WPNet: Wallpaper Recommendation with Deep Convolutional Neural Networks.

The recommendation quality of new users plays an increasingly important role in recommender systems. Collaborative Filtering cannot handle the cold-start problem, while the content-based approach sometimes can achieve recommendation with new items. To recommend in the wallpaper field, this paper proposes a content-based recommender system and extracts the features of wallpaper via the deep learning approach. The first part of the recommendation model is the convolution layers, and the model takes the output of full connection layer as features to employ. In order to improve the scalability, the model adopts deep neural network as non-linear dimension reduction method to reduce the image features. Taking the recommended results into account, this paper compares the feature similarities of user images and those in the image library. Finally, the model sorts them via cosine similarity, and presents the recommendation results using Top-K list. In the experiment, our model is trained with selected wallpapers on MIRFLICKR dataset, and uses VGG on ImageNet for feature extraction. The experimental results indicate that WPNet will have higher hit rates with different K if the image division of some wallpapers can be improved, and achieve a better performance in less time under the recommendations of new items.