2013-Learning Deep Structured Semantic Models for Web Search using Clickthrough Data

Table of Contents

1. Problem Definition
2. Previous Methods & new Methods & mainly contribution
3. Application in RecSys
4. Ref
5. TODO & Questions & Further Reading

DSSM was initial invented for NLP word-matching task, recent years, it is widely used in candidate generation stage of recommendation system.

Problem Definition

1. Main objective:
   1. web document ranking task
2. Main metrics (only offline evaluation available)
   1. NDCG@1, 3, 5

Previous Methods & new Methods & mainly contribution

1. Main contributions
   1. learning from clickthrough data
   2. DNN supervised learning
   3. word hashing: n-gram based word hashing

Application in RecSys

1. Pros:
   1. fast: item-embedding calculated offline; user-embedding calculated in real-time
2. Cons:
   1. no interaction between item and user features

Ref

1. 为什么现在推荐系统喜欢用双塔模型？双塔模型相较于单塔有什么优缺点？ - 数据拾光者的回答 - 知乎
2. 推荐系统(十八) 大厂实践经验学习：双塔模型 - 缄默笔记的文章 - 知乎
3. 读透Learning Deep Structured Semantic Models for Web Search using Click through Data

TODO & Questions & Further Reading

1. Latent Semantic Model: intend to map a query to its relevant documents at the semantic level where keyword-based matching often fails
2. get cosine similarity, then apply softmax -> yes
3. word hashing: letter n-gram
4. to read
   1. Sampling-Bias-Corrected Neural Modeling for Large Corpus Item Recommendations: from google 2019, tried to removed the influence of popular item to negative sampling
      1. in-batch softmax ??就是利用batch内样本互相做彼此的负样本，来构建softmax损失
   2. Embedding-based Retrieval in Facebook Search: from facebook 2022
   3. deepmatch DSSM example

• Previous
  2021-Hotel2vec, Learning Hotel Embeddings from User Click Sessions with Side Information
• Next
  2016-Deep Neural Networks for YouTube Recommendations