# Noise Contrastive Estimation for Scalable Linear Models for One-Class Collaborative Filtering

**Authors:** Ga Wu, Maksims Volkovs, Chee Loong Soon, Scott Sanner, Himanshu Rai  
**arXiv:** [1811.00697](https://arxiv.org/abs/1811.00697)  
**Date:** 2018-11-02  
**Summary page:** https://www.summarizepaper.com/en/paper/1811.00697v1/

## Key points

- The paper proposes a novel method to improve the accuracy of one-class collaborative filtering methods.
- Existing scalable methods such as PLRec use fast randomized SVD to embed items into a latent space, but these often exhibit popularity bias that skews the ability to accurately embed niche items.
- The authors leverage insights from NCE to derive a closed-form, efficiently computable "depopularized" embedding called NCE-PLRec.
- NCE-PLRec uses improved item embedding of NCE while correcting for its popularity unbiasing in final recommendations through linear regression to learn personalized user models.
- Empirically, NCE-PLRec outperforms state-of-the-art methods as well as various ablations of itself on large-scale recommendation datasets.
- The proposed method provides uniform recommendations across all levels of popularity and addresses the issue of popularity bias in item embeddings.

## Summary

The paper titled "Noise Contrastive Estimation for Scalable Linear Models for One-Class Collaborative Filtering" proposes a novel method to improve the accuracy of one-class collaborative filtering methods. The existing scalable methods such as Projected Linear Recommendation (PLRec) use fast randomized SVD to embed items into a latent space, followed by linear regression methods to learn personalized recommendation models per user. However, these naive SVD embedding methods often exhibit a popularity bias that skews the ability to accurately embed niche items. To address this issue, the authors leverage insights from Noise Contrastive Estimation (NCE) to derive a closed-form, efficiently computable "depopularized" embedding. The proposed method is called NCE-PLRec and it uses the improved item embedding of NCE while correcting for its popularity unbiasing in final recommendations through linear regression to learn personalized user models. While this method may not be ideal for direct recommendation using methods like PureSVD since popularity still plays an important role in recommendation, an analysis of the recommendation popularity distribution demonstrates that NCE-PLRec uniformly distributes its recommendations over the entire popularity spectrum while other methods exhibit distinct biases towards specific popularity subranges, thus artificially restricting their recommendations. Empirically, NCE-PLRec outperforms state-of-the-art methods as well as various ablations of itself on a variety of large-scale recommendation datasets. The authors also provide insights into how their method can be used in practice and suggest future research directions. Overall, this paper presents a significant contribution towards improving the accuracy and scalability of one-class collaborative filtering methods by addressing the issue of popularity bias in item embeddings and providing uniform recommendations across all levels of popularity.