APPLICATION OF CONVOLUTIONAL NEURAL NETWORKS IN E-COMMERCE MARKETING STRATEGY OPTIMIZATION

Authors :

Omar Al-Hassan and Yasmin Al-Nuaimi

Address :

Department of Computer Networks, American University in Dubai, UAE

Department of Artificial Intelligence, Zayed University,UAE

Abstract :

E-commerce platforms collect massive amounts of data, including product images, user interactions and purchase activity. These are usually underutilized in marketing decision-making processes. The fast rise in e-commerce has brought further challenges in optimising marketing strategies toward customer attraction, interaction, and retention. Inefficient data-driven strategies have caused suboptimal efforts in marketing, thus resulting in lower satisfaction among customers and reduced conversion rates. Conventional marketing techniques can rarely take advantage of massive data generated by e-commerce, resulting in low efficiency in customer targeting and campaign effectiveness. This paper presents a novel method called ECMS-CNN that applies a Convolutional Neural Network for optimizing E-Commerce Marketing Strategies focused on three key areas: product image analysis, personalized recommendations, and dynamic ad targeting. The CNN model would be trained on large datasets such as Amazon Product Reviews to identify patterns in product images, customer preference, and purchase behavior. Further, the model will be integrated into a marketing pipeline to automate product tagging, personalized recommendations, and real-time ad targeting, all supported by visual and behavioral data. The application of CNN significantly enhances marketing strategies for effectiveness and efficiency. These experiments demonstrate the accuracy of personalized product recommendations increased by up to 20%, and ad targeting became more efficient by 25%. Additionally, product image analysis with the application of CNNs enabled simplification in the tagging process by taking 30% less manual effort and providing more relevant suggestions to the customers. These results demonstrate the role of CNNs in increasing conversion rates, Mean Absolute Error, and Click-Through Rate on e-commerce platforms.

1.Introduction

Buying and selling products and services online, or "e-commerce," necessitates the transmission of both financial and personal information. This is at the head of new technology-based changes within marketing strategies, where enabling product information facilities improves decision-making [1]. The e-marketing platform is a part of business operations that, in the present times, is used to launch products, develop consumer loyalty, and disseminate information. Under such circumstances, the development of business strategies online, in these times of digital media, is certainly an opportunity for entrepreneurs to market their products online to fulfil various human needs, from primary to secondary needs [2]. It becomes quite difficult to attract and hold the attention of consumers. Consumer attention is becoming increasingly limited in the contemporary media environment's information explosion [3]. Efficient e-commerce marketing management can boost product and brand awareness and competitiveness, but all marketing efforts targeting the same demographic will use the same data obtained from the same platform [4]. While the market scale and commercial format of e-commerce become more complete, the importance of customer service perception has been increasingly prominent. Efficient logistics distribution capability and high-quality logistics services will be one of the focuses in the next round of competition among e-commerce giants [5].

A well-articulated digital marketing strategy will help a firm in solving several problems, such as:

Daily business objectives and priorities are established.

Evaluate the means at one's disposal for achieving the objectives.

Align activities of multiple organizational units to create cohesive brand management across diverse digital landscapes.

Selection of appropriate partner and management of their work [6].

Figure 1 shows the Key Elements in E-Commerce Marketing Strategies.

With rapid growth in information technology, e-commerce is one of the online services that may speed up and increase sales. Customers can access the E-commerce service and order from different locations [7]. Besides the advantages, the e-commerce sector also faces some challenges. The major issues facing E-commerce are identity theft, payment fraud, and other cybercrimes. Due to the tendency of cyber-attacks, e-commerce sites lead to data breaches and customer loss of trust [8].

Advanced pattern recognition capabilities, especially in big datasets with high dimensionality, are some of how CNNs are changing marketing strategies in e-commerce. CNNs can tag product images with attributes like colour or style, increasing catalogue accuracy and speeding up the process. They also enrich personalized recommendations by analysing visual and behavioral data to align products with customer preferences for higher satisfaction and conversion rates. Through dynamic ad targeting, CNNs predict that the product or ad will likely engage the users in interaction by improving ad relevance and return on investment. In addition, with the integration of CNN with NLP, extensive sentiment analyses of customer reviews provide valuable insights into businesses to refine product offerings and enhance customer experiences.

