Network activity feed: finding needles in a haystack

Social networks have evolved over the last decade into an omni-popular phenomenon that revolutionized both the online and offline interactions. They are used for a variety of purposes and are fast becoming the place to share and discover news, activities, and content of interest. Facebook alone reports more than 1 billion users, each having on average 130 friends and connected to 80 communities, and spending on Facebook less than one hour a day. The volume of the generated content of potential interest is, thus, overwhelming and ever growing, but the time spent on the social networks is fairly limited. How can users stay abreast of the activities of interest given this severe information overload? Activity feed is a simple mechanism deployed nowadays by many social networks, which performs information filtering on the users' behalf. Typically, activity feed encompasses reverse chronologically ordered items corresponding to activities carried out by direct friends and followees. However, activity feed can hardly cope with the volume and diversity of the activities. In order to alleviate information overload, simplify content discovery, and sustain user engagement, there is a need to personalise the activity feed, i.e., identify items of a particular interest and relevance for the user and filter out irrelevant items. The feed personalisation task can be naturally represented as a top-K recommendation problem. Let us denote by N the set of items corresponding to activities that can potentially be included in the feed, e.g., all the activities carried out since the user's last visit. Hence, the personalisation task aims at selecting and recommending a smaller set of items, K (|K|≪|N|), corresponding to activities of the highest relevance for the user. Essentially, the recommendation process entails scoring all the |N| candidate items and selecting |K| top-scoring items. What information can facilitate the item scoring? When interacting with a social network, users typically leave very little explicit feedback, primarily their 'likes'. There is a moderate amount of strong implicit user-to-user feedback, e.g., friending and direct communication (messages and comments), and abundance of weak implicit user-to-activity feedback, such as content viewing and contribution, community membership, and event participation. Finally, there is some self-reported and often unreliable information pertaining to user demographics, location, preferences, skills, or interests. How can all this this information be modelled, fused, mined, and eventually leveraged for scoring and recommending activity feed items? This problem has been investigated from different angles in the recent years [1-10]. In this talk, we will overview most prominent works into the personalisation of the activity feed. These works proposed a spectrum of algorithmic approaches and evaluated them with numerous social networks of a highly heterogeneous nature. We will summarise the main components underpinning these approaches, overview the obtained findings, discuss their advantages and shortcomings, survey their combinations, analyse evaluation metrics and methodologies, and, finally, identify gaps calling for further research.