Ecommerce-recommendation-system

A machine learning–powered Product Recommendation System built to help e-commerce platforms deliver personalized product suggestions.
By analyzing browsing patterns, ratings, and purchase behavior, the system uses collaborative filtering and content-based filtering to recommend items users are most likely to enjoy.
The goal: improve user experience and boost conversions for online sellers.

---

❄️ Cold Start Problem

The cold start problem occurs when a recommender system lacks enough data to make meaningful suggestions. It appears in two main forms:

Types of Cold Start

• User Cold Start: A new user has no interaction history.
• Item Cold Start: A newly added item has no ratings or interactions.

---

🔧 How to Handle the Cold Start Problem

1. Rank-Based Recommendations

Suggest popular or highly rated items for new users when no personalized data exists.

2. Content-Based Filtering

Recommend new items by analyzing their attributes or metadata, matching them to users with similar tastes.

3. Hybrid Methods

Combine collaborative filtering, content-based filtering, and popularity-based methods for more robust performance.

---

📝 Examples

• User Cold Start: Use collaborative filtering to find similar users and recommend items they liked.
• Item Cold Start: Use content-based filtering to match the item’s features with user preferences.
• Hybrid Approach: Blend methods to reduce limitations of any single strategy.

---

📊 Dataset

The project uses an Amazon electronics ratings dataset.
Since the data has no headers, each user and product is assigned a unique numeric ID to avoid bias and preserve anonymity.

• Dataset link:
  https://www.kaggle.com/datasets/vibivij/amazon-electronics-rating-datasetrecommendation/download?datasetVersionNumber=1

---

🚀 Approach

1. Rank-Based Product Recommendation

Objective

• Recommend products with the highest number of ratings
• Provide popular items for new users (Cold Start Problem)
• Improve initial recommendations

Output

• Top 5 products with at least 50 or 100 interactions

Method

• Calculate average rating per product
• Count total number of ratings
• Combine metrics into a DataFrame and sort by rating
• Build a function to return the top n products based on interaction thresholds

---

2. Similarity-Based Collaborative Filtering

Objective

• Provide personalized recommendations using behavior of similar users

Output

• Top 5 recommended products based on similar user interactions

Method

• Convert user_id from object to integer (0–1539)
• Create a function to find similar users:
  • Compute cosine similarity for the target user
  • Sort users by similarity score
  • Exclude the target user
• Create a recommendation function:
  • Identify products the target user has already interacted with
  • Recommend products highly interacted with by similar users but not the target user

---

3. Model-Based Collaborative Filtering (SVD)

Objective

• Generate personalized suggestions using matrix factorization
• Address scalability and data sparsity issues in user–item matrices

Output

• Top 5 product recommendations per user

Method

• Convert the interaction matrix to a CSR (Compressed Sparse Row) matrix
• Apply Singular Value Decomposition (SVD) to reduce data into 50 latent factors
• Reconstruct predicted ratings using: U × Σ × Vᵀ
• Store predicted ratings in a DataFrame aligned with the original matrix
• Build a recommendation function to:
  • Combine actual and predicted ratings for a user
  • Filter out rated items
  • Sort by predicted rating
  • Return the top recommended products

Model Evaluation

• Calculate average actual vs predicted ratings
• Store in an rmse_df DataFrame
• Compute RMSE to measure accuracy of the SVD model