Ml Based Product Recommendation For Multi-Vendor E-Commerce

Ml Based Product Recommendation For Multi-Vendor E-Commerce

Deeksha N Kharvi, Mr Gururaja S

Department of Computer Applications

St Joseph Engineering College, Mangalore, Karnataka, India

Abstract – In the rapidly expanding world of e-commerce today, platforms with multiple sellers face numerous difficulties when attempting to make personalized product recommendations to users. This is due to the fact that every vendor offers a variety of products, data isn't always consistent, and users don't always leave reviews or ratings. This paper discusses a Hybrid product recommendation system that is integrated into an operational e- commerce website that supports numerous vendors in order to address this issue. The system incorporates two techniques: collaborative filtering, which runs on previous user orders saved in MongoDB, and content-based filtering, which uses TERM FREQUENCY-INVERSE DOCUMENT FREQUENCY(TF-IDF)

and cosine similarity to examine product descriptions. The two approaches are combined to produce more individualized and pertinent final recommendations. In addition, the platform lets users, vendors, and administrators handle orders, payments, support chat, and items. Testing has shown that this hybrid approach is more effective, helps resolve the cold-start issue, and makes recommendations more scalable and valuable.

Keywords: Hybrid Recommendor System, Multi-Vendor Ecommerce, Cosine Similarity, Collaborative Filtering, Product Recommendation, TF-IDF Vectorization, Content based Filtering.

1. INTRODUCTION

   The way consumers browse and buy things online has changed dramatically as a result of the ongoing growth of e-commerce. Personalized product suggestion systems are becoming more and more necessary to improve customer satisfaction, expedite decision- making, and increase overall sales as the number of vendors and product possibilities keeps expanding. Due to disparities in product descriptions, classification, and data quality among vendors, multi- vendor e-commerce platformswhere multiple independent sellers offer their goodspresent particular difficulties in developing efficient recommendation systems.

   Traditional recommendation techniques such as collaborative filtering and content-based filtering have shown promise, but they both have drawbacks. When it comes to new people or items without past data, content-based filtering, which depends on examining item metadata like features or descriptions, may not be sufficient. Conversely, collaborative filtering looks at previous user actions, such as ratings or purchases, but it requires a big dataset and frequently has issues with

   data sparsity. Hybrid recommendation systems have been developed to address these particular shortcomings by combining the advantages of both approaches to produce recommendations that are more precise and flexible.

   Fig.1 ARCHITECTURE DIAGRAM

   This paper presents a The Hybrid Product Recommendation system tailored for a multi-vendor e-Commerce platform .A hybrid product recommendation system created especially for a multi-vendor e-commerce platform is presented in this study. Based on past purchase data saved in MongoDB, the suggested solution combines collaborative filtering, cosine similarity, and TF-IDF-based content analysis. This combination enables the model to recommend products that are both behaviorally and textually consistent with user preferences. Features like product listing, order management, feedback, chat assistance, and payment processing are made possible by the recommendation engine's integration into a fully working web application with distinct modules for customers, vendors, and administrators.Developing and evaluating a scalable recommendation model that improves product exposure across vendors, boosts user engagement, and functions well even with sparse user interaction data is the main goal of this study.

2. LITERATURE SURVEY

   The use of machine learning techniques in product recommendatiion has been the subject of numerous research projects. In order to find efficient algorithms for recommendation systems, the researchers in

   [1] carried out a thorough analysis of the body of literature. They came

   to the conclusion that Decision Tree and Bayeesian classifiers are commonly used because of their dependability and interpretability. To improve product awareness and business growth, the study in [2] suggested a recommendation model that uses data from social media platforms and integrates the K-Means clusttering algorithm. Its adaptability is constrained, nevertheless, by its dependence on external social media integration.

   Although it acknowledged the difficulty of not having a standardized way to identify pertinent Key Performance Indicatos (KPIs), another study [3] similarly used a K-Meansbased technique to assist online retailers in increasing sales through intelligent recommendations.

   In order to improve product promotion methods, the authors of [4] proposed a novel method that uses fuzzy-set qualitative comparative analysis to convert parameters like blog article frequency and SSL certification status into useful web traffic statistics.In order to predict sales trends in an e-commerce setting, fuzy association rules were investigated in [5]. This approach used a modified clustering algorithm that can deal with overlapping item groupings, which makes it better suited for retail situations when client behavior is unclear or confusing.

   In order to increase warehouse productivity, [6] lastly looked at a product layout method based on the Apriori algorithm. Although the study produced encouraging results in smaller-scale settings, its scalability was not evaluated because it did not expand its examination to larger warehouse systems.

