Skip to main content
Open Access Publications from the University of California


UCLA Electronic Theses and Dissertations bannerUCLA

An improvement on E-Commerce Search Engine for Automotive Parts Industry Using Siamese Neural Network with Triplet Loss and Contrastive Loss by Training Catalog Item Embeddings


This paper indicates the use of siamese neural network with triplet and contrastive loss function to improve the customers’ online search experience of automotive parts products inspired by the DoorDash’s siamese network implementation article for their item catalog[Ram19]. These networks have neural network architectures that consist of sub-networks which are identical copies of each other, and are structured to learn more effectively when dealing with imbalanced data (even with a very few quantities of some labeled data included) than traditional deep neural networks that solve classification problems. These models are chosen because of the large number of product categories and imbalance data in auto parts industry. After words embeddings are passed through LSTM encoder which compresses word embeddings into sentence embeddings, these siamese models can provide the advantage of semantic sentence embeddings as outputs by adjusting the parameters and weights of embeddings by finding the similarity of each sentence. Lastly, comparing various text mining processes and hyperparameter tuning shows the best performing model with accuracy as 88.05%.

Main Content
For improved accessibility of PDF content, download the file to your device.
Current View