Abstract:Sequential recommendation requires understanding the dynamic patterns of users' behaviors, contexts, and preferences from their historical interactions. Most existing works focus on modeling user-item interactions only from the item level, ignoring that they are driven by latent shopping intentions (e.g., ballpoint pens, miniatures, etc). The detection of the underlying shopping intentions of users based on their historical interactions is a crucial aspect for e-commerce platforms, such as Amazon, to enhance the convenience and efficiency of their customers' shopping experiences. Despite its significance, the area of main shopping intention detection remains under-investigated in the academic literature. To fill this gap, we propose a graph-regularized stochastic Transformer method, G-STO. By considering intentions as sets of products and user preferences as compositions of intentions, we model both of them as stochastic Gaussian embeddings in the latent representation space. Instead of training the stochastic representations from scratch, we develop a global intention relational graph as prior knowledge for regularization, allowing relevant shopping intentions to be distributionally close. Finally, we feed the newly regularized stochastic embeddings into Transformer-based models to encode sequential information from the intention transitions. We evaluate our main shopping intention identification model on three different real-world datasets, where G-STO achieves significantly superior performances to the baselines by 18.08% in Hit@1, 7.01% in Hit@10, and 6.11% in NDCG@10 on average.
What problem does this paper attempt to address?
### Problems the Paper Attempts to Solve
This paper aims to address the issue of identifying users' primary shopping intentions in sequential recommendation systems. Specifically, most existing sequential recommendation algorithms only model the interaction between users and items at the item level, ignoring that these interactions are driven by underlying shopping intentions (e.g., ballpoint pens, miniature models, etc.). Identifying these underlying shopping intentions is crucial for e-commerce platforms (such as Amazon) to enhance user experience. However, research in this area is relatively scarce. To this end, the authors propose a graph-regularized stochastic Transformer method (G-STO), which models intentions as sets of products and user preferences as combinations of intentions, representing them as stochastic Gaussian embeddings in a latent space. The G-STO method introduces a global intention relationship graph as prior knowledge, making related shopping intentions closer in distribution. Finally, the regularized stochastic embeddings are fed into a Transformer-based model to encode sequential information in intention transitions.
### Main Contributions
1. **Task Innovation**: To the best of the authors' knowledge, this is the first work focusing on identifying primary shopping intentions using only intention-level data. This helps in user understanding and improves the performance of downstream tasks (such as item-level recommendation and page optimization ranking).
2. **Gaussian Distribution Representation**: Describing intentions as Gaussian distributions, using stochastic representations to reflect the high-level characteristics of intentions, the collaborative transitivity between intentions, and the uncertainty of user preferences.
3. **Graph Regularization**: Introducing a shopping intention relationship graph as prior knowledge and proposing a novel graph regularizer to constrain stochastic representations in distribution-based methods.
4. **Experimental Proof**: Developed three different real-world Amazon datasets covering long-term, short-term, and purchase-related user cases. Experimental results show that G-STO significantly outperforms existing baseline models on the three datasets, with an average improvement of 18.08% in Hit@1, 7.01% in Hit@10, and 6.11% in NDCG@10.
### Solution
1. **Stochastic Embedding Layer**: Representing each intention as an elliptical Gaussian distribution, including mean and covariance embeddings, to capture the high-level semantics of shopping intentions.
2. **Intention Relationship Graph Regularizer**: Constructing an intention relationship graph, using a Graph Convolutional Network (GCN) to propagate node features, transferring knowledge from common nodes to uncommon nodes, thus alleviating the cold start problem.
3. **Mean and Covariance Transformer**: Encoding sequential information in users' historical interactions to generate stochastic representations of user preferences.
4. **Wasserstein Distance**: Used to measure the difference between intentions and user preferences, combined with Bayesian Personalized Ranking (BPR) loss for model training.
### Experimental Validation
1. **Benchmark Datasets**: Created three benchmark datasets simulating long-term shopping scenarios, frequent user intention interactions within 24 hours, and purchase-related user cases.
2. **Performance Evaluation**: Evaluated the performance of G-STO using multiple evaluation metrics (such as Hit@1, Hit@10, NDCG@10) and compared it with existing baseline models.
### Conclusion
G-STO effectively addresses the cold start problem and the modeling of relationships between intentions in primary shopping intention identification by introducing graph regularization and stochastic representations, significantly enhancing the performance of recommendation systems.