International Journal on Recent and Innovation Trends in Computing and Communication

ISSN: 2321-8169 Volume: 11 Issue: 9

Article Received: 25 July 2023 Revised: 12 September 2023 Accepted: 30 September 2023
___________________________________________________________________________________________________________________

A Model for Predicting E-Commerce Product Returns

Using Hybrid CNN-GRU
A.Arulmurugan1, Arul N2, Meera Alphy3, Vidya Rajasekaran*4, Kuncham Sreenivasa Rao5, Karumuri Sri Rama
Murthy6
1 Department of Computing Technologies, School of Computing
1 SRM Institute of science and Technology, kattankulathur , chennai,Tamilnadu, India-603203
1 arulmura1@srmist.edu.in

2 Department of Computer Science and Engineering

2 St.Peter’s
Engineering College,Hyderabad, India
2 arulthala82@gmail.com

3 Department of Computer Science and Engineering,

3 Mahatma Gandhi Institute of Technology, Gandipet, Hyderabad,India
3 drmeeraalphy@gmail.com

4
Department of Information Technology
4
B.S. Abdur Rahman Crescent Institute of Science and Technology, Chennai, India
*4
Corresponding author e-mail: vidyarajesh23@gmail.com
5 Department
of Computer Science and Engineering,
5
Faculty of Science and Technology, (IcfaiTech), ICFAI Foundation for Higher Education, Hyderabad, India
5 ksrao517@gmail.com

6 Department of Computer Science and Engineering,

6 VNR Vignana Jyothi Institute of Engineering and Technology, Hyderabad,India
6 sriram.jntuphd@gmail.com

Abstract—The aim of this work is to predict the product return rate of e-commerce using deep learning algorithms. A novel hybrid model
combining Convolutional Neural Network(CNNs) and Gated Recurrent Units (GRUs) is proposed to predict e-commerce product return rates.
The developed model is trained and validated on a large e-commerce dataset with features like consumer demographic informations, product
details, transaction and product return details, and consumer feedbacks. The return history of every product is learned and products least and
most sold is analyzed. Based on all this analysis the return of products in the future is predicted. The proposed work demonstrates that the hybrid
CNN-GRU model outperforms conventional models and standalone CNN and GRU architectures with an accuracy of 83%. Also focus is made
on understanding the features influencing the product returnsthat help firms to make data driven decisions and minimize product returns . This
type of predictive models can be applied to enhance the business strategies in making informed decisions enhancing the overall satisfaction of
the consumer and improving the revenue of the business.
Keywords-E-Commerce; Return To Origin; Prediction Model; Deep Learning; Convolutional Neural Network; Gated Recurrent Units.

I. INTRODUCTION learning techniques to overcome the challenges and

understanding the complex data patterns.
The e-commerce industry has undergone an unprecedented One of the main issues focused in this paper is the return of
growth and tremendous revolution and is developed into a products. The product returns affect the revenue of the
vivacious and sternly competitive business by changing the business, the operational effectiveness, and customer
way consumers obtain goods and services. The online shopping satisfaction need to be concentrated and handling them is one
continues to expand in popularity throughout the world and of the major issues that e-commerce enterprises face in recent
change the shopping experience of the consumers, the e- days. The vast growth of the e-commerce industry has been
commerce businesses must make measures to balance the fueled by the ease, accessibility, and wide range of goods
possibilities of the problems in order to assure the customer provided by online retailers. This online market faces several
prospect and uphold the business development. The e- challenges and several set of difficulties along with its
commerce businesses face several challenges to make their numerous benefits. One of the major concerns of all the issues
business successful and several issues affect the revenue of the is product return. Product returns stand out among these issues
business. The returns disturb the supply chain operations as a crucial problem that may have an extensive pressure on
increasing the financial and logistical burden on the retailers. both the monetary constancy and consumer relations of e-
The e-commerce sector makes use of predictive and machine commerce businesses.

