dried food! Information diffusion prediction based on sequence hypergraph neural network

Click on the blue words

Pay attention to our

AI TIME Welcome to everyone AI Fans join in ！

Diffusion cascade prediction is the key to understanding the spread of information on social networks . Most methods typically focus on the order or structure of infected users in a single cascade , Therefore, global users and cascading dependencies are ignored , Limited predictive performance . The current strategy of introducing social networks can only obtain the social homogeneity between users , Not enough to describe their preferences .

In order to solve the above problems , We propose a new information diffusion prediction method , Called memory enhanced sequence hypergraph attention network （MS-HGAT）.

say concretely , In terms of learning users' global dependencies , We not only take advantage of their friendship , Also consider their interactions at the cascade level . Besides , In order to dynamically capture user preferences , We divide the diffusion hypergraph into a series of subgraphs based on timestamp , Construct hypergraph attention network to learn sequence hypergraph , And the gating fusion strategy is used to connect them .

Besides , A memory enhanced embedded search module is also proposed , It is used to capture the learned user representation to a specific cascade embedding space , So as to highlight the sequence information inside the cascade .

In this issue AI TIME PhD studio , We invite you to Xi'an Jiaotong University 2019 I'm a direct doctoral student —— Sun Ling , Bring us report sharing 《 Information diffusion prediction based on sequence hypergraph neural network 》.

Sun Ling ：

Xi'an Jiaotong University 2019 I'm a direct doctoral student , Under the guidance of Professor Rao yuan . The main research direction is information communication modeling for social media 、 False information detection and intervention methods , Relevant work is published in AAAI、IJCAI、ACL Wait for the International Conference .

Research background

Information diffusion prediction

Given ： Historical diffusion data （ Cascade tree ： complex 、 Hard to get , Cascade sequence ）

Mission ：

●  At the micro level ：

Research motivation

Representation based learning ：

The current problems are all based on the fixed propagation mode , But in the real information scenario, the predefined patterns may not be followed . There are two main problems in the current method ：

1. Most studies ignore the global dependence of users in the process of information consumption

Our solution ： At the same time, mining the global friendship relationship and global diffusion interaction relationship

There may be cascading interactions between different users in the above networks .

2. Few models can learn more about the dynamic relationship between users and cascades

Our solution ： Set up a series of hypergraphs to dynamically learn the interaction between users and cascades in a specific time interval .

An edge in a hypergraph may contain multiple points , We construct a series of hypergraphs to represent the interaction of users at the global level in different time intervals and learn their dynamic preferences .

MS-HGAT Model

The first part is the static learning of user relationship network ;

The second part is to dynamically use a series of propagating hypergraphs to learn their dynamic interaction behavior ; The memory enhanced embedded search module stores the static and dynamic learned user representations in the memory block for subsequent search .

MS-HGAT

Module1： Users statically rely on learning

The second module is the main innovation of this study , It mainly studies the dynamic interaction of users .

First step , Because of users 1,2,3 At the same time, it participates in the super edge , According to this 3 Users and the root on the super edge （root） Distance between users , Calculate the attention coefficient , And the weighted aggregation is the representation of super edge .

The second step , It is a process of Super Edge aggregation back to point . Our first step converged to the edge , Learned 3 The characteristics of bar cascade , And then back to the point . That is, the edge features learned in the global scope are returned to the nodes . We can finally learn the user's interaction preferences . however , This hypergraph learns short-term user interaction behavior , Instead of calculating extra attention, the average aggregation is calculated directly .

The third step , Is an update . That is, point to edge aggregation again , Update the representation of the superedge , Then it is stored in the memory module .

The memory module mainly stores the features of users and hyperedges , So as to facilitate subsequent search .

The hypergraph of global diffusion is divided into several subgraphs according to the timestamp , Subgraphs are connected through an adaptive fusion gate mechanism .

We aggregate the input and output of the previous moment as the input of the next stage .

Finally, perform embedded search , Given a specific cascade , Find the static and dynamic representation features in the memory module . In the aspect of dynamic search, it is necessary to fuse the results of user angle and cascade angle .

We learn static and dynamic cascading representations in interactive networks , But the graph neural network can not learn the sequence relationship between the participating users .

therefore , We introduced transformer Self attention decoder module in , Learn more about the interaction within the cascade .

Last , We also use these two representations through adaptive fusion methods , The resulting output is our final cascaded representation , According to this expression , We predict which users will join the cascade next .

experiment

● RQ1：MS-HGAT Whether it is superior to the most advanced information diffusion prediction methods ？

● RQ2： How the quantity and quality of training sets affect the prediction performance of the model ？

● RQ3： How user relationships and our learning strategies affect MS-HGAT Prediction performance of ？

We changed it K value , Look at the model in K Performance when taking different values .

We can also see , It's better to use global users and interactions than just cascading to make predictions .

The above experimental results also show that , Our model has more in-depth feature capture .

We adjusted the maximum length of the trainable couplet , It can be seen from experiments that our model can achieve good results on any length of cascades .

In the ablation experiment , We prune the modules and analyze their impact .

First, we removed the social network graph , The second is to remove the diffusion network graph , The third and fourth remove the user's memory module and the cascading memory module respectively , The fifth one converts all the gated fusion mechanisms into a simple form of splicing .

Future work direction

● Use hypergraphs to describe tree cascades , Not just sequence .

●  Combine information content with diffusion characteristics .

●  Consider the diverse user behavior in social platforms （ for example “ give the thumbs-up ” and “ Comment on ”）

carry

Wake up

Thesis title ：

MS-HGAT: Memory-enhanced Sequential Hypergraph Attention Network for Information Diffusion Prediction

Thesis link ：

https://www.aaai.org/AAAI22Papers/AAAI-4362.SunL.pdf

Click on “ Read the original ”, You can watch this playback

Arrangement ： Lin   be

author ： Grandchildren   Water caltrop

Excellent articles in the past are recommended

Remember to pay attention to us ！ There is new knowledge every day ！

  About AI TIME 

AI TIME From 2019 year , It aims to carry forward the spirit of scientific speculation , Invite people from all walks of life to the theory of artificial intelligence 、 Explore the essence of algorithm and scenario application , Strengthen the collision of ideas , Link the world AI scholars 、 Industry experts and enthusiasts , I hope in the form of debate , Explore the contradiction between artificial intelligence and human future , Explore the future of artificial intelligence .

so far ,AI TIME Has invited 600 Many speakers at home and abroad , Held more than 300 An event , super 210 10000 people watch .

I know you.

Looking at

Oh

~

Click on Read the original   View playback ！