Source-linked AI summary
CosRec: 2D Convolutional Neural Networks for Sequential Recommendation
An Yan, Shuo Cheng, Wang-Cheng Kang, Mengting Wan, Julian McAuley
TL;DR
Sequential recommendation models can struggle to uncover complex relationships in users’ ordered histories, especially when useful interactions are nonadjacent. CosRec addresses this gap by encoding item pairs into a three-way tensor and applying a 2D CNN. On two real-world datasets, it significantly outperforms recent sequential approaches.
Problem
Existing sequential models can be constrained by ordered chain structures and struggle to capture complex or nonadjacent relationships in users’ interaction histories.
Method
CosRec encodes previous item embeddings into pairwise representations and applies a 2D convolutional network to extract sequential features.
Results
CosRec outperforms all evaluated baselines on MovieLens and Gowalla, including a 25.0% improvement in MAP on ML-1M.
Takeaways & Limitations
CosRec is presented as a generic model for sequential recommendation that captures high-order and nonadjacent item relations.
Abstract
from arXiv · showhide
Sequential patterns play an important role in building modern recommender systems. To this end, several recommender systems have been built on top of Markov Chains and Recurrent Models (among others). Although these sequential models have proven successful at a range of tasks, they still struggle to uncover complex relationships nested in user purchase histories. In this paper, we argue that modeling pairwise relationships directly leads to an efficient representation of sequential features and captures complex item correlations. Specifically, we propose a 2D convolutional network for sequential recommendation (CosRec). It encodes a sequence of items into a three-way tensor; learns local features using 2D convolutional filters; and aggregates high-order interactions in a feedforward manner. Quantitative results on two public datasets show that our method outperforms both conventional methods and recent sequence-based approaches, achieving state-of-the-art performance on various evaluation metrics.
1 INTRODUCTION
Sequential recommendation predicts future user behavior from ordered action histories, but chain-structured models can miss nonadjacent item relationships. CosRec relaxes this constraint with pairwise encoding and 2D convolutions to capture skip behavior and complex correlations.
- Sequential recommendation predicts users’ future behavior from historical action sequences while incorporating the order of past actions.
- Markov Chains use preceding actions, whereas RNNs and CNNs can incorporate users’ complete interaction sequences.
- Ordered item representations preserve locally concentrated dynamics but remain constrained by the one-directional chain structure.
- CosRec introduces pairwise encoding so nonadjacent items can interact, allowing 2D convolutional kernels to capture skip behavior.The framework can use small filters such as 3 × 3 and adapt to shallow or deep structures.
2 RELATED WORK
Prior sequential recommendation methods include Markov Chains, recurrent networks, convolutions, and self-attention. CosRec differs from nearby convolutional approaches by applying standard 2D CNNs to pairwise representations of previous items.
- Sequential recommendation has used Markov Chains, RNNs, TCNs, and self-attention to model patterns among previously consumed items.
- Caser and NextItNet use convolutions, but their filter structures differ from CosRec’s standard 2D convolution.
- CosRec is presented as the first 2D CNN-based approach for next-item recommendation.
- Other CNN recommendation models process images, documents, or user-item outer products, whereas CosRec captures sequential patterns in previous items.
3 METHOD
CosRec embeds users and items, converts previous items into pairwise representations, and applies a 2D CNN to extract sequential features. The resulting sequence representation is combined with user preferences to predict the next item.
- 3 METHOD: The sequential recommendation task predicts a user’s next item from the sequence of previously interacted items.
- 3 METHOD: CosRec comprises an embedding look-up layer, pairwise encoding module, and 2D convolution module.
- 3.1 Embedding Look-up Layer: Items and users are embedded into matrices with latent dimensionality d, and prior item embeddings are retrieved for each user and time step.
- 3.2 Pairwise Encoding: The pairwise encoding module creates a three-way tensor that permits flexible interactions among items.
- 3.2 Pairwise Encoding: The encoded tensor matches the shape of an image feature map, enabling standard CNN architectures to be adapted for sequential recommendation.
- 3.3 2D Convolutions: A 2D CNN uses 1 × 1 layers to enrich representations and 3 × 3 layers to aggregate sequential features and extract deeper relations.
- 3.3 2D Convolutions: The convolutional blocks and fully connected layer produce the final sequential feature vector v(u,t), which is combined with the user embedding to generate item probability scores.
- 3.5 Comparison with Existing CNN-based Approaches: CosRec addresses Caser’s handling of irrelevant interactions and limited channel-wise representation by using pairwise representations and 2D kernels.
4 EXPERIMENTS
CosRec is evaluated on MovieLens and Gowalla under the Caser protocol using top-N ranking metrics. It outperforms the compared baselines and its filters capture both recent/local and nonadjacent item patterns.
- Datasets: Evaluation uses MovieLens and Gowalla benchmark datasets under the protocol used for Caser.MovieLens-1M and Gowalla are the two datasets used in the experiments.
- Metrics: Performance is reported with MAP, Precision@N, and Recall@N for N = 1, 5, and 10.
- Performance Comparison: The comparison includes popularity, non-sequential ranking, Markov-chain, and other sequential recommendation baselines.The listed baselines include PopRec, BPR, FMC, and FPMC.
- Visualization: Figure 3 shows filters emphasizing recent items in one case and scattered nonadjacent item patterns in another.Darker grid values indicate larger weights for pairwise item encodings; the filters use kernel size 5 and are trained on ML-1M.
- Performance Comparison: CosRec outperforms all baselines on both datasets, with a 25.0% MAP improvement on ML-1M.The authors attribute the particularly large ML-1M improvement to its density and richer sequential signals.
5 CONCLUSION
The conclusion presents CosRec as a 2D CNN framework for sequential recommendation. It encodes item sequences into pairwise representations and extracts sequential features with 2D convolution, outperforming recent sequential approaches on two real-world datasets.
- CosRec encodes sequences into pairwise representations and uses a 2D CNN to extract sequential features.
- Experiments on two real-world datasets show that CosRec significantly outperforms recent sequential approaches.The authors present it as a generic model for sequential recommendation tasks.