ERN: Emerging Researchers National Conference in STEM

Recommending Composite Items with Graph Neural Networks

Undergraduate #110
Discipline: Data Science
Subcategory: Computer Science & Information Systems
Session: 3
Room: Exhibit Hall

Naomi C Rohrbaugh - University of Pittsburgh
Co-Author(s): Edgar Ceh-Varela, Eastern New Mexico University, Portales NM


Recommender systems (RS) help users to deal with the problem of information overload. These systems suggest items to users based on their preferences. In some real-life situations, the user’s preference for a particular item could be influenced by the entity (e.g., restaurant) that sells or produces the item. For example, an item could be perfectly suited to a user’s preferences but is produced at a venue with poor service or lousy ambiance, aspects that are not to the user’s liking. On the other hand, a venue might have the perfect vibe for a user, but the items produced at that venue do not match the user’s preferences. The recommendation of an item would be improved by considering the user’s preferences for both item and venue. Therefore, the user’s preference for the entity’s characteristics also needs to be considered. This research presents a model for composite recommendations (i.e., recommending items with a “Has-A’ relationship”). The model will account for a user’s preference for both of the related items. As a basis, we use metapath2vec, a graph neural network method, to obtain node embeddings from a heterogeneous graph. This representation learning method maintains the structural relationship of a network with nodes of different types (i.e., user and two different items). The node embeddings for entities with a composite relationship are aggregated during recommendation to account for the user’s preference of these entities. Our proposed model was tested with a real-world dataset collected from BeerAdvocate.com composed of users, breweries, and beers from the state of California, USA. The dataset includes user-beer and user-brewery interactions (i.e., reviews), as well as beer-brewery interactions. The composite recommender system recommends a tuple to each user. To make this recommendation, we form a heterogeneous graph with users, beers, and breweries as heterogeneous nodes and user-beer, user-brewery, and beer-brewery as heterogeneous edges. For each of these nodes, we obtain their representation. To account for a user’s brewery preferences, we use a dot product between the user and a matrix of brewery embedding vectors, calculating the similarity between the user and breweries. The most similar brewery vectors are aggregated with the user’s vector, effectively combining their embeddings. We then use this new aggregated vector to find appropriate beers for the user with that same generalization of matrix factorization method involving a dot product. This method simultaneously considers a user’s preference for both item and vendor to provide the user with the best possible recommendation set. A baseline model that only considers a user’s preference for beers was used for comparison. We use as a metric the Hit Ratio to analyze the performance of our model’s recommendation and compare it to the baseline. The results show that our proposed approach obtains better results than a baseline model, which does not consider the composite relationship between two items. The use of feature aggregation to account for a user’s vendor preferences in selecting an item improves the recommendation by considering all aspects of the recommended set. The composite recommender system’s significant out-performance of the baseline model provides support for future use of feature aggregation in recommender systems.

Funder Acknowledgement(s): This research was part of an REU program funded by the NSF.

Faculty Advisor: Edgar Ceh-Varela, Eduardo.Ceh@enmu.edu

Role: I began by extracting the beeradvocate dataset using various web-scraping techniques. My research mentor directed me to the metapath2vec method as I was searching for a way to do representation learning with heterogeneous GNNs. I manipulated my dataset into the correct form to use with the metapath2vec algorithm. I got node representations and used them to create a composite recommender system as described in the abstract. My mentor directed me to some good ways to do a feature aggregation, and after trying some out I found a suitable way to implement the composite recommender system. I also implemented a baseline method for comparison. I then tested the system with different parameters to maximize performance. I compared results to the baseline to make my final conclusions about the efficacy of the method. In this project, my mentor directed me to some ideas and helped out when I ran into difficulties. I handled the implementation of the project and made data specific decisions.