[Paper Review] Variational Autoencoders for Collaborative Filtering

Dawen Liang, Rahul G. Krishnan+2 more|arXiv (Cornell University)|Feb 16, 2018

Recommender Systems and TechniquesComputer Science46 references66 citations

TL;DR

The paper extends variational autoencoders to collaborative filtering for implicit feedback using a multinomial likelihood and a beta-regularized objective, achieving state-of-the-art results on real datasets.

ABSTRACT

We extend variational autoencoders (VAEs) to collaborative filtering for implicit feedback. This non-linear probabilistic model enables us to go beyond the limited modeling capacity of linear factor models which still largely dominate collaborative filtering research.We introduce a generative model with multinomial likelihood and use Bayesian inference for parameter estimation. Despite widespread use in language modeling and economics, the multinomial likelihood receives less attention in the recommender systems literature. We introduce a different regularization parameter for the learning objective, which proves to be crucial for achieving competitive performance. Remarkably, there is an efficient way to tune the parameter using annealing. The resulting model and learning algorithm has information-theoretic connections to maximum entropy discrimination and the information bottleneck principle. Empirically, we show that the proposed approach significantly outperforms several state-of-the-art baselines, including two recently-proposed neural network approaches, on several real-world datasets. We also provide extended experiments comparing the multinomial likelihood with other commonly used likelihood functions in the latent factor collaborative filtering literature and show favorable results. Finally, we identify the pros and cons of employing a principled Bayesian inference approach and characterize settings where it provides the most significant improvements.

Motivation & Objective

• Extend variational autoencoders to collaborative filtering for implicit feedback.
• Adopt a multinomial likelihood to better model click data and ranking loss.
• Introduce a beta-regularized ELBO with annealing to improve performance.
• Compare multinomial VAE to Gaussian and logistic baselines and analyze Bayesian inference benefits.

Proposed method

• Generative model: latent z_u ~ N(0,I) mapped via a non-linear function f_theta to a probability over I items, then x_u ~ Mult(N_u, pi(z_u)).
• Inference: amortized variational posterior q_phi(z_u|x_u) ~ N(mu_phi(x_u), diag(sigma_phi^2(x_u))).
• Optimization: maximize the evidence lower bound ELBO with a reparameterization trick for gradients; introduce beta to regularize KL term.
• Beta-regularized objective: L_beta = E_q[log p_theta(x_u|z_u)] - beta * KL(q_phi(z_u|x_u)||p(z_u)); anneal beta from 0 to 1 to improve training.
• Prediction: rank items by unnormalized multinomial probabilities using z_u = mu_phi(x_u) (or z from the encoder in Mult-DAE).
• Notes: address computational considerations for large item sets and relate approach to autoencoder taxonomy.

Experimental results

Research questions

• RQ1 Does a multinomial likelihood improve ranking-oriented evaluation over Gaussian or logistic likelihoods in implicit-feedback data?
• RQ2 Can a beta-regularized VAE with KL annealing yield better ranking metrics than standard VAE training in collaborative filtering?
• RQ3 How does Mult-VAE-pr compare to denoising autoencoders and other neural CF methods on real-world datasets?
• RQ4 What are the trade-offs of explicit uncertainty modeling (Mult-VAE-pr) vs deterministic encoders (Mult-DAE) in recommendations?

Key findings

• Mult-VAE-pr achieves state-of-the-art results on several real-world datasets compared to baselines including recent neural-network approaches.
• The multinomial likelihood consistently outperforms Gaussian and logistic likelihoods for implicit feedback in this setting.
• Both Mult-VAE-pr and Mult-DAE provide competitive performance, with explicit uncertainty modeling offering advantages in certain regimes.
• Beta annealing is crucial for avoiding underutilization of the latent space and yields significant performance gains; best results typically obtained by annealing then selecting the peak validation metric.
• The approach connects to information-theoretic ideas (information bottleneck, maximum-entropy discrimination) and remains robust under data sparsity.