ADS Capstone Chronicles Revised

15

According to the RFM values of each customer, 7 types of spenders are separated. 1) Champions 2) Potential Loyalist 3) Loyal Customers 4) At-Risk 5) Hibernating 6) Big Spender 7) Others. To make a good product recommendation system, retail companies should know their customer better in terms of their buying behavior. As customers evolve, the multi classification model helps to adapt according to the real time data. It will automatically relabel the customers with their corresponding sections and shows necessary recommended products to buy. It may help the company to give different marketing strategies to retain the customers. So, building an efficient multiclass model can help to boost the recommendation system. It will increase the conversion and retention rate of the customer. ​ 4.4.1​ Selection of modeling techniques for recommendation. For this analysis, three different recommendation models were utilized to optimize product recommendations: Singular Value Decomposition (SVD), Non-Negative Matrix Factorization (NMF), and K-Nearest Neighbors (KNN). Each model offers unique advantages and operates based on distinct mathematical principles, allowing us to explore various dimensions of user-item interactions. The following sections provide a detailed overview of each model, including the equations they use and the metrics employed for evaluation.​ SVD is a powerful matrix factorization technique used in collaborative filtering for recommendation systems. It decomposes the user-item interaction matrix into three smaller matrices: (Bhatt et al., 2023). The , Σ, equation is given by equation (1), where A is the original interaction matrix, U represents the user features, Σ is a diagonal matrix of singular values, and denotes the item features. ​ (1) = Σ SVD excels at capturing latent factors that influence user preferences, making it highly effective for personalized recommendations. Key

metrics for evaluating SVD include Root Mean Square Error (RMSE), which measures the accuracy of predicted ratings against actual ratings, and precision@k and recall@k, which assess the relevance of the top-k recommended items (Zhang et al., 2009).​ NMF is another matrix factorization technique that decomposes the user-item interaction matrix into two non-negative matrices W and H, such as A≈ WH (Van Otten et al., 2023). Here, A is the original interaction matrix, W represents user latent features, and H represents item latent features. The non-negativity constraint makes NMF particularly useful for interpreting the components of user-item interactions, as all values remain in the original data's domain. The objective function is to minimize the Frobenius norm of the difference between A and WH, typically expressed in equation (2). Metrics include reconstruction error, which measures how well the factorized matrices approximate the original matrix, and RMSE for prediction accuracy, alongside precision@k and recall@k for recommendation relevance. KNN is a straightforward and intuitive algorithm used for recommendation systems. It operates by finding the k nearest neighbors to a user based on their interaction history and recommending items that these neighbors have interacted with. The similarity between users or items can be computed using metrics such as cosine similarity or Pearson correlation. The primary equation for KNN in this context is similar to shown in equation (3)​, where and are ratings given , , by users u and v to items i, and ​and ​are the average ratings of users u and v (Srivastava et al., 2024). for evaluating NMF ​​ (2) , || − ||

59