CN115618124A - Propagation popularity prediction method for bidirectional social influence learning - Google Patents
Propagation popularity prediction method for bidirectional social influence learning Download PDFInfo
- Publication number
- CN115618124A CN115618124A CN202211196722.XA CN202211196722A CN115618124A CN 115618124 A CN115618124 A CN 115618124A CN 202211196722 A CN202211196722 A CN 202211196722A CN 115618124 A CN115618124 A CN 115618124A
- Authority
- CN
- China
- Prior art keywords
- graph
- node
- time
- attention
- propagation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 230000002457 bidirectional effect Effects 0.000 title claims abstract description 48
- 238000000034 method Methods 0.000 title claims abstract description 37
- 230000007246 mechanism Effects 0.000 claims abstract description 42
- 238000012549 training Methods 0.000 claims abstract description 18
- 239000013598 vector Substances 0.000 claims description 34
- 238000004364 calculation method Methods 0.000 claims description 26
- 230000008447 perception Effects 0.000 claims description 26
- 238000013528 artificial neural network Methods 0.000 claims description 22
- 238000011176 pooling Methods 0.000 claims description 18
- 230000009466 transformation Effects 0.000 claims description 13
- 239000011159 matrix material Substances 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 6
- 230000002776 aggregation Effects 0.000 claims description 6
- 238000004220 aggregation Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 5
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 230000003044 adaptive effect Effects 0.000 claims description 3
- 238000012512 characterization method Methods 0.000 claims description 3
- 238000013135 deep learning Methods 0.000 claims description 3
- 230000000644 propagated effect Effects 0.000 claims description 3
- 230000002123 temporal effect Effects 0.000 claims description 3
- 238000013519 translation Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 2
- 230000036962 time dependent Effects 0.000 claims description 2
- 238000010276 construction Methods 0.000 claims 1
- 230000005540 biological transmission Effects 0.000 abstract description 7
- 238000010586 diagram Methods 0.000 description 6
- 230000008569 process Effects 0.000 description 5
- 230000006399 behavior Effects 0.000 description 3
- 230000036961 partial effect Effects 0.000 description 3
- 230000003068 static effect Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000002860 competitive effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000002829 reductive effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000007480 spreading Effects 0.000 description 1
- 238000003892 spreading Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/01—Social networking
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Economics (AREA)
- General Health & Medical Sciences (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a propagation popularity prediction method for bidirectional social influence learning. The method comprises the following steps: collecting information transmission network data and constructing graph sequence data of information transmission; constructing a model based on a graph sequence attention network; constructing a macroscopic dynamic popularity prediction model based on the combination of a graph sequence attention network and a decoder; inputting the graph sequence data into a macroscopic dynamic popularity prediction model based on the combination of a graph sequence attention network and a decoder for iterative training; and inputting the graph sequence data of information propagation, and predicting the popularity of the information propagation at different moments by using the model obtained by training. The invention applies a multilevel bidirectional attention mechanism to learn the space-time characteristics of the transmission dynamic graph, effectively learns the local and global dependency relationship of the transmission network subgraphs at different moments, considers the bidirectional social influence information of the transmission subgraphs at the previous and next moments on the structure and has better usability.
Description
Technical Field
The invention relates to the field of social network information propagation research, in particular to a propagation popularity prediction method for bidirectional social influence learning.
Background
The online social network provides a platform channel for information transmission, meets the requirements of users for expressing self, transmitting innovation ideas and transmitting scientific and technical information, and also provides a channel for enterprise competitive product analysis and marketing planning innovation. The social network is convenient for the masses to know the information issued by local government affair handling organizations and enterprises and public institutions in time, and the information gap existing in the online social network is reduced. The information dissemination popularity prediction can help us to mine and recommend microblogs or articles with high quality and capable of arousing wide user interests. By modeling and predicting the dynamic changes of the propagation popularity, the social driving factors of the propagation behaviors among the users can be understood. The method has the advantages that the macroscopic popularity growth trend of information dissemination is researched, the future dissemination popularity is predicted in the early stage of information dissemination, and the method has important social significance and commercial value.
