Modeling impression discounting in large-scale recommender systems

Abstract

Recommender systems have become very important for many online activities, such as watching movies, shopping for products, and connecting with friends on social networks. User behavioral analysis and user feedback (both explicit and implicit) modeling are crucial for the improvement of any online recommender system. Widely adopted recommender systems at LinkedIn such as "People You May Know" and "Endorsements" are evolving by analyzing user behaviors on impressed recommendation items.In this paper, we address modeling impression discounting of recommended items, that is, how to model user's no-action feedback on impressed recommended items. The main contributions of this paper include (1) large-scale analysis of impression data from LinkedIn and KDD Cup; (2) novel anti-noise regression techniques, and its application to learn four different impression discounting functions including linear decay, inverse decay, exponential decay, and quadratic decay; (3) applying these impression discounting functions to LinkedIn's "People You May Know" and "Endorsements" recommender systems.

User Fatigue in Online News Recommendation

“…However, the main focus of this piece of research work was on how to estimate CTR, which is quite different with the theme of our paper. In a recent study detailed in [12], Lee et al followed the work in [2] by experimenting with the LinkedIn person recommendation data. Instead of estimating CTR, the authors proposed to optimize for a different metric, i.e., Conversion Rate, by taking into consideration of the impression discounting factor.…”

ContentWise Impressions

“…Impression datasets have been used by several articles [1,2,7,13,14,20]. They can be classified into two categories: private datasets, collected by the authors of the article but, to the best of our knowledge, not made accessible to the community, and non-redistributable datasets, made accessible only to the participants of a challenge under a non-redistribute clause.…”

“…Examples of private datasets are LinkedIn PYMK Impressions and LinkedIn Skill Endorsement Impressions. Both were used to model impressions discounting on large-scale Recommender System (RS) [14] and contain users registered on the LinkedIn 5 platform. More specifically, LinkedIn PYMK Impressions was used to recommend possible new user connections, and Linkedin Skill Endorsement was used to recommend skill endorsement of known users.…”

“…More specifically, LinkedIn PYMK Impressions was used to recommend possible new user connections, and Linkedin Skill Endorsement was used to recommend skill endorsement of known users. Impressions in these datasets were present as a list of users, and a list of user-skill tuples, respectively [14].…”

“…A recent emerging trend is to provide the impressions, i.e., what was recommended to the user alongside the user interactions. Recent articles, also from industry, propose algorithms that leverage impressions showing they can improve the recommendation quality [7,14,20].…”