AI-POWERED RECOMMENDER SYSTEM USING DEEP LEARNING
FOR RETAIL TRANSACTION
Team Members:
Babyshalini C - 920821104009
Fahmitha Sirin N – 920821104019
Madhumitha J – 920821104031
Susmitha N – 920821104056
Abstract
The exponential growth of online retail, personalized product recommendations
have become a critical element in enhancing customer experience and boosting
sales. Traditional rule-based or collaborative filtering methods often fail to adapt
to dynamic consumer behavior. This project proposes an AI-powered
recommender system using deep learning techniques, comparing Feedforward
Neural Networks (FNN), Neural Collaborative Filtering (NCF), and Deep
Matrix Factorization (DMF). These models are trained on a transactional dataset
and optimized for high recommendation accuracy. The system integrates an end-
to-end pipeline, from data ingestion and transformation to model selection,
evaluation, and visualization. A Flask-based backend supports model inference
and React.js frontend offers interactive visual dashboards. The system achieves
over 90% accuracy and adapts dynamically to user and product profiles. Future
enhancements include real-time recommendation, user feedback loops, and
integration of advanced models like BERT4Rec.
Keywords: Recommender System, Deep Learning, FNN, NeuralCF, DMF,
Product Recommendation, Retail Analytics, Flask, React, Personalized AI
Introduction
The modern retail industry heavily depends on intelligent recommendation
systems to personalize customer journeys and improve engagement. Traditional
approaches often struggle with sparse data, cold-start problems, and inability to
capture complex user-product interactions. Deep learning offers a scalable and
adaptable alternative, learning latent features directly from large datasets.
This project presents a deep learning-driven recommendation engine built for
retail transactional data. It evaluates three powerful neural architectures — FNN,
NCF, and DMF — to determine the most effective model for personalized
product suggestions. Each model is trained using PyTorch and evaluated based
on RMSE, MAE, and Precision-Recall metrics. A robust backend using Flask
handles data requests and model inference, while React.js visualizes the training,
performance, and recommended products.
Proposed Work
FIG 1 : System Architecture
Fig 1 shows illustrates the system pipeline from data acquisition, preprocessing,
and model training to recommendation inference and web visualization.
Dataset Collection
Retail transaction data was sourced from a structured CSV format with features
such as customer name, product list, payment method, season, city, promotion
type, and total cost. These features are critical for modeling purchase behavior
and trends.
Data Preprocessing
The raw data underwent ETL processes — cleaning, normalization, encoding
categorical features, and generating product-user interaction matrices. Missing
values were imputed, and one-hot encoding was applied to features like
customer category and season. The final dataset was split into training and
testing sets.
Model Architectures
Feedforward Neural Network (FNN):
Utilizes dense layers to learn global patterns in user-product interactions.
Neural Collaborative Filtering (NCF):
Combines generalized matrix factorization with MLP to capture both linear and
nonlinear relations.
Deep Matrix Factorization (DMF):
Learns user and item embeddings via separate dense layers and combines them
for prediction.
Each model is trained with ReLU activations, dropout regularization, and Adam
optimizer.
Model Training
All models were trained on PyTorch using GPU acceleration with a learning rate
of 0.001 and batch size of 128. Training was monitored using validation loss and
evaluation metrics such as RMSE and MAE. Results are stored in a database for
comparison.
Flask Backend & React Frontend
Flask APIs manage model training, result serving, and recommendation
generation. The React.js frontend fetches data via REST APIs and displays
training curves, comparison tables, and top recommended products for each
customer. Chart.js is used for visual analytics.
Results
Experimental Results
The models were trained on transactional data over multiple epochs. The
performance was evaluated using:
Root Mean Squared Error (RMSE)
Mean Absolute Error (MAE)
Precision & Recall
Training Time & Model Complexity
Top product recommendations were generated for test users, showcasing
adaptability of each model.
Performance Comparison
Model RMSE MAE Precision Recall Accuracy
FNN 1.13 1.02 89% 86% 90.2%
NCF 0.97 0.89 91% 88% 92.6%
DMF 0.94 0.85 93% 90% 93.8%
Conclusion
The AI-powered recommender system successfully demonstrates how deep
learning models can outperform traditional methods in retail personalization. By
training on transactional datasets and dynamically choosing among FNN, NCF,
or DMF, the system achieves high accuracy and robust recommendations.
Deployed through a modern Flask-React interface, it enables seamless
interaction and real-time analytics. This project contributes to the growing
application of AI in retail intelligence and opens up future opportunities like
real-time learning, customer segmentation, and hybrid modeling with NLP
techniques.
STUDENT SIGNATURE GUIDE SIGNATURE
Baby Shalini C Dr.M.Indra Devi ME., Ph.D,
Fahmitha Sirin N
Madhumitha J
Susmitha N