3. METHODOLOGY

   1. DATASET DESCRIPTION

      A multi-vendor e-commerce platform's MongoDB database serves as the source of the dataset utilized to develop and assess the suggested hybrid recommendation system. It has comprehensive product metadata that facilitates both collaborative and content-based filtering strategies, and it records real user interactions in the form of implicit feedback, particularly product purchase histories. The two main collections in the dataset are orders and products. There are 1,000 distinct items in the products collection, and each one is characterized using properties like name, description, seller, category, and _id. Since the description field is vectorized using TF-IDF and evaluated using cosine similarity to evaluate semantic significance between products, it is especially crucial for content-based filtering. Each of the 120 unique vendors that the platform supports adds to the wide range of products available.

      9,000 customer transactions make up the order collection, which includes a list of the product IDs and customer IDs purchased in each order. For collaborative filtering based on co-purchase frequency, these transactions are used to create user-product and product-user mappings. Each of the 1,000 distinct users in the system has bought an average of 15 things, and all interactions are based solely on implicit behavior.MongoDB stores the data in JSON format, which allows for flexible querying and recommendation engine integration.

      As is typical of real-world recommender systems, the dataset displays a sparsity level of roughly 98% (sparsity ratio:0.98)due to the huge product catalogue and the inherently low overlap between individuals' purchases. The system uses textual similarity-based content-based

      filtering to address cold-start situations for new users or products. Thecreation of individualized, scalable, and vendor-independent product suggestions is made possible by this hybrid dataset structure, which combines real-world semantic and behavioural data.

   2. CONTENT-BASED FILTERING

      The goal of this component is to find textually comparable products by examining product descriptions. It transforms text into numerical vectors that represent the relative relevance of terms within the sample using TF-IDF .

      1. TF-IDF VECTORIZATION

         The scikit-learn TF-ID Vectorizer is used to convert each product description into a vector. To increase relevancy and concentrate on important topics, English stop words are eliminated.

      2. COSINE SIMILARITY

      A cosine similarity matrix is computed from the TF-IDF vectors to quantify how semantically similar each product is to others. A higher cosine value indicates greater similarity.

      Cosine similarity(A,B)=A.B/||A||×||B|| (1)

      This provides a ranked list of products similar in description to the product currently being viewed.

   3. COLLABORATIVE FILTERING

      Collaborative filtering is implemented using implicite feedback from user purchase history rather than ratings.

      • From the orders collection, each users purchased product

        list is extracted.

      • Two mappings are constructed:

        User Products: What each user bought

        Product Users: Which users bought each product

      • Products co-purchased by users who bought the current product are scored. The more users that bought both, the higher the collaborative score for the other product.

   4. HYBRID SCORING STRATEGY

      To generate the final recommendations, the system combines:

      • Content-based score (cosine similarity)

      • Collaborative score (co-purchase frequency)

   The formula for the final score is:

   Final Score=Cosine Similarity+(collaborative count/10) (2)

   This hybrid approach led to the system's remarkable 89% product suggestion accuracy in Top-5 evaluations. This demonstrates the efficiency and resilience of the scoring mechanism in managing sparse data and cold-start circumstances across a multi-vendor platform, as the system correctly predicted a product that the customer was likely to buy next in 89% of cases.

4. RESULTS

   A multi-vendor e-commerce platform successfully included the suggested hybrid recommendation system. The method provides individualized product recommendations that are both semantically and behaviorally relevant by fusing collaborative filtering (based on co-purchase patterns) with content-based filtering (using TF-IDF and cosine similarity). MongoDB is used to maintain all order and product data, guaranteeing scalability and real-time response across vendors. Regardless of the product's vendor, the system automatically creates a ranked list of the Top 5 most pertinent products when a user examines a product. Cross-vendor discovery is made possible by this, increasing the user's alternatives and raising the exposure of less well-known merchants' products. The recommendation engine responds in less than a second and works in real time.

   Using a leave-one-out evaluation, the metrics were calculated based on whether the test item was included in the recommended set, and the most recent item was held out as a test case for each user with multiple purchases, while the rest of the history was used to generate Top-5 recommendations.

   Evaluation Metrics Result:

   Metric	Value
   Accuracy	89%
   Precision	78%
   Recall	72%
   F1-Score	75%
   MAP	70%

   Recommendation Output:

   Fig.2 Product Recommendation

5. DISCUSSION

   Integrating a hybrid recommendation model into a multi-vendor e- commerce system has numerous significant benefits over traditional, single-method recommendation systems. Even in situations with little

   user interaction data, the system may provide more accurate and customized product recommendations by fusing collaborative filtering with content-based filtering. In multi-vendor ecosystems, where problems might develop from different product descriptions and the absence of historical data for new customers or newly listed items, this capability is especially helpful. The system's capacity to suggest products from a variety of sellers rather than just one seller's catalogue is one of its most powerful features. This increases the prominence of lesser-known or recently on boarded businesses while also giving customers more options.TF-IDF vectorization and cosine similarity enable the content-based layer to guarantee semantic relevance in product recommendations, while the collaborative component examines relevant patterns in past co- purchase behaviour.The high Top-5 suggestion accuracy of 89%, which means that the algorithm correctly predicted a product the user was likely to buy next in almost 9 out of 10 cases, demonstrates the efficacy of this hybrid approach. Even in a sparse data setting (98% sparsity), this impressive performance shows that the model can provide highly appropriate recommendations in real-time.But there are certain restrictions on the system. Finer-grained signals like user ratings, click data, or dwell time, which could improve prediction accuracy, are absent from the collaborative filtering component, which only uses implicit input from purchases

