[interesting reading] deepinf: social influence prediction with deep learning

1. Preface

I saw an interesting paper on the official account ：

pdf Address ：https://arxiv.org/pdf/1807.05560.pdf

Code address ：https://github.com/xptree/DeepInf

2. Reading notes

DeepInf: Social Influence Prediction with Deep Learning

2.1 Introduction of research points

In the abstract, the author also puts forward that the research of this paper is actually “social influence prediction”, That is, social impact prediction .

• Conventional “ Social impact forecast ” The method relies on manually defined rules to extract user and network features , However, these methods are limited by personal knowledge in relevant professional fields .
• therefore , In this paper, a deep neural network is designed “DeepInf”, To learn the user's potential feature representation , And then predict the social impact .
• To be specific , This paper designs some strategies to integrate the network structure 、 User characteristics .

Read here , So the traditional “ Social impact forecast ” What are the methods ？ Let's take a look at the method of comparison first ：

• Logistic Regression（LR）, Logical regression .
• Support Vector Machine (SVM), Support vector machine .
• PSCN, This is what the author said in the original ：“As we model social influence locality prediction as a graph classification problem, we compare our framework with the state-of-the-art graph classification models, PSCN[34].”, translate ： When we model social impact location prediction as a graph classification problem , We combine our framework with the most advanced graph classification model PSCN [34] Compare .

So why can these three baseline methods become the “ Tradition ” Social network prediction methods ？
The author divides them into two categories , stay LR and SVM in , Three types of features are considered （Vertex, Embedding, Ego), Here's the picture ：

about Vertex and Embedding Well understood. , That is, node characteristics and DeepWalk64 Embedded representation of dimensions , So in Ego Medium active neighbors What does that mean? ？ Continue to look at the relevant references given ：

• Group formation in large social networks: membership, growth, and evolution.

It is also mentioned in this paper ,“Of those communities which had at least 1 post, we selected the 700 most active communities along with 300 at random from the others with at least 1 post.”
That is, select the most active number of predefined numbers （ Post ） And random selection .

That is the third feature Ego In fact, it is defined by the author himself “ Most active ”.

Back to the topic , That is to say LR and SVM The above three features （Vertex, Embedding, Ego) Training data as a classifier , And then carry out classification training .

be aware ,PSCN From thesis ：“Learning convolutional neural networks for graphs, ICML’2016”, It is also a classification study considering the use of graph networks .

2.2 Related work

As we know from the previous section , It can be said that this article （DeepInf: Social Influence Prediction with Deep Learning） In fact, the study of is a study of graph classification , The method used is graph depth neural network . So and “Social Influence Prediction” What is the connection between ？ Let's start reading the introduction with this question .

Social impact ：“refers to the phenomenon that a person’s emotions, opinions, or behaviors are affected by others.”
“there is little doubt that social influence has become a prevalent, yet complex force that drives our social decisions, making a clear need for methodologies to characterize, understand, and quantify the underlying mechanisms and dynamics of social influence.”

Simply speaking , Social influence is the perception of others 、 emotional 、 The impact of the decision , And in the literature [26, 32, 42, 43] It is studied in .
Author's goal ：“We aim to predict the action status of a user given the action statuses of her near neighbors and her local structural information.”

The specific process is ：DeepInf, to represent both influence dynamics and network structures into a latent space. To predict the action status of a user v, we first sample her local neighbors through random walks with restart. After obtaining a local network as shown in Figure 1, we leverage both graph convolution and attention techniques to learn latent predictive signals.

Use random walk to get user structure characteristics , About social impact , In the second section of this paper .

2.3 Social impact

2.3.1 r-neighbors

That is, the shortest path is less than or equal to r A collection of nodes ：

And the subgraph formed by the set of the above nodes , be called r-ego netwrok, Described as ：

2.3.2 Social Action

The description here is very advanced , Make a note of ：

2.3.3 Social Influence Locality

According to the above two definitions are introduced , Here's the introduction of “Social Influence Locality” The concept of . Because in “Social Action” The forwarding behavior representation of time series is introduced in , Therefore, “Social Influence Locality” The concept of time series is also introduced into this concept ：
The probability of activation at the next moment is ：

Suppose there is N An example , Then the goal of the overall social impact prediction problem is ：

2.4 DeepInf

• Step one ： Neighbor sampling , Use BFS Extract node v Of r-ego The Internet , Expressed as $G_v^r$. But the size of this network may be due to “ Little world ” Characteristic and especially large , To solve this problem , Control the size , It can be understood that r-ego Perform secondary sampling on the network ; The sampling process samples according to the proportion of edge weights （ That is, biased ）, In addition, there is a certain probability that the random walk will start again at each time step ,

• Step two ： Neural network model , The main purpose is to integrate the structural attributes and behavior states of nodes , Finally, it is used to predict the forwarding behavior of users （ namely ：0 perhaps 1）.

be aware （d） The feature of the is composed of two parts ,（ Whether to activate , Whether it is ego） And its network embeddedness .

The middle part in the above figure is common , I won't introduce . Focus on the final comparison ：

It becomes a classification problem , That is, whether to forward .

The final comparative experiment also cites logistic regression 、 Support vector machine, etc , Classify them , Then compare the classification results .

3. summary

According to the previous reading , We know that actually the author of the paper “DeepInf: Social Influence Prediction with Deep Learning”
The problem to be solved is to embed possible influencing factors , Then predict whether the user will forward the message .

Personally feel the problem “Social Influence Prediction” Very advanced .