Projects

  • ENSURE - looks at ways of improving the transparency and decision support for recommender systems (like Amazon and Spotify), in recommendation scenarios that contain both surprising recommendations and trade-offs.
  • SuSPECT -  smaller project for the center for education and learning (CEL) looking at issues relating to fake news in education and problem-based learning: SuSPECT: Scaffolding Student PErspectives for Critical Thinking.

I am on the outlook for enthusiastic PhD and Masters students to work on topics relating to the transparency of intelligent systems, and usability/interface issues (e.g. diversity, novelty) in recommender systems. Here are some topics I would like to collaborate on:


Interfaces for serendipitous discovery in recommender systems

Recommender Systems help users to find relevant items from a vast amount of information. In order to provide tailored recommendations, these systems have to learn from users' behaviour with the aim of discovering their preferences. One of the potential strengths of a recommender system is helping a user to explore the long tail – items that are particularly relevant to them but are not necessarily very popular or interesting to other people. However, the risk with tailoring too well to a user leads to a form of over-fitting that colloquially is called the filter bubble. In this bubble users do not discover items outside their comfort zone. While they may discover new things, they do not encounter anything unexpected. This thesis will focus on ways of helping users explore the search space in a way that helps them break out of the filter bubble. That is, to increase the coverage of items in a catalog that they review, and the number of serendipitous (felicitous and novel) discoveries they make. Simultaneously, the interface will aim to maintain a similar level of accuracy to a baseline system without causing cognitive overload on the user.


Handling low confidence recommendations

Recommender Systems help users to find relevant items from a vast amount of information. In order to provide tailored recommendations, these systems have to learn from users' behaviour with the aim of discovering their preferences. Many times is hard to make reliable predictions. There are several reasons why this may be the case: an item may have few ratings (item cold start problem), or a user may not have entered many ratings (user cold start). Also, polarizing items may be more risky recommendations. A typical example of this is the movie “Napoleon Dynamite” - the movie has been rated more than two million times in the Netflix database, and the ratings are disproportionately one or five stars. The thesis will investigate how to deal with low confidence recommendations. Omitting them from recommendations or demoting them in lists of recommendations is one possibility. E.g. the confidence measure can be used to help decide among several items with the same expected rating. A recommender system can also combine different recommendation techniques based on the confidence each has when predicting a particular user-item pair. However, whichever method is used it should not overly demote good long-tail movies, e.g. items that are less popular but are highly relevant. This would mean working against the strength of recommender systems of supplying felicitous and unexpected discovery of items the user likes. In cases where low confidence items are still chosen for promotion, an explanation can be offered to the user in order to mitigate the potential effect of the occasional inaccurate recommendation. However, this needs to be done in a way that increases user trust and has a positive effect on recommendation accuracy.


Using recommendations to reduce churn

Churn rate, when applied to a customer base, refers to the proportion of contractual customers, or subscribers, who leave a supplier during a given time period. One way to minimize the churn rate is to create services that add value compared to competitors. This thesis will focus on a) identifying customers that have a high impact on others in the customer base, b) supplying these high-impact customers with added-value in terms of personalized content. This thesis will use network analysis methods to identify individuals who are highly connected and interact frequently with other users, but are at risk for churn. A particularly relevant group are users who are gate-keepers, highly connected within a group of users, but less so with the larger customer-base. Having identified such individuals, the thesis will use recommender systems algorithms to suggest content (information, and other people) that could be valuable to these people. This information will be taken from a subset of users that have some similarity (using e.g. content-based or collaborative filtering recommendation methods) to the target user, but are slightly different. This approach benefits both the user and the service provider. For the service provider, it helps minimize churn by strengthening connections between users in the user-base, by e.g. increasing the value of free calls within the same provider. For the users, it helps them discover novel but interesting content and connections. This in turn will not only decrease the risk of churn, but also addresses an issue often found in current recommender systems which is over-tailoring. Over-personalization has been shown to result in polarization of views, and in networks in particular it has been demonstrated that we have a tendency to self-filter. A test bed using publicly available (mined) Twitter data will be used to create simulations of this system. The system will be evaluated using both offline (performance of efficiency in simulations, coverage and diversity of recommendations), and online (with users in terms of perception and acceptance of recommendations).


Persuasive strategies for supporting collaborative learning

Advances in communications and information technology create new opportunities for organizations to build and manage virtual teams in academic settings. Team work is often also an educational process whereby team members learn from each other by sharing knowledge, building consensus, and jointly creating (or revising) mental models. Such teams are composed of collaborators with unique skills, who must collaborate to accomplish challenging tasks. While growth in the virtual collaboration sector is still growing rapidly, it has been noted that the uptake of collaboration services has reached a point where it is less to do with the ability of current technology, and more to do with the reluctance of people to collaborate in this way. For example, knowledge is often situated, and team members are either unaware of the resources available, or they do not understand their value in relation to a problem currently at hand. Even in co-located teams, team members may not be aware of shared resources and their value, underutilizing the resources they need in order to solve problems. This thesis will address this problem by investigating the benefit of recommendations in both co-located and distributed teams. That is, whether suggestions of relevant information, resources, or people, can improve the quality of communication in distributed teams. In addition, it will investigate different persuasive strategies for using the information suggested by recommendations.

 

 

Nava Tintarev