Advancing Context-Aware Recommender Systems: A Deep Context-Based Factorization Machines Approach

Abstract

Context-aware recommender systems (CARS) aim to offer personalized recommendations by incorporating user contextual information through analysis.By analyzing these contextual cues, CARS can better understand the preferences and needs of users in different situations, thereby improving the relevance and effectiveness of the recommendations they provide.However, integrating contextual information into a recommendation system presents challenges due to the potential increase in the sparsity and dimensionality.Recent studies have demonstrated that representing user context as a latent vector can effectively address these kinds of issues.In fact, models such as Factorization Machines (FMs) have been widely used due to their effectiveness and their ability to tackle sparsity and to reduce feature space into a condensed latent space.Despite these advantages, FMs encounter limitations when dealing with higher-order feature interactions, since the model's design, primarily focused on second-order interactions.Furthermore, a significant drawback of FMs is their inability to distinguish between different contexts effectively.By utilizing a uniform latent space to model interactions across all features, FMs overlook the nuanced differences that distinct contexts bring to the interactions.This article introduces a CARS model called Deep Context-Based Factorization Machines (DeepCBFM).The DeepCBFM combines the power of deep learning with an extended version of Factorization Machines (FMs) to model non-linear feature interactions among user, item, and contextual dimensions.Additionally, it addresses specific shortcomings of FMs with the goal of enhancing recommendation accuracy.We implemented our method using two datasets that incorporate contextual information, each having distinct context dimensions.Experimental findings demonstrate that the DeepCBFM model surpasses baseline models, thereby validating its efficacy.