User-specific Models for Filtering Distracting Messages on Mobile Instant Messaging

Abstract

The Internet has contributed to a huge transition in lifestyle for most people worldwide. Also, Mobile Instant Messaging (MIM) has been one of the prevalent modes of peer-to-peer communication in the last decade (e.g. WhatsApp, Facebook Messenger). WhatsApp, by comparison, has been shown to lead to poor academic performance related to WhatsApp addiction. It follows that somehow the harms should be controlled to use MIM to our benefit. A machine learning model which shall be able to classify messages as distracting or non-distracting could help by keeping the messaging systems professional and thus enhancing productivity. Online learning has been used in this paper to gradually adapt a model. In this work, multinomial Naive Bayes, linear SVM, perceptron and passive-aggressive models, have been attempted in conjugation with stochastic gradient descent to obtain online models. All of them provide F1-scores better than 0.86, with Multinomial Naive Bayes providing the best F1-score of 0.89. This work paves the way for user-specific models which can be learnt on the fly using user-feedback for users of social networking and instant messaging apps.