[PhD proposal 2020] Modelling dynamics in graphs: application to social recommendation
Where: KIWI team, LORIA lab (wwww.loria.fr), Université de Lorraine
Supervisors: Geoffray Bonnin & Armelle Brun (firstname.lastname@example.org, email@example.com)
Starting: October 2020
Recommender systems (RS) is a highly active domain, that aims at providing personalized recommendations to users, within a large set of items.
Recommendations are varied : movies, music, cultural items, e-commerce, pedagogical resources, even users. The goal of the recommendations is also diverse: short, mid or long term and the recipient of the recommendations may be alone or a group.
Challenges are manifold: algorithms complexity, neutrality, explainability, inclusivity, the adoption of the recommendations by the users, their impact, etc.
Collaborative Filtering remains the most popular approach in RS (Jannach et al., 2012). It assumes that users who shared preferences of behaviours in the past will also have similar behaviour or preferences in the future. So, Collaborative Filtering recommends a user some items that have been appreciated in the past by similar users.
However, users are not independent, their preferences or behaviour may be influenced by the others (including by the community they belong to, whether it is explicit or implicit). This influence has been modelled by using propagation mechanisms, by relying on trust between users [Jamali et Ester, 2010],similarity of preferences [Wang et al., 2018], etc. Most of these approaches rely on the representation of data with graphs.
Problem and goal
Most of the recommender systems consider the recommendation task as a static and local one.
Local?Given an active user, the goal is to recommend him the best items, i.e. those that will satisfy him. However, this recommendation may also have an impact on this user’s community, or on the entire set of users (due to the influence between users). The traditional point of view is thus local.
Static?Given this active user, the goal is also to satisfy him at the moment where the recommendation is made, but the impact of this recommendation through time is not considered. The recommendation is static.
The problem of interest in this PhD thesis is thus two-fold. First, it aims at considering recommendation as a dynamic problem, especially from the temporal point of view and model the impact of a recommendation through time. Second, it aims at considering recommendation as a global problem and model the impact of a recommendation on a community of users (or the entire set of users).
Given both problems, it will be possible to focus on the recommendation task with constraints: time, number of recommendations, etc. For example, minimising the number of recommendations, while guaranteeing that the goal is reached, or planning to reach a goal in a limited time.
These problems require to totally rebuild of recommendation algorithms.
We can consider that data is represented under the form of a graph, a social graph, probably temporal, and that recommendation can be either recommendation of items, or a recommendation of links (other users).
The scientific challenge is the following: given a temporal social graph at a time t, given the goal (a graph) that we aim at reaching at the latest at time t+n(this graph is probably underspecified), what recommendations should be made, to reach this goal? The model that will be proposed will consider time constraints, number of recommendations, propagation in the graph, etc.
To answer this question, the work will inspire from several domains: temporal graph modelling, graph mining, propagation in graphs, link prediction, recommender systems, etc.
[Jamali et Ester, 2010] Jamali, M., & Ester, M. (2010, September). A matrix factorization technique with trust propagation for recommendation in social networks. In Proceedings of the fourth ACM conference on Recommender systems(pp. 135-142).
[Jannach et al., 2012] D. Jannach, M. Zanker, M. Ge et M. Gröning, «Recommender systems in computer science and information systems – a landscape of research,» chez Proc. EC-WEB, 2012
[Wang et al., 2018] Wang, H., Zhang, F., Wang, J., Zhao, M., Li, W., Xie, X., & Guo, M. (2018, October). Ripplenet: Propagating user preferences on the knowledge graph for recommender systems. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management(pp. 417-426).
[Lee et Lee, 2015] Lee, K., & Lee, K. (2015). Escaping your comfort zone: A graph-based recommender system for finding novel recommendations among relevant items. Expert Systems with Applications, 42(10), 4851-4858.