3615
IJRITCC | September 2023, Available @ http://www.ijritcc.org
International Journal on Recent and Innovation Trends in Computing and Communication
ISSN: 2321-8169 Volume: 11 Issue: 9
Article Received: 25 July 2023 Revised: 12 September 2023 Accepted: 30 September 2023
___________________________________________________________________________________________________________________

The product return issues faced by e-commerce industries

are driving the online merchants to develop strategies and This hybrid model is proposed to address the urgent
effectively handle product return issues [1]. Small return rates demand for superior prediction models in e-commerce. This
are not a big deal but high return rates greatly affect the hybrid model adopts a holistic viewpoint, acknowledging that a
businesses productivity as well as its buyer’s contentment. The variety of elements, including visual features, textual content,
development of prognostic models that can forecast return rates and temporal dynamics, affect return rates. Our approach
are very helpful to firms in this situation. strives to give a full understanding of the complicated nature of
Product returns are typical in the retail industry, but they e-commerce product returns by merging CNNs to extract
may be a complicated problem in the world of online shopping. spatial characteristics from product photos and GRUs to
The easiness of online shopping, which enables customers to analyze sequential data, which includes customer reviews and
buy things without feeling the product, provides a setting where product descriptions. Fig. 2 depicts the GRU architecture used
refunds are a necessary component of the business model. in the hybrid model.
Excessive return rates can provide serious financial and
logistical issues for e-commerce companies, even though they Hidden Hidden
are a necessary part of consumer rights and customer-centric State 1 State 2
practices. Because there is no actual touch with the items
before purchase, product returns pose a complicated problem in
the e-commerce environment. Customers frequently rely on X + X + Output
product descriptions, photos, and reviews to make wise
purchases. The lack of in-person inspection, however, can lead
Update Gate -1
to a greater percentage of product returns compared to
traditional brick-and-mortar retail, even with these facilities. E- tanh
commerce companies must strike a balance between assuring Sigmoid
Function
customer pleasure, preserving operational efficiency, and
lowering the financial burden of excessive returns. There are
several research papers proposed for predicting the product Reset Gate X +
returns [2] [3] [4].
The complex world of e-commerce product returns has
given rise to predictive modeling as a useful tool. These models Input
enable e-commerce companies to forecast return rates and
obtain a better knowledge of the variables affecting customer
decisions to return goods. These data enable businesses to fine- Figure 2. GRU Architecture
tune their product descriptions, better inventory management,
and improve overall consumer experiences, eventually resulting The hybrid model presented in this study acknowledges the
to lower return rates and higher profitability. Predictive models need for a varied strategy to comprehend product returns in e-
that can foresee and comprehend product return rates have commerce. It requires the integration of several data sources to
developed into indispensable tools for e-commerce businesses develop a comprehensive knowledge of the elements
in order to properly manage this dilemma [5]. By decreasing influencing returns rather than being restricted to just one form
returns and boosting customer satisfaction, these models give of data, such as customer evaluations or product photos.
organizations the ability to manage inventory proactively, Therefore, our model aims to take advantage of the advantages
improve product descriptions, and enhance customer service. of GRUs to analyze sequential data, such as customer reviews
There are several models proposed using CNN algorithm and product descriptions, and CNNs to extract spatial
for predictions [16] [17] as shown in Fig. 1. But the information from product photos.
development of hybrid models is more accurate compared to
The next sections of this article will present an in-depth
the conventional ones. Using a hybrid model that unites the
examination of the hybrid CNN-GRU model, focusing on its
strength of CNN and GRU, this research proposes a unique
architecture, the training methods used, and the assessment
method for forecasting e-commerce product return rates.
outcomes. Our goal is to demonstrate the unique model's higher
prediction skills in comparison to conventional approaches and
Input Layer Convolutional Convolutional standalone CNN or GRU architectures. As a result, we give e-
Layer 1 Layer 2 commerce businesses with a strong tool for making data-driven
choices, improving operational efficiency, and ultimately
Sub-sampling lowering return rates while increasing consumer happiness.
Pooling Pooling The work proposes a significant step in the ongoing
Layer1 Layer2 attempts of finding the difficulties in anticipating the return of
products of e-commerce businesses. A promising approach is
Sub-sampling developed which enables the online firms to continuously
Fully mitigate the risk by applying the proposed hybrid CNN-GRU
connected methodology.
Layer

Figure 1. Classical CNN model.

