Efficient Algorithms for Recommendation Systems
Much of the data collected and used today is to understand human behavior, which is very important from both a societal as well as a business perspective. Everyday, we interact and observe each other's decisions and body language to collect enough data to understand people around us.
These impressions that we both make and get form the basis for some sort of classification of different people around us. Based on this classification, we act and react in a certain way around different profiles of people. In a similar manner, the data of our "digital selves" and our interactions and activities using our digital devices reveal interesting properties of our profiles.
So how can this kind of insight of different profiles be useful when profile data is available in a digital format? Recommender systems, or expert systems, are actually examples of how data is used to build profiles and act based on it. Recommender systems consist of two main parts. The first part is classification and segmentation of people, markets, products, and services. For instance, people can be partitioned based on different attributes such as age, gender, geographical region, language, martial status, etc. Also, our digital traces such as our web-surfing activities may also contribute to the classification task of different user or customer profiles. But the attributes we take into account should be relevant to the market, product, and/or the service in question. Based on this insight of different profiles and segmentations, a recommender system can build a model to be able to predict the behavior of users and customers.
Companies like Amazon, Netflix, and Spotify have recommender systems that are based on this model. Amazon has profiles based on both people's personal attributes as well as their shopping history. People with similar profiles and shopping histories get similar recommendations. Netflix has rating systems for shows, series, and movies. For instance, if Alice and Bob liked similar movies, a movie that Alice liked will be recommended to Bob. Spotify's "Discover Weekly" functionality works in a similar manner. Depending on the listener's taste (music consumption history), the recommender system would suggest relevant music content.
Our research considers novel efficient algorithms that can solve some of the problems that state of the art faces today. The first problem is to guarantee a certain performance based on a given amount of information provided to the recommendation system. The second problem we try address the problem of explainable AI, where we automatically extract the essential relationships of items and events and provide an explanation and meaning of the extracted information. The third problem that we are approaching is to design efficient algorithms that are able to provide recommendations efficiently that are implementable on devices as part of the goal of having what is known as edge AI.
The purpose of the project is to develop expert systems that can provide high-precision recommendations and provide explanations as to why the recommendations have been given, as part of an "Explainable AI".
The aim with the project is to develop recommendation systems that have higher performance than state of the art, and develop AI systems that can provide reasonable explanations for certain recommendations with proper information.
RISE role in project