In this paper, we introduce a new task called Logical Session Complex Query Answering (LS-CQA) that tackles the challenge of answering complex user queries on multi-relational hypergraph data. This task builds upon the existing Complex Query Answering (CQA) problem and expands it to include sessions as ordered hyperedges of items. Our proposed approach leverages Atten-Mixer, a pure attention-based method, to aggregate embeddings from different layers and conduct session-level reasoning. The results show that LS-CQA demonstrates strong compositional generalization capabilities and outperforms other transformer-based methods by 1.28 to 3.22 in MRR on three datasets.
Introduction
Have you ever wondered how to systematically answer complex user queries on multi-relational data? Look no further! In this article, we propose a new task called Logical Session Complex Query Answering (LS-CQA) that tackles just that challenge. LS-CQA builds upon the existing Complex Query Answering (CQA) problem and expands it to include sessions as ordered hyperedges of items. By leveraging Atten-Mixer, a pure attention-based method, we can aggregate embeddings from different layers and conduct session-level reasoning.
What is LS-CQA?
LS-CQA is the task of answering complex user queries on multi-relational hypergraph data, where sessions are treated as ordered hyperedges of items. Think of it like a complex recipe that requires you to gather ingredients from different parts of the kitchen and combine them in a specific order to create something delicious. In LS-CQA, we need to answer logical queries on an aggregated hypergraph of sessions, items, and attributes.
Why is LS-CQA important?
With the growing amount of multi-relational data, there is a rising demand for effective query answering methods that can handle complex user intentions. LS-CQA fills this gap by providing a formal framework for tackling complex queries on hypergraph data. By leveraging Atten-Mixer, we can improve the performance of GNN models and provide more accurate recommendations to users.
How does LS-CQA work?
Our proposed approach consists of two stages: session encoding and query answering. In the first stage, we encode each session as a vector using Atten-Mixer, which captures the global context and item orders. In the second stage, we conduct query answering on an aggregated hypergraph of sessions, items, and attributes to generate the final output.
What are the results?
We conducted ablation studies and evaluations on three datasets and found that LS-CQA demonstrates strong compositional generalization capabilities. The results show that LS-CQA outperforms other transformer-based methods by 1.28 to 3.22 in MRR. This means that LS-CQA can handle complex user queries more effectively and provide more accurate recommendations to users.
Conclusion
In this article, we proposed the Logical Session Complex Query Answering (LS-CQA) task, which tackles the challenge of answering complex user queries on multi-relational hypergraph data. By leveraging Atten-Mixer, we can aggregate embeddings from different layers and conduct session-level reasoning to improve the performance of GNN models. The results show that LS-CQA demonstrates strong compositional generalization capabilities and outperforms other transformer-based methods by 1.28 to 3.22 in MRR on three datasets. This work has significant implications for improving recommendation systems and handling complex user queries in various applications.
