[recommendation system] the classic technical architecture of the data flow of the recommendation system + the most complete evolution map of the top 10 deep learning CTR models of Microsoft, Alibaba,

One 、 recommend 、 advertisement 、 The difference between search systems ？

1.1 Differences in fundamental issues

• advertisement ： The goal of advertising algorithm is to directly increase the revenue of the company
• Search for ： Response to questions about efficient access to information about search terms
• recommend ： Increase user participation , Improve user stickiness and retention

1.2 The difference between optimization objectives

• advertisement ： forecast CTR and CVR, Reverse the value of traffic
• Search for ： Value being able to recall the correct answer
• recommend ： The recommended algorithm has different goals , The video category is more inclined to the video playback market , News forecasts CTR Click through rate , Estimated customer unit price of e-commerce

1.3 Differences in the model itself

classical Attention Recommended model

Fused sequence structure DSIN

Google Play Search two tower model of

Two 、 Recommend the technical architecture of the system

2.1 Data section

Mainstream big data architecture

• Batch Architecture
• Streaming architecture
• Lambda framework
• Kappa framework
• Unified framework

2.2 The model part

Recall layer ：

• Embedding
• Local sensitive hash
• Hot item recall
• Social relations recall
• Recall of fresh items

Sort layer ：

• Collaborative filtering class model
• LR、FM、MLR
• Composite class model
• Deep learning model

Supplementary strategies and algorithms ：

• diversity
• The real time
• Popularity
• Freshness

2.3 Technical architecture diagram of recommended system data flow

• Client real-time features ： Often use the client to collect time 、 place 、 Recommend contextual features such as scenes , Then let these features follow http Requests arrive at the server side together , Participate in model prediction .

• Quasi real time feature processing of stream processing platform ： The so-called stream processing platform , Is to stream logs mini batch A quasi real-time computing platform for processing , The features calculated by the stream processing platform can be immediately stored in the feature database for the use of the recommended system model , Although it is impossible to change user results according to user behavior in real time , But the minute level delay can basically ensure that the user's recommendation results are affected by the previous behavior in quasi real time .

• Full feature processing of distributed batch processing platform ： As the data finally arrives HDFS Mainly distributed storage system .Spark The equal distribution computing platform can finally calculate and extract the full amount of features . At this stage, we also focus on the data of multiple data sources join And delay signal combination .

3、 ... and 、 The evolution trend of deep learning recommendation model

3.1 Pre deep learning era CTR The evolution of prediction model

3.2 Google 、 Ali 、 Microsoft, etc 10 Deep learning CTR The most complete evolution map of the model 【 recommend 、 advertisement 、 Search area 】

3.3 CTR Basic model structure

3.4 DIN The Internet

3.5 DIEN The Internet

3.6 How to calculate according to the historical behavior data of users CTR？

• Mode one ： Consider the impact of all behavioral records , utilize average pooling take embedding vector Average the number of people who form this user user vector

• Mode two ： Use time decay, Let recent behavior have a greater impact , Doing it average pooling Adjust the weight according to the time

• Mode three ： introduce attention Mechanism , Add different weights to different behavioral interests