[Paper Review] Basket Recommendation with Multi-Intent Translation Graph Neural Network

Zhiwei Liu, Xiaohan Li+4 more|arXiv (Cornell University)|Oct 22, 2020

Recommender Systems and Techniques38 references4 citations

TL;DR

This paper proposes MITGNN, a multi-intent translation graph neural network that models basket recommendation by learning multiple latent intents through translation-based entity modeling and graph neural networks. It outperforms existing methods on both transductive and inductive settings by capturing complex, multi-relational item correlations across baskets.

ABSTRACT

The problem of basket recommendation~(BR) is to recommend a ranking list of items to the current basket. Existing methods solve this problem by assuming the items within the same basket are correlated by one semantic relation, thus optimizing the item embeddings. However, this assumption breaks when there exist multiple intents within a basket. For example, assuming a basket contains \{ extit{bread, cereal, yogurt, soap, detergent}\} where \{ extit{bread, cereal, yogurt}\} are correlated through the "breakfast" intent, while \{ extit{soap, detergent}\} are of "cleaning" intent, ignoring multiple relations among the items spoils the ability of the model to learn the embeddings. To resolve this issue, it is required to discover the intents within the basket. However, retrieving a multi-intent pattern is rather challenging, as intents are latent within the basket. Additionally, intents within the basket may also be correlated. Moreover, discovering a multi-intent pattern requires modeling high-order interactions, as the intents across different baskets are also correlated. To this end, we propose a new framework named as extbf{M}ulti- extbf{I}ntent extbf{T}ranslation extbf{G}raph extbf{N}eural extbf{N}etwork~({ extbf{MITGNN}}). MITGNN models $T$ intents as tail entities translated from one corresponding basket embedding via $T$ relation vectors. The relation vectors are learned through multi-head aggregators to handle user and item information. Additionally, MITGNN propagates multiple intents across our defined basket graph to learn the embeddings of users and items by aggregating neighbors. Extensive experiments on two real-world datasets prove the effectiveness of our proposed model on both transductive and inductive BR. The code is available online at https://github.com/JimLiu96/MITGNN.

Motivation & Objective

• Address the limitation of existing basket recommendation models that assume single semantic relations among items within baskets.
• Model complex, multi-intent patterns where items in a basket may belong to multiple latent intents (e.g., breakfast, cleaning).
• Capture high-order interactions across baskets by propagating multiple intents through a user-defined basket graph.
• Enable effective inductive learning by training dedicated aggregators for unseen baskets, improving generalization.
• Improve recommendation performance by jointly modeling intent relations and user-item interactions via a type-guided attention mechanism.

Proposed method

• Model each basket as a head entity and learn T distinct intent vectors as tail entities via translation: q_i = h + r_i, where r_i is a learnable relation vector.
• Use multi-head attention mechanisms to aggregate user and item information into relation vectors, capturing diverse semantic relations.
• Construct a basket graph where nodes represent baskets and edges represent shared items or intents, enabling message passing across baskets.
• Apply multi-layer GNN propagation to aggregate neighborhood information, learning enriched embeddings for users, items, and intents.
• Integrate type-guided attention to selectively aggregate hidden intents into user-guided and item-guided basket embeddings for downstream recommendation.
• Train the model end-to-end using a contrastive learning objective to optimize both intent discovery and recommendation performance.

Experimental results

Research questions

• RQ1 Can modeling multiple latent intents within a basket improve basket recommendation performance compared to single-relation assumptions?
• RQ2 How do high-order interactions across baskets influence the quality of learned intent and embedding representations?
• RQ3 What is the optimal number of intents and layers for the GNN architecture in the context of basket recommendation?
• RQ4 Does the proposed MITGNN framework generalize well to inductive settings where test baskets are unseen during training?
• RQ5 How do relation vectors and attention mechanisms contribute to capturing complementary and correlated intents across baskets?

Key findings

• MITGNN achieves state-of-the-art performance on both transductive and inductive basket recommendation tasks across two real-world datasets (Instacart and Walmart).
• Performance improves significantly when modeling multiple intents (T > 1) compared to single-intent modeling (T = 1), with optimal T = 3 on Instacart and T = 4 on Walmart.
• The model achieves peak performance at L = 3 layers, indicating that higher-order message passing enhances representation learning, but L = 4 leads to performance degradation due to over-smoothing.
• Inductive BR performance is higher than transductive BR on Walmart data, as the inductive setting retains 87.5% of ground-truth items per basket, making prediction easier.
• MITGNN outperforms strong baselines like Triple2Vec, BPR-MF, and DREAM, especially in inductive settings, due to its learned aggregators and multi-intent modeling.
• The type-guided attention mechanism effectively captures user- and item-specific intent patterns, enhancing the quality of user and item embeddings.