| Summary: | 碩士 === 國立臺灣大學 === 資訊工程學研究所 === 107 === In recent years, collaborative filtering methods based on matrix factorization techniques have achieved great success in recommender systems, while cold-start and data sparsity have not solved well. Item reviews written by users have a large amount of information, hence it has been a significant trend to involve review information into rating prediction as one of the dominant solutions. In recent research, it has been shown that deep learning methods can improve the performance of rating prediction over traditional methods when review text is available.
A model named Deep Cooperative Neural Networks (DeepCoNN) has been pro-posed, which aims to build user and item latent representations using review text. It consists of two parallel convolution neural networks, where one neural network learns user latent features using the reviews written by the user, and the other neu-ral network learns item latent features using the reviews for the item. Finally, a shared layer combines user representations and item representations as the input of factorization machines.
In this thesis, I first implement the DeepCoNN model by tensorflow tools. Then I propose an idea that I extend DeepCoNN by changing its shared layer from factor-ization machines to a neural prediction layer. The neural prediction layer is based on neural collaborative filtering techniques. Hence the extended model I proposed combines deep learning and collaborative filtering.
In the experiments, the extended DeepCoNN model outperforms the original DeepCoNN model in Amazon public datasets.
|
|---|
