Project: Dish Recommendation System with Collaborative Filtering and Clustering

A machine learning project exploring two models—collaborative filtering and ingredient-based clustering—for restaurant dish recommendations.

Overview

This project implements two distinct models for dish recommendation using real-world user rating datasets:

  • Model 1: Memory-based collaborative filtering (user-user)
  • Model 2: Hard clustering of dishes (ingredient-based, K-medoids)

Both models were evaluated on their ability to predict user ratings and generate ranked recommendations, using real datasets of user–dish ratings and dish ingredient features.

Features

  • Handles sparse and noisy real-world data
  • Compares collaborative and content-based approaches
  • Evaluation on multiple ranking metrics
  • Command-line interface for reproducible experiments

Running the Code

Model 1 (Collaborative Filtering)

python3 MODELPART1.py dishes.csv user_ratings_train.json user_ratings_test.json

  • dishes.csv: Contains dish metadata (dish IDs, ingredients)
  • user_ratings_train.json: User–dish ratings for training
  • user_ratings_test.json: User–dish ratings for testing

Model 2 (K-Medoids Clustering)

python3 MODELPART2.py dishes.csv user_ratings_test.json 5

  • The final number is the number of clusters (e.g., 5).

Methodology

Model 1: Collaborative Filtering

  • Supervised learning using train/test split
  • Computes per-user average ratings
  • Predicts ratings for test users based on similar users’ history (cosine similarity on rating vectors)
  • If no overlapping ratings, defaults to the user’s average rating from training data
  • Clamps predicted ratings to the valid interval [1, 5]
  • Metrics computed: MAE, Precision@n, Recall@n

Observations

  • Works well with dense data; struggles with sparse user–item overlap.
  • Average MAE ≈ 0.6 for typical datasets.
  • Performance (especially recall) improves as training data grows.
  • Limited by lack of shared history between users.

Model 2: Ingredient-based Clustering

  • Content-based: Dishes are clustered by ingredient similarity (binary vectors, custom similarity metric).
  • Uses K-Medoids (robust to outliers, binary-friendly) with customizable number of clusters.
  • Clusters are evaluated for tightness and separation (WC_SSE, BC_SSE).
  • User’s test ratings are predicted as the mean rating within the assigned cluster.
  • Recommendations made by selecting highest-rated clusters for each user.
  • Metrics computed: MAE, Precision@n, Recall@n

Observations

  • MAE values improve substantially as K increases (more clusters = finer grouping).
  • Model outperforms a mode-rating baseline for all datasets.
  • Recommendations are better aligned with actual user taste trends.
  • The mean as a cluster label provides best results.

Results Summary

Metric Model 1 (500) Model 1 (2000) Model 2, K=50 (500) Model 2, K=50 (2000)
MAE 0.63 0.58 0.34 0.34
Prec@10 0.89 0.91 0.95 0.95
Recall@10 0.43 0.44 0.46 0.46
  • Model 2 consistently outperforms Model 1 on all metrics, especially with more clusters and data.

Insights

  • Collaborative filtering struggles with sparse user–dish overlap, especially in cold-start scenarios.
  • Clustering by dish ingredients (content-based) captures meaningful food similarities, producing better predictions and recommendations.
  • Higher cluster counts (K) generally improve model performance by capturing finer-grained food preferences.

Please see this class report for additional insights:Report

How to Use

  • Plug in your own CSV/JSON files with dish and user-rating data.
  • Choose the model appropriate for your data density (collaborative vs. content-based).
  • Tune the number of clusters (K) for best results in Model 2.

Code

Link