Content

Authors: Hoang-Vu Dang, Dung Ngo

Abstract: Many recommender systems operate on transactional data, of which the most widely available measure is the level of interaction between users and items. In real-world applications the recommendation process is often further divided into two stages: candidate generation, which selects a number of suitable items from the whole inventory, and ranking, which determines the order of suitability among the candidates. In this paper we define two separate tests to represent the candidate generation and ranking tasks, and evaluate a diverse set of old and new collaborative filtering methods on these tests. The results on three different datasets show that model accuracy on the two tasks do not correlate strongly with each other, thus in practical applications the model choice for each task should be considered independently. Furthermore we introduce an attentional autoencoder for implicit recommendation, based on the Autorec and Transformer architectures, trained with a variational objective. The proposed model yields consistently superior results to the other evaluated methods in both the candidate generation and ranking tasks.

Published in: CIIS 2019: Proceedings of the 2019 2nd International Conference on Computational Intelligence and Intelligent Systems

Date of Conference: November 2019

Conference Location: Bangkok Thailand

DOI: https://doi.org/10.1145/3372422.3372423