Sentiment Analysis for Cold Start Items
Aspect-based opinion mining techniques try to extract aspects (e.g., zoom, battery life, etc. for a digital camera) and estimate their ratings (as stars) from customer reviews. Current models are usually trained at the item level which is fine when the item has been reviewed extensively. However, in real-life datasets more than 90% of items have less than 10 reviews, so called cold start items. In this presentation, we introduce a novel model based on Latent Dirichlet Allocation to address the cold start problem. Our experiments on three real-life datasets demonstrate the improved effectiveness of the proposed model.