This study's primary contribution is

• To automate product image tagging, CNNs categorize e-commerce images efficiently, reducing manual input and improving catalogue management.

• To enhance recommendations, CNNs combine visual and behavioural data, increasing personalized product suggestion accuracy and relevancy.

• CNNs analyze user interactions to optimize ad targeting, creating dynamic ads that improve conversion rates and marketing return on investment.

• To extract sentiment insights, a CNN approach analyzes images and text in reviews, revealing deeper customer preferences and performance feedback.

• To evaluate these methods, large datasets like Amazon Reviews show improved recommendations, ad targeting, and customer satisfaction metrics.

The remaining portion of the paper follows this structure: Section 3 details the suggested methodology's overall design. Section 4 contrasts the suggested methodology's outcomes with those of more conventional approaches. Part 5 concludes the research.

2. Literature Survey

To construct a deep attention layer, Sethi V. et al. [9] suggested the LCNA model, which combined CNN with Long-Short-Term Memory Networks, to improve recommendation systems for e-commerce. It resolves challenges such as cold start and sparse datasets based on semantic ranking and collaborative filtering. The results depict the enhanced metrics MAE, RMSE, and accuracy on the Amazon electronics dataset, proving the model's effectiveness in generating valid recommendations.

Tang, H et al. [10] proposed that the DL technique is appropriate in an e-commerce recommender system because of its effectiveness. It is also recommended for "information overload," effective marketing strategy optimization, and a deep understanding of customers' preferences using advanced data characterization and analysis. Improved personalized recommendations, better user engagement, and the formulation of enterprise marketing strategies in a structured manner result in enhanced marketing outcomes and user satisfaction.

Yasir, A. A. et al. [11] made the development and optimization of the sales prediction model in the e-commerce industry using CNN and RNN. The CNN model detects the spatial patterns of the sales data, while the RNN model detects the temporal patterns. Evaluation results show that CNN and RNN perform very well with low loss values. Application of the models shows that both models are reliable in predicting sales based on historical and external data.

Jha, A. et al. [12] proposed a ML-based model for optimising e-commerce advertising campaigns, focusing on Amazon. This used K-means clustering, a Decision Tree Classifier, and probabilistic methods to analyze campaign performance and predict profitability. The results improved the Return on Ad Spend (ROAS) prediction accuracy from 87.98% to 90.92%, showing the positive relationship between keyword bids and profitability, enhancing decision-making for ad campaign strategies.

Devita M. et al. [13] discussed Shopee's marketing strategy in international business, emphasizing public relationship marketing and the ability to retain consumer loyalty. They adopted a descriptive method and qualitative approach based on secondary data collection through a review of related literature to validate the findings. The result of this paper indicated that Shopee can combine push and pull strategies that build consumer loyalty and raise brand awareness. Some of the strategies adopted by Shopee, such as Shopee Streamers Academy, contribute positively toward the perception of the public and the sales record.

Vavliakis. K. N. et al. [14] have developed a hybrid model of LSTM (Long short-term memory) and ARIMA (autoregressive integrated moving average) networks for sales forecasting. In such a manner, it leverages the strengths of ARIMA in data linearity and the capability of an LSTM to analyze nonlinear residuals to improve prediction accuracy within e-commerce. Based on this, the model subsequently captured prominent improvements of 5.82%, 5.29%, and 11.44% on Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Square Error (MSE), respectively, over those of ARIMA, LSTM, and Zhang's model, thereby proving efficient for different sales scenarios.

