Identifying Retweetable Tweets with a Personalized Global Classifier

TL;DR

This paper introduces a personalized global classifier for predicting which tweets users are likely to retweet, combining data-driven training with individual user preferences to improve relevance.

Contribution

The paper presents a novel personalized classification approach trained on large-scale data, effectively balancing generality and individual user customization for retweet prediction.

Findings

Achieved F1 score of approximately 0.9 with 10 features.

Effective personalization with a global classifier trained on 130K tweets.

Utilized diverse features including content, novelty, similarity, and influence.

Abstract

In this paper we present a method to identify tweets that a user may find interesting enough to retweet. The method is based on a global, but personalized classifier, which is trained on data from several users, represented in terms of user-specific features. Thus, the method is trained on a sufficient volume of data, while also being able to make personalized decisions, i.e., the same post received by two different users may lead to different classification decisions. Experimenting with a collection of approx.\ 130K tweets received by 122 journalists, we train a logistic regression classifier, using a wide variety of features: the content of each tweet, its novelty, its text similarity to tweets previously posted or retweeted by the recipient or sender of the tweet, the network influence of the author and sender, and their past interactions. Our system obtains F1 approx. 0.9 using only 10 features and 5K training instances.