CNN with depthwise separable convolutions and combined kernels for rating prediction

• Published in: Expert systems with applications 2021-05, Vol.170, p.114528, Article 114528
• Main Authors: Khan, Zahid Younas, Niu, Zhendong
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Convolutional neural networks; Deep learning; Depthwise separable convolutions; E-learning; Feature extraction; Kernels; Machine learning; Product reviews; Rating prediction; Reviews
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2020.114528

•A deep learning model using CNN with depthwise separable convolutions on review text.•These convolutions with multiple parallel kernels better infer the hidden features.•The features obtained in this way give better user/item repersentation;context-aware.•Experimental results validates the claim and outperfroms the baseline alternatives. Recently, deep learning based techniques exploiting reviews are extensively studied for rating prediction and result in good performance. Some studies consider word level review information along with attention mechanism to capture the most influential content, thus making the methods even complex. Deep neural networks with Depthwise Separable Convolutions have made significant progress in the area of image and video analysis. The success of these techniques encourages adopting them towards improvements in rating prediction using review text. In this paper, we present a novel CNN based architecture with Depthwise Separable Convolutions and Combined Kernels (CNN-DSCK) for rating prediction exploiting product reviews. In the proposed method, we use two parallel CNNs with Depthwise Separable Convolutions to extract semantic features from the text reviews of users and items using different kernels in parallel and then select the important information from these features through pooling. Finally concatenate the pooling information obtained from different kernels in each network. The features obtained through each network are then fused and the most relevant higher-order features are extracted through fully connected dense layer at the top of network. Extensive experiments on real-world datasets demonstrate that CNN-DSCK significantly outperforms state of the art baseline models.