Beyond Matrix Factorization: Using hybrid features for user-business recommendations

Categories: Data Analysis, Business & Sales | Pricing: Free | Official Website ↗

Yelp's engineering blog details a hybrid recommendation model that combines content-based and collaborative filtering for improved user-business suggestions.

This blog post from Yelp Engineering describes a new hybrid approach for user-to-business recommendations, moving beyond traditional matrix factorization. The previous system, based on Spark's Alternating Least Squares (ALS), struggled with 'tail users' (those with few interactions) and couldn't incorporate content-based features like reviews or user segments. The new model addresses these limitations by integrating various features and optimizing for different objectives. The hybrid model uses a supervised learning-to-rank technique, specifically XGBoost with a rank:ndcg objective, to combine interaction features (from matrix factorization) and content-based features (e.g., business categories, review ratings, text-based similarity from encoded reviews). This approach significantly improved recommendation coverage and performance, doubling the number of users for whom recommendations could be generated and enhancing results for all users. The development prioritized cost-efficiency and an iterative design, building on existing infrastructure and leveraging well-supported ML techniques.

Key Features

Hybrid recommendation model combining collaborative and content-based features
Utilizes XGBoost with rank:ndcg objective for personalized ranking
Incorporates text-based similarity from business reviews using Universal Sentence Encoder
Addresses cold-start problem for 'tail users'
Scalable prediction system with recall steps based on user location and product constraints
Negative sampling strategy with a recall step for training

Pros

Doubled the number of users for whom recommendations could be made
Vastly improved performance for all users compared to matrix factorization
Effectively handles the cold-start problem for users with few interactions
Allows incorporation of diverse content-based features
Cost-efficient and iterative development approach
Improved offline evaluation metrics (NDCG, MAP)

Cons

Negative sampling for implicit interactions is complex and prone to bias
Feature computation and XGBoost evaluation are more expensive than simple dot-products
Requires careful definition of 'groups' for ranking tasks to ensure personalization
Scaling predictions for millions of users and businesses remains an intensive job
Relies on specific recall strategies for training and prediction to manage bias and scale

Use Cases

Personalized push notifications
Email engagement campaigns
Home feed content curation
Collections recommendations
User-business matching in local service directories

Best For

Machine learning engineers
Data scientists
E-commerce platforms
Content platforms
Recommendation system developers

Platforms: Web

Watch demo on YouTube ↗

View full Beyond Matrix Factorization: Using hybrid features for user-business recommendations profile on Tools-Radar | Browse Data Analysis tools | Alternatives to Beyond Matrix Factorization: Using hybrid features for user-business recommendations

Tools-Radar is a free directory of 10,000+ AI tools — discover, compare, and choose the right AI software for your needs. Visit tools-radar.com