Matrix Factorization vs. Deep Learning Autoencoders: A Performance Analysis in E-commerce Personalization
Rishabh Sharma, Abhinav Mishra · 2024
The development of sophisticated recommendation algorithms has taken place due to the evolution of online shopping platforms that aim to offer personalized content to users to improve their engagement and satisfaction. This research paper examines the efficacy of two prominent collaborative filtering techniques: traditional matrix factorization and different types of deep learning-based autoencoders are being discussed within the aspect of the e-commerce recommendation system. The comparative approach of this research shall help evaluate the efficiency of each approach toward problem resolutions such as data scarcity, scalability, and popularity fluctuation. The demonstrated value of the matrix factorization model (as the consequence of its broad use and the proven history of success with processing big datasets) makes it the initial functional model choice; the autoencoder model assumes a stand-in position to prove its capability to approximate complex, non-linear dependencies between data entries which other models may underline. The evaluation of the models was based on mean absolute error (MAE), root mean square error (RMSE), and normalized discounted cumulative gain (NDCG) across an available e-commerce dataset divided into training and test sets that are open to the public. The result shows that the “deep learning-based autoencoder model” approach gets better scores than the “matrix factorization-based model” and achieves less MAE and RMSE scores and higher NDCG value which means that it has more ability to make exact and relevant recommendations than others. According to the results, the autoencoders can emphatically increase the dimensions of the quality of the recommendations, delivering to the users a highly customized fulfillment during their shopping. This study not only gives some indication of the practical implications of advanced machine learning techniques that can be successfully used in improving the recommendation systems but also suggests lines for future research, e.g. the optimizations and tunings of the deep learning models in eclectic contexts.