How ginoshun/steam-game-recommendation-system works
Project Type and Purpose
• This is a machine learning-based recommendation system focused on Steam games.
• The system leverages various models (LSTM, ResNet, NCF) to process user data and game information, and to generate personalized game recommendations.
• It is driven by experiments including different training approaches (with balanced focal loss vs. unbalanced) and provides model evaluation and outcome reporting.
Analyzing the File Structure
• Top-level directories include:
• model/: Contains both “balanced(best)” models and “unbalanced(old)” models, representing different experiments.
• outcome/: Stores the results including plots, evaluations, and performance metrics for each type of model (LSTM, ResNet, NCF).
• The file structure shows a clear separation of concerns – models (trained weights), notebooks (workflow scripts), and outcomes (evaluation results).
Core Components
a. Data Preparation & Ingestion:
• Notebook component (prepare.ipynb) that visualizes and readies the dataset (sourced from Kaggle).
• headers.txt may be used for API calls or data collection.
b. Model Training and Experimentation:
• Notebook: train.ipynb for overall training orchestration.
• Separate model directories for different experiments:
• model/balanced(best) with the best models using focal loss balancing.
• model/unbalanced(old) representing earlier, unbalanced training.
• This component illustrates the experimentation with various architectures: LSTM, ResNet, and NCF.
c. Model Evaluation and Analysis:
• Notebook: model_evaluation.ipynb is responsible for computing evaluation metrics (accuracy, confusion matrices, ROC curves) for each model.
• outcome/ directory shows subcomponents for each model type (LSTM, NCF, ResNet) depicting related performance outcomes.
Interactions and Data Flow
• Data flows from the data source (Kaggle dataset) → Data Preparation (prepare.ipynb) → Model Training (train.ipynb) which outputs trained models stored in the model/ directory.
• After training, evaluation processes are executed (model_evaluation.ipynb) reading the trained .h5 models and generating performance outcomes stored in outcome/.
Architectural Patterns and Design Principles
• The project follows a typical machine learning workflow pattern where each stage (data preprocessing, training, evaluation) is modularized.
• There is a clear separation between experimental configurations (balanced vs. unbalanced) which aids in model selection and performance benchmarking.
• Each ML model type (LSTM, ResNet, NCF) is treated as a separate “component” under the training and evaluation phase.