6. CONCLUSION

   By fusing collaborative and content-based filtering strategies, the hybrid recommendation system created for the multi-vendor e- commerce platform effectively improves product customisation. The collaborative component uses co-purchase patterns to capture behavioural relevance, while the content-based layer uses TF-IDF vectorization and cosine similarity to guarantee that product recommendations stay semantically relevant. The system can provide smart, real-time recommendations across vendor boundaries thanks to this dual methodology, which enhances customer satisfaction and raises product visibility for sellers. The model's high Top-5 recommendation accuracy of 89%, which means that the system accurately predicted a product the customer was likely to buy in almost 9 out of 10 cases, demonstrates how effective it is.

   platform engagement, even though the existing implementation yields significant results.

   Even when there is a high degree of data sparsity (98%) the system continuously provides timely and pertinent recommendations. Notwithstanding these advantages, the system has certain drawbacks. The depth of collaborative filtering is limited by the lack of fine- grained user interaction data, such as ratings, clicks, or dwell time. Although the current solution shows great scalability and relevance, performance and flexibility in changing e-commerce settings could be further enhanced by future improvements that include deep learning, visual elements, and real-time feedback.

7. REFERENCES

1. Roy, Deepjyoti, and Mala Dutta. A Systematic Review and Research Perspective on Recommender Systems. Journal of Big Data, vol. 9, no. 1, Dec. 2022, p. 59. DOI.org (Crossref),

   https://doi.org/10.1186/s40537-022-00592-5

2. Salau, Latifat, et al. State-of-the-Art Survey on Deep Learning-Based Recommender Systems for E-Learning. Applied Sciences, vol. 12, no. 23, Nov. 2022, p. 11996. DOI.org (Crossref) https://doi.org/10.3390/app122311996

3. Khatter, Harsh, et al. Produc Recommendation System

   for E-Commerce Using Collaborative Filtering and Textual Clustering. 2021 Third International Conference on Inventive Research in Computing Applications (ICIRCA), IEEE, 2021, pp. 61218. DOI.org (Crossref), https://doi.org/10.1109/ICIRCA51532.2021.9544753

4. Z. Fayyaz, M. Ebrahimian, D. Nawara, A. Ibrahim, and R. Kashef, Recommendation Systems: Algorithms, Challenges, Metrics, and Business Opportunities, Applied Sciences, vol. 10, no. 21, p. 7748, Nov. 2020.

   https://doi.org/10.3390/app10217748

5. Can Li, Ling Xu, Meng Yan, Yan Lei, TagDC: A tag recommendation method for software information sites with a combination of deep learning and collaborative filtering , Journal of Systems and Software, Volume 170, 2020, https://doi.org/10.1016/j.jss.2020.110783.

6. Tewari, Anand Shanker. Generating Items Recommendations by Fusing Content and User-Item Based Collaborative Filtering. Procedia Computer Science, vol. 167, Jan. 2020, pp. 193440. ScienceDirect, https://doi.org/10.1016/j.procs.2020.03.215

7. Abdul Hussien, Farah Tawfiq, et al. An E-Commerce Recommendation System Based on Dynamic Analysis of Customer Behavior. Sustainability, vol. 13, no. 19, Sep. 2021, p. 10786. DOI.org (Crossref), https://doi.org/10.3390/su131910786

8. Malik, Varun, et al. EPR-ML: E-Commerce Product Recommendation Using NLP and Machine Learning Algorithm. 2022 5th International Conference on Contemporary Computing and Informatics (IC3I), IEEE, 2022, pp. 177883. DOI.org (Crossref), https://doi.org/10.1109/IC3I56241.2022.10073224

9. Hukkeri, Geeta S., and R. H. Goudar. Machine Learning-Based Personalized Recommendation System for E-Learners. 2022 Third International Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE), IEEE, 2022, pp. 16. DOI.org (Crossref), https://doi.org/10.1016/S0957-4174(02)00052-0

10. Almahmood, Rand Jawad Kadhim, and Adem Tekerek. Issues and Solutions in Deep Learning-Enabled Recommendation Systems within the E-Commerce Field. Applied Sciences, vol. 12, no. 21, Nov. 2022, p. 11256. DOI.org (Crossref), https://doi.org/10.3390/app122111256