Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Shallow AutoEncoding Recommender with Cold Start Handling via Side Features

Edward DongBo Cui, Lu Zhang, William Ping-hsun Lee

TL;DR

This work addresses the persistent cold-start problems in recommender systems by extending the EASE framework to incorporate both user and item side information through FEASE. The authors develop FEASE-U for user cold-start, FEASE-I and FEASE-I-Prior for item cold-start with content priors, and FEASE-Prior for joint handling, all deriving closed-form solutions to enable efficient, scalable training. Across Netflix, MovieLens, and Amazon Books, FEASE variants demonstrate improved accuracy and robustness over baselines, with controlled trade-offs between warm-item performance and cold-item coverage, and the ability to adjust the prior strength to tailor cold-start handling. The approach offers a practical, interpretable baseline that remains competitive in warm settings and provides a foundation for integrating richer side information and dynamic, real-time recommendations in industrial contexts.

Abstract

User and item cold starts present significant challenges in industrial applications of recommendation systems. Supplementing user-item interaction data with metadata is a common solution-but often at the cost of introducing additional biases. In this work, we introduce an augmented EASE model that seamlessly integrates both user and item side information to address these cold start issues. Our straightforward, autoencoder-based method produces a closed-form solution that leverages rich content signals for cold items while refining user representations in data-sparse environments. Importantly, our method strikes a balance by effectively recommending cold start items and handling cold start users without incurring extra bias, and it maintains strong performance in warm settings. Experimental results demonstrate improved recommendation accuracy and robustness compared to previous collaborative filtering approaches. Moreover, our model serves as a strong baseline for future comparative studies.

Shallow AutoEncoding Recommender with Cold Start Handling via Side Features

TL;DR

This work addresses the persistent cold-start problems in recommender systems by extending the EASE framework to incorporate both user and item side information through FEASE. The authors develop FEASE-U for user cold-start, FEASE-I and FEASE-I-Prior for item cold-start with content priors, and FEASE-Prior for joint handling, all deriving closed-form solutions to enable efficient, scalable training. Across Netflix, MovieLens, and Amazon Books, FEASE variants demonstrate improved accuracy and robustness over baselines, with controlled trade-offs between warm-item performance and cold-item coverage, and the ability to adjust the prior strength to tailor cold-start handling. The approach offers a practical, interpretable baseline that remains competitive in warm settings and provides a foundation for integrating richer side information and dynamic, real-time recommendations in industrial contexts.

Abstract

User and item cold starts present significant challenges in industrial applications of recommendation systems. Supplementing user-item interaction data with metadata is a common solution-but often at the cost of introducing additional biases. In this work, we introduce an augmented EASE model that seamlessly integrates both user and item side information to address these cold start issues. Our straightforward, autoencoder-based method produces a closed-form solution that leverages rich content signals for cold items while refining user representations in data-sparse environments. Importantly, our method strikes a balance by effectively recommending cold start items and handling cold start users without incurring extra bias, and it maintains strong performance in warm settings. Experimental results demonstrate improved recommendation accuracy and robustness compared to previous collaborative filtering approaches. Moreover, our model serves as a strong baseline for future comparative studies.

Paper Structure

This paper contains 24 sections, 37 equations, 2 figures, 17 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. EASE and SLIM
  4. AutoEncoder
  5. Neighborhood-based Collaborative Filtering
  6. EASE Model with Item Side Information
  7. Methodology
  8. User Cold Start
  9. Item Cold Start
  10. Simultaneous User and Item Cold Start Handling
  11. Similarity Model
  12. Implementation in Python
  13. Experimental Results
  14. Datasets
  15. Baseline Models
  16. ...and 9 more sections

Figures (2)

  • Figure 1: Learned weights from EASE (a) and FEASE-U (b) models on Netflix data.
  • Figure 2: Model performance comparison between joint optimization (i.e. models with "-Prior" suffix) and cold item weight replacement on Netflix data. a) FEASE-I (solid line) vs. FEASE-I-Prior (dashed line). b) FEASE (solid line) vs. FEASE-Prior (dashed line). Panel (a) shows FEASE-I (solid line) versus FEASE-I-Prior (dashed line), and panel (b) compares FEASE (solid line) to FEASE-Prior (dashed line). In both cases, as the weight $\delta$ increases, overall model performance (measured by HitRatio@20 on the y-axis) declines, while the exposure of cold items (measured by ColdItemHR@20 on the x-axis) improves. Notably, the strategy of directly replacing cold item weights can sustain cold item recommendation performance without significantly harming overall accuracy.