The information cascade propagation track in the social network forms a dynamic information cascade propagation network, the cascade propagation network records the dynamic growth process of the information, the network structure of the cascade propagation network also changes with time, and the propagation scale (popularity) in each time period also changes with time. The information or microblog publisher induces the user to generate forwarding behavior depending on the attraction of the information content and the social influence of the information or microblog publisher, the forwarding behavior of the user also influences the propagation growth, and the social influence in the information propagation has bidirectional property. The social network propagation has the characteristic of bidirectional social influence, and the bidirectional inspiration helps us to learn forward structural information and consider backward structural information when establishing a propagation model. In the prior art, the DeepHawkes model [ Cao Q, shen H, cen K, et al. The method comprises the steps of Bridging the gap between prediction and interpretation of information broadcasts, learning path feature vectors on a propagation path, predicting the static popularity of information propagation by using the time sequence features of a recurrent neural network learning Hooke process, and modeling the dynamic characteristics of propagation subgraphs at each moment.
The traditional graph neural network GNN is mostly only suitable for static graph structure learning and cannot model a dynamic graph network, while an information propagation network, an academic introduction network, a social relation network and the like are dynamic graphs, and the structure of the graphs changes along with time, so that the existing graph neural network is difficult to directly apply to the dynamic graph network.
Disclosure of Invention
The invention aims to provide a propagation popularity prediction method based on bidirectional social influence learning aiming at the defects of the prior art.
The purpose of the invention is realized by at least one of the following technical solutions.
The propagation popularity prediction method for the bidirectional social influence learning comprises the following steps:
s1, collecting information propagation network data, and constructing graph sequence data of information propagation, wherein the graph sequence data comprises a graph adjacent matrix, a node representation feature and time sequence representation feature data;
s2, constructing a model based on a graph sequence attention network, wherein the model comprises a graph deformation module and a deformation coding module based on bidirectional local graph attention at a node level and a bidirectional graph sequence attention mechanism at a graph level;
s3, constructing a macroscopic dynamic popularity prediction model based on the combination of the graph sequence attention network and a decoder, wherein the macroscopic dynamic popularity prediction model comprises the graph sequence attention network and a time sequence perception attention decoder;
s4, inputting the graph sequence data into a macroscopic dynamic popularity prediction model based on the combination of a graph sequence attention network and a decoder for iterative training;
and S5, inputting graph sequence data of information propagation, and predicting the popularity of the information propagation at different moments by using a macro dynamic popularity prediction model obtained by training and based on the combination of a graph sequence attention network and a decoder.
Further, in step S1, information propagation network data is collected, information is propagated in the social network to form a propagation network, and a post m issued by a user u is forwarded by a user v at time t to form a propagation quadruple data (v, u, t, m); at time i, the publisher and all the forwarders of a post m form a propagation network G i (ii) a Modeling a propagation network as a sequence of graphs, the propagation network G at all times i Construct a graph sequence data, denoted as { G } 0 ，G 1 ，...，G T ，G T+1 ，G T+2 A. }, T =0,1,2.., T, where T is the observation time, the node features and the subgraph structure of the corresponding subgraph snapshot at time i are denoted as
And
N i is subgraph G of time i i F is the initial characteristic dimension of the node, and the time sequence characteristic of the time i is recorded as
The time sequence characteristics, the node characteristics and the sub-graph structure at all the moments respectively form time sequence representation characteristics, node representation characteristics and graph adjacency matrix data of graph sequence data.
Further, in step S2, attention based on the bipartite local graph is specifically as follows:
features of a node in a propagation subgraph
Not only has a relation with the current neighbor, but also is influenced by the subgraphs of the forward and backward moments, and the calculation formula is as follows:
wherein f is combine Is a method of combining the current node characteristics and the previous and subsequent time node characteristics,
representing the characteristics of the propagation subgraph of node v at time t.
