Abstract
The explosive growth of online retail has made personalised product recommendation a key driver of revenue and customer satisfaction. This synopsis presents an E-Commerce Recommendation Engine that combines two machine-learning techniques — SVD-based Collaborative Filtering for personalised top-N product recommendations and K-Means++ Clustering for behavioural customer segmentation — within an interactive Flask web application.
The system ingests transactional data (customer ratings, purchase histories, demographics), factorises a sparse user–item rating matrix using Singular Value Decomposition with 50 latent factors, and reconstructs full predicted ratings for all user–product pairs. Simultaneously, seven RFM-inspired features per customer are standardised and clustered; the optimal cluster count is selected via combined Elbow and Silhouette analysis, consistently yielding a Silhouette Score above 0.60. Four semantic customer segments emerge: Premium Customers, Frequent Buyers, Budget Buyers, and Regular Customers.
A RESTful Flask backend exposes API endpoints for real-time recommendations, segment queries, similar-product retrieval, and an analytics dashboard. The system achieves an RMSE of approximately 0.85 on observed ratings with Precision@10 ~0.40 and sub-10 ms inference latency, demonstrating a complete, deployable ML pipeline from raw data to live web service.

Introduction
E-commerce platforms now host millions of products, yet a typical user interacts with fewer than 0.01% of available items per session. Recommendation engines bridge this gap by predicting which products a specific user is most likely to purchase or rate highly, surfacing them before the user has to search. Industry benchmarks confirm that well-designed recommendation systems drive 35–38% of total revenue at leading platforms such as Amazon and Netflix.
This project implements the two most industrially relevant approaches to recommendation: Collaborative Filtering (CF), which identifies user–item affinity patterns from collective behaviour, and K-Means Customer Segmentation, which groups users by purchase behaviour for segment-level personalisation and marketing intelligence. The two models share the same transactional dataset and are served together by a single Flask application, creating a unified recommendation and analytics platform.
The technology stack is entirely open-source — Python, scikit-learn, scipy, pandas, and Flask — making the solution cost-effective and straightforward to deploy on any cloud or on-premise server. A modular architecture ensures that the ALS model, segmentation model, and web interface can be independently updated without disrupting the other components.

Objective
• Design and implement a personalised product recommendation engine using SVD-based Collaborative Filtering that generates top-N recommendations for any registered user in under 10 ms.
• Segment the customer base into behavioural groups using K-Means++ Clustering with automatic optimal-K selection (Elbow + Silhouette), achieving a Silhouette Score ≥ 0.60.
• Build a RESTful Flask web application with dedicated API endpoints for recommendations, segment queries, similar-product retrieval, and an analytics dashboard.
• Evaluate model performance using RMSE, Precision@10, Recall@10, and Silhouette Score, and visualise all results through auto-generated charts (10+ graphs).
• Demonstrate a complete, reproducible ML pipeline — from raw CSV data ingestion and feature engineering through model training, serialisation, and live web inference.

Block Diagram