RevNet is a novel deep neural network architecture that utilises user reviews in a social recommender system, of which, its main goal is rating prediction.
Recommender systems have become increasingly popular in the online domain and play a critical role in suggesting information of interest to users. Various techniques have been explored while implementing these systems, such as collaborative filtering, content-based filtering, and preference-based solutions. In recent times, personal data has become ubiquitous and has prompted researchers to explore avenues that use user reviews on items of interest and social relationships to improve output recommendations. Furthermore, to the foregoing, the immense interest in Neural Networks has provided a platform for applying deep learning techniques to improve existing recommender system solutions. In this paper, we present a novel deep neural network framework (RevNet) for social recommendations through the utilization of user reviews. The model captures the user-user and user-item spaces typically found in social recommender systems, as well as adds another dimension in the form of a user-review space. In this regard, in the user-review space, we aim to capture the reviews provided by users on items. Thus, we propose to employ the use of doc2vec representations of user reviews on items which will be incorporated into a neural network for ratings prediction. Evaluation of the proposed solution on a real-world dataset, shows that the inclusion of user reviews in a social recommender system is effective
Recommender Systems have been around for quite some time. These systems provide personalised service support to users by learning their previous behaviours and predicting their current preferences [1]. Moreover, there has been a rapid increase in web technologies that employ recommendation systems to enhance a users' experience. The driving factor behind this is that it solves a major problem surrounding today's times, information overload. Recommender systems can be used to overcome this phenomenon by automating the decision-making process and providing a set of informed recommendations, thereby enhancing a user's experience [1]. The incorporation of social relations into recommender systems has been backed by social theories, where it is argued that people are influenced by their social connections, which ultimately leads to them having similar preferences [2] [3] [4].
As depicted in the figure below, a user is involved in both the item and social spaces of a social recommender system. The user-item space denotes the interactions between a user and items, whereas the user-user space captures the friendships amongst users.
Furthermore, the aim of our work is to utilise user reviews in a social recommender system. Thus, we propose to add a user-review space to learn item latent factors. As shown in the figure below, an item may have multiple reviews from many users.
The proposed model contains three main components i.e., user modelling, item modelling and rating prediction. Furthermore, we derive latent representations from the user modelling and item modelling components, via their respective spaces. User modelling seeks to understand users better, thus we facilitate this via the user-item and user-user spaces. Regarding user-item interactions, we propose appropriate embeddings to capture a user’s opinion on an item. Furthermore, as in a typical social recommender system, the social relationships are considered. This is done to model users from the social perspective. The outputs obtained from each of these spaces represent item user latent factors (hIi) and social user latent factors (hSi), respectively. We choose not to combine the obtained user latent factors into a single user latent factor representation. The second component of the proposed architecture is item modelling. This is done through the consideration of text reviews and user opinion on items. Thus, we receive a latent factor representation (hRi) for the items. Lastly, to facilitate rating prediction, we combine user and item modelling latent factor representations through concatenation, which is thereafter fed into subsequent neural network layers.
Python: 3.9Keras: 2.5.0Gensim: 4.0.1
- Argparse
- Matplotlib
- NLTK
- Numpy
- Pandas
- Pydot
- Sci-kit Learn
- Tensorflow
The Yelp dataset was used. In particular, we are interested in the user, business and reviews dataset files. Ensure that the dataset files are stored in the parent directory ../.
-
User Dataset File: https://drive.google.com/file/d/1P2EIB6fJFqSXL6R3NQ1HW6uO9tL_VCHO/view?usp=sharing
-
Business Dataset File: https://drive.google.com/file/d/1Ccg_4U4Md_bjUE8AtaWbzGOaldGZ0v0b/view?usp=sharing
-
Review Dataset File: https://drive.google.com/file/d/13SNDU5v3d3TDNluINVTeKDcP1VXQA3dm/view?usp=sharing
-
Alternatively, you can download the complete dataset here:https://www.yelp.com/dataset/download
The code is structured in a way that allows for model hyper parameter variation. To run the model with a default set of parameters, execute the following command into your machines terminal
python application.py
To run the model with custom hyper paramaters, the following format should be adhered to
python application.py --parameter1 <value>python application.py --epochs 50 --batch_size 128 --d2v_vec_size 256
- The training of the model may cause substantial memory usage in the case that you experiment with the
--usersand--reviewshyper paramaters. As such, ensure that you have in excess of 8GB of RAM when deviating from the default parameter values.
Refer to the paper within the Report directory
- [1] Q. Zhang, J. Lu, and Y. Jin, “Artificial intelligence in recommender systems,” Complex & Intelligent Systems, vol. 7, no. 1, pp. 439–457, Nov. 2020, doi: 10.1007/s40747-020-00212-w.
- [2] W. Fan, Y. Ma, D. Yin, J. Wang, J. Tang, and Q. Li, “Deep Social Collaborative Filtering,” doi: 10.1145/3298689.3347011.
- [3] Eytan Bakshy, Itamar Rosenn, Cameron Marlow, and Lada Adamic. 2012. The role of social networks in information diffusion. In World Wide Web conference. 519–528.
- [4] Wenqi Fan, Tyler Derr, Yao Ma, Jianping Wang, Jiliang Tang, and Qing Li. 2019. Deep Adversarial Social Recommendation. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, IJCAI-19.