2019-Neural Product Retrieval at Walmart.com

Table of Contents

1. Problem Definition
2. Methods
3. TODO & Questions & Further Reading

Problem Definition

1. Mainly focus on product retrieval in Walmart -> match the query and products
   1. To solve the problem of product retrieval, the key is to assess the relevance between a query and a product. I
2. Two stage model
   1. retrieval
   2. re-ranking

Methods

1. All implementation based on tensorflow ranking
2. Pair-wise loss has a better performance than pointwise loss
3. Our neural retrieval model in Section 3 is closely related to CLSM and DSSM.
4. In this section we propose three different models:
   1. A simple bag-of-token model based directly on tokens in query and product,
   2. a neural text model which can incorporate semantic similarity between query and product, and
   3. lastly a customized neural text model which is more optimized for retrieval.
5. Labels: Note that though we are optimizing for order rate, we did not use order rate as label for training because order data is sparser compared with click data.
6. Position bias
   1. a simple click model

TODO & Questions & Further Reading

1. TF ranking
2. An Introduction to Neural Information Retrieval

• Previous
  2020-Improving Deep Learning For Airbnb Search
• Next
  2023- Learning To Rank Diversely