CSRS

Abstract

The effectiveness of recommendation systems is improving with the incorporation of richer context. The frequentist recommendation methods such as Markov models are not efficient in simultaneous use of context and preference sequences over items due to state space explosion. On the other end, abstractionist models such as Matrix Factorization where each item or user is represented as a set of abstract features are difficult to explain. An example in this case is, recommending Web Based Trainings (WBTs) to employees, similar to Massively Open Online Courses (MOOCs), wherein use of the sequence information in the recommendation of WBTs would help the user to gradually build expertise in their area of interest. For training recommendation, it is important to identify the held expertise level in technical area, that represents a state and possible sequences of trainings that represent transitions in terms of real world entities such as trainings and associated features. Alternatively the model can estimate expertise as a mixture over a tractable set of latent interests in terms of trainings completed, contextual features such as the training sequences, keywords and user profile. To the best of our knowledge, the state-of-the-art recommendation methods do not consider both explicit context and sequence information in a single model. In this paper, we propose a Context and Sequence Aware Recommendation System (CSRS) based on latent topic modelling framework, identifying topic-memberships for items, contextual features as well as for user interests. We demonstrate benefits of incorporating both context and sequence of items for recommendation on three real world datasets.