Shobana J. et al. [15] proposed a framework for detecting probable client loss in the B2C E-commerce business using ML-based support vector machines and producing personalized retention strategies using hybrid recommendation techniques. The nonlinear classification capability of the SVM is utilized to detect Churn. In contrast, hybrid recommendation techniques such as Collaborative, Content-Based, Knowledge-Based and Demographic Techniques generate tailored retention actions. The combination of SVM with the hybrid recommendation model demonstrated higher churn prediction and personalized retention performance than the individual models while improving the hit rate, coverage rate, lift degree, and precision rate.

Kedi, W. E. et al. [16] recommended including machine learning software in enhancing small and medium-sized enterprises (SME) social media marketing campaigns. The approach can analyzed user data and automate the process while giving a better understanding of the targeted audience. It utilized data and algorithm-driven insights to enhance better targeting and engagement of the audience. This enabled SMEs to obtain better marketing output, more engaged customers, and a higher return on their campaign investment.

3. Proposed Methodology

a. CNN Layer

The main structure of the convolutional neural network and a stack of these basic structures can also be modified based on different tasks. To optimize a marketing strategy (MS), researchers first run the input strategy through a convolution operation; the output of the prior layer is used as the input for the subsequent layer. This marketing strategy feature information will be continuously acted on by the convolutional operation of the multi-layer network, which, during this process, will gradually learn some features. Therefore, when measuring semantic similarity, it is essential to use the general characteristics of the network. Figure 2 shows the Primary Convolutional Neural Network Architecture.

Each layer in a CNN used for the marketing of e-commerce plays an important role in process optimization, such as the recommendation of products or ad targeting. The raw data of the input sets, such as product images or customer interaction, is sent through an input layer to prepare it to feed into the system. The convolutional layer then applies filters to identify key features in product images, such as colour or style, which could be used to catalogue or make recommendations. Non-linearity is introduced with the help of the activation layer, which helps the network learn complex patterns in visual and behavioural data to improve personalization and targeting. The pooling layer reduces the dimensionality of the data but retains the most important features, such as specific product attributes, simplifying computation. The fully connected layer integrates all learned features for final decisions, like product classification or ad recommendations. In the output layer, the outcome gets generated based on which products to recommend or which ads to target individual customers based on their behaviour and preferences.

b. Overall Working Process of the Proposed ECMS-CNN Methodology

The E-Commerce Marketing Strategy using Convolutional Neural Networks (ECMS-CNN) methodology) upgrades e-commerce marketing activities by using CNN to process image and behavioral data. Figure 3 shows the proposed method's overall working process and involves the following steps.

a) Data Collection

The model's input is the images from the datasets composing the Amazon Product Reviews dataset, which include images of products, user interactions, purchasing habits, and textual reviews.

b) Data Pre-Processing

Success with CNNs depends hugely upon the quality of the input data. Here, the data ranges from image data to textual and behavioural data in the form of reviews, browsing history, and purchase history. Preprocessing will involve:

Image data Preprocessing: For example, all the product images should be resized to a fixed size, say 224 × 224 pixels, for consistency and normalization of pixel values between 0 and 1. This is achieved using Equation 1.

Data augmentation is a technique that involves rotating and flipping training data to increase its diversity and prevent data overfitting.

c) CNN Model Architecture

The CNN model architecture will process both behavioural and textual inputs from images to handle the tasks of product tagging, personalized recommendation, and ad targeting in e-commerce marketing.

Convolutional layer: Convolutional layers apply filters (kernels) to the input image to select features, which may be in the form of edges, textures, or shapes. The mathematical definition of the Convolutional operation is given in equation 3.

where I is the input image, K is the convolutional kernel, b is the bias term, Con is the result of doing the convolution on the feature map.

Pooling Layer: The feature maps' spatial dimensions are reduced by applying pooling layers after each convolutional layer. The network becomes more resilient to minor input image shifts and distortions, and computational complexity is reduced as a result. Max pooling, the most used pooling operation, is defined by equation 4 and uses to reduce dimensionality while keeping the most relevant information; it does this by selecting the maximum value from a feature map region.

Fully connected Layer: After the feature maps have been downsampled by convolution and pooling, a fully linked layer or layers receive the output for final classification or regression tasks. The final convolutional layer's output will be a 2D feature map that needs to be flattened to a 1D vector to pass into the fully connected layers. Flattening converts the 2D matrix of pixel values into a long vector. In the fully connected layer, like a conventional neural network, every neuron connects to every neuron in the layer below it. Eq. 5 calculates the value of each neuron in the complete connection layer.

where σ is the activation function, W_c are the weight for the connections, x is the input feature vector, and b is the bias term.

d) CNN Training

The most common way of training a CNN on a preprocessed dataset is by the mini-batch stochastic gradient descent, a variant of gradient descent. This iterative process essentially consists of four major steps: first, in a forward pass, the network is passed a batch of images to make some predictions; second, a loss function that compares these predictions to the true labels and makes some error calculations. Then, the backward pass will proceed with the computation of the gradients of the loss concerning model parameters by a process called backpropagation. The resulting gradients are subsequently used to alter the weights and biases of the network in such a way as to minimize the loss. The whole process is repeated across many batches and epochs until performance or satisfaction with the model results are converged. The categorical cross-entropy in equation 5 is utilized in classification tasks, such as tagging products, by allowing the determination of a loss function.

where C is the number of classes, z_i is the true label and p_i is the predicted probability.

e) Product Image Analysis

Feature extraction is a continuous process through CNN, starting from the first layer. A CNN, while most vital in later layers, learns to detect features from simple edges and textures in a progressive manner to more complex patterns. To make the network more resistant to minor input translations, pooling layers reduce the spatial dimensions of the feature maps. These fully connected layers combine features from across the image to enable high-level reasoning. Before the fully connected layers, the output of the last convolutional layer is the usual choice for a general feature vector. Let this feature vector be referred to as F and computed in equation 6.

where X is the input image.

f) Personalized Recommendations

Analysing User Behaviour Data using CNN: Therefore, the data from the user behaviour has to be transformed into a format suitable for the inputs of a CNN. The data used are usually 2D or 3D representations of user interactions. More precisely, the creation of a user-product matrix with users as rows and products as columns takes place, cells containing interaction data such as viewings, purchases, or ratings. This acts as an image-like input for the CNN. This architecture will take a matrix as input, CNN that extracts the patterns in user-product interaction, pooling layer that reduces the dimensionality and captures the important features, and fully connected layers that combine the features for final predictions. The CNN is trained on historical data to predict user interactions with products.

Extracting Patterns in User Preferences and Purchasing History: Once trained, CNN extracts meaningful patterns from user data most proficiently. This usually involves a few fundamental aspects: Feature extraction uses the outputs of intermediate layers as feature representations that represent large complex patterns in user behaviour. CNN also performs latent factor analysis through non-linear dimensionality reduction that reveals the hidden factors driving user preferences. Time information encoded in the input contributes to the network's ability to pick out temporal patterns, offering insight into the evolving user preference. Cross-product relationships can also be extracted from convolutional layers to get insight into complementary or substitute product associations. The extracted patterns contribute toward a comprehensive pattern of user preference and purchasing behaviour, the foundation for deriving personalized product recommendations.

Generating Personalized Product Recommendations: It consists of a few major steps: calculating the similarity between user and product feature vectors using methods such as cosine similarity. Then ranking products based on these scores, the top N products in this ranked list would act as the recommended list. For better results, CNN-based features can be combined with collaborative filtering methods. Available metadata will provide initial feature vectors for new users or items, thus solving the cold start problem. Some products with relatively low similarity scores are included on purpose in order to introduce some diversity and uncertainty to recommendations. Real-time updates: every time users interact with the site, the system updates the user feature vectors in real time. Continuous A/B testing measures and optimizes different recommendation strategies with metrics like click-through rate, conversion rate, and user satisfaction.

g) Dynamic Ad Targeting

Dynamic Ad Targeting is a smart real-time strategy for user ad delivery based on their immediate behaviour and foreseen intentions. This involves a Convolutional Neural Network that analyzes users' browsing patterns in real-time and continuously so that the system can infer their current interests and intent. This real-time interpretation allows the system to consciously make decisions about selecting ads for display, which will be highly relevant to the user's current mindset and browsing context. The selection will be dynamic and responsive, heightening the possibility of a user's engagement and conversion since this aligns with their immediate needs and interests.

Click-through prediction is a more fine-grained type of ad targeting. Given a particular user and ad, it predicts the probability of the user clicking on that ad. Equation 7 gives an example of logistic regression for click-through prediction.

where P(A│u,p) is the probability of the event A (user clicking the ad) given u (user data) and p (product data). x is the feature vector, and w is the weight vector.

The ad targeting system follows an advanced procedure: feature extraction, whereby a CNN scans product images for features within the images and then combines them with user behavioural data into a single input vector. This input vector feeds into the prediction model to predict the probability of a user clicking an ad: a weighted sum followed by the sigmoid. The probability of the output would range between 0 and 1, in which higher probabilities encourage showing the ad. Importantly, the system is designed with a feedback loop, observing user interactions with the displayed ads. It is used to continuously refine and improve its predictive powers to ensure truer and more effective targeting of advertisements.

h) Integrating with the Marketing Pipeline

It is made from a holistic approach to integrate the CNN model into the marketing pipeline of the e-commerce platform with API integrations and robust data pipelines. RESTful APIs enable real-time communications between the CNN model and the platform, allowing instant recommendations and ad targeting based on user behaviour and product data. Meanwhile, it engineers ongoing data pipelines that keep the model constantly fed with new information, such as recent user interactions and newly added products. This two-point integration strategy ensures the CNN model stays current and responsive to constant delivery relevant in timely recommendation and targeted advertising, further enhancing the overall effectiveness of the e-commerce platform marketing efforts.

4. Results and discussion

a. Dataset Description

These datasets on Amazon product reviews are a rich repository of consumer sentiment and feedback about the manifold products sold through their platform. Common information in these datasets includes product identifiers, review text, star ratings, helpfulness votes, and sometimes even demographics about reviewers, among other metadata. Amazon product review datasets are used for various purposes by researchers and businesses, including sentiment analysis, product recommendation systems, market trend analysis, and understanding consumer preferences and behavior. This helps them gauge valuable insights from the reviews regarding the characteristics, quality, and levels of satisfaction of the products and trending issues in the market with large volumes of reviews. The various applications of these datasets range from machine learning algorithms to natural language processing techniques, including classification of sentiment, identification of key topics and themes, and building predictive models for product success or failure.

b. Performance Metrics

This section compares traditional models such as LCNA-LSTM [9], Support Vector Machine (SVM) [15], and ARIMA and LSTM [14] with the suggested ECMS-CNN technique. While conventional approaches use SVM, long short-term memories (LSTMs), and recurrent neural networks (RNNs) to solve data overload and forecast sales, ECMS-CNN improves marketing strategies by analyzing product images, making tailored recommendations, and targeting ads. The performance measurements, which include root mean squared error (RMSE), mean absolute error (MAE), click-through rate (CTR), and conversion rate, allow a thorough comparison between ECMS-CNN and traditional approaches. These metrics measure the effectiveness of ad targeting, the accuracy of recommendations, and the overall influence on user interaction and purchase conversions.

a) Mean Absolute Error (MAE)

The average amount of the mistakes between the expected and actual values in a dataset, without taking their direction into consideration, is what this statistic measures. It measures how accurately a model predicts outcomes such as customer purchases, product preferences, or the likelihood of ad clicks, comparing predicted values to actual user behaviour. It can be calculated by equation 8.

where x_i is the actual observed value (e.g., actual purchases, clicks), x_i^' is the predicted value from the model, and n is the number of data points (users, sessions, or product recommendations). Table 1 highlights the MAE values for product recommendations, ad targeting, and sales forecasting tasks.

The suggested ECMS-CNN approach is compared to three traditional methods in the table for important e-commerce tasks, including product recommendations, dynamic ad targeting, and sales forecasting: LCNA-LSTM CNN, DL Techniques, and the Sales Prediction Model. Each method's Mean Absolute Error (MAE) value is included so that their accuracy can be quantitatively compared. By contrasting it with other methods, this comparison allows a thorough assessment of the suggested ECMS-CNN approach to optimizing different parts of e-commerce marketing campaigns.

b) Root Mean Squared Error (RMSE)

This metric is frequently employed to assess the precision of prediction models. In other words, it determines how dispersed the residuals (errors in predictions) are. It can be calculated as in equation 9.

where x_k the actual observed value, x_k^' predicted value from the model. Table 2 shows the RMSE analysis between the proposed ECMS-CNN and traditional methods.

ECMS-CNN outperforms conventional models again, with RMSEs indicating it can withstand bigger prediction mistakes better. As a result of ECMS-CNN's capacity to identify complicated patterns from big datasets, both the average error and the substantial deviations from the real values are reduced, leading to a noticeable improvement in ad targeting and sales forecasting. This means that ECMS-CNN is better at avoiding big mistakes and has better overall accuracy.

c) Click-Through Rate (CTR)

It is an essential performance indicator utilized in digital marketing, particularly in e-commerce, to determine the success of social media advertising and email marketing initiatives. In other words, it's the percentage of people who see an ad, page, or email and then click on a specific link. CTR measures the proportion of impressions that lead to a click. Each display of the advertisement or content constitutes one impression. It can be achieved by equation 10.

Figure 4 compares the performance of various models with respect to Click-Through Rate (CTR): ECMS-CNN, LCNA-LSTM, ARIMA & LSTM, and SVM. With a greater CTR, ECMS-CNN achieves better ad relevance to users' interests and improved targeting using advanced image and behavioral analysis, as seen in the figure, surpassing the older models. The capacity of ECMS-CNN to extract rich information from visual and behavioral information enhances its ability to dynamically alter the targeting of ads in real-time, leading to higher user engagement and superior CTR performance.

d) Conversion Rate

It is a critical metric in e-commerce that measures the percentage of visitors who complete a desired action. In e-commerce, this typically means making a purchase, but it can also include signing up for a newsletter, creating an account, or adding items to a cart. Equation 11 achieves this.

It measures the effectiveness of a website and its related marketing in turning visitors into paying customers. The Conversion Rate directly affects profitability because it is directly related to the revenue generation and the return on investment (ROI) for marketing campaigns.

Figure 5 compares the conversion rates for the same models. One way to evaluate an online store's performance is to look at its conversion rate. The ECMS-CNN method outperforms existing models regarding conversion rates, which could be attributed to its ability to deliver more tailored suggestions and ad targeting. The capacity of ECMS-CNN to thoroughly examine consumer actions, purchasing patterns, and product pictures leads to extremely pertinent suggestions and variable ad placements that inspire buyers to make a purchase, increasing the conversion rate.

5. Conclusion

This research presents ECMS-CNN, a new method using Convolutional Neural Networks to improve e-commerce marketing tactics. Its primary goals are to analyze product images, provide individualized suggestions, and implement dynamic ad targeting. ECMS-CNN demonstrates lower MAE and RMSE in tasks like product recommendations, ad targeting, and sales forecasting compared to established models like LCNA-LSTM, ARIMA & LSTM, and SVM across several performance criteria. The method's improved conversion rates and greater CTR suggest better ad targeting and user engagement. By utilizing CNNs for product picture analysis, ECMS-CNN can provide more relevant suggestions to customers and decrease manual effort in product labelling by 30%. Full integration of visual and behavioral data provides a comprehensive view of consumer preferences and marketing strategy optimization. With these updates, e-commerce systems may boost consumer happiness, conversion rates, and overall performance. Potential areas for future research include domain-specific applications, the integration of natural language processing methods, the examination of long-term effects on consumer loyalty, and the investigation of transfer learning. A state-of-the-art solution to the ever-changing e-commerce scene, ECMS-CNN is an enormous step forward in optimising e-commerce marketing strategies with machine learning.