Understanding Inter-Session Intentions via Complex Logical Reasoning
Jiaxin Bai, Chen Luo, Zheng Li, Qingyu Yin, Yangqiu Song
TL;DR
This paper introduces a new task called logical session complex query answering (LS-CQA) to understand intricate user intentions across multiple sessions, and proposes a transformer-based model, LSGT, that achieves state-of-the-art results.
Contribution
The paper formulates LS-CQA as a hypergraph-based complex query answering task and develops the LSGT model to effectively capture logical relationships across sessions.
Findings
LSGT achieves state-of-the-art performance on three datasets.
The model effectively captures complex logical relationships in user sessions.
The approach demonstrates improved recommendation accuracy for intricate user intentions.
Abstract
Understanding user intentions is essential for improving product recommendations, navigation suggestions, and query reformulations. However, user intentions can be intricate, involving multiple sessions and attribute requirements connected by logical operators such as And, Or, and Not. For instance, a user may search for Nike or Adidas running shoes across various sessions, with a preference for purple. In another example, a user may have purchased a mattress in a previous session and is now looking for a matching bed frame without intending to buy another mattress. Existing research on session understanding has not adequately addressed making product or attribute recommendations for such complex intentions. In this paper, we present the task of logical session complex query answering (LS-CQA), where sessions are treated as hyperedges of items, and we frame the problem of complex…
Peer Reviews
Decision·Submitted to ICLR 2024
1. The motivation that incorporates logical session query answering into product recommendation to model user intent is novel. 2. The experimental results demonstrate the effectiveness of the proposed LSGT. 3. The authors theoretically justify the expressiveness and operator-wise permutation invariance of LSGT.
1. There are some obvious typos. Authors should scrutinize the writing of the paper. (1) In the 5th line of section 4.3, the formula after “The edge feature is denoted as” lacks a proper superscript. (2) In Table 5, the first word “Predicti” in explanation of query type 2p should be “Predict”. (3) In Table 5, the word “prodict” in explanation of query type ip should be “product”. (4) In the 2nd line below Figure 5, the word “descibed” should be “described”. 2. In Figure 5, the query structure of
**S1.** The item recommendation task is framed as a complex logical query. While the task per se is not new (LogiRec [2] originally introduced it with projection and intersection operators), this paper extends it to unions and negations and to hypergraphs. **S2.** Evaluation includes several baselines (that show, on the other hand, that the proposed approach only marginally outperforms existing models, but more on that in W2)
Starting from the claimed contributions: **W1. Task.** The formulated task of Logical Session Query Answering is essentially query answering over hypergraphs. Sessions are n-ary edges, and other relations form 2-ary edges, so the hypergraph has edges of different arity. The temporal aspect of items in session hyperedges (that items follow each other in one session) seems to be of little use as the best-performing models are not using this information anyway. I would recommend the authors to foc
- This paper proposes the task of Logical Session Query Answering (LSQA), providing an novel paradigm for enhancing applications like session-based recommendation and query recommendation by understanding the logical structures of users' latent intents. - The paper provides a theoretical analysis on the expressiveness of the proposed Logical Session Graph Transformer (LSGT) model. - The paper innovatively build a unified representation model for items, sessions and logical operators using hyperg
- Though the proposed task is novel, the proposed technical solution LSGT relies on existing hypergraph structures and transformer architeactures. Such designs have limited differences compared to existing sequential models and graph models. This lower the technical contribution of this paper. - The evaluation part could be enhanced with more diverse experiments to conduct a more comprehensive empirical study, such as ablation study, hyperparameter study, case study on the generated queries, and
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsSemantic Web and Ontologies · Advanced Graph Neural Networks · Data Management and Algorithms
MethodsAttention Is All You Need · Laplacian EigenMap · Linear Layer · Dense Connections · Position-Wise Feed-Forward Layer · Absolute Position Encodings · Laplacian Positional Encodings · Adam · Residual Connection · Dropout