Context Relevancy Assessment in Tensor Factorization-based Recommender Systems

• Published in: 2020 3rd International Conference on Communication System, Computing and IT Applications (CSCITA) pp. 141 - 145
• Main Authors: Patil, Vandana A., Jayaswal, Deepak J.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.04.2020
• Subjects: Context aware Recommendation System (CARS); Context-aware services; Correlation; Data Sparsity; Dimensionality reduction; Feature extraction; Filtering; Matrix Factorization (MF); Tensor Factorization (TF); Tensors; Training
• DOI: 10.1109/CSCITA47329.2020.9137778

Recommender systems (RSs) are modern data filtering tools aiming to generate personalized recommendations using different approaches. The data sparsity problem raised due to unavailability of sufficient rating information is observed as a major challenge in front of RSs as they are highly dependent on the explicit rating data. Factorization based dimensionality reduction approach provides a favorable solution to this sparsity problem as it solves the missing data problem using matrix completion. Context aware recommender systems are advanced RSs, which take into consideration various contexts e.g. age, location, gender etc. while generating recommendations. The data sparsity problem in context aware RSs is handled through Tensor Factorization (TF) based techniques, which is the generalization of Matrix Factorization approach. In this work, the importance of incorporating adequate number of relevant contexts into Tensor factorization based RSs is highlighted. Too less contexts lack in adding the true flavor of context awareness while too many degrades the performance of Context aware Recommender System (CARS) by adding noise. We propose a fusion based techniques built on Information Gain and correlation based feature selection technique to derive the list of relevant contexts. We further determine the optimum number of relevant contexts to maintain the trade-off between accuracy and computational complexity. The experiments are performed on two context aware datasets from different domains to generalize the findings.