Social Influence Prediction Project Presentation Insights

Presentation for Project Social Influence Prediction Li Ziyi Wu Shushu Yue Ye Liu Du Presentation will be started right now

Background 01 Related information about influence prediction Framework Contents 02 Framework of the whole project Experiment Setup Contents of the whole presentation 03 We set up our experiments with large-scale real-world datasets Future Work 04 Future work of the project

Chapter 1 Background Related information about influence prediction

Background Social Influence Prediction Social Activities Social and information networking activities such as on Facebook, Twitter, WeChat, and Weibo have become an indispensable part of our everyday life. So a social influence prediction for each user is critical for a variety of applications such as online recommendation and advertising.

Social Influence Prediction Social Influence Prediction Aggregated/global social influence User-level social influence Prediction target: cascade size within a time-frame Prediction target: action status of a user

Chapter 2 Framework Framework of the whole project

Formulation Social Influence Prediction u, v: the two nodes of the edge; w: the weight of the edge; r: the relationship of the two nodes; E = {(u, v, w, r)} V = {(v, s)} v: the node; s: the state of the node;

Formulation Social Influence Prediction v v ?? 0,1 : a binary action status of user u :

Formulation Social Influence Prediction 0 : the action states of its neighbors So the activation probability of v: Hence, this social influence prediction can be formulated as a binary classification problem: min

Model Framework Social Influence Prediction Instance Normalization Embedding Layer Output Layer GCN Layer Raw Input mini-batch of size B xv yv v Loss avv avu u xu yu NetSMF

Network Embedding Social Influence Prediction Given an undirected and weighted network ? = of n vertices, ? as the edge set of m edges, and ? as the adjacency matrix, the goal is to learn a function ? ??that maps each vertex to a d-dimensional (? ?) vector that captures its structural properties, e.g., community structures. ?,?,? with ? as the vertex set

Network Embedding DeepWalk 1. 2. It generates several vertex sequences by random walks over a network It applies the skip-gram model on the generated vertex sequences to learn the latent representations for each vertex. Commonly, skip-gram is parameterized with the context window size T and the number of negative samples b. R

Network Embedding NetMF A theoretical study reveals that DeepWalk essentially factorizes a matrix derived from the random walk process. More formally, it proves that when the length of random walks goes to infinity, DeepWalk implicitly and asymptotically factorizes the following matrix: log (???(?) ?) ? Note that the matrix as above would be ill-defined if there exist a pair of vertices unreachable in T hops, because log(0) = . So NetMF uses the logarithm truncated at point one, that is, trunc_log (?) = max (0,log (?)). Thus, NetMF targets to factorize the matrix tran_log (???(?) ?) ? *2018. Network Embedding as Matrix Factorization: Unifying DeepWalk, LINE, PTE, and node2vec. In WSDM 18. ACM, 459 467.

Network Embedding NetSMF 1. Random-Walk Molynomial Sparsification 2. Construct a NetMF Matrix Sparsifier 3. Truncated Singular Value Decomposition

GCN Based Network Encoding Social Influence Prediction What is Graph Convolutional Network (GCN) ? A generalization of neural network models to graph(or networks) structured datasets The goal is to learn a function of signals/features on a graph G = (V, E) which takes as input : A feature description xifor every node i; summarized in a N D feature matrix X ( N: number of nodes, D: number of input features) A representative description of the graph structure in matrix form; typically in the form of an adjacency matrix A (or some function thereof) And produces a node-level output Z (an N F feature matrix, where F is the number of output features per node) *Thomas N Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. ICLR 17 (2017).

GCN Based Network Encoding Social Influence Prediction GCN layer-wise propagation rule Each layer can be written as a non-linear function Consider a very simple form is for every node, we sum up all the feature vectors of all neighboring nodes but not the node itself A is typically not normalized and therefore the multiplication with A will completely change the scale of the feature vectors. Using a symmetric normalization : D is the diagonal node degree matrix.

GCN Based Network Encoding Social Influence Prediction GCN : a two layer example The forward model then : where

Chapter 3 Experiment Setup We set up our experiments with large-scale real-world datasets to quantitatively evaluate the proposed framework.

Datasets Social Influence Prediction Our experiments are conducted on four social networks from different domains OAG, Digg, Twitter, and Weibo. OAG Digg Twitter Weibo |V| 953,675 279,630 456,626 1,776,950 |E| 4,151,463 1,548,126 12,508,413 308,489,739

Datasets Social Influence Prediction Digg OAG Open Academic Graph dataset is generated by linking two large academic graphs: Microsoft Academic Graph and Aminer. The social network is defined to be the co-author network, and the social action is defined to be citation behaviors a researcher cites a paper from the above conferences. We are interested in how one s citation behaviors are influenced by her collaborators. Digg is a news aggregator which allows people to vote web content, up or down. The dataset contains data about stories promoted to Digg s front page over a period of a month in 2009. For each story, it contains the list of all Digg users who have voted for the story up to the time of data collection and the time stamp of each vote. The voters friendship links are also retrieved. Weibo Twitte rThe Twitter dataset was built after monitoring the spreading processes on Twitter before, during and after the announcement of the discovery of a new particle with the features of the elusive Higgs boson on Jul. 4th, 2012. The social network is defined to be the Twitter friendship network, and the social action is defined to be whether a user retweets Higgs related tweets. Weibo is the most popular Chinese microblogging service. The complete dataset contains the directed following networks and tweets (posting logs) of 1,776,950 users between Sep. 28th, 2012 and Oct. 29th, 2012. The social action is defined as retweeting behaviors in Weibo a user forwards (retweets) a post (tweet).

Evaluation Metric Social Influence Prediction Parameter Sensitivity Prediction Performance AUC: Area Under Curve (TPRate-FPRate) We analyze several hyper-parameters in our model and test how different hyper- parameter choices can influence prediction performance. ?? Prec.: ???? = ??+?? ?? Rec.: ?????? = ??+?? 2 F1: ?1 = 1 1 ????.+ ???.

Comparison Methods Social Influence Prediction PSCN 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. For each graph, PSCN selects w vertices according to a user-defined ranking function, e.g., degree and betweenness centrality. Then for each selected vertex, it assembles its top k near neighbors according to breadth-first search order. For each graph, The above process constructs a vertex sequence of length w k with F channels, where F is the number of features for each vertex. Finally, PSCN applies 1-dimensional convolutional layers on it. AUC: Prec: Rec: F1: PSCN 0.8130 0.4717 0.7322 0.5738 GCN 0.8299 0.5019 0.7259 0.5934

Chapter 4 Future Work Future work of the project

Future Work Social Influence Prediction More Datasets Other types of data There re many excellent datasets about social influence for training. Also, some special relationships involving edges in the network can be studied in the future work in other datasets. And our model performs well on the existing datasets, and we will continue to test the performance on more datasets. Some more relationships are out there for further mining as well

Social Influence Prediction Thanks for Listening Li Ziyi Wu Shushu Yue Ye Liu Du That s all for today s presentation thank you!

Social Influence Prediction Q&A Li Ziyi Wu Shushu Yue Ye Liu Du