3616
IJRITCC | September 2023, Available @ http://www.ijritcc.org
International Journal on Recent and Innovation Trends in Computing and Communication
ISSN: 2321-8169 Volume: 11 Issue: 9
Article Received: 25 July 2023 Revised: 12 September 2023 Accepted: 30 September 2023
___________________________________________________________________________________________________________________

II. LITERATURE REVIEW III. PROPOSED METHODOLOGY

The authors in [6] explore the challenges occurred due to The proposed model uses the strengths of both the CNNs
the increased rate of product returns from the consumers back and GRUs. CNN extracts information influencing consumer’s
to the origin in e-commerce firms. The primary focus of the decision. The GRUs processes the data on consumer reviews,
work is to explore strategies to limit the product returns. product descriptions, and product return details. The below
The work in [7] investigates the vendor’s strategy to Fig. 3 denotes the architectural representation of the CNN-
enhance the sales by integrating the product return costs. It GRU proposed model.
applies both centralized and de-centralized channels with an
extensive cost agreement.
The work in paper [8] highlights the return logistics and Preprocessing
their issues with consumer satisfaction. The paper states the (Data cleaning, Handling outliers and
Input Missing values, Removing
importance of businesses to adapt new techniques in handling
Features Duplicates, Normalization)
challenges in returns of products. A systematic approach is
developed for handling product returns, with minimized costs.
The authors also surveyed several companies facing product
returns and their challenges. Model Optimization
(Forward Selection Wrapper Method)
The work concentrates on e-commerce products quality
management through sentiment analysis [9]. The insights on
products reviews are extracted and deep learning techniques are
applied to it. The developed deep learning model helps to
explore the opinion of the consumers using deep learning CNN GRU
algorithms resulting in good accuracy level.
The authors in [10] address the issues of the existing
forecasting model to handle dynamic changes and non-linearity
Model Validation
of the time-series data. The author has proposed a hybrid model
using deep neural networks to overcome these challenges. The
developed hybrid model uses a feature extraction module to
extract the trends and patterns using time-series data of varying
times. The author developed a three-dimensional sales Detected Value Actual Value
forecasting model and it outperforms the existing model and
adapts the new techniques and results in improved sales
forecasting capabilities.
There are several research papers proposed by combining Resultant score
GRU with other algorithms. The inventory management for
forecasting future sales is focused in this work [11]. Their Figure 3. Architecture of the proposed model based on CNN-GRU for
proposed hybrid approach using GRU outstand the other product return prediction
prediction models. The further usage of posterior test confirms
the effectiveness and accuracy of the proposed model. The A. Data Preprocessing
research work in [12] proposes building load forecasting The data preprocessing involves several stages starting
model using Convolutional Neural Network and GRU. The from data cleaning to remove the outliers and inconsistencies in
proposed model outperforms the existing ones with lower the dataset. It is a crucial step that aims to find and rectify the
prediction losses. The hybrid approach results superior to errors and enhance the accuracy and reliability of the dataset.
conventional models. The paper in [13] introduces a novel The dataset is checked for missing values and the missing
classifier for Human Activity Recognition. A hybrid of CNN values were removed. The duplicates were removed. Then data
along with GRU is proposed to provide a solution. The type conversion as numeric format is carried out to make the
developed model integrates with adaptive user interfaces and dataset suitable for the model. The final dataset is normalized
to make sure that the values are in a similar scale to fit the
pattern based surveillance that outperforms similar structured
developed model.
architectures. The work in [14] proposes a model to predict
traffic patterns in Intelligent Transport Systems (ITS). A B. Hybrid Model Architectural Design
comprehensive algorithm is proposed for integrating the The CNN and GRU model is combined to develop a hybrid
heterogeneous data using exploratory data analysis tools. Then model for comprehensive understanding of the factors leading
several deep learning algorithms are applied and hybrid to product returns. The below Fig. 3 represent the architectural
combinations are developed and LSTM-GRU results in high block of the hybrid model.
accuracy compared to the proposed models. This work [15]
discusses the applications of deep learning algorithms and the C. Feature Extraction
prediction of stock market indices in finance sector. Several The process of creating a subset of pertinent and
models are proposed and the hybrid model outperforms the instructive features from a larger set of data is termed as
other models in terms of accuracy and error rates. feature extraction. Feature extraction is performed to reduce
the dimensionality of the proposed hybrid model and improve
the models efficiency. It also leads to better predictive

