Recommendation systems are a subclass of information filtering system that seeks to predict the "preference" a user would give to an item.They are primarily used in commercial applications.In this paper,we approach the recommendation systems using the Transformers - firstly, we implement a Behaviour Sequential Transformer model which translates the user behavior into sequences and predicts a rating for each target item. For our second approach we leveraged the Matrix Factorization and enhanced it using BERT embeddings from the items, to create a hybrid approach by combining Content-based filtering with Collaborative filtering approach. Further, we built a K-Nearest Neighbor model to capture the change in item embeddings and used the newly generated embed
Datasets Our experimental evaluation utilizes two diverse datasets that capture user-item interactions across different domains, enabling comprehensive testing of our approaches on varying feature spaces and dimensionalities:
- Food.com Recipes Dataset
Scale: 230,186 unique recipes with 1,125,284 user-recipe interactions Features: Rich metadata including recipe tags, ingredient lists, and detailed descriptions Domain: Culinary recommendations with complex feature relationships Interaction Type: User ratings capturing preference signals
- Restaurant Data with Consumer Ratings
Scale: 1,161 user-restaurant interactions Features: Multi-dimensional attributes including cuisine types, payment methods, parking availability, and amenity information Domain: Restaurant recommendations with categorical and binary features Interaction Type: Consumer ratings reflecting dining experiences
The selection of these datasets enables evaluation across both large-scale (Food.com) and smaller, feature-rich (Restaurant) scenarios, providing insights into model performance under different data characteristics and sparsity conditions. Key Contributions This work advances recommendation system research through:
Novel application of Transformer architectures for sequential behavior modeling in recommendations Hybrid approach combining deep contextual embeddings (BERT) with traditional collaborative filtering Dynamic embedding adaptation using KNN for capturing temporal changes in item representations Comprehensive evaluation across domains with varying scale and feature complexity