DeepDive: Deep Latent Factor Model for Enhancing Diversity in Recommender Systems

Most collaborative filtering techniques concentrate on increasing the accuracy of business-to-customer recommender systems. Emphasis on accuracy alone leads to repetitive recommendations based on user's past preferences; such predictions pose a problem from both business and user's perspective as they fail to recommend niche items and maintain the user's interest. Incorporating diversity in recommendations can overcome these issues. Most prior studies include diversity by randomizing the item-set predicted by the collaborating filtering technique. These techniques do not have control over the accuracy vs. diversity trade-off; one needs to be mindful that a drastic loss in accuracy is not acceptable from the recommender system. Our work proposes a deep latent factor model with a diversity cost/penalty that allows us to control the trade-off between diversity and accuracy. Experimental results obtained with the Movielens dataset demonstrate the superior performance of our proposed method in providing relevant, novel, and diverse recommendations compared to state-of-the-art techniques; with a slight drop in accuracy, our proposed method provides an improvement in different established measures of diversity.