Latent factor models for collaborative filtering

Abstract

The enormous growth in online availability of information content has made Recommender Systems (RS) an integral part of most online portals and e-commerce sites. Most websites and service portals, be it movie rental services, online shopping or travel package providers, offer some form of recommendations to users. These recommendations provide the users more clarity, that too expeditiously and accurately in limiting (shortlisting) the items/information they need to search through, thereby improving the customer's experience. The direct link between customer's satisfaction and revenue of e-commerce sites induce widespread interest of both, academia and industry, in the design of efficient recommender systems.

The current de-facto approach for RS design is Collaborative Filtering (CF). CF techniques use the ratings provided by users, to a subset of the items in the repository, to make future recommendations. However, the rating information is hard to acquire; often a user has rated less than 5% of the items. Thus, the biggest challenge in recommender system design is to infer users’ preference from this extremely limited predilection information. The lack of adequate (explicit) preference information has motivated several works to augment the rating data with auxiliary information such as user’s demographics, trust networks, and item tags. Further, the scale of the problem, i.e. the amount of the data to be processed (selecting few items out of hundreds and thousands of items for an equally large number of users) adds another dimension to the concerns surrounding the design of a good RS. There have been several developments in the field of RS design over the past decades. However, the difficulty in achieving the desired accuracy and effectiveness in recommendations leaves considerable scope for improvement.

In this work, we model effective recommendation strategies, using optimization centric frameworks, by exploiting reliable and readily available information, to address several pertinent issues concerning RS design. Our proposed recommendation strategies are built on the principals of latent factor models (LFM). LFM are constructed on the belief that a user’s choice for an item is governed by a handful of factors – the latent factors. For example, in the case of movies, these factors may be genre, director, language while for hotels it can be price and location.

Our first contribution targets improvement in prediction accuracy as well the speed of processing by suggesting modifications to the standard LFM frameworks. We develop a more intuitive model, supported by effective algorithm design, which better captures the underlying structure of the rating database while ensuring a reduction in run time compared to standard CF techniques. In the next step, we build upon these proposed frameworks to address the problem of lack of collaborative data, especially for cold start (new) users and items, by making use of readily available user and item metadata - item category and user demographics. Our suggested frameworks make use of available metadata to add additional constraints in the standard models; thereby presenting a comprehensive strategy to improve prediction accuracy in both warm (existing users/items for which rating data is available) and cold start scenario.

Although, high recommendation accuracy is the hallmark of a good RS, over-emphasis on accuracy compromises on variety and leads to monotony. Our next set of models aims to address this concern and promote diversity and novelty in recommendations. Most existing works, targeting diversity, build ad-hoc exploratory models relying heavily on heuristic formulations. In the proposed work, we modify the latent factor model to formulate a joint optimization strategy to establish accuracy-diversity balance; our models yield superior results than existing works.

The last contribution of this work is to explore the use of another representation learning tool for collaborative filtering – Autoencoder (AE). Conventional AE based designs, use only the rating information; lack of adequate data hampers the performance of these structures, thus, they do not perform as well as conventional LFM based designs. In this work, we propose a modification of the standard autoencoder – the Supervised Autoencoder – which can jointly accommodate information from multiple sources resulting in better performance than existing architectures.