3617
IJRITCC | September 2023, Available @ http://www.ijritcc.org
International Journal on Recent and Innovation Trends in Computing and Communication
ISSN: 2321-8169 Volume: 11 Issue: 9
Article Received: 25 July 2023 Revised: 12 September 2023 Accepted: 30 September 2023
___________________________________________________________________________________________________________________

performance. In wrapper method the feature selection is randomly permutes its values in the dataset. So there is an
carried out to evaluate the different subsets of features by interrupt between the feature variables and the target one. The
training and evaluating it. The idea behind applying forward selected features are shuffled in a random manner keeping the
technique is that they are capable of finding feature subsets for other features remain unchanged. A new score for the
specific models and datasets. Also the interactions between the performance metric of the model is evaluated using the dataset
features are considered and also provide a method to optimize using the perturbed feature. The impact of the individual
the performance of the feature selection process. features can be learnt using permutation feature importance
model. The permutation explains the features influencing the
D. Model Training and Validation
models behavior and decision.
The process of model training, testing and validating are
vital components of the predictive model process. IV. RESULTS AND DISCUSSION
The dataset is split as 80% for training the predictive model, The model is proposed by using hybrid CNN-GRU for
10% for validating the model and remaining 10% is for testing finding product return rates. The performance of the developed
the model. The model is trained on the training dataset and model is evaluated using the training, validation and test
learns from features and target variables to make accurate dataset. The optimization is carried out through Adam
predictions. Validation is done to optimize the model. The algorithm. The experiment results are shown in the Table 1. It
performance of the proposed hybrid model is evaluated using is observed that the hybrid CNN- GRU model outperforms the
evaluation metrics scores. Finally the hybrid model is other models.
evaluated on the test dataset to evaluate the unbiased
estimation of the performance on unseen data and generalize TABLE 1: COMPARISON TABLE OF PERFORMANCE
it. The overall process flow diagram of the proposed hybrid EVALUATION OF THE PROPOSED MODELS
model is represented in Fig.4.
Evaluate Models Accuracy Sensitivity Specificity Precision
Input Layer GRU 76.32 72.64 73.53 75.67
Individual CNN 72.16 72.12 71.35 72.15
Training
Convolutional Layer and Test CNN–
86.77 87.14 87.58 85.21
Set GRU
CNN Layer

Max Pooling GRU 80.11 72.13 75.46 77.56

Cross CNN 76.16 73.24 77.17 76.02
Validation CNN –
Convolutional Layer 87.24 87.14 87.01 87.28
GRU

Max Pooling The dataset is split as 80:10:10 for the training, Validating
and testing data. CNN-GRU model has resulted with an
accuracy of 87.24%. The resulting accuracy is comparatively
Flatten better than the other proposed models. The results of
evaluation metrics are illustrated in Table 1 and from that it
GRU Layer can be finalized that hybrid CNN-GRU model outstands
compared to the other ones as shown in Table 1.
Drop Out
100
Dense Layers 80

