Description
Date depot: 20 février 2020
Titre: Temporal Prediction and Control of User Behaviour in a Social Platform
Encadrant :
Nicolas BASKIOTIS (ISIR (EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
Social media and the way information is circulated through them are now more complex than ever. The monetization of user online activity by social media providers with the help of multivariable recommendation algorithms has resulted in a complicated online environment that constantly evolves, in a way that cannot be easily predicted or controlled. In an effort to better understand how social media work, the following questions are raised:
Q1: What patterns can we detect in the user activity (posting & interaction), in terms of time, frequency and subject?
Q2: How can we accurately predict these patterns?
Q3: How can we take advantage of these patterns to measure the circumstances under which a post is more likely to circulate?
Q4: How much does the Newsfeed & posting behavior of other users (e.g. their friends) affect their own profile and the likelihood to interact with a post they see on their Newsfeed?
Q5: How can we incorporate all the above in a compact Newsfeed recommendation system to maximize a given goal (such as post engagement/interaction)?
As an answer to these questions we plan to explore the following:
(A) Time-series analysis of user profiles in social networks. A user’s profile can be analyzed in two parts: (i) their posting behavior and how it evolves over time, both in frequency and subject. Study the user posting patterns (e.g. which time during the day) also in relation to the state of their Newsfeed. Make predictions of these patterns (ARIMA models), (ii) their Newsfeed interaction behavior (e.g. when do they interact with a post, which subjects, source, etc.). Explore how posting behavior relates to user interaction.
(B) Control of user posting behavior. Explore which parameters influence post visibility the most. Find optimal opportunities to post that can maximize post visibility and diffusion, given also the posting and interaction behaviour of others.
(C) Newsfeed control. Suggest Newsfeed recommendation algorithms to present a post to the most suitable user profile so as to maximize a given objective. This objective can be user engagement in the platform (time spent, commenting, liking), click-through probability on links (for ads), etc. The recommendation policy will depend on two things: (i) the user interests and built profile (e.g. by the time-series analysis in part (A)), (ii) the set of posts produced between two consecutive user visits. We will use tools form Reinforcement Learning and Recommender Systems.
Doctorant.e: Papanastasiou Effrosyni