Proactive Return Prediction in Online Fashion Retail Using Heterogeneous Graph Neural Networks

Round 1

Reviewer 1 Report

1. Regarding the contributions made by the paper, the authors state that the proposed model presents a way of complex encapsulation of customer preferences, product attributes and order characteristics, with direct effects on yield prediction, but it is not very clear what this method consists of /method. I recommend the authors to provide detailed information about the proposed method/modality and the advantages/benefits it brings, by applying it within the proposed model.

2. Since the effectiveness of the proposed model depends on the volume of analyzed data, and the training of the neural network is closely correlated with it, I recommend the authors to create and manage a database in which to store historical data related to features such as: Product ID, ID- customer ID, order details: quantity, price, coupons used, payment method), product attributes: color, size, product group, and quantity of returned products. Such a database would be very useful in terms of implementing the presented algorithms and creating distinct applications that would allow multi-criteria querying of the database in order to provide operational and efficient results related to the strategic decisions of retailers in the online product trade of fashion, relative to the prediction of the volume of returned products.

Author Response

Firstly, our proposed method entails a meticulously designed graph neural network aimed at proactively predicting customer returns. In Section 3, we offer a comprehensive exposition of our innovative Customer-Order-Product heterogeneous graph neural networks. Additionally, in Section 3.2, we delve into the advantages of our prediction method, highlighting how our Customer-Order-Product Heterogeneous Graph Neural Networks (COP-HGNN) leverage three distinct types of nodes and edges to capture nuanced information:“Our Customer-Order-Product Heterogeneous Graph Neural Networks (COP-HGNN) for return prediction leverage three types of nodes and edges to capture nuanced information, enabling a deeper understanding of the underlying structures. By incorporating diverse node and edge types, our model is better equipped to discern intricate patterns”. Furthermore, in the conclusion, we further explained that“By training neural networks, hidden feature vectors can be constructed for customers，orders, and products, enabling the prediction of customer return behavior for a specific product in a new order. This approach allows each node of the graph to embed into a feature vector and utilize edges with features to transmit information in the graph, capturing a comprehensive graph structure”.

Secondly, the database provided for this study aligns with the type as described. Constructed from data sourced from various channels, including fundamental customer and product information, order-level transactions, and return records, our database forms the foundation upon which our neural networks are trained to predict future returns. In this revision, we have included a sentence elucidating how the database was assembled for training our neural network.

Reviewer 2 Report

The development of the COP-HGNN model represents a significant contribution to the field of online retail analytics. Your approach to addressing the high return rates in online fashion retail is innovative and timely.

Even though the methodology is well-structured and scientifically sound, I would encourage you to provide more detailed descriptions or clarifications for the readers in certain parts of your paper, such as the data preprocessing steps or the choice of your specific model parameters (e.g. embedding dimension and learning rate). Adding a brief rationale behind the selection of the Adam optimization algorithm as one of your chosen research methods would help readers engage more easily with your paper.

Consider adding a discussion on the model's limitations, such as potential biases that might be present in the dataset or the model’s performance in varying market conditions.

Author Response

Thank you very much for taking the time to review our manuscript. Your feedback has been invaluable in improving the quality of our work. In response to your suggestions, we have made significant enhancements in this revision, particularly in providing further details on our data preprocessing steps.

Our data preprocessing encompasses a series of essential steps including data cleaning, transformation, feature engineering, normalization, and data splitting. During the data cleaning phase, we meticulously identified and removed abnormal return records where the quantity of returned products exceeded the amount ordered. Additionally, we employed forward and backward filling methods to impute missing values in the product price column. In the data transformation stage, we implemented measures to capture seasonal effects by converting the date of the order into two seasonal features: the day of the week and the month of the year. This allowed us to effectively incorporate seasonal variations into our analysis of product returns. Furthermore, in the feature engineering stage, we introduced additional features such as the number of products in the same order, the order value, and the average order value. These features were specifically designed to capture the impact of various order characteristics on return behavior, thus enhancing the predictive capabilities of our model. Lastly, we partitioned the data into training, validation, and test sets, and standardized the numerical features using a standardized normalization method based on the training dataset. This ensured consistency and reliability in our analysis while mitigating the risk of overfitting.

To optimize the proposed HGNN network, we select the embedding dimension E and learning rate 'lr' of the Adam optimizer as hyperparameters. Due to the independence of these two hyperparameters, we employed a grid search approach to optimize them separately. The range of search space for embedding dimension E was set to be [8, 12, 16, 24, and 32], while the learning rate's search range was [0.1, 0.005, 0.001, and 0.0005]. We used the validation set to evaluate the hyperparameters and the corresponding optimal training steps were ultimately determined by embedding vector dimension E of 12, learning rate of 0.001, and 25 training steps.

We have included a statement elucidating our choice of the Adam optimization algorithm: "The Adam optimization algorithm is favored for optimizing graph neural networks because of its adaptive learning rate, memory efficiency, robustness to sparse gradients, and faster convergence speed."

In the last paragraph of the conclusion, we have provided a discussion on the potential limitation of the proposed paper.

"In addition to highlighting the strengths of our graph neural network model in proactively predicting product returns, it is essential to acknowledge its limitations. One significant limitation pertains to the potential biases inherent in the dataset used for training and evaluation. Despite efforts to ensure data quality and integrity, biases such as sampling bias or selection bias may still exist, potentially influencing the model's predictions. Additionally, the performance of our model may vary under different market conditions or contexts not fully captured by the dataset. Factors such as changes in consumer preferences, economic fluctuations, or unforeseen external events could impact the predictive accuracy of the model. Moreover, the effectiveness of proactive prediction strategies may be influenced by the dynamic nature of markets and evolving customer behaviors, warranting further exploration and adaptation of the model over time. Recognizing these limitations is essential for interpreting the model's predictions accurately and for guiding future research efforts aimed at enhancing the robustness and generalizability of predictive models in real-world scenarios."

Reviewer 3 Report

Presented paper “Proactive Return Prediction in Online Fashion Retail using Heterogeneous Graph Neural Networks” investigates the online fashion retailers and impact of returns to the carbon footprint of e-commerce and other contributions. The authors propose model intricately encapsulates customer preferences, product attributes, and order characteristics, providing a holistic approach to return prediction. They have real-world data sourced from an online fashion retail platform to demonstrates superior predictive accuracy on return behaviour of repeat customers, compared to conventional machine learning techniques.

References cited in the manuscript are enough and present analysis of the problem.

All data in the figures and tables is clear and well visible as per journal requirements.

Despite of the above I have the following remarks to the authors reviewing this article:

1. Statista.com is not included in reference list. Once the authors cite source of information they are obliged to present all requisitions in order to be verified from the readers.

2. The authors have to cite correctly the sources of all formulas.

3. The proposed manuscript content many abbreviations. It would be easier for the readers to have all of them in a separate section. Such section or table would improve text acceptance. 

Author Response

Thank you very much for dedicating your time to review our manuscript. Your insightful comments have been instrumental in refining our work. In this revised version, we have addressed all the concerns you raised.

Firstly, we have incorporated a hyperlink to the data source from Statista.com in the revised paper.

Secondly, we have conducted a comprehensive review of all citations within the text, specifically focusing on ensuring accurate citation of all formulas as per your feedback.

Thirdly, in response to your suggestion, we have included a table in Appendix A dedicated to an abbreviation list encompassing all the abbreviations used throughout the paper. We believe this addition not only enhances the clarity of our manuscript but also facilitates easier comprehension for readers.

Once again, we sincerely appreciate your valuable feedback and constructive criticism, which have significantly contributed to the refinement of our manuscript.