HIT: Learning a Hierarchical Tree-Based Model with Variable-Length Layers for Recommendation Systems

Abstract

Large-scale industrial recommendation systems (RS) usually confront computational problems due to the enormous corpus size. Hence, an efficient indexing structure is a practical solution to retrieve and recommend the most relevant items within a limited response time. The existing approaches that adopted embedding or tree-based index structures cannot handle the long-tail phenomenon. To address this issue, we propose a HIerarchical Tree-based model with variable-length layers (HIT) for recommendation systems. HIT consists of a hierarchical tree index structure and a user preference prediction model. It can fully exploit all the training data by dynamically adjusting the lengths of layers in its tree index structure, which can effectively alleviate the long-tail problem. To assess the models’ resistance against the long-tail problem, we further define two types of equilibrium under our index structure. To satisfy the equilibrium, we propose a corresponding hierarchical tree learning algorithm. Furthermore, for those items with a rare appearance in the training data, on which the learning algorithm would fail, we design a dedicated bandit layer to solve them. Extensive experiments on three large-scale real-world datasets show that HIT can significantly outperform the existing methods in terms of efficient recommendations on items with different frequencies.