Data Completeness Impact on Deep Learning Based Explainable Recommender Systems

Recommendation systems become an important part in helping users find the most fit items in many domains, such as healthcare, transportation, agriculture, media, and e-commerce. The development of an explainable recommendation system was claimed to add more value to improve user satisfaction. Various studies have been conducted regarding the model to solve real world issues. Yet, the impact of data quality in building an explainable recommendation system is still scarce. In this paper, we investigate whether the completeness of data used in building explainable recommendation system will impact the performance of the recommender systems and quality of the explanation. We use the Yelp and MovieLens dataset and train a deep learning explainable recommendation system model, Co-Attentive Multi-task Learning (CAML), on various amounts of data by reducing the data records using algorithm technique to achieve Missing Completely at Random (MCAR) and Missing at Random (MAR), and by eliminating selected features on each dataset. Finally, we evaluated the outcomes based on Root Mean Square Error (RMSE) for rating evaluation and Bilingual Evaluation Understudy (BLEU) & Recall-Oriented Understudy for Gisting Evaluation (ROUGE) for explanation evaluation. Our experiments conclude with Pearson Correlation Coefficient between completeness of data with the evaluation result in each of the experiment. We found out that different types of data reduction and dataset impact differently to the level of the performance of the ratings and explanation.