Fully Connected 60 Accuracy

Layer
40 Sensitivity
Final Prediction
Results 20 Specificity
Figure 4. Flow diagram of the proposed model 0 Precision
E. Feature Attribution GRU CNN CNN– GRU CNN CNN
The process of determining the impact of the features in GRU –
predicting the outcome is termed as feature attribution. It is GRU
used in extracting the most relevant features contributing to
the model prediction. We apply permutation feature Figure 5: Performance results of proposed models
importance technique for evaluating the importance of the The graphical representation based on the evaluation
unique features of the proposed hybrid model. The calculated metrics like precision, specificity, sensitivity, and accuracy is
performance metrics of the hybrid model is used as a baseline represented in Fig.5. It clearly states that CNN-GRU hybrid
for this model. At a particular time individual feature model competently overcomes the alternative proposed models

3618
IJRITCC | September 2023, Available @ http://www.ijritcc.org
International Journal on Recent and Innovation Trends in Computing and Communication
ISSN: 2321-8169 Volume: 11 Issue: 9
Article Received: 25 July 2023 Revised: 12 September 2023 Accepted: 30 September 2023
___________________________________________________________________________________________________________________

with enhanced accuracy score of 87.24% and the other models Deep-learning model using hybrid adaptive trend estimated
CNN and GRU are lesser than 80%. series for modelling and forecasting sales. Annals of
Operations Research. 10.1007/s10479-022-04838-6.
[11] Chunqiang Lu, Gang Shang, Liyun Xu, Huan
V. CONCLUSION Shao and Beikun Zhang , Sales Volume Forecast of Typical
The necessity of product return rate prediction cannot be Auto Parts Based on BiGRU: A Case Study, E3S Web
overstated as e-commerce businesses continue to thrive and Conf., 409 (2023) 04008. DOI:
https://doi.org/10.1051/e3sconf/202340904008
expand. The proposed approach has the potential to empower [12] Ming-Chuan Chiu, Hsin-Wei Hsu, Ke-Sin Chen, Chih-Yuan
online e-commerce retailers by allowing them to make hands- Wen,A hybrid CNN-GRU based probabilistic model for
on steps to lesser the return rates, to improve customer load forecasting from individual household to commercial
happiness, and restructure their business processes. The model building, Energy Reports,Volume 9, Supplement
10,2023,Pages 94-105,ISSN 2352-
serves as an intended benefit in the competitive e-commerce 4847,https://doi.org/10.1016/j.egyr.2023.05.090.
business zone in addition to being a forecasting system. The [13] Dua N, Singh S, Semwal V and Challa S. (2022). Inception
CNN-GRU model acts as a base for prospect research and inspired CNN-GRU hybrid network for human activity
development. Integrating of further data sources, research of recognition. Multimedia Tools and
alternate deep learning architectures, and more advanced Applications. 10.1007/s11042-021-11885-x. 82:4. (5369-
5403). Online publication date: 1-Feb-2023.
approach for model analysis and optimization may be [14] Zafar N, Haq IU, Chughtai J-u-R, Shafiq O. Applying
incorporated in future revisions. In an active e-commerce world Hybrid Lstm-Gru Model Based on Heterogeneous Data
where consumer experience and operational effectiveness are Sources for Traffic Speed Prediction in Urban
vital, the hybrid CNN-GRU approach is a solid choice. It is Areas. Sensors. 2022; 22(9):3348.
ready to support online businesses in their efforts to prosper, https://doi.org/10.3390/s22093348
[15] Song H, Choi H. Forecasting Stock Market Indices Using
provide a faultless shopping experience, and progress their the Recurrent Neural Network Based Hybrid Models: CNN-
return handling measures. LSTM, GRU-CNN, and Ensemble Models. Applied
Sciences. 2023; 13(7):4644.
https://doi.org/10.3390/app13074644
REFERENCES [16] Liu, Xia. (2022). E-Commerce Precision Marketing Model
Based on Convolutional Neural Network. Scientific
[1] Priya Ambilkar, Vishwas Dohale, Angappa Gunasekaran & Programming. 2022. 1-11. 10.1155/2022/4000171.
Vijay Bilolikar (2022) Product returns management: a [17] Z. Zhuo, S. Chen, and K. Y. Chau, “A new model of
comprehensive review and future research manufacturer's optimal product supply strategy in the
agenda, International Journal of Production context of precision marketing: based on real demand
Research, 60:12, 3920- pattern,” Mathematical Problems in Engineering, vol. 59,
3944, DOI: 10.1080/00207543.2021.1933645 no. 02, pp. 62–98, 2020.
[2] Li Y, Martínez-López FJ, Feng C, Chen Y. Green [18] Adrian MICU & Marius GERU & Alexandru CAPATINA
Communication for More Package-Free Ecommerce & Constantin AVRAM & Robert RUSU & Andrei
Returns. Journal of Theoretical and Applied Electronic Alexandru PANAIT, 2019. "Leveraging e-Commerce
Commerce Research. 2022; 17(4):1450-1472. Performance through Machine Learning
https://doi.org/10.3390/jtaer17040073 Algorithms," Economics and Applied Informatics, "Dunarea
[3] M. Radhi and G. Zhang, “Optimal cross-channel return de Jos" University of Galati, Faculty of Economics and
policy in dual-channel retailing systems,” International Business Administration, issue 2, pages 162-171. DOI:
Journal of Production Economics, vol. 210, pp. 184–198, https://doi.org/10.35219/eai1584040947
2019. [19] Petroșanu D-M, Pîrjan A, Căruţaşu G, Tăbușcă A, Zirra D-
[4] Qingyun Xu, Zhen Shao, Lin Zhang, Yi He, Optimal L, Perju-Mitran A. E-Commerce Sales Revenues
livestream selling strategy with buy-online-and-return-in- Forecasting by Means of Dynamically Designing,
store, Electronic Commerce Research and Developing and Validating a Directed Acyclic Graph
Applications,Volume 61,2023,101307,ISSN 1567-4223 (DAG) Network for Deep Learning. Electronics. 2022;
https://doi.org/10.1016/j.elerap.2023.101307 11(18):2940. https://doi.org/10.3390/electronics11182940
[5] Rajasekaran, V., & Priyadarshini, R. (2021). An e- [20] Chen X, Long Z. E-Commerce Enterprises Financial Risk
Commerce prototype for predicting the product return Prediction Based on FA-PSO-LSTM Neural Network Deep
phenomenon using optimization and regression Learning Model. Sustainability. 2023; 15(7):5882.
techniques. Communications in Computer and Information https://doi.org/10.3390/su15075882
Science, 230-240. https://doi.org/10.1007/978-3-030-88244-
0_22
[6] Han, H. (2019) Review and Prospect on Return Problems of
E-Commerce Platform. Open Journal of Business and
Management, 7, 837-847. doi: 10.4236/ojbm.2019.72057.
[7] Bieniek, M. Returns handling in e-commerce: How to avoid
demand negativity in supply chain contracts with
returns?. Electron Commer Res (2023).
https://doi.org/10.1007/s10660-023-09689-2
[8] Dobroselskyi, M., Madleňák, R., & Laitkep, D. (2021).
Analysis of return logistics in e-Commerce companies on
the example of the Slovak Republic. Transportation
Research Procedia, 55, 318-
325. https://doi.org/10.1016/j.trpro.2021.06.037
[9] Yi Liu, Jiahuan Lu, Jie Yang, Feng Mao. Sentiment analysis
for e-commerce product reviews by deep learning model of
Bert-BiGRU-Softmax[J]. Mathematical Biosciences and
Engineering, 2020, 17(6): 7819-7837. doi:
10.3934/mbe.2020398
[10] Efat, Md & Hájek, Petr & Abedin, M. & Azad, Rahat &
Jaber, Md & Aditya, Shuvra & Hassan, M. Kabir. (2022).

3619
IJRITCC | September 2023, Available @ http://www.ijritcc.org