Knowledge Recommendation Based on Item Response Theory

In Knowledge Management System (KMS), the users have to spend much time searching for knowledge items because of the overabundance of information. For improving users' satisfaction, several approaches have been proposed to recommend knowledge by using user feedback, especially collaborative filtering algorithm. But the user feedback easily suffers from noise which affects the accuracy of recommendation. In this paper, we propose a new knowledge recommendation method that aims at extracting latent trait from user feedback and substituting latent trait to user feedback in recommendation model. First, motivated by the Item Response Theory (IRT), we view each feedback as a user's detailed response to an item, and assume the response is jointly determined by the ability of users and the difficulty of items. Second, we suppose that the user feedback obeys Gaussian distribution, in which the latent trait represents the users' ability of comprehending the knowledge items and is jointly determined by user feedback and parameters of model. The parameter mean of Gaussian distribution denotes the difficulty of items, and its parameter variance denotes the ability of users. Last, we recommend knowledge item by Matrix Factorization (MF) method, in which we minimize the squared error between the latent trait and its predictive value. The predictive value is the inner product value of the two latent matrix factorized from the known matrix. Extensive experiments on call center dataset show the effectiveness of the proposed solution by comparing the state-of-the-art methods of MF.