Cold Start Thread Recommendation as Extreme Multi-label Classification

Kishaloy Halder, Lahari Poddar, Min-Yen Kan

Abstract: In public online discussion forums, the large user base and frequent posts can create challenges for recommending threads to users. Importantly, traditional recommender systems, based on collaborative filtering, are not capable of handlingnever-seen-before items (threads). We can view this task as a form of Extreme Multi-label Classification (XMLC), where for a newly-posted thread, we predict the set of users (labels) who will want to respond to it. Selecting a subset of users from the set of all users in the community poses significant challenges due to scalability, and sparsity. We propose a neural network architecture to solve thisnew thread recommendation task. Our architecture uses stacked bi-directional Gated Recurrent Units (GRU) for text encoding along with cluster sensitive attention for exploiting correlations among the large label space. Experimental evaluation with four datasets from different domains show that our model outperforms both the state-of-the-art recommendation systems as well as other XMLC approaches for this task in terms of MRR, Recall, and NDCG.