Evolving context-aware recommender systems with users in mind

A context-aware recommender system (CARS) utilizes users’ context to provide personalized services. Contextual information can be derived from sensors in order improve the accuracy of recommendations. In this work, we focus on CARSs with high-dimensional contextual that typically impacts recommendation model, for example, by increasing model’s dimensionality and sparsity. Generating accurate recommendations is not enough constitute a useful user’s perspective, since use some may cause problems, such as draining battery, raising privacy concerns, more. Previous studies suggested reducing amount utilized using domain knowledge select most suitable information. This approach only applicable when set contexts small handle sufficient preventing Moreover, hand-crafted represent an optimal features process. Another compress into denser latent space, but limit ability explain recommended items users or compromise their trust. paper, present multi-step selecting low-dimensional subsets incorporating them explicitly within CARSs. At core our novel feature selection algorithm based genetic algorithms, which outperforms state-of-the-art reduction CARS algorithms improving interpretability. Over course evolution, thousands diverse are generated; deep model produced each subset, stacked together. The resulting uses interpretable, explicit features. Our includes mechanism tuning different underlying affect user battery consumption. We evaluated two datasets smartphones. An empirical analysis results confirms proposed models while transparency interpretability user. addition results, several cases, examples methodology how researchers, experts modelers tweak various concerns • algorithm-based systems. means addressing usage, privacy, etc. Accurate interpretable take account. State-of-the-art qualitative quantitative datasets.