RecKross: A Novel Recommender System with k-Cross Kernel Net

Authors: Abhishrut Vaidya, Niladri Chatterjee[1]

Summary

This paper introduces RecKross, a novel collaborative filtering model designed for personalized recommendations, particularly in data-constrained environments[1]. It conceptualizes the recommendation task as a matrix completion problem, proposing a new architecture that combines a 2D Kernel layer for multi-dimensional feature extraction and a k-Cross Kernel layer to improve collaborative filtering[1]. The model achieves state-of-the-art (SOTA) performance on benchmark datasets like MovieLens and Douban, demonstrating significant improvements in training speed and effectiveness in handling the cold start problem without needing any extra side information[1].

Key Innovations

2D Kernel Layer

An extension of standard kernelized networks, this layer adds an extra dimension to capture more complex, non-linear relationships and multi-dimensional latent features from the user-item interaction matrix[1]. It uses reparameterization (the “kernel trick”) to sparsify network weights, which reduces the model’s complexity and improves training efficiency[1].

k-Cross Kernel Layer

This layer is specifically designed to enhance collaborative filtering by identifying similarities between users and items. It uses two distinct kernels—a horizontal kernel and a vertical kernel—that are convoluted across the user-item matrix[1].

• The horizontal kernel captures item correlations among different users.
• The vertical kernel captures user correlations across different items[1].

How It Works: Model Architecture

RecKross employs an AutoEncoder-like structure that processes a user-item rating matrix. The optimal configuration found during experiments is a three-layer network[1]:

1. Input Layer: A 2D Kernel layer processes the initial user-item interaction matrix.
2. Hidden Layer: A k-Cross Kernel layer performs the core collaborative filtering task and captures neighbor information efficiently.
3. Output Layer: A final 2D Kernel layer reconstructs the rating matrix, predicting the missing values[1].

Key Advantages

• Superior Performance: Outperforms previous state-of-the-art models on the ML-1M and ML-100K datasets in terms of Root Mean Squared Error (RMSE)[1].
• High Efficiency: Trains significantly faster than competing models like GLocal-K. For instance, on the ML-1M dataset, RecKross took approximately 3,032 seconds to train, compared to GLocal-K’s 7,880 seconds on the same hardware[1].
• Cold Start Effectiveness: Demonstrates better performance than strong baselines in highly sparse data settings, making it highly effective for new users and items with limited interaction data[1].
• No Side Information Required: Achieves top results using only the user-item interaction matrix, without needing additional data like user demographics or item attributes[1].

Performance Highlights

RecKross was evaluated against several baseline models on three datasets. It achieved the lowest RMSE (lower is better) on the MovieLens-1M dataset, indicating its superior prediction accuracy[1].

The model also achieved the best performance on the ML-100K dataset with an RMSE of 0.8910 and was highly competitive on the Douban dataset[1].