The Bayesian Mallows for Clicking Data (BMCD) (Liu et al., 2019) is a probabilistic recommender system. The research so far is based on assuming that when a user clicks on an item, she makes that choice among all other items, and hence prefers by the clicked item to all other items. This is mainly due to available data being limited to what a user clicks on and when. A more realistic scenario is that a user chooses to click on one item out of a ranked list of items presented to them, called a slate, which results from e.g. a search query or a recommendation system. Recently, the FINN.no Slates Dataset (Eide et al., 2021) was published, containing information on what slates each user has been presented (either by search queries or recommendation), and on what the user chose to click on (if any) for each slate. This make it possible to use more realistic modeling assumptions in the recommender system.
This project will adapt the BMCD to slate data, and analyse the performance by applying it to the FINN.no Slates Data.
The dataset also contains attributes on the items. The current BMCD does not incorporate features for items or users, and modeling and implementing this could also be interesting.
References
-
Liu, Q., Reiner, A. H., Frigessi, A., & Scheel, I. (2019). Diverse personalized recommendations with uncertainty from implicit preference data with the Bayesian Mallows model. Knowledge-Based Systems, 186, 104960. DOI: 10.1016/j.knosys.2019.104960
-
Eide, S., Leslie, D. S., Frigessi, A., Rishaug, J., Jenssen, H., & Verrewaere, S. (2021, September). FINN. no Slates Dataset: A new Sequential Dataset Logging Interactions, all Viewed Items and Click Responses/No-Click for Recommender Systems Research. In Fifteenth ACM Conference on Recommender Systems (pp. 556-558). DOI: https://doi-org.ezproxy.universitetetioslo.no/10.1145/3460231.3474607