Enhanced Recommender Systems with Deep Neural Networks
- Publication Type:
- Thesis
- Issue Date:
- 2022
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
The main contributions of this thesis are as follows:
To model the conception of fashion and visual factors on the fashion recommendation task, we propose a cross-domain recommendation method based on visual collocation knowledge transfer. First, we extract visual collocation knowledge of fashion items from images on a popular fashion website and even street photography and incorporate the learnt knowledge to the recommender system through transfer learning. By collecting cross-domain information and updating visual collocation knowledge, the accuracy of clothing recommendation is improved.
To overcome the difficulty of accurately extracting latent features of new users and non-popular products, we propose a recommendation method based on deep graph convolutional neural network. We learn the deep representation of users and items from the high-order similarity between users and items. It inherits the advantages of graph convolutional neural network to quickly combine local information on the graph, so that we can obtain the node embedding which consists of the node's information, neighbors' information, and local structure information. At the same time, we propose an information propagation method based on the attention mechanism, which can effectively alleviate the over-smoothing problem when the graph convolutional neural network is too deep.
To solve the problem that the user's preference is affected by the environment and changes with time, we propose a recommendation method based on the user's long-term and short-term preference. In one session, the products browsed by the user have a certain continuity. This method models the user's current shopping intention through the items that the user has browsed in the current session. At the same time, the method also combines the user's long-term stable preferences contained in the user's historical records to provide users with in-time recommendations. The method can quickly adapt to the changes of the user's current interests caused by changes of the context and improve users' stickiness to shopping websites.
To solve the problem of data sparsity, we propose a recommendation method based on generative adversarial strategy. The algorithm generates a user's latent feature vector by training a generator network with a denoising autoencoder, which generates recommendations for the user accordingly, while training a discriminant network to distinguish the recommendation prediction generated by the generating network from the user's real transaction records. The adversarial training between the discriminating network and the generating network helps to push recommendation predictions closer to the real transaction records.
Please use this identifier to cite or link to this item: