Enabling “Untact” Culture via Online Product Recommendations: An Optimized Graph-CNN based Approach

Abstract

The COVID-19 pandemic is swiftly changing our behaviors toward online channels across the globe. Cultural patterns of working, thinking, shopping, and use of technology are changing accordingly. Customers are seeking convenience in online shopping. It is the peak time to assist the digital marketplace with right kind of tools and technologies that uses the strategy of click and collect. Session-based recommendation systems have the potential to be equally useful for both the customers and the service providers. These frameworks can foresee customer’s inclinations and interests, by investigating authentic information on their conduct and activities. Various methods exist and are pertinent in various situations. We propose a product recommendation system that uses a graph convolutional neural network (GCN)-based approach to recommend products to users by analyzing their previous interactions. Unlike other conventional techniques, GCN is not widely explored in recommendation systems. Therefore, we propose a variation of GCN that uses optimization strategy for better representation of graphs. Our model uses session-based data to generate patterns. The input patterns are encoded and passed to embedding layer. GCN uses the session graphs as input. The experiments on data show that the optimized GCN (OpGCN) was able to achieve higher prediction rate with around 93% accuracy as compared with simple GCN (around 88%).

1. Introduction

After the COVID-19 outbreak and dire need for social distancing, there has been a tremendous upsurge in the demands for online services and technologies. In South Korea, “untact” has become a key plank after the country’s economy was severely affected by the pandemic. Untact is a Konglish term which means “no contact” and it envisions a way of interaction that involves excessive online connections among people and replaces humans with machines [1,2]. Due to COVID-19, there is a sudden shift of users from offline to online mediums, resulting in high demands for online services such as online shopping, online food delivery, online classes, virtual meetings, and online medical consultation, etc., [3]. In South Korea, with the expanding untact culture, there was recorded 15.02% hike in the usage of food delivery apps from 2.73 million users to 3.14 million users just in the span of 2–3 weeks. Similarly, in the same period of time, the number of online shopping users increased by 7% [4]. All the biggest shopping marts such as Homeplus, Lottemart, and Emart have witnessed a sudden rise in the online members in a very short period, requiring them to ramp up their delivery services. This expansion of online customers is not just limited to shopping and food delivery, Netflix has also witnessed 12.8% increase in users just in the first half of February 2020 [4].

Recommendation systems enable multiple online businesses and web applications to select and acquire interesting information from the systems. Some popular applications are e-commerce, retail, media, banking, telecom, utilities, etc. Therefore, the current pandemic situation has also accelerated the need for recommendation systems which enable users to make decisions in a relatively short time. There are other reasons as well for the demand for recommendation systems such as the rapidly growing internet applications generate a huge amount of data that can easily overload the users with an abundance of information. Recommendation systems become fundamental for redeeming users from such tiresomeness by alleviating the challenges caused by information overload. These recommendation systems effortlessly become a reason for the increase in sales. For example, Amazon and Netflix are taking advantage of recommendation systems by witnessing a 35% increase of purchases and a 75% increase of views on Amazon and Netflix, respectively [5].

Recommendation systems are a specific type of information filtering, which makes use of user’s past activities and user comparisons to produce a rundown of information list that is customized according to the end user’s requirements and preferences.

The majority of existing recommendation frameworks depends on the user’s historical data. Nonetheless, in numerous situations, the identity of the user might be unknown and only the actions of the user in the ongoing session are known. There are some session-based recommendation techniques proposed to handle these scenarios where the system can predict the next action of the user, e.g., in case of product recommendation the next click of the user on an item can be predicted based on his previous sequences of clicks.

In this paper, we propose a product recommendation system for an e-shopping mall known as ejeju mall. This recommendation system uses the history data of user i.e., click sequences and recommend the item that the user is most likely to purchase. The state-of-the-art methods for recommendation systems assumes that user preferences and history data is completely known. However, in the real world, user preferences are very dynamic and can alter based on the user’s mood. Therefore, the smart and intelligent modeling of user behaviors is required. We propose a Graph Convolutional Neural Network-based recommendation model that overcomes the limitations of previously proposed methods and it also considers the network structure. The proposed method optimizes the network structure for better representation of nodes.

We propose a session-based optimized Graph Convolutional Neural Network (GCN) model that aims to capture and represent user behaviors in a more dynamic way. The sequence of clicks for each user in a session is recorded. These sequences are used to generate session graphs. These session graphs are directed graphs that represent the order of the clicks with time t and the last node represents the item that the user has purchased in a session. These session graphs are then passed through the embedding layer where embeddings for each graph are generated; these embedded graphs are used as input for the GCN layer. Figure 1 presents the idea of the session graphs built on the user’s clicks order. The left side of the figure shows M users and sequences of clicks for each user. The products are represented by the colored circles, i.e., v1, v2, … vn. The directed arrows between products represent the order in which they are clicked by the user. We can also refer to them as user’s interaction patterns with different products.

The right side of Figure 1 shows the session graphs generated based on the click sequences of each user. In these session graphs, the chronological order of the clicks is preserved. The number of paths to an item is also important. If there are multiple paths to a product vi, it means it’s been clicked multiple times so the users might be more interested in this product. This process is repeated for all the users and sessions. Consequently, the model aims to capture the complex interactions between products, multiple sessions of a user, and sessions of multiple users.

We have different data processing units; first, the raw data is processed to remove the anomalies and brings the data format in a consistent format. Then different other features are derived from this data such as the user’s location, clicks per session, and purchased items per session, etc. These input features are then combined to generate the complete dataset for GCN processing. We used two different GCN for experiments. The first GCN in a basic GCN model used in previous studies, and then we optimized the graph structure in the other implementation of GCN. The comparisons of the results show that based on the data the network structure should be optimized accordingly to enhance the performance of the model.

In the rest of this paper, we first perform a review of the literature related to this work. The next section explains the proposed methodology in detail. After this experimental details and results are presented and the paper is concluded with some discussions and concluding remarks.

2. Related Works

In this section, we present the literature review of the related works. We have divided the related works into three sub-sections. In Section 2.1, we present related works on recommendation systems; Section 2.2 presents research related to deep learning in product-based recommendation systems and Section 2.3 briefly presents work done on GCN in recommendation systems.

2.1. Recommendation Systems

Product-based recommendation systems are tools that provide suggestions to users about items they might be interested in [6]. Collaborative filtering is one the effective approaches in building recommendation systems. When recommendation is based on the preference of the user it is known as the memory-based collaborative filtering and representation of knowledge, graphs and rules are known as model-based collaborative filtering [7,8]. The other approach is the content-based filtering where the system compares between the content of the item and the profile of the user. Recommendation is based on the previous activities of the user [9,10]. The hybrid recommendation system approach is the combination of both the above approaches.

E-commerce sites are emerging every day and it has become a part of everyone’s daily life. Hassle free shopping in the comfort of home with personalized recommendations is the main advantage of online shopping. Various algorithms have been developed to enhance the recommendation mechanism that combines both implicit and explicit feedback of the customer using matrix co-factorization [11]. Another factor that can impact customers’ choice are reviews by people they trust. In this case it’s important to collect the information about the customer’s social network and to recommend products that might enhance the customer’s trust and interest for a particular product [12]. One of the challenges of building a recommendation system is the huge and diversified dataset. A hybrid approach with neural network framework named as outer product based neural collaborative filtering (ONCF) and convolutional neural collaborative filtering (ConvNCF) along with data mining techniques have been used by researchers to overcome data challenges [13,14,15].

Generative adversarial networks (GAN), named as RecommenderGAN [16], have been one of the frameworks to develop product recommendation systems that generates samples depending on the combination of viewed and bought products. When developing for small scaled retailers with limited dataset and processing power, GAN’s and data or rule mining have proved to be of excellent capability [16,17,18].

2.2. Deep Learning in Product-Based Recommendation Systems

Deep learning has excellent accomplishments in various field of applications. Autoencoder-based approaches are one of the techniques that have contributed well in product-based recommendation [19]. Collaborative filtering for product-based recommendation suffers from cold start and incomplete cold start problems where rating records are either not available or available in much less quantity. To deal with this problem deep learning neural networks can be used for extracting content features [20,21]. Vectors of the user rating and item rating are generally treated as input to the deep neural networks and batch normalization is used for eliminate overfitting [22]. According to [23], movies and e-commerce are the most common domain for which recommendation systems are built.

In [24], the author has developed a tag-aware recommendation system where the user profiles are represented by tags and deep neural networks extracts detailed feature from every layer of the tag. With the same approach, to capture every detail and interest of the user attention-aware recommendation systems were developed using memory components [25]. In addition to it, co-attention network considered the visual information like images to improve the performance of the recommendation system [26]. Convolutional neural network (CNN), recurrent neural network (RNN) and auto-encoders are commonly used techniques for recommending top-N products to individual user and is commonly termed as personal recommendation [27,28,29]. To overcome the problem of negative sampling in CNN and RNN, GAN is generally used while designing recommendation system [30].

2.3. Graph Convolutional Networks for Recommendation Systems

Graph neural networks in general have similar architectures and work on a structured dataset. Recently many researchers have started identifying the efficiency of using graph convolutional networks over traditional networks. Sessions are an important attribute while considering online shopping. In [31,32], researchers worked on session-based graphs where every click was labeled as a sequence. Auto-regressive moving average were implemented for combining a graph convolutional layer into the graph neural network to develop the recommendation framework. Frameworks have also been developed for anonymous users from where we get a limited amount of information about users’ preferences using a deep neural network [33]. We have used Graph-CNN that produced better results in this case.

Matrix operations on graph networks for personalized recommendation using both implicit and explicit feedbacks [34,35,36] are widely used, but in [37], the author has defined a graph attention network that overcomes the challenges faced by the spectral-based graph neural networks. In [38,39,40,41], the authors implemented semi-supervised algorithms on graph structured data. The authors claimed that an approximation of the first order for graph convolutions can extract hidden information and requires low computation time. In this paper we propose an optimized graph convolutional neural network that represents graphs in a more convenient way for product recommendations.

3. Data

In this section, we present and explain the data we used for our experiments. It primarily includes data type, its size, its source and preprocessing. Data selection is an important component of the experimental process. Therefore, we have spent a lot of time on scrutinizing and preprocessing the data. As the aim of this study is to recommend items to users based on their previous click patterns, we have gathered data for an online shopping mart. A description of the data source is given in

Table 1. Data Description and Examples.

Data	Description	Examples
acces_ip	IP address of the user	210.54.66.88
access_date	The date when user is active and browsing	11/02/2020
access_time	The exact time when user clicked a product	17:45:46
access_page	URL of the page where user is currently active	/shop/list.php?ca_id=30
item_id	Unique identification number (ID) of the clicked item	202020
item_name	Name of the clicked product	VERSACE
Item_type	Type of the product clicked	Purchased, clicked or visited

below.

3.1. User’s Clicks Data

We have users’ click history data on different items which includes the previously clicked items so far by the user in one session. User’s clicks are used to understand the behavior of a user in terms of his pattern of exploring different online products. We also refer this data as clicks sequence of a user. This data was then processed to discover unique clicks, purchased items per user and then all the data was encoded into unique item IDs (identification number).

3.2. User’s Personal Data

Besides the click’s sequences for each user, we have the user’s IP address referred as access_ip, the date and time when user clicked a product referred as access_date and access_time respectively. From this access_ip, we derived the location of each user. Similarly, we used access_date and access_time to generate different sessions.

3.3. Product Data

As the recommendation system is aiming to suggest different products to users, the dataset contains some product features as well. These features include product ID (referred as item_id), product name (referred as item_name), and product click type (referred as item_type).

4. Proposed Optimized Graph-CNN-based Approach for Online Product Recommendation

In this section, we explain the proposed session-based recommendation system. First, we present the problem formulation and then present the conceptual framework of the proposed system. This is followed by a complete architectural diagram and explanation. In this section we also define some useful notations and system parameters.

4.1. Problem Definition

This project aims at recommending users with next most likely purchasable product to be clicked. Session-based recommendation aims to predict which item a user would like to click next, only based upon his/her current interaction sequence. A session is referred to a time period where a set of items (e.g., products in our scenario) are clicked or collected or visited by users. Session-based recommendation aims at predicting the future actions of a user in a given session. For example, by having some partial data related to a session (e.g., some history or pattern of clicks), session-based product recommendation system will recommend an item which a user will likely to click next.

Here we give a formulation of the session-based recommendation problem as below.

There are two important factors to be considered, the products and the order in which they are clicked. All the unique products in all sessions are represented in Equation (1):

$$V=\left\{v1, v2, v3,\dots ., v\left|V\right|\right\}$$

(1)

Similarly, the order of their clicks is represented in time order as shown in Equation (2):

$$C=\left\{c1, c2, c3,\dots , cn\right\}$$

(2)

where $cn$ is the latest clicked item and c1 is the very first item clicked by a user in a session. All the clicked items in Equation (1) belong to $V$ which represents a particular item clicked by a user at a certain time. For example, if we have a clicked item in C named as $ct$, it represents an item in $V$ clicked by user at time t. Therefore, the objective of our proposed model is to predict the next possible product click that is most likely to be purchased i.e., $cn$+1. It can be formulated using Equation (3):

$$Ct=\left\{ c1, c2, c3,\dots \dots , ct-1, ct\right\}(1\le t<n))$$

(3)

Precisely, a ranked list of all the possible items that can be clicked and purchased by the user in a session is generated by the model. It shows that the output is the probability of a product to be purchased from the complete list of products in the dataset i.e., V. Hence, we can say that:

$$\widehat{y}i=\left\{\widehat{y}1, \widehat{y}2, \widehat{y}3,\dots , \widehat{y}\left|\mathrm{V}\right|\right\}$$

(4)

where, $\widehat{y}i$ represents the recommendation score of products vi. As, the recommendation system usually aims at suggesting more than one product to the user therefore, we select top k products form Equation (4) for recommendation.

Now, we express this problem in terms of graphs. First, we use the generic representation of a graph in terms of vertices and edges and define an attributed graph as shown in Equation (5):

$$G=\left(V, E, X\right)$$

(5)

where, V is the set of nodes and each node represents a unique product as shown in Equation (1). E is the set of edges between nodes i.e., clicks sequence or order between different products also shown in Equation (2). The attributed graph differs from the traditional graph as it includes an additional parameter as X. X is the feature matrix that represents the features of all the products. Next, we build an adjacency matrix A which also represents the network topology. The entries in the adjacency matrix are either 0 or 1. If there is an edge between two products then the entry will be 1 otherwise it will be 0. The degree of all the nodes is represented in the degree matrix D.

4.2. Session-Based Recommendations Model

The proposed framework is a recommendation framework that uses users’ online clicks on different products and generate the interactions among them by using graph structure. The Figure 2 below shows the conceptual framework of the proposed model. First, we have real data of an online Korean shopping mart, this data can be divided into three different types such as, user data, item data and general data. The user data contains user related features like his IP address, and purchased items etc. The Item data contains item related features such as, ID and name. All other features are added to the general data category.

The data is prepared and processed to use as input for graph representation. Here, the sessions are generated based on clicks sequences of users in chronological order. From these session graphs, adjacency matrix, feature matrix and other general data such as node degree matrix, node frequency etc. are generated. Adjacency matrix represents the relationship between products. The final task is pattern classification and prediction, where the GCN classifier is used to classify the patterns and predict items that are most likely to be purchased given some number of clicks. Finally, the top K products are recommended to user based on their probability of being clicked next.

4.3. System Architecture

A detailed view of the whole procedure and system architecture is shown in Figure 3. There are multiple layers in the architecture. The first layer deals with the input data. After basic processing of the data, it is passed to the data analysis unit, where the data is thoroughly analyzed to understand the data behavior and find some useful insights. After the analysis, some additional features are derived from the raw data. These features include user click history, items purchased per user, items clicked per user etc. the complete data is used to generate sessions.

These generated sessions are used in the next layer as input where the click sequences are transformed into session graphs for each user. To build the session graphs, we followed the method similar to SR-GNN (Session-based Recommendation with Graph Neural Networks), in which the graph structure is based on user’s clicks sequences. These session graphs are then passed to the embedding layer that generates embeddings for each session graph. Each product is embedded in an integrated embedded space and denoted as a node vector. The resulting embedding table can be a representation of latent attributes of the products that is used to depict their fundamental characteristics. Concurrently, the adjacency matrix is used to save the relationships between products in terms of edges. Therefore, there are chances that many edges can occur repeatedly. To solve this issue, edges are normalized by assigning them a normalized weight. This normalized weight is calculated as shown in Equations (6) and (7):

$$Edge\_Frequenc{y}_{\left(x,y\right)}={\displaystyle \sum }_{i=1}^{n}Ei\left(x, y\right)$$

(6)

$$Weigh{t}_{normalized}=\frac{Edge\_Frequenc{y}_{\left(x,y\right)}}{de{g}_x}$$

(7)

where, Ei is an edge between two products x and y. to calculate the occurrences of this edge we check its frequency as Edge_F $requency$_(x,y). In Equation (7), $de{g}_x$ is the degree of the starting node in the edge.

Once, the session graphs as Gs = (Vs, Es) based on Equation (5) along with their respective embeddings are generated, now each node or product v can be mapped to X i.e., the feature matrix containing features of each product. After that E is calculated to generate the adjacency matrix represented as A. We can express the graph as a combination of products with feature and adjacency matrix i.e., G = (V, X, A). GCN works by consistently updating node’s state to its the neighboring nodes and then at some point it converges.

As the dataset used in this work is very large and it contains a large amount of nodes the computational cost can be very high. The generated graphs have lots of features and nodes and it will take a lot of computational time to update the node state to the neighboring nodes. Therefore, we have introduced an additional module that helps in optimizing the network structure and maximizes the weights of product nodes that are purchased by the customer. We optimize the network structure by reducing or removing the nodes that are rarely accessed and are given lowest weightage. We also used Chebyshev polynomial method that is mostly used to reduce the computational complexity of an algorithm [42].

The GCN has L number of layers and each layer is responsible to propagate the information between the product nodes. According to the GCN introduced in [43], the neighboring nodes are treated equally, however, we have introduced an additional module to optimize the network structure and that is responsible to update the weights of nodes based on their likelihood of being purchased. As shown below the Y includes all the products that are being purchased by users. Hence, the objective function tries to maximize the weight of the products that belong to Y referred as t in the Equation (8):

$$Objective function=Maximize\left(Weigh{t}_t\right)$$

(8)

where Weight is the weight for the products being purchased. After this the prediction layer comes, at this level the we compute the probability $\widehat{p}$ or score for each product given that session embedding h_t as shown in the Equation (9) below,

$$\widehat{p}=h_t T\left[xi\right]$$

(9)

where xi denotes the clicked item at time T. After this we calculate the output vector by passing through the softmax layer as shown in Equation (10):

$$\widehat{y}=softmax \left(\widehat{p}\right)$$

(10)

where $\widehat{y}$ represents the probability of recommending product i to be clicked and most likely to be purchased. For reducing the training and testing loss and optimizing, we used the most widely used cross entropy loss. The final optimization or loss function of our model is shown in Equation (11) below:

$$Optimizatio{n}_{Functiom}\left(Loss\right)={\displaystyle \sum }_{i=1}^n{y}_i \log ({\widehat{y}}_i)+(1-y_i) \log \left(1-\widehat{y_i}\right)+\lambda |\left|\theta \right|{|}^2$$

(11)

where the ground truth product is represented by its one-hot vector encoding $\widehat{y}$ all the learnable parameters set is represented by θ. The last factor λ is a regularization parameter on θ, which we use to overcome the challenge of overfitting.

The prediction layer presents the probabilities of all the items to be purchased next. We select the top k products with highest prediction score and recommend them to the user.

4.4. Data Processing

The data processing part as shown in Figure 4 is the most important part of the whole system. This processing unit enables us to understand the underlying patterns of data, basic data structures, data anomalies, regular and irregular behaviors of data. Primarily, there are two tasks of data processing, data analysis and data preparation for experiments. Data analysis learns the basic data patterns, structures and statistics of data such as unique IPs, sessions, clicks per product, sales per user, and sales per session etc. Data analysis is followed by task 2, i.e., data preparation. Here, tasks such as filtering data and generating session patterns are performed. The data filtration process is very important step in data preparation task as here we take crucial decisions such as what records are to be kept and what records should be discarded. The discarded records are usually responsible for anomalies or irregular patterns in data. Finally the session sequences are generated that are based on the patterns of click sequences.

In Figure 5, the process of data categorization is shown. The raw data has specific fields that reflect the product type or category. We generate different categories of products based on the access pages. There are some clicks on the home page, then user can click on any category of product which is referred as a list in the data, it means user has chosen a generic product category e.g., Food. The list or category of product can be nested e.g., Food can be categorized into Fruits, Vegetables, Meat or Snacks, etc. Each category can have multiple products in it.

Then user can click on any specific product in that particular category referred as an item, e.g., user can click apple from fruits category which is a subcategory of Food. User can click on any item and just visit it e.g., just go through the description, or he can purchase that item.

5. Implementation and Testing Environment

In this section, we present the implementation environment in detail. This section includes experimental setup details, explanation of data collection process and data description, and the model structure. First, we describe the dataset used for experiments and the parameter selection. We compared the GCN model with our optimized GCN. We also compared the results with other baseline model and SR-GCN model.

5.1. Experimental Setup

The experimental setup is summarized in

Table 2. System’s components and specification.

System Component	Specification
Operating System	Ubuntu 18.04.1 LTS
Memory	32 Gb
Language	Python
GPU	Nvidia GForce 1080
Language Version	3.6.1
APIs	Tensorflow 1.13,
Key Liberates	Numpy, skearn, Pandas
Encoding	EUC-KR

. The core system components include a long-term support (LTS) version of Ubuntu (Version 18.04.1, Canonical, London, United Kingdom, 2018) as an operating system, with 32 Gb memory, and graphics processing unit (GPU) is GeForce 1080 (Nvidia, Santa Clara, CA, United States). The implementation is done in Python language along with Tensorflow API (Version 1.13, Google Inc., Mountain View, CA, United States, 2019).

5.2. Experimental Data

The details for the experimental data are presented in

Table 3. Characteristics of data used for experiments.

Experimental Data	Details
Total no. of Records	294894
Data Duration	(1 year) 2019-2020
Total No. of Unique Users	51386

. For all experiments, we divided the data into 60% training data 20% validation data and 20% testing data. The total number of 294894 records. The raw data is collected in log files as shown in Figure 6 which we processed to generate session patterns.

This data is very recent and is of around 1 year form Jun 2019 to July 2020. There are in total 51386 unique IP addresses in the data.

For experimental purposes we have used three different kind of settings. The data records are built accordingly. We have experimented on different lengths of click sequences. For this purpose, we have generated three types of records each containing different numbers of user clicks sequence lengths. The minimum size of all click sequences is considered to be 3 and the maximum size is varied till the maximum number of clicks in a session by any user i.e. 40. The following

Table 4. Data characteristics based on click sequence length.

Experimental Data	Clicks Sequence Length (Min to Max)
	3 to 15	3 to 25	3 to 40
Total rows	294894	294894	294894
Discarded rows (len < Min)	1440	2596	2321
Useful Rows	293454	292298	292573
Expansion rows (>Max)	3013	3013	3013

describes the data setting in detail.

We have performed experiments based on the GCN, the optimized GCN. The parameters are shown in the

Table 5. Item status type vs click frequency.

Item Status Type	Clicked Frequency
Clicked Items	244533
Visited description	42630
Purchased	7701

below.

5.3. Evaluation Metrics

Here, we list down the evaluation metrics used to access the system performance:

(1) Accuracy: We calculated accuracy of the model by using following formula as shown in Equation (12):

$$Accuracy=1-\frac{||Y,-,\widehat{Y}||F}{||Y|| F}$$

(12)

(2) Root Mean Square Error (RMSE): We calculated RMSE by using following Equation (13):

$$\mathrm{RMSE}=\sqrt{\frac{1}{MN} {\displaystyle \sum }_{j=1}^M{{\displaystyle \sum }}_{i=1}^N {\left(y{i}^j-{\widehat{yi}}^j\right)}^2}$$

(13)

where $y{i}^j and {\widehat{yi}}^j$ are subsets of Y & $\widehat{Y}$ and represent the actual user click data and predicted click of the jth time sample in the ith session respectively. M is total time samples, and N is number of products. RMSE is specifically used to estimate the prediction error. The smaller the value of RMSE is the better is prediction rate or score according to Equation (14):

$$Prediction rate \propto \frac{1}{\mathrm{RMSE} }$$

(14)

while accuracy is used to detect the precision of predictions and it has an opposite effect than RMSE on prediction rate as shown in Equation (15). The higher the value of accuracy is, the better is prediction rate:

$$Prediction rate \propto Accuracy$$

(15)

(3) Mean Reciprocal Rank (MRR@10): MRR is defined as an average of reciprocal ranks of the correctly recommended products. In our case we used top 10 products i.e., the reciprocal rank becomes 0 when the actual rank surpasses 10. The purpose of using MRR is to capture the order of recommendation ranking. Therefore, a correct recommendation in the top of the ranking list if it has a larger value of MRR.

5.4. Model Parameters

Here, we present the model parameters used while GCN and optimized GCN training. These parameters mainly include learning rate, training epochs, batch size, hidden units’ size etc.

As shown in

Table 6. Model parameters and their values.

Model Parameters	Value Range
Learning rate	0.01, 0.001, 0.0001
Number of Epochs	1000 to 5000
Batch size	25, 32,45
Hidden units	8, 16, 32, 64, 100, 128

, we experimented with different learning rates as 0.01, 0.001 and 0.0001. We also varied the number of epochs between 1000 to 5000. The used different batch sizes as 25, 32 and 45. Finally choosing the number of hidden unit is crucial as different sizes can have different impact on the prediction accuracy. Therefore, we experimented with different numbers of hidden units such as 8, 16, 32, 64, 100 and 128.

6. Results

This section presents the results and analysis for online product recommendations based on user’s previous click sequences and purchased items. In Section 6.1, we present the testing errors. In Section 6.2, we report on the accuracy of the results of the proposed model in comparison with other models.

6.1. Impact of Hidden Units Selction

As selection of hidden units can impact on different parameters of system’s performance therefore, we performed some experiments by varying the number of hidden units.

The graph in Figure 7 shows the RMSE estimated for both GCN and optimized GCN by varying number of hidden units. The number of hidden units are important and right selection of hidden units can enhance the prediction effect. As we can see, the optimal number of hidden units in our scenario are 100 as both models show their lowest RMSE value when the hidden units selected are 100. The RMSE of OpGCN is vividly very less as compared with simple GCN.

The graph in Figure 8 shows the accuracy estimated for both GCN referred as Acc-GCN and optimized GCN referred as Acc_OpGCN by varying number of hidden units. Here also we can observe that the optimal number of hidden units are 100 as the accuracy recorded at this point is highest for both the algorithms. The highest accuracy achieved by GCN is around 0.89 and for optimized GCN is around 0.93.

In summary, the prediction results are observed to be better if the hidden units are set to 100. We can also observe that when we increase the number of hidden units, the RMSE first starts to drop and then after some point it starts increasing again. Similar pattern can be seen in prediction accuracy. This happens when the hidden unit is bigger than a certain number that increases the model’s complexity and imposes additional computational challenges. And as a result, overfitting on training data happens. Therefore, we used optimal value of 100 for our remaining experiments.

6.2. Accuracy Vs. Number of Epochs

Here, we compare the training and testing accuracies on different number of epochs for both the models.

Figure 9 and Figure 10 show that the optimized GCN succeeded in achieving higher accuracy of prediction both in training and testing as compared with a simple GCN model. The accuracy of GCN is reffered as Acc_GCN_train and Acc_GCN_test for training and test data respectively. Similarly, the accuracy for OpGCN is referred as Acc_OpGCN_train and Acc_OpGCN_test for training and test data respectively. This proves that the optimization of nodes structure does impact the overall performance of the system. OpGCN achieved around 92% prediction accuracy while GCN achieved around 88%.

6.3. Training and Testing Loss

Here, we compare the training and testing loss on different number of epochs for both the models.

The graph in Figure 11 shows the training and testing loss comparison of the predicted results for both the simple GCN model and the proposed variation of GCN, i.e., the optimized GCN model. The testing loss of GCN is higher as compared with the training and testing loss of the optimized GCN model.

6.4. Impact of Click Sequence Length

As a user can click as many products as he wants, therefore, it is very important to find the optimal length of a click sequence for better recommendations.

Therefore, we generate different records based on the maximum length of a click sequence. The maximum length of clicks recorded in one session are 40. We vary the maximum length of the click sequence i.e., 15, 25 and 40. The graph in Figure 12 shows that the model’s performance is enhanced when the number of clicks are between 15 to 20. The accuracy starts decreasing when the length of the clicks increases i.e., for 40 clicks the model’s performance is very poor.

6.5. Comparisons with Baseline Methods

We compare our model with other state of the art methods to evaluate the overall performance of the proposed method. The overall performance is measured in terms of accuracy, MRR@10 and RMSE. We also used other datasets for experiments. These data sets include a citation network and a product network named as Cora [44] and Diginetica [45], respectively.

We used the following baseline models to compare the performance of our proposed method:

(1)

POP [46]: This model ranks items based on their popularity and recommends the n-most popular items to the user.

(2)

HRM [47]: This model introduces both linear and non-linear pooling operation to combine the transaction history of the user.

(3)

NARM [48]: This model uses a global and local recurrent neural network based recommender that captures the sequential behaviors of users.

Table 7. Performance comparison of different baseline methods on different datasets.

Datasets	Type	Evaluation Metrics	POP	HRM	NARM	OpGCN
Cora	Citation Network	Accuracy	0.76	0.84	0.81	0.93
		RMSE	6.50	5.42	5.81	4.52
		MRR@10	21.87	33.45	29.56	34.72
Diginetica	Product Network	Accuracy	0.73	0.85	0.83	0.92
		RMSE	6.12	5.36	5.11	4.50
		MRR@10	23.85	31.55	29.01	36.78
Our dataset	Product Network	Accuracy	0.72	0.83	0.82	0.93
		RMSE	5.51	5.92	5.35	4.50
		MRR@10	22.57	29.74	31.67	37.91

shows the values for the evaluation metrics for Diginetica, Cora and our dataset. According to the evaluation results, our proposed model OpGCN achieves an average accuracy of around 93%, RMSE score of 0.08 and mean reciprocal rank of 36.47 that depicts our model performs better than the existing models.

7. Conclusions

We propose a model based on graph-CNN for session-based product recommendations in online marketplaces. The proposed model comprehensively emphasizes the customer’s interests and previous click patterns. After extensive experiments and parameter tuning, our model learns the complex relationships between product data, network structure and as well as the temporal information. With the simple GCN model the system finds it difficult to get more than 80% prediction accuracy. Also, without fine tuning the model’s structure, it becomes computationally rich and complex. That affects the performance of the model in terms of low prediction rates and model overfitting. Therefore, we propose another variation by introducing an optimization module. This module is incorporated with the simple GCN model, it learns the optimal values for each node and then updates the weights accordingly. During the process, records with less than three clicks are discarded along with the records that show some abnormal patterns.

Specifically, we first process the log data files, the random data records are converted to user click sequences. These click sequences of users are converted to anonymous user sessions. Then directed graphs are constructed from these sessions referred as session graphs. The products are encoded and are later converted into nodes. Each node represents an item. These session graphs are embedded, and we use them as input to the GCN model. It basically takes click sequences in form of session graph and each session graph has a destination node i.e., a final product bought by the user. If a user has multiple items bought in the same session, we generate separate paths for each destination node. This way all the dependencies in the data are learnt. We use an objective function that maximized the weights of our desired nodes and hence, minimize the weights of undesired nodes. This way the graph structures become less computationally complex. Then the prediction layer predicts the likelihood of each product to be clicked and purchased by the user. The recommendation module selects top k items for the user.

We have different sizes of session graphs based on the maximum number of clicks made by a user in a session. In experiments, we tune the maximum clicks and evaluate the performance of the model. We have learnt that optimal size of click sequence is 15 to 20 clicks. If number of clicks increases beyond 20, the performance of the systems starts degrading. This means that we find the best possible structure of the graph with appropriate number of nodes that are the best representative of the whole data. If a system is successful in optimizing the graph structure, it can drastically improve the performance of the recommendation system. We have figured out the importance of the graph structure and how it can uplift or enhance a systems performance.

Therefore, we performed experiments with different settings, we generated graph structures with different tuning parameters such as the nodes selection, fixing sequence length, optimal weights of the nodes etc. The experiments show that our proposed variation performed better on the same data with same experimental settings. The network structure is the key part, based on different decisions on data representation, records generation, session building and graph representation. In future, we will focus on more variations of GCN model and on the effective network structure. The network structure can be highly altered by changing the data records, graph structure such as directed or undirected, and session records etc. We will also be focusing on hybrid approaches of GCN with other state of the art approaches in order to handle large scale data and for representation learning to achieve more effective and interpretable recommendations.

8. Contributions and Future Directions

In this paper we have illustrated the significance of online marketplaces that are playing their parts in enabling untact culture, especially after the COVID-19 pandemic. Recommender systems are one key component of this whole cyberworld chain as they are equally effective and lucrative for both customers and business owners. Therefore, there is a dire need for effective and intelligent recommendation systems that can enhance the online buying process helping customers with their decision-making processes. We propose a model based on graph structures for making product recommendations to users. The model is based on GCN which is a relatively new approach for recommendation systems and requires a lot more to discover by varying the network structure. We contribute by proposing an optimized variation to it. This enables an effective network structure and hence improves the performance. We also experiment with different lengths of user click sequences and find the optimal number of clicks. It is essential for a recommendation system to assess the user behavior proficiently to make effective recommendations. The purpose of experimenting with different click sequence lengths was to get an understanding of how many clicks it can take for an effective recommendation. As there are cases where a user makes too many clicks in a single session, using all those clicks might not be a smart design choice as there can be hidden click sequences inside a pattern. Therefore, an optimal click sequence length is an important factor to be considered when making design choices for the model.

There are numerous promising opportunities for future work. The proposed model can be applied in other recommendation systems such as movies, books or online product recommendations where we can first learn a user’s click pattern and then can add other context vectors such as price, season, brand, range, reviews, and location and so on. It can also be interesting to see how user behaviors and preferences change or develop overtime under specific circumstances. We would like to evaluate our proposed model in a more realistic setting, e.g., examine how users interact or make preferences based on a specific condition, situation or feature. Recommendation systems with better context understanding and awareness are expected to generate effective recommendations by considering all the relevant contextual features of users and environment as well. Currently, though we are using user clicks as a primary form of input, it would be an interesting addition to extend the features list by including product based and user profile based features. As, recommendation systems are to provide efficient services to users by learning their characteristics and behaviors, therefore, user profile based features, environmental factors, product features, and other relevant contextual features can be an interesting addition to the model in the future.

Moreover, for the current coronavirus pandemic, we would like to evaluate our model on large data collected from different online shops during this period and critically analyze its impact on the habitual shift of customers form offline to online platforms.