The social influence of the neighbor node of the node on the node in the attention of the bidirectional local graph is specifically as follows:
initial feature x of node v v Is converted into conversion characteristics through a neural network layerSign
The calculation formula is as follows:
wherein, W l Are parameters of the neural network layer and,
f represents x v L layers of initialized feature dimension, F l Is a layer transformation characteristic
The LeaklyReLU is an activation function;
transformation characteristics of node v after transformation through neural network layer
And translation characteristics of neighbor node u of node v
And carrying out aggregation, wherein u belongs to N (v), and N (v) is a neighbor node set of the node v, and the method comprises the following steps:
wherein, f prop For calculating attention values, e.g.
Is the attention mechanism sharing neural network parameters, e uv The attention value of the node v for the neighbor node u of the node v,
to calculate the attention value of node v to node u.
The weight calculation mode of the neighbor node u of the node v acting on the node v is as follows:
wherein alpha is uv Is the attention coefficient of the neighbor node u of the node v to the node v after normalization, the node k is one of the neighbor node set of the node v, e kv The attention value of the node v is the neighbor node k of the node v.
Compute node v aggregates updated transformation features
The calculation method is as follows:
where σ is an activation function
The characteristics of a node in the propagation subgraph are also affected by the subgraphs at the forward and backward moments, the subnetwork G at the previous moment t-1 t-1 Social impact of intermediate node u on node v
The calculation method is as follows:
wherein u ∈ V t-1 ，V t-1 F is a neighbor node set of a node v in a sub-network at the previous time t-1, a calculation method of node aggregation coefficients at the current time and the forward time is set as an Euclidean space of node characteristics,
n is the dimension of the feature, sub-network G of the later time t +1 t+1 Social impact of intermediate node u on node v
The calculation method is as follows:
wherein u ∈ V t+1 ，V t+1 Which is the set of neighbor nodes for node v in the sub-network at the later time t + 1.
Further, in step S2, the attention mechanism based on the bipartite graph sequence is as follows:
generating a time sequence-related co-production growth dependence relationship between a subgraph and a non-adjacent subgraph at a certain time of a non-adjacent subgraph, and learning the bidirectional long dependence of the subgraph on the time sequence by adopting a multi-head self-attention mechanism; the calculation formula of the space-time characteristic Q integrating the propagation structure and the time sequence characteristic is as follows:
wherein,
is the characteristic of the propagation subgraph snapshot node characteristic at the moment t after the average pooling, tp t For the time-series representation of the time t, X t For the sub-graph node feature at time t, sumPool is the sum pooling operation, W t Is a Linear neural network parameter;
the space-time characteristics Q are learned through a multi-head self-attention mechanism, so that the graph sequence attention network model can deeply fuse propagation structures and time sequence characteristics, and the formula is as follows:
wherein,
is the ith attention head, has the same K, V and Q and comes from the same space-time characteristic input,
is a scaling factor; operating a plurality of multi-head self-attention machines to learn the sequence characteristics of the graph respectively, finally connecting the output characteristics of the c multi-head self-attention machines together and outputting a group of characteristic vectors
F o Is the dimension of the group feature vector H, c is the number of subgraphs; each sub-graph corresponds to a feature vector H, calculated by the formula, where W attn Is a multi-head attention mechanism weight parameter:
the multi-head self-attention mechanism can learn a group of influence weights of the level of the bipartite graph of the subgraph and the subgraphs at other moments, and node features of the subgraphs at different moments are fused.
