Should I Stay or Should I Go? Improving Event Recommendation in the Social Web

TL;DR

This study investigates whether incorporating social and contextual features like reachability, reputation, and friends' participation improves event recommendation accuracy on the Social Web, achieving a 4.1% error reduction.

Contribution

It demonstrates that adding social and contextual features to event recommendation models enhances prediction accuracy, with optimal results from a specific feature combination.

Findings

Additional factors improve prediction accuracy by 4.1%.

Combining content and social features yields best results.

Linear regression effectively models user interest based on these features.

Abstract

This paper focuses on the recommendation of events in the Social Web, and addresses the problem of finding if, and to which extent, certain features, which are peculiar to events, are relevant in predicting the users' interests and should thereby be taken into account in recommendation. We consider in particular three "additional" features that are usually shown to users within social networking environments: reachability from the user location, the reputation of the event in the community, and the participation of the user's friends. Our study is aimed at evaluating whether adding this information to the description of the event type and topic, and including in the user profile the information on the relevance of these factors, can improve our capability to predict the user's interest. We approached the problem by carrying out two surveys with users, who were asked to express %with a score their interest in a number of events. We then trained, by means of linear regression, a scoring function defined as a linear combination of the different factors, whose goal was to predict the user scores. We repeated this experiment under different hypotheses on the additional factors, in order to assess their relevance by comparing the predictive capabilities of the resulting functions. The compared results of our experiments show that additional factors, if properly weighted, can improve the prediction accuracy with an error reduction of 4.1%. The best results were obtained by combining content-based factors and additional factors in a proportion of approximately 10:4.