Binary Latent Representations for Efficient Ranking: Empirical Assessment
arXiv:1706.07479
Abstract
Large-scale recommender systems often face severe latency and storage constraints at prediction time. These are particularly acute when the number of items that could be recommended is large, and calculating predictions for the full set is computationally intensive. In an attempt to relax these constraints, we train recommendation models that use binary rather than real-valued user and item representations, and show that while they are substantially faster to evaluate, the gains in speed come at a large cost in accuracy. In our Movielens 1M experiments, we show that reducing the latent dimensionality of traditional models offers a more attractive accuracy/speed trade-off than using binary representations.
References in corpus (6)
- Adam: A Method for Stochastic Optimization
- BPR: Bayesian Personalized Ranking from Implicit Feedback
- Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1
- Session-based Recommendations with Recurrent Neural Networks
- Asymmetric LSH (ALSH) for Sublinear Time Maximum Inner Product Search (MIPS)
- Faster Population Counts Using AVX2 Instructions