HeteLFX: Heterogeneous recommendation with latent feature extraction

Abstract

This study proposes a heterogeneous recommendation model that does not rely on data sharing. Previous studies have predominantly focused on nested homogeneous domains that share data. However, this approach encounters issues as it could lead to diminished recommendation performance when there is a scarcity of redundant data within these domains. To overcome these challenges, we propose the HeteLFX model, which extracts and bridges the latent features (LF) of each domain. This model resolves the problems by leveraging the metainformation of domain items to generate an LF. LF is extracted for each domain, and bridges are established based on the relevance of the latent knowledge, thereby enabling heterogeneous recommendations. The efficacy of the HeteLFX model was assessed by comparing it with four other heterogeneous recommendation systems, which are variants of X-Map and NX-Map. The results revealed that the HeteLFX model improved performance by reducing the mean absolute error (MAE) by approximately 0.3, thereby underscoring the superiority of the model. Additionally, HeteLFX reduced the MAE by up to approximately 0.45, depending on the relevance of the data within the domain.