switching recommender system for music platform

Capstone project in recommender systems for MIT applied data science program.

Applied Data Science Program by MIT Professional Education with Great Learning.

Overview

This project consisted on developing a music recommendation system to suggest the top 10 songs for users based on their preferences and listening history. It tackles challenges such as overwhelming content, sparse data, and the cold-start problem. The proposed solution involves a hybrid recommendation system that adapts to user behavior, categorizing users into three groups: new, moderately active, and highly active. This approach balances exploration and engagement, improving user retention, increasing time spent on the platform, and offering personalized music suggestions while addressing scalability and computational costs.

For this project, I was provided with the data from Taste Profile Subset released by the Echo Nest as part of the Million Song Dataset to fit my models.

Problem Statement

The music industry faces overwhelming content volume, where users interact with only ~7% of songs, leading to sparse data and cold-start problems. Without strong recommendations, platforms risk losing users to competitors.

Goals

  • Retain users and increase time spent on the platform.
  • Provide top-10 personalized recommendations for all user types.
  • Balance engagement, exploration, and operational cost.

Approach

  • Data Cleaning: Processed raw user–song interaction logs, handled missing values, removed noise, and ensured consistency in play counts.
  • Exploratory Data Analysis (EDA): Examined user behavior patterns, song popularity distributions, and sparsity issues to guide model selection.
  • Model Fitting: Trained three models tailored to user groups: popularity-based, content-based, and collaborative filtering (SVD).
  • Model Improvement: Tuned hyperparameters, optimized similarity aggregation, and validated performance with offline experiments.
  • Model Comparison: Benchmarked approaches using Precision@k, Recall@k, F1@k, and RMSE to identify trade-offs in accuracy, diversity, and scalability.
  • Final Solution: Designed a switching hybrid recommender system that dynamically selects the best model depending on user play history, addressing both cold-start and active-user scenarios.
Exploratory Data Analysis of the music data set.

Tools and Technologies

  • Programming & Modeling: Python, NumPy, Pandas, Scikit-learn, SciPy.
  • Recommendation Methods: Popularity-based ranking, Content-based filtering (similarity matrices), Collaborative filtering with SVD (Matrix Factorization).
  • Evaluation: Precision@k, Recall@k, F1@k, RMSE.
  • Databases & Vector Stores: Pinecone (for embedding and similarity search).

Solution

The solution is a switching hybrid recommender system. Users are divided into three groups by their play count:

  • Group A: No history → Popularity-based recommendations.
  • Group B: Few plays (< median) → Content-based recommendations.
  • Group C: Active users (≥ median) → Collaborative filtering with Matrix Factorization (SVD).

New songs with zero plays are handled by content-based similarity, ensuring catalog freshness.

The users are splitted into three groups, and for each group I use a different approach to recommend songs.

Group A: Popularity Model

  • Recommends top-10 most popular tracks (above 75th percentile of play count).
  • Designed for cold-start users with no history.

Group B: Content-Based Model

  • Aggregates similarity scores from top-N user songs to rank unheard tracks.
  • Offline similarity matrix computation.
  • Limitation: lacks diversity and collaborative effects.

Group C: Collaborative Filtering

  • Uses Matrix Factorization (SVD) to capture latent features.
  • Achieved F1@k = 0.502, Precision@k = 0.415, Recall@k = 0.635, RMSE = 1.0141.
  • Handles sparse data well; computationally expensive.

Results

  • Switching strategy addressed cold start (Groups A, B).
  • Collaborative filtering achieved strong accuracy and engagement for active users.
  • Balanced exploration and personalization, reducing churn and increasing session time.
  • Business modeling showed potential monthly benefits ($3,952.50) vs. costs ($2,882).
Comparison of the metrics from the different models tested.
What-if analysis of the proposed solution in terms of costs.

Suggested Next Steps

  • Add onboarding questionnaire for Group A to capture preferences earlier.
  • Dynamic thresholding for group classification (ɳ).
  • Incorporate weighted hybrid models to smooth transitions between groups.
  • Track additional metrics (play depth, skip rate).
  • Invest in scalable infrastructure, CI/CD pipelines, and continuous retraining.