Bandit algorithms for real-time data capture on large social medias

TL;DR

This paper develops and compares several bandit algorithms for real-time social media data collection, focusing on selecting relevant user accounts under resource constraints to maximize information quality.

Contribution

It introduces multiple bandit models, including contextual and non-stationary variants, for dynamic user selection in social media data capture.

Findings

Models outperform baseline methods on artificial datasets.

Contextual bandits improve relevance of captured data.

Latent space models capture complex user interactions.

Abstract

We study the problem of real time data capture on social media. Due to the different limitations imposed by those media, but also to the very large amount of information, it is impossible to collect all the data produced by social networks such as Twitter. Therefore, to be able to gather enough relevant information related to a predefined need, it is necessary to focus on a subset of the information sources. In this work, we focus on user-centered data capture and consider each account of a social network as a source that can be listened to at each iteration of a data capture process, in order to collect the corresponding produced contents. This process, whose aim is to maximize the quality of the information gathered, is constrained by the number of users that can be monitored simultaneously. The problem of selecting a subset of accounts to listen to over time is a sequential decision problem under constraints, which we formalize as a bandit problem with multiple selections. Therefore, we propose several bandit models to identify the most relevant users in real time. First, we study of the case of the stochastic bandit, in which each user corresponds to a stationary distribution. Then, we introduce two contextual bandit models, one stationary and the other non stationary, in which the utility of each user can be estimated by assuming some underlying structure in the reward space. The first approach introduces the notion of profile, which corresponds to the average behavior of a user. The second approach takes into account the activity of a user in order to predict his future behavior. Finally, we are interested in models that are able to tackle complex temporal dependencies between users, with the use of a latent space within which the information transits from one iteration to the other. Each of the proposed approaches is validated on both artificial and real datasets.