Ranking social media news feeds: A comparative study of Personalized and Non-Personalized prediction models

Abstract

Ranking news feed updates by relevance has been proposed to help social media users catch up with the content they may nd inter- esting. For this matter, a single non-personalized model has been used to predict the relevance for all users. However, as user interests and pref- erences are di erent, we believe that using a personalized model for each user is crucial to re ne the ranking. In this work, to predict the relevance of news feed updates and improve user experience, we use the random forest algorithm to train and introduce a personalized prediction model for each user. Then, we compare personalized and non-personalized mod- els according to six criteria: (1) the overall prediction performance; (2) the amount of data in the training set; (3) the cold-start problem; (4) the incorporation of user preferences over time; (5) the model ne-tuning; and (6) the personalization of feature importance for users. Experimen- tal results on Twitter show that a single non-personalized model for all users is easy to manage and ne-tune, is less likely to over t, and it ad- dresses the problem of cold-start and inactive users. On the other hand, the personalized models we introduce allow personalized feature impor- tance, take into consideration the preferences of each user, and allow to track changes in user preferences over time. Furthermore, personalized models give a higher prediction accuracy than non-personalized models.