Today, I share an article published by Baidu on SIGIR 2021 In terms of click through rate estimation Short Research Paper The paper .
Deep neural network (DNN) The model has been widely used in online advertising click through rate (CTR) forecast .CTR The training framework usually consists of embedded layer and multi-layer perceptron (MLP) form . In Baidu search advertising system ( Also known as phoenix nest ,Phoenix Nest), A new generation CTR The training platform has become PaddleBox, One is based on GPU Parameter server system .
In this paper , The author introduces Baidu's recently updated CTR Training framework , A multitask neural network called gating enhancement (GemNN). Specially , They developed a multi task learning model based on neural network for CTR forecast , It gradually reduces candidate advertisements in a coarse-grained to fine-grained manner , The parameters between upstream tasks and downstream tasks are allowed to be shared , So as to improve the training efficiency . Besides , The author is also embedding layers and MLP A gating mechanism is introduced between , Used to learn feature interaction and control transfer to MLP Layer information flow .
The author is in Baidu PaddleBox The platform deploys the model scheme , Considerable improvements have been observed in both offline and online assessments , It is now part of Baidu's advertising system .
Multi-task NN Model With Parameter Sharing
Address of thesis ：https://arxiv.org/pdf/2007.03519.pdf