Standard Collaborative Filtering (CF) algorithms make use of interactionsbetween users and items in the form of implicit or explicit ratings alone forgenerating recommendations. Similarity among users or items is calculatedpurely based on rating overlap in this case,without considering explicitproperties of users or items involved, limiting their applicability in domainswith very sparse rating spaces. In many domains such as movies, news orelectronic commerce recommenders, considerable contextual data in text formdescribing item properties is available along with the rating data, which couldbe utilized to improve recommendation quality.In this paper, we propose a novelapproach to improve standard CF based recommenders by utilizing latentDirichlet allocation (LDA) to learn latent properties of items, expressed interms of topic proportions, derived from their textual description. We inferuser's topic preferences or persona in the same latent space,based on herhistorical ratings. While computing similarity between users, we make use of acombined similarity measure involving rating overlap as well as similarity inthe latent topic space. This approach alleviates sparsity problem as it allowscalculation of similarity between users even if they have not rated any itemsin common. Our experiments on multiple public datasets indicate that theproposed hybrid approach significantly outperforms standard user Based and itemBased CF recommenders in terms of classification accuracy metrics such asprecision, recall and f-measure.
展开▼
