Anzeige
Erschienen in:
2014 | OriginalPaper | Buchkapitel
Deep Learning of Semantic Word Representations to Implement a Content-Based Recommender for the RecSys Challenge’14
verfasst von : Omar U. Florez
Erschienen in: Semantic Web Evaluation Challenge
Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by
Abstract
In this paper, we will discuss a recommender system that exploits the semantics regularities captured by a Recurrent Neural Network (RNN) in text documents. Many information retrieval systems treat words as binary vectors under the classic bag-of-words model; however there is not a notion of semantic similarity between words when describing a document in the resulting feature space. Recent advances in neural networks have shown that continuous word vectors can be learned as a probability distribution over the words of a document [3, 4]. Surprisingly, researchers have found that algebraic operations on this new representation captures semantic regularities in language [5]. For example, \(Intel + Pentium - Google\) results in word vectors associated to \(\{Search, Android, Phones\}\).
We used this deep learning approach to discover the continuous features describing content of documents with vectors of semantic words and fitted a linear regression model to approximate user preferences for documents. Our submission to the RecSys Challenge’14 obtained a RMSE of \(0.902\) and ranked 6th for Task 1. Interestingly enough, our approach provided better vector representations than LDA, LSA, and PCA for modeling the content of book abstracts, which are well-known techniques currently used to implement content-based recommender systems in the recommendation community.