Further, in step S2, the model based on the graph sequence attention network includes a graph deformation module and a deformation coding module;
the graph deformation module comprises a local attention layer of a bipartite graph, a graph pooling layer and a sequence attention layer of the bipartite graph, and the deformation coding module comprises Add + Norm layer and Feed Forward layer; the Add + Norm layer and the Feed Forward layer are the same as the transform Encoder Block. The Add + Norm is subjected to batch normalization processing after a residual connecting network, so that the generalization capability of the model can be enhanced, and the neural network learning characterization capability is prevented from declining along with the layer number. After the graph deformation coding module completes the calculation, a group of graph sequence characteristics H are output gtb Feature H of sequence of images gtb Performing dynamic Pooling (Adaptive Pooling), which is a common operation in deep learning, and fusing features at different moments to output a feature H representing the sequence of the whole graph gtb To train a regression model.
Further, in step S3, the macro dynamic popularity prediction model based on the combination of the graph sequence attention network and the decoder comprises a graph sequence attention network module and a decoder module;
in the structure of a Decoder, a time sequence perception information is added to a graph sequence Attention network to enhance the time-dependent learning of information propagation by a model, the basic structure of a time sequence perception Attention Decoder TTA (Temporal Aware attachment) Decoder Block is close to a Decoder in a transform, a time sequence perception part is added on the basis, the time sequence characteristics of a current time sub-graph are obtained by representing and learning the information representing the position and the current time, and the learning characteristics are closely related to the time sequence through the self-Attention mechanism of the Decoder.
Further, step S4 includes the steps of:
s4.1, a Graph transformation module (Graph transform Block) receives Graph sequence data comprising a Graph adjacency matrix, a node representation characteristic and a time sequence representation characteristic as input;
s4.2, the local attention layer of the bipartite graph of the graph deformation module pools all node features in a propagation subgraph and outputs a feature vector corresponding to each subgraph
Learning node features including structural information of mutual influence in the forward and backward directions;
s4.3, running a plurality of multi-head self-attention mechanisms on the bidirectional graph sequence attention layer of the graph deformation module to respectively learn graph sequence characteristics, and for the characteristic vector of each sub-graph
The self-attention mechanism can learn a group of influence weights of bipartite graph levels of the subgraph and the subgraphs at other moments, finally, output features of the c multi-head self-attention mechanism are connected together, and a group of feature vectors are output
F o Is the dimension of the output feature of the multi-head self-attention mechanism, and the set feature vector H comprises the feature vector H corresponding to the cascade subgraphs of each moment i i Dynamically pooling the group feature vector H, and outputting a vector representing the features of the whole image series;
s4.4, the time sequence perception attention decoder takes the output vector of the image deformation module as V, K and Q, and the output of the time i-1 (Power embedding) at the time on the time sequence perception attention decoder
The Position Encoding (Position Encoding) and the time sequence representation feature are added, and the output of the time sequence perception attention decoder time i is
S4.5, in the model training stage, shifting (Shift) is carried out on the propagation popularity, the propagation popularity whole sequence is shifted to the left by one bit, the true value of the propagation popularity at the previous moment i-1 becomes the input of the time sequence perception attention decoder of the model at the moment i, and the TeacherForcing training is carried out.
Further, in step S5, the predicted value of the propagation popularity at the current time is output by using the predicted value of the propagation popularity at the previous time as an input of the time-series perceptual attention decoder of the model at the current time.
Compared with the prior art, the invention has the following advantages and beneficial effects:
the propagation popularity prediction method for bidirectional social influence learning provided by the invention aims at the problem that the existing propagation model can only model unidirectional social influence in propagation, bidirectional social influence information of sub-graphs at the previous and later moments is not considered, a graph sequence attention neural network with bidirectional propagation characteristic learning capability is provided, and a dynamic propagation network is modeled into a graph sequence learning problem. The model integrates a multidimensional attention mechanism and a multidimensional bidirectional graph attention mechanism, and in the attention of a bidirectional graph partial graph at a node level, node characteristics at each moment simultaneously pay attention to sub-graph information at front and rear moments; in a self-attention mechanism of a bipartite graph sequence at a graph level, long dependency of propagation sub-graph features at different moments is learned, the problem of common long dependency in information propagation time sequence features is solved, and mutual dependency features of a propagation network in a local structure and a global time sequence are learned respectively.
