Product Recommendation SystemUsing Autoencoders

Abstract

The rapid expansion of e-commerce has made recommendation systems essential for helping users navigate vast product catalogs and discover items that match their preferences. Traditional approaches, such as collaborative filtering and content-based filtering, often struggle with challenges like the cold-start problem, where new users or items lack sufficient data, and scalability issues in large datasets. Deep learning techniques, particularly autoencoders, have emerged as powerful tools to address these limitations by learning compact, latent representations of data in an unsupervised manner This paper presents a product recommendation system that leverages an autoencoder to generate recommendations based on product metadata, including descriptions, categories, stores, and prices, from the Amazon Electronics dataset.
Our approach involves preprocessing the dataset to create a feature matrix using TF-IDF for text descriptions, one-hot encoding for categorical attributes, and Min-Max Scaling for prices The autoencoder, consisting of an encoder and decoder with dense layers and ReLU activations, compresses this feature matrix into a lower-dimensional latent space and reconstructs it to capture essential patterns Bayesian Optimization is employed to tune hyperparameters, such as the number of layers, units, and learning rate, ensuring optimal model performance The model is trained with Mean Squared Error (MSE) as the loss function and Cosine Similarity as an additional metric, using early stopping to prevent overfitting.The system was evaluated on a subset of 50,000 products, Visualizations of training and validation loss curves confirmed model convergence, while cosine similarity trends validated the quality of learned representations.
This work highlights the potential of autoencoder-based recommendation systems in e-commerce, offering accurate and scalable solutions. Future improvements could incorporate user interaction data or explore advanced architectures like variational autoencoders to enhance personalization. These results underscore the system's effectiveness in addressing traditional recommendation challenges, paving the way for more robust e-commerce solutions.