Reproduce “Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding” on OpenResearch

About this paper

Jiaxi Tang, Ke Wang

The ConvolutionAl Sequence Embedding Recommendation Model (Caser) applies convolutional neural networks to model item sequences as 'latent images,' effectively capturing both union-level sequential patterns and skip behaviors. It consistently outperformed state-of-the-art baselines across four real-world datasets for personalized Top-N sequential recommendation.

Key insights

• Convolutional Neural Networks (CNNs) can effectively model sequential user behavior by treating item sequences as 'images' in latent space, capturing patterns beyond simple item-to-item transitions.
• Explicitly modeling 'union-level' sequential patterns (combinations of past items) and 'skip behaviors' (influential items not immediately preceding) is crucial for accurate sequential recommendation and can be achieved through CNN filters and pooling.
• A unified framework that integrates both short-term sequential patterns and long-term general user preferences significantly improves recommendation accuracy across varied datasets.
• Vertical CNN filters can implicitly learn to emphasize more recent past actions in a sequence, aligning with the recency effect in sequential behavior.

Summary by alphaXiv.