Contact Information

  • Institute: University of Cergy-Pontoise
  • Address: Bureau 583, Bâtiment A, Site Saint Martin, 2 av. Adolphe Chauvin, Pontoise 95000 France
  • Email: aikaterini.tzompanaki [at] u-cergy [dot] fr

Title: Explaining Recommender Systems via Why-Not questions

Subject

With the abundance of information in the Big Data era, it is evident that internet users are likely to be unaware of what to search for, or what they would find interesting in various domains, from listening to music to retrieving health information. As a solution to this deficiency, recommender systems are built to aid users locate items that they are most probable to like. Especially in the e-commerce domain, recommendations are a key tool used to advertise products and increase sales. The two main categories of recommender systems are the content-based and the score-based categories. A content-based recommender [1] exploits similarities among the properties of the items. Score-based (also known as collaborative filtering) recommenders make recommendations based on the item scores given by similar - to the target user - users (also called peers)[2].

One important problem from the user perspective is to be able to assess the relevance of the recommendations and to understand why certain items are recommended, or why other items are not. While the literature is rich in approaches explaining recommended items (see [3] for a survey), this is not the case for missing items. A possible reason is that why questions are straight-forward to pose by a user and explanations are easier to provide [4]. Nevertheless, individual users of the system, or companies that sell their items on e-commerce websites, may also be surprised not to find certain items in the recommendations. If such cases cannot be well explained too, the user may lose their trust to the system, and companies may decide to stop using the related site to promote their products.

Explaining missing answers (expressed in Why-Not questions) is a problem addressed in various scientific fields, notably in the databases domain [5]. The answer to a Why-Not question, i.e, a Why-Not explanation, can either be found in the source data, in the transformation(s) used to derive the results (in our context the relevance function), or in both. Even though explanations to Why questions are also located in the same sources (deriving the provenance of the resulting item), explaining Why-Not questions is a more complex problem, with a wider search space of solutions (deriving the Why-Not provenance of missing results, which typically is not unique).

In this project, we aim to

  • 1) Analyze Why-Not questions on recommender systems. The recommendations provided by a system do not come with any knowledge of the underlying database content, or the set of its users. So, in this first objective, we want to identify all the possible and meaningful ways for asking a Why-Not question, for example: ‘Why not Titanic?’, or ‘Why not any drama movies?’. The outcome of this study will be a taxonomy based on which we will analyze possible and meaningful types of explanations, for example: ‘Titanic is not well rated by similar users’, or ‘There are no drama movies in the database’. This first study will serve as a guide on producing the Why-Not explainable recommender.
  • 2) Create a prototype of a Why-Not explainable recommender. Once the first step completed, we aim to design and implement a Why-Not explanations module on top of a recommender system, starting with collaborative filtering. To this end, we have to explore the different input variables for culprits, and design an algorithm to compute the most relevant – out of the possibly numerous- explanations to the final user. As a complementary feature, our prototype will provide suggestions for updating either the recommender (useful for improving the system) or the input data (useful for product redesign for the company-clients of the system).

Requirements and skills

The candidate should have solid knowledge of the Java or Python programming language. Familiarity with databases, recommender systems, or machine learning is desired.

Duration and Location

The internship has a duration of up to 6 months, starting after February 2020 (negotiable). The successful candidate will be a member of the ETIS lab, located in Cergy-Pontoise, only a half an hour RER (urban train) transport from Paris.

Interested candidates are requested to send a detailed CV (including projects), one recommendation letter and bachelor and master transcripts to Katerina Tzompanaki at atzompan@u-cergy.fr. We will accept complete applications until the 30th of January 2020, or until the position is filled.

Supervisors

Katerina Tzompanaki, Associate Professor, University of Cergy-Pontoise, France, Email: aikaterini.tzompanaki at u-cergy.fr
Kostas Stefanidis, Associate Professor, University of Tampere, Finland, Email: konstantinos.stefanidis at tuni.fi

References

  • [1] Lops, Pasquale, Marco De Gemmis, and Giovanni Semeraro. "Content-based recommender systems: State of the art and trends." In Recommender systems handbook, pp. 73-105. Springer, Boston, MA, 2011.
  • [2] Ben, Dan Frankowski, Jon Herlocker, and Shilad Sen. "Collaborative filtering recommender systems." In The adaptive web, pp. 291-324. Springer, Berlin, Heidelberg, 2007.
  • [3] Tintarev, Nava, and Judith Masthoff. "A survey of explanations in recommender systems." In 2007 IEEE 23rd international conference on data engineering workshop, pp. 801-810. IEEE, 2007.
  • [4] Lim, Brian Y., Anind K. Dey, and Daniel Avrahami. "Why and why not explanations improve the intelligibility of context-aware intelligent systems." In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pp. 2119-2128. ACM, 2009.
  • [5] Bidoit, Nicole, Melanie Herschel, and Aikaterini Tzompanaki. "Efficient computation of polynomial explanations of why-not questions." In Proceedings of the 24th ACM International on Conference on Information and Knowledge Management, pp. 713-722. ACM, 2015.
84,1 Bot