Tag recommendation model using feature learning via word embedding

Abstract

Tag recommendation models serve as extracting metadata for target objects like images, videos and Web pages. However, these models tackle cold start problem due to absence of initial tags. To improve tag quality in tag recommendation services, most of previous works exploit the statistical properties such as co-occurrence patterns or term frequency to predict the candidate tags to a target object. Yet, these tag recommendation methods fail to be effective when initial tags are absent or low quality texts are available for objects. Recently, sentence modeling via word embeddings achieves successes in many natural language processing tasks. Therefore, this paper aims to introduce a novel tag recommendation algorithm that can analyze the relation between words in a text associated with target object using word embedding. In fact, we involve grammatical relations between words in a text or sentence with focus on feature learning methods. Skip-gram model is used to optimize feature values and learn the representation vector of words for tag recommendation. Our method shows improvements to previous research methods with gains of up to 10 percent in precision using real data from Movielens dataset.