Predicting Online Item-Choice Behavior: A Shape-Restricted Regression Approach
Round 1
Reviewer 1 Report
In this paper, the authors address limitations in Graph Attention Networks (GATs), a popular method for graph embedding. While GATs effectively utilize attention mechanisms to aggregate first-order neighborhood node information, they struggle to consider high-order neighborhood nodes and structural information. Simply incorporating high-order neighborhood node information into traditional GATs may lead to over-smoothing.
To overcome this challenge, the authors propose a joint attention graph embedding model that incorporates similar networks and structural correlation. They introduce the concept of structural correlation to the original graph attentional method, taking into account not only node content features but also joint node topological structure features when calculating the attention score.
There are some concerns about this version:
- Notations are a mess and some of them are wrong. Please correcting them accordingly.
- Too many places of equations are empty in this Word document, it would be better to submit a PDF document.
- Eq.(13) is wrong.
- The presentation of the manuscript should be improved totally. There are too many grammar issues. For example, three dataset...
I will review the revised version since the current version is not readable.
- The presentation of the manuscript should be improved totally. There are too many grammar issues. For example, three dataset...
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Overall, this paper is well written, with a clear motivation and well-designed experiments. The authors proposed using PV sequences to predict user behavior on an e-commerce website, which is interesting and beneficial to real E-business platforms. The research method is presented clearly, and the experimental results suggest the proposal's effectiveness.
Some minor concerns are as follows:
1. PV sequence is only one of the factors that may impact user behavior on e-commerce websites. So, in general, there should be a well-formed model consisting of all the necessary factors for predicting user behavior. Unfortunately, this paper lacks discussions on other factors. Please add additional content to explain why the model does not involve other factors.
2. The ground truth in the experiment should be described first. For example, how did you calculate the F1 score?
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The authors suggest a new method for predicting item choice probabilities from user page views history, treated as time series. The method they suggest is based on QP global optimization and on representing page views as transitive directed graphs. The method's descriptions and proofs are sound to the best of my knowledge. Experimental evaluation is performed and the method is compared to several baselines.
General comments: what is missing is a survey (and possibly a comparison) of modern neural methods, especially graph neural networks such as GCN. While the proposed method is unsupervised in its core, such a comparison should be performed, and GNN results can be treated as a topline. For instance, these works are relevant:
Zhu, G., Wu, Z., Wang, Y., Cao, S. and Cao, J., 2019. Online purc hase decisions for tourism e-commerce. Electronic Commerce Research and Applications, 38, p.100887. Liu, Z., Wang, X., Li, Y., Yao, L., An, J., Bai, L. and Lim, E.P., 2022. Face to purchase: Predicting consumer choices with structured facial and behavioral traits embedding. Knowledge-Based Systems, 235, p.107665. Simple node embeddings in the graph can have high impact on predicting purchases. Detailed comments: l.126 User-item choice probabilities change over time. Is time a parameter in this setup? eq.2-5 Time (i.e., purchase sequence order) is not expressed in this model. l.158 Is the dot at the end of the formula necessary? Does it express a dot operator of some sort? If not, it is just confusing here. l.165-156 If there is an Up operation and Move operation that descreases v_t, then why isn't there a Down operator? It is necessary when the changes are not symmetric. Some explanation is needed. l.180-183 While it is very simple, a short proof of the fact that this relation is a partial order relation is needed. Table 4: What type of ANN? What are its parameters? It is unclear from the table.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
This paper investigates the correlation between user pageview (PV) histories and their item selection patterns on an e-commerce platform. To predict item choices accurately, the paper proposes a shape-restricted optimization model. This model enforces item-choice probability estimation for various PV sequences while adhering to monotonicity constraints. These constraints utilize partial orders based on PV sequence characteristics like recency and frequency. To enhance computational efficiency, the study introduces algorithms to remove redundant constraints via partial order transitivity. Empirical evaluations on real-world clickstream data showcase the model's superior predictive performance compared to both a leading optimization model and conventional machine learning approaches.
The article is well-written, with a clear presentation, solid theoretical foundation, and experimental support. Therefore, I recommend its publication.
