Recommender System New Frontiers

Introduction

Recommender system has evolved over the years and now being widely used in almost all the services where plenty of options are available for user. Collaborative filtering based methods currently dominate the recommendation space and has been widely used. While current recommendation systems are good at coming up with relevant recommendations, they do have shortcomings. Future recommendation systems are expected to solve the shortcoming of current systems and also improve user experience and usefulness of the system.

Shortcomings of Current Systems

Most of the recommender systems makes recommendation based on users past behavior and similarity of users or items. As the items shown to user are selected by the system, user get to see only content that the system thinks user will like. This results in ‘Filter Bubble’ in which the user is separated from information that disagrees with their viewpoints, effectively isolating them in their own cultural or ideological bubbles. The current recommender systems also does not account for change in human behavior or choices. In addition current recommender systems are subjected to exploitation and used to spread fake news and to swing user opinion by distorting truth.

Future Recommender System

Future recommender system would be more closer to user, be context driven, will account for human nature and be sensible of impact of recommendations on society, politics and culture. One of the methods being tried to come up with such a system is to create a model that represents user and build a system that provides more personalized recommendations. Such a system would be able to predict change in user behavior, choices and adjust recommendation accordingly. Such a system would also be aware of implicit context and would provide information at right time.

Modeling users and their behavior is an area under active development. While modeling user especially their behavior remains a challenge, companies are trying to build systems that accounts for some aspects of human behavior. For example amazon is trying to simulate window shopping experience online after realizing that such an experience is lacking when items are listed by their traditional recommendation system. A recommender system for window shopping needs to use different metric compared to collaborative filtering system in order to come up with items to be listed to user. Metric such as freshness, diversity, seasonality, visual appeal and relevance to user style are used to come up with recommendation for a window shopping list.

There are ample work being done on context based recommendation. Many systems are currently making use of explicit context and trying to become better at getting the implicit context right. For example Google Now with access to many user data does a pretty good job of getting implicit context. Google Now try to determine the context even before a user issues a query and refines the context once user provides a query.

The future recommender system would also take advantage of development in deep learning to model user and to predict context. Future recommender system also need to consider the impact of recommendations and eliminate exploitation of the system. Making recommendation system more closer to user makes it less vulnerable to attacks/exploitation. Machine learning is used to understand user better and detect bots and not trustworthy users and use this information to limit exploitation.

My Viewpoint

Current recommendation systems mostly provide recommendations on a specific product category. While some systems provide suggestion for related add-on item, rarely the systems provide recommendation over broad range of items. For example when user looks for a birthday dress other items that are relevant to the occasion such as birthday cake, decorations, gifts are not usually shown. It would be novel when a user looks for birthday dress for a kid and gets recommendations not only for dress but also for an investment plan for college education of kid. I would love to see such a recommender system that takes holistic view and provides recommendation from multiple categories. Such a system would be required to identify context, item of interest and other categories which are relevant to either the context or the item of interest. Each product category would have its own recommendation system and the recommendations will come from multiple recommendation systems. Then a sorter would take recommendations from multiple systems, sort them based on user query and implicit/explicit context and deliver recommendations from multiple categories in appropriate ratio.

References