The propagation popularity prediction method for bidirectional social influence learning combines the graph sequence attention neural network and the self-encoder model, dynamically models the propagation popularity as a decoding process, deeply fuses the space-time propagation characteristics by applying a decoder with time sequence perception, and fuses the static propagation popularity and the dynamic propagation popularity into the same frame.
Drawings
FIG. 1 is a flow chart of a method for predicting popularity of bi-directional social impact learning propagation in an embodiment of the present invention;
FIG. 2 is a schematic partial attention diagram of a bipartite graph according to an embodiment of the invention;
FIG. 3 is a schematic attention diagram of a sequence of bipartite graphs in an embodiment of the invention;
FIG. 4 is a schematic diagram of a graph sequence attention network in an embodiment of the present invention;
FIG. 5 is a block diagram of a map morphing module in an embodiment of the present invention;
FIG. 6 is a block diagram of a timing aware decoder TAA according to an embodiment of the present invention;
fig. 7 is a diagram of a graph sequence attention network and decoder in combination with a predictive macro dynamic popularity framework in an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to examples and drawings, but the present invention is not limited thereto.
Example 1:
the method for predicting the propagation popularity of the bidirectional social influence learning, as shown in fig. 1, comprises the following steps:
s1, collecting information propagation network data, and constructing graph sequence data of information propagation, wherein the graph sequence data comprises a graph adjacent matrix, a node representation feature and time sequence representation feature data;
collecting information propagation network data, wherein information is propagated in a social network to form a propagation network, and a post m issued by a user u is forwarded by a user v at a time t to form propagation four-tuple data (v, u, t, m); at time i, the publisher and all the forwarders of a post m form a propagation network G i (ii) a Modeling a propagation network as a sequence of graphs, the propagation network G at all times i Construct a graph sequence data, denoted as { G } 0 ，G 1 ，...，G T ，G T+1 ，G T+2 Said, T =0,1,2, T, where T is the observation time instant, the node features and the subgraph structure of the corresponding subgraph snapshot at time instant i being denoted as observation time instant
And
ni is the number of nodes of the subgraph Gi of the time i, F is the initial characteristic dimension of the nodes, and the time sequence characteristic of the time i is recorded as
The time sequence characteristics, the node characteristics and the sub-graph structure at all the moments respectively form time sequence representation characteristics, node representation characteristics and graph adjacency matrix data of graph sequence data.
In this embodiment, the microblog transmission is collectedBroadcasting network data, and constructing microblog propagation popularity map sequence data; in this embodiment, there are 134279 pieces of microblog data collected, including documents [ Huang Z, wang Z, zhuY, et al.prediction of Cascade Structure and Outbranches Recurrence in Microblogs]And literature [ Cao Q, shen H, cen K, et al. Bridging the gap between prediction and interpretation of information cassettes]A collection of (1). The observation time is set to be 1 hour, the microblog information propagation popularity is predicted every 0.5 hour, and the model learning rate is set to be 1e -5 。
S2, constructing a model based on a graph sequence attention network, wherein the model comprises a graph deformation module and a deformation coding module based on bidirectional local graph attention at a node level and a bidirectional graph sequence attention mechanism at a graph level;
as shown in fig. 2, attention is specifically given as follows based on the bidirectional partial graph:
features of a node in a propagation subgraph
Not only has a relation with the current neighbor, but also is influenced by the subgraphs of the forward and backward moments, and the calculation formula is as follows:
wherein f is combine Is a method of combining the current node characteristics and the previous and subsequent time node characteristics,
representing the characteristics of the propagation subgraph of node v at time t.
The social influence of the neighbor node of the node on the node in the attention of the bidirectional local graph is specifically as follows:
initial feature x of node v v Is converted into conversion characteristics through a neural network layer
The calculation formula is as follows:
wherein, W l Are parameters of the neural network layer and,
f represents x v L layers of (1) initialize the feature dimension, F l Is a layer transformation characteristic
The LeaklyReLU is an activation function; in this example, F =32,f l ＝64。
Transformation characteristics of node v after transformation through neural network layer
And translation characteristics of neighbor node u of node v
And carrying out aggregation, wherein u belongs to N (v), and N (v) is a neighbor node set of the node v, and the method comprises the following steps:
wherein f is prop In order to calculate the attention value, in the present embodiment,
is the attention mechanism sharing neural network parameters, e uv The attention value of the node v for the neighbor node u of the node v,
to calculate the attention value of node v to node u.
The weight calculation mode of the neighbor node u of the node v acting on the node v is as follows:
wherein alpha is uv Is the attention coefficient of the neighbor node u of the node v to the node v after normalization, the node k is one of the neighbor node set of the node v, e kv The attention value of a neighbor node k of the node v to the node v;
compute node v aggregates updated transformation features
The calculation method is as follows:
where σ is the activation function, note
The characteristics of a node in the propagation subgraph are also affected by the subgraphs at the forward and backward moments, the subnetwork G at the previous moment t-1 t-1 Social impact of intermediate node u on node v
The calculation method is as follows:
wherein u ∈ V t-1 ，V t-1 F is a neighbor node set of a node v in a sub-network at the previous time t-1, a calculation method of node aggregation coefficients at the current time and the forward time is set as an Euclidean space of node characteristics,
n is the dimension of the feature, sub-network G at a later time t +1 t+1 Social impact of intermediate node u on node v
The calculation method is as follows:
wherein u ∈ V t+1 ，V t+1 Which is the set of neighbor nodes of node v in the subnetwork t +1 at a later time.
In this embodiment, n =64.
As shown in fig. 3, the attention mechanism based on the bipartite graph sequence is as follows:
generating a time sequence correlation co-production growth dependence relationship between the subgraph and a non-adjacent subgraph at a certain moment of non-neighbor subgraph, and learning the bidirectional long dependence of the subgraph on the time sequence by adopting a multi-head self-attention mechanism; the calculation formula of the space-time characteristic Q fusing the propagation structure and the time sequence characteristic is as follows:
wherein,
is the characteristic of the propagation subgraph snapshot node characteristic at the moment t after average pooling, tp t For time-series representation of the time t, X t For the sub-graph node feature at time t, sumPool is the sum pooling operation, W t Is a Linear neural network parameter;
the space-time characteristics Q are learned through a multi-head self-attention mechanism, so that the graph sequence attention network model can deeply fuse propagation structures and time sequence characteristics, and the formula is as follows:
wherein,
is the ith attention head, has the same K, V and Q and comes from the same space-time characteristic input,
is a scaling factor; operating a plurality of multi-head self-attention machines to learn the sequence characteristics of the graph respectively, finally connecting the output characteristics of the c multi-head self-attention machines together and outputting a group of characteristic vectors
F o Is the dimension of the group feature vector H, c is the number of subgraphs; each sub-graph corresponds to a feature vector H, calculated by the formula, where W attn Is a multi-head attention mechanism weight parameter:
the multi-head self-attention mechanism can learn a group of influence weights of the bipartite graph level of the subgraph and subgraphs at other moments, and node characteristics of the subgraphs at different moments are fused.
As shown in fig. 4, the model based on the graph sequence attention network includes a graph deformation module and a deformation coding module; as shown in fig. 5, the graph deformation module includes a local attention layer of a bidirectional graph, a graph pooling layer, and a sequence attention layer of the bidirectional graph, and the deformation encoding module includes an Add + Norm layer and a Feed Forward layer; the Add + Norm layer and the Feed Forward layer are the same as the transform Encoder Block. Add + Norm adds a batch normalization after a residual connecting network, can strengthen the generalization ability of the model, prevent godThe characterization ability via web learning declines with the number of layers. After the graph deformation coding module completes the calculation, a group of graph sequence characteristics H are output gtb Feature H of sequence of images gtb Performing dynamic Pooling (Adaptive Pooling), which is a common operation in deep learning, and fusing features at different moments to output a feature H representing the sequence of the whole graph gtb To train a regression model.
S3, constructing a macroscopic dynamic popularity prediction model based on the combination of the graph sequence attention network and a decoder, wherein the macroscopic dynamic popularity prediction model comprises the graph sequence attention network and a time sequence perception attention decoder;
as shown in fig. 7, the macro dynamic popularity prediction model based on the combination of the graph sequence attention network and the decoder comprises a graph sequence attention network module and a decoder module;
as shown in fig. 6, in the Decoder structure, the time-series perception information is added to the graph sequence Attention network to enhance the model to learn the information propagation timeliness, the basic structure of the time-series perception Attention Decoder TTA (Temporal Aware attachment) Decoder Block is close to the Decoder in the transform, a time-series perception part is added on the basis, the presentation position information and the current time information are learned to obtain the time-series characteristics of the current time subgraph, and the learning characteristics and the time series are closely related through the self-Attention mechanism of the Decoder.
S4, inputting the graph sequence data of 1 hour observation time into a macroscopic dynamic popularity prediction model based on the combination of a graph sequence attention network and a decoder for iterative training, and comprising the following steps of:
s4.1, a Graph deformation module (Graph Transformer Block) receives Graph sequence data comprising a Graph adjacency matrix, a node representation feature and a time sequence representation feature as input;
s4.2, the local attention layer of the bipartite graph of the graph deformation module pools all node features in a propagation subgraph and outputs a feature vector corresponding to each subgraph
Learning node features including structural information of mutual influence in the forward and backward directions;
s4.3, running a plurality of multi-head self-attention mechanisms on the bidirectional graph sequence attention layer of the graph deformation module to respectively learn graph sequence characteristics, and for the characteristic vector of each sub-graph
The self-attention mechanism can learn a group of influence weights of a bipartite graph level about the subgraph and the subgraph at other moments, finally, output features of the c multi-head self-attention mechanism are connected together, and a group feature vector is output
F o Is the dimension of the output feature of the multi-head self-attention mechanism, and the group feature vector H comprises a feature vector H corresponding to the cascade subgraph of each moment i i Dynamically pooling the group characteristic vector H, and outputting a vector representing the characteristics of the whole image series;
s4.4, the time sequence perception attention decoder takes the output vector of the graph deformation module as V, K and Q, and the time sequence perception attention decoder outputs the output at the moment i-1 (Power embedding)
The Position Encoding (Position Encoding) and the time sequence representation characteristics are added, and the time sequence perception attention decoder time i is output as
S4.5, in the model training stage, shifting (Shift) is carried out on the propagation popularity, the whole sequence of the propagation popularity is moved to the left by one bit, the real value of the propagation popularity at the last moment i-1 is used as the input of the time sequence perception attention decoder of the model at the moment i with the probability of 0.5, and the Teacher Forcing training is carried out.
And S5, inputting the graph sequence data of microblog propagation, and predicting the popularity of the microblog propagation at different moments by using a macroscopic dynamic popularity prediction model obtained by training and based on the combination of a graph sequence attention network and a decoder.
And adopting the predicted value of the propagation popularity at the previous moment as the input of a time sequence perception attention decoder of the model at the current moment, and outputting the predicted value of the propagation popularity at the current moment.
Example 2:
the difference between the embodiment and the embodiment 1 is that a method for predicting popularity of an academic citation network based on bidirectional social influence learning is provided;
step S1, collecting academic quotation network data and constructing academic quotation network spreading popularity graph sequence data; in this embodiment, 68433 pieces of academic citation network data are collected. The observation time is set to 5 years, the propagation popularity of the academic quotation network is predicted every 180 days, and the model learning rate is set to 1e -5 。
In the step S2, the process is carried out,
in step S4, the graph sequence data of 5-year observation time is input into a macroscopic dynamic popularity prediction model based on the