models Used: Linear Regression, Random Forest, Gradient Boosting, XGBoost.
• Best Performing Model: Random Forest for likes and XGBoost for dislikes.
• Evaluation Metrics: RMSE, MAE.

User Interface

An interactive UI was developed using Gradio. Users can input problem attributes and receive real-time predictions. Features include:

• Predicting related topics
• Predicting problem difficulty
• Predicting acceptance rate
• Predicting likes and dislikes

Future Work

• Enhance neural network models with deep learning architectures
• Implement a recommendation system for personalized problem suggestions
• Improve NLP-based embeddings using transformer models (e.g., BERT)

Links

• Dataset: Kaggle - Leetcode Problem Dataset
• User Interface: Gradio UI

Contributors

• Noa Fishman
• Rachel Belokopytov