Similar Product Identification System in Amazon (2020)
I was trying to learn more about identifying similar products using ML, when I came across this paper by a few data scientist in Amazon. Here are some highlights from the paper:
Product Embeddings
Embeddings are a common way to extract features from image and text. They are obtained by passing an observation (e.g. image or text) through a large pre-trained model, and extract one of the last few layers as embeddings. In this paper, they used:
- AlexNet for image feature
- Word2Vec trained on Google News for text features
After extraction, each embeddings are normalized and concatenated together for the next step.
SiameseNets Deep Learning Architecture
A Siamese neural network is used to train the model. During training, 2 items will be compared against each other, and a L2 distance will be computed between them. A pariwise loss will then be computed, and back-propagated to optimize the network. We need two types of labels here: positive - 2 products that are similar to each other, negative - 2 products that are not similar to each other.
Labelled data
I was interested to know whether there are any data-driven approach to mine labels automatically. In this paper, they tried to use customer view-to-purchase data.
- For positive pairs, the most recent viewed products and the purchased product right after that were selected
- For negative pairs, producted viewed N slots before the purchaesd product were selected
Hand-labelled dataset still performed better than the data-driven method. But a combination of both seems to produce the best result.
Other approaches
- In the Shopee - Price Match Guarantee Kaggle competition, embeddings and ArcFace were used