“A revolutionary click prediction model” – paper abstract

Time:2022-5-23

1. Foreword****

For the current single ad Click prediction, we can get the paired correlation between elements, but we can’t know the local correlation between elements. Moreover, for different time intervals, the existing processing methods of sequential click prediction do not have a good effect. For this problem, this paper proposes a novel model, convolutional click prediction model (CCPM), which is based on convolutional neural network. CCPM can capture the local characteristics of input elements, whether in a single click event or sequential clicks within a certain time interval.

2、Introduction

Matrix factorization (MF) and factorization machines (FM) have been widely used in recommendation systems. However, MF and FM only capture low-order combined features in a single click advertisement, ignoring high-order combined features. In order to make full use of the information of historical sequential clicks, a prediction model based on RNN is proposed. This model mainly takes the historical records viewed by users as a sequence, and divides the historical click sequence of users in different time intervals. However, in the real scenario, users’ interest in goods will change over time, and RNN model is limited in the current scenario.

In order to mine important semantic features in complex scenes, DNN is a good choice. The convolution layer and pooling layer in CNN can well find local important features. CCPM is the first to use CNN to improve the accuracy of click prediction.

3、CCPM

Structure diagram:

In a single ad impression, there are many elements: element = (user; query; ad, impression time, site category, device type, etc). Whether a user clicks an ad is related to the user’s history ad impression. A series of ad impressions form sequentia ad impression. The click probability is predicted based on the above two cases.

CCPM consists of revolutionary layers + flexible p-max pooling layers.

3.1、Convolution Layer

Each sample has n features. For each feature, use embedding to obtain the vector EI ∈ RD with a fixed length of D, and then construct a matrix s ∈ rdxn. After obtaining the S matrix, CNN can be applied.

3.2、Flexible p-Max Pooling

Since the input length is variable, in order to reduce this impact, the parameters of the corresponding pool layer should also be flexible and variable. The definition PI:

Where, l represents the number of layers of the convolution layer, n represents the input length (characteristic number), and PI represents the parameters of the ith pool layer. For example, when n = 18 and L = 3, the corresponding values of P are P1 = 16, P2 = 6 and P3 = 3 respectively.

In sequential click prediction, the value of P is very important. It can not only select the most important local features, but also save the relative order of these features.

3.3、Feature Maps

Each sample can be represented by a matrix, that is, feature map. Then, each feature map is convoluted and pooled, and the final result is connected to the full connection layer.

This work adoptsCC agreementThe author, link and reprint of this article must be indicated

article!! Started on my blogStray_Camel(^U^)ノ~YO

Recommended Today

06 cubemx + keil + Proteus simulation STM32 – GPIO (V)

The example of this paper refers to the development example of STM32 single chip microcomputer – based on Proteus virtual simulation and Hal / ll librarySource code:https://github.com/LanLinnet/STM33F103R6 Project requirements Realize matrix keyboard scanning. When any button is pressed, the nixie tube will immediately display the corresponding key value of the currently pressed button. hardware design […]