Combining Extreme Multi-label Classification and Principal Label Space Transformation for Cold Start Thread Recommendation

Abstract

The recommendation system has been widely used in various areas, e.g., entertainment, education, and travel. However, this technique faces two main challenges which are Cold Start and High-Dimensionality problems. The cold start happens when the system does not have enough profile of new users; therefore, the system cannot recommend products to them. The second issue comes from the fact that there are a lot of distinct products or users to be recommended. Recently, Extreme Multi-label Classification (XMLC) has been applied to the recommendation system and addressed the Cold Start issue. However, the previous method still has a high-dimensionality issue. In this paper, we proposed a new approach, namely XMLC-PAO, which integrated label space reduction with XMLC. In more details, we transformed the recommendation problem to XMLC and applied Singular Value Decomposition (SVD) to generate reducing operator of label space (products’ or users’ label space). For the feature space, Deep Learning technique has been used to extract features from texts. From the experiments with Stackoverflow online forums dataset, we have found that the XMLC-PAO showed better performance in terms of RECALL@M and NDCG@M when the dimensions were reduced to 50% and 80% of the original size.