Hybrid E-Recommendation System for Multi-Shop Environmentⁱ

2.1. Literature Review

Recommendation systems, as a field, have evolved rapidly, with several works addressing critical challenges, such as data sparsity, cold-start, and personalization, etc. These works are also closely related to this research, which employs a hybrid recommendation system (HRS) to improve poorly integrated and fragmented markets. In line with the overarching aim of our research, we reviewed some studies that offer insight and contribute to research goals.

Hasan and Ferdous (2024) highlight the effectiveness of HRSs that combine text-to-number transformation, matrix factorization techniques such as ALS, and cosine similarity for precise recommendations. Their study focuses on preprocessing, similarity calculations, and latent factor modeling, using the TMDB 5000 dataset and RMSE measures for evaluation. Their study further demonstrates how combining multiple methods in a hybrid approach improves the efficiency and accuracy of recommendations, especially in addressing data scarcity and enhancing system robustness [6].

Mouhiha et al. (2024) explore the combination of CF and CBF to address their individual limitations, such as the problem of cold start and data scarcity. Their study indicates that a HRS effectively improves accuracy and user satisfaction. Also, they describe various hybrid methods that tailor recommendations to user preferences by utilizing both user-item interactions and item specificity [7].

Loukili et al. (2023) investigate recommendation systems designed for e-commerce, using the FP-growth algorithm to analyze purchase patterns and generate personalized recommendations. Their study describes the decision-making challenges posed by the large number of product options. The rating metrics, including accuracy and conversions generated, also demonstrate how robust recommender systems have a direct impact on increasing sales and customer attraction, and highlight practical applications in real-world e-commerce platforms [8].

Cherkaoui et al. (2024) discusses using machine learning to predict customer behavior and improve strategies for marketing. This is achieved through the use of Apriori algorithms to mine the association rules and more sophisticated techniques, such as neural networks that are used to analyze user data to enhance understanding of the behavior of the users. Furthermore, their system concentrates on data segmentation, target estimation, and measures of effectiveness, thereby assisting businesses with realistic recommendations and improving the relevance of the goods recommended to users in highly competitive markets [9].

Ozturk et al. (2024) present a CrossGR model integrated with Graph Isomorphism Networks (GINs), facilitating the function of cross-market recommendation systems even with data sparsity and the limitations posed by the single markets. By employing GINConv layers and multi-layer perceptrons, the model derives insights into the typically complex user-item engagements, thus generating effective recommendations for different markets. Their novel concept in recommendation systems has demonstrated significant improvements in metrics, such as NDCG^@10 and HR^@10, suggesting the efficacy of graph-based learning approaches in recommendation systems [10].

The reviewed literature highlights various approaches, such as (hybrid recommendation models, machine learning approaches, graph-based methods, and NLP sentiment analysis). While these studies address issues such as data scarcity, personalization, and cold start issues, they do not specifically focus on fragmented market integration, a major challenge in many parts of the world, such as the Kurdistan Region.

This thesis aims to addressee the gap by combining different techniques such as CBF and CF, designed to enhance market recommendations in a complex multi-store environment. By collecting product information from multiple markets, the proposed system provides consumers with a unified shopping experience, overcoming the problems of fragmentation in the domestic e-commerce sector. Customers can also access multiple stores and products through a single application, which reduces customer search time and storage, especially on smartphones.

2.1.4. Machine learning

Machine learning (ML) is a subpart of artificial intelligence (AI) that focuses on developing algorithms and models that enable computers to learn from data and make predictions or decisions without being explicitly programmed [13]. ML has revolutionized the field of E-recommendation systems, making the systems more intelligent, more accurate in personal recommendations and enabling them to understand changes faster. E-commerce has expanded significantly due to its benefits, which has created a very good environment for the use of machine learning as a branch of AI. Due to the expansion of available information, there is a need for personalized experiences, addressing fraud detection and security challenges, enhancing supply chain optimization capabilities, and recognizing the importance of customer sentiment analysis [14]. Table I shows some ML models and algorithms that are used for E-RSs.

TABLE I: Machine learning models and algorithms used for E-RSs

3.1. Data Collection and Preparation

Figure 1 show the data collection and preparation process for five laptop shops and 1000 user ratings for shops and products. The product datasets include information about products from different shops. Each product dataset includes features such as (market name, brand name, processor type, processor brand, generations, RAM, storage capacity in GB, have SSD, HDD, graphics capacity, price, display size, display type, and color). The User rating dataset includes customer ratings. This allows the system to generate more accurate recommendations based on user preferences. It contains the user ID, shop ID, product ID, and the rating given by the consumer to the shops and products.

Fig. 1. Data collection and preparation process.

Figure 1 also illustrates the main preprocessing steps applied. Data preparation and preprocessing began by combining the datasets from five shops into a single dataset. Subsequently, data cleaning (such as removing duplicates, punctuation, and handling missing values), normalization (e.g., for price and rating), transformation (such as converting the rating matrix to a tabular format), and the selection of key features for building the e-recommendation system based on the features Kurdish users consider important when purchasing laptops were performed. In the final step, the rating dataset was divided into a training set (80%) and a testing set (20%) to evaluate model performance effectively.

3.2. Term Frequency-Inverse Document Frequency (TF-IDF) and Cosine Similarity

TF-IDF is a text analysis method used to calculate the relevance of words in a document when compared to a collection of documents. IDF reduces the weight of frequent words that exist in multiple documents, while TF measures the frequency of a word’s occurrence in a document. The TF-IDF equation is:

TF-IDF = TF×IDF

Where is calculated as the number of times a word appears in a text or document divided by the total number of terms in that document, and IDF is computed as:

Where N is the total number of documents, and df is the number of documents containing that word. TF-IDF helps in filtering out less important words while highlighting meaningful ones in recommendation systems.

A measure called cosine similarity is used to evaluate how similar two text vectors are to each other as well as providing a quantitative measure of document similarity by calculating the cosine of the angle formed by two vectors in n-dimensional space. The formula is:

A and B are document vectors, and the numerator’s dot product is divided by the product of their scales. Recommendation systems use both methods to detect product similarities based on descriptions in texts [15].

3.3. Singular Value Decomposition (SVD)

SVD is a matrix factorization algorithm that converts a matrix into three matrices for efficient representation and dimensionality reduction. This technique has found extensive applications across various areas, such as recommendation systems, where it has been shown to recommend new products and encourage repeat purchases, resulting in an improved user experience [16].

SVD in a recommendation system:

Input: A matrix where rows are users, columns are items (like movies), and the values are ratings.

SVD splits the matrix into three smaller matrices:

A = U ⋅S ⋅V^T

U: Describes the users.

S: Contains the importance of features.

V_T: Describes the items.

Using this, we can predict missing values in the original matrix, such as movies a user hasn’t rated yet.

3.4. K-Nearest Neighbors (K-NN)

K-NN is an instance-based learning algorithm that is widely used in classification and regression problems [17]. Because it is so easy and supports large-dimensional data, it is widely used for text classification, image recognition, and medical diagnosis. Although algorithm performance may depend significantly on the distance used, it has been shown to achieve competitive results in most real-world applications. K-NN for recommendation system is a neat application, in which you can detect similar users or objects from their characteristics and propose them recommendations.

3.5. CBF Process

We combined multiple product attributes (such as price, brand, RAM, screen size, etc.) into a single row for each product. This row represents a general description of the product.

Fig. 2. Content-based filtering process.

3.6. CF Process

The user ratings are normalized using normalization techniques to ensure all ratings are within the range of 0 to 1. This process reduces bias arising from different rating scales across users.

Fig. 3. Collaborative filter process.

3.7. The Flowchart of the Proposed System

In Figure 4, the flowchart shows the user workflow in the E-RS. To begin with, the user can either create an account by registering or log in to their present account. Alternatively, they can search the system anonymously, limiting the level of personalization. Once a user is engaged, they can start browsing products. The system uses a combination of recommendation techniques. For example, if a user searches for “RAM 32GB”, the system uses CBF to analyze user search queries and compares them with product descriptions using techniques such as TF-IDF and cosine similarity, which helps the user to find laptops in multiple shops with similar attributes such as RAM, storage, and price. The user search result is ordered based on similarity score. For example, if shop A has a high similarity score, it will be shown at the top for the user. The system also makes use of CF for registered users to analyze past user behavior, such as ratings and compares it to the behavior of other users with similar interests. Algorithms, such as K-NN and SVD are used to detect these similarities and predict the shop and products that the consumer might like.

Fig. 4. Flowchart of the proposed system.

Finally, the hybrid recommendation system combines the strengths of both CBF and CF. This approach offers a more nuanced and accurate set of recommendations by taking into account both product characteristics and user ratings.

4.1. CBF Implementation and Result

Figure 5 illustrates the top five recommended Apple products retrieved using CBF. The system computed the similarity scores using TF-IDF and cosine similarity. Higher scores indicate stronger textual similarity to the user’s search. This figure shows that the most relevant match (0.773) is from the “Shop W” market, while other competitive options appear in “Shop R” and “Shop M_A”.

Fig. 5. Content-based filter similarity score result.

The brief code sample in the screenshot above presents a fundamental CBF code structure. The CBF component relied on TF-IDF to convert textual product descriptions into numerical vectors that were then used to calculate cosine similarity. Cosine similarity measures the angle between two vectors in multidimensional space which provides a measure to find the similarity between user search and product attributes. These detailed calculations ensure that the best ranking related to user search is provided.

To evaluate the CBF approach in a real-world user-facing system, we consider an example of a user search with loosely defined features (“8 GB RAM core i5 gen 11 white hp”). As observed, this query exhibits characteristics of a user search with loosely defined features. The search results for this query were available from five distinct vendors. Figure 6 presents the results from a selection of these shops.

Fig. 6. User search example and content-based filter result.

4.2. CF Implementation and Result

The model needs to be trained and tested after it is built. The dataset was partitioned into a testing set (20%) and a training set (80%). For evaluating the models, metrics, such as root mean squared error (RMSE), mean absolute error (MAE), and mean squared error (MSE) were used. The below screenshot for the sample code shows how SVD and K-NN are combined to improve shop and product recommendations.

In addition to these results in table 2, the response time was also tested to evaluate the real-time performance of the hybrid proposal model. The test yielded an average prediction response time of 0.0020 s, or 2 ms, demonstrating the system’s efficient and near-instantaneous shop recommendation capability. This outcome suggests that the model is effectively optimized for real-time applications, enabling it to handle user demands with minimal latency.

TABLE II: Cross-validation result

4.3. Hybrid E-Recommendation System Result

By bringing together both CBF and CF into one system, we create a hybrid recommendation model that takes the best of both worlds while reducing each method’s weaknesses. The HE-RS model combines the feature-based approach from CBF, which uses product attributes to find similarities, and the behavior-based insights from CF, which look at user ratings and patterns to offer more personalized and robust recommendations. This means that even if a user hasn’t rated many shops and products, CBF can still suggest items based on product features, while CF can step in when product details alone aren’t enough to capture user preferences. Contextual and behavioral data are both included in the final recommendation outcome, which produces a system that can effectively handle problems, such as data lacking, new users, and modifying preferences. Furthermore, this hybrid approach not only improves the accuracy of recommendations but also creates a more intuitive and user-friendly shopping experience by continuously matching product recommendations and shop recommendations to changing consumer preferences. In Figure 7, what is on the right side of the system is the result of a customer’s recommendation based on the rating they gave, in which machine learning algorithms play a role. The list on the left shows the results of the customer’s search and the results according to the similarity of the user’s search and products. In all stores, products are recommended by the recommender system based on CBF.

Fig. 7. Hybrid e-recommendation screen result.

4.4. Discussion

Table 2 is shows the hybrid model (SVD with K-NN) gives better recommendations than the SVD model alone. Hybrid model has a lower RMSE (0.1435) and MAE (0.1147) compared to the SVD model, indicating that the hybrid approach makes more accurate predictions. SVD finds hidden patterns in the data, while K-NN improves prediction by comparing similar users. Thus, combining both methods can reduce errors and give better recommendations. The results of this study are presented in two ways that reflect real-world applicability. One of the important aspects that distinguishes this research from other studies is the development of a comprehensive system that integrates multiple recommendation techniques to enhance performance. The evaluation and results suggest that the prototype system is suitable for further development and optimization, and that a centralized recommendation platform can serve as a practical solution for mitigating challenges in fragmented markets. Instead of requiring each shop to develop its own application, which causes data dispersion, businesses can simply create an account within a unified system tailored to their specific industry, improving accessibility and user experience.

4.5. Comparative Analysis

To validate the performance of our hybrid recommendation system, we compare it with recent models that also apply hybrid or advanced recommendation techniques. As shown in Table III, our system combines SVD and K-NN for CF with TF-IDF and Cosine Similarity for CBF. The CF achieves a low RMSE of 0.14 and MAE of 0.11 on a sparse, custom laptop dataset. Furthermore, the CBF achieves successful performance in responding to customer searches. Compared to other studies, our system demonstrates higher precision and adaptability in market-specific recommendations.

TABLE III: Comparison of HE-RS with other studies

5.1. Future Work

Finally, our research demonstrates that a HE-RS offers a means to address market fragmentation, provides effective recommendations for customers, and generates increased opportunities for businesses.