What is it about?
Item catalog size can be a decisive limiting factor for the practical use of recommenders. In many cases, the best models are too computationally expensive to be used on real-world datasets, and simpler models must be used instead. In this work, we show how to extend the applicable range of one of the most popular models for collaborative filtering -- EASE^R, which also suffers from poor scalability to domains with large item sets. The key is exploiting the inherent sparsity of the modeled problem to find a sparse, strongly compressed approximation of its otherwise prohibitively large parameter matrix. This allows us to decrease 1) training time, 2) training memory requirements, and 3) the final model size, all by several orders of magnitude and without quality degradation.
Featured Image
Photo by Bastian Riccardi on Unsplash
Why is it important?
The paper shifts the paradigm of large-scale collaborative filtering. Since the relations modeled in these tasks are fundamentally sparse, techniques tailored to sparse structures uncover colossal efficiency gains. As a result, we extend the EASE^R model to 100x larger datasets than previously possible. On the other hand, we provide a cheap yet very accurate compressed approximation of EASE^R on medium-sized datasets. This combination of efficiency, accuracy, and compression opens new possibilities for researchers and industry applications. Additionally, the techniques presented in this paper apply to many other problems that rest on finding a cheap approximation of a very large inverse covariance (or precision) matrix.
Perspectives
We really love the EASE^R algorithm as it is brilliantly simple, easy to implement, easy to explain, and yet very powerful. By this work, we aimed to address its only major weakness: scalability w.r.t. volume of items. Indeed, being a student or a junior researcher playing with RS algorithms, you quickly reach the limits of your hardware while training EASE^R. Already some tens of thousands of items may be prohibitive for a standard laptop. SANSA gives you sufficient scalability to evaluate principally the same model on orders of magnitude larger datasets. The same benefit may be grasped by industry practitioners, who may easily scale an EASE-like model for millions of items.
Ladislav Peška
Charles University
At GLAMI we recommend millions of items to millions of users, it would not be feasible to compute EASE on this scale of data. Thanks to this contribution we are able to train models on even larger scales and bring more relevant recommendations to our users.
Radek Bartyzal
Our work demonstrates that sparsity of user-item interactions can be used to our advantage for building a very cheap recommender system even for very large tasks. I believe this idea translates beyond the realm of recommender systems --- there are other domains where dominant, strong relationships between datapoints are sparse. Where else can we exploit this sparsity? Exploring this could unveil cost-effective solutions in other fields.
Martin Spišák
GLAMI Group
The biggest positive of working on this model, for me, was connecting the area of recommender systems with numerical calculations on sparse data. The end product is a very intuitive model that is easy to work with and easy to tinker with in production.
Antonín Hoskovec
Czech Technical University in Prague
Read the Original
This page is a summary of: Scalable Approximate NonSymmetric Autoencoder for Collaborative Filtering, September 2023, ACM (Association for Computing Machinery),
DOI: 10.1145/3604915.3608827.
You can read the full text:
Resources
Contributors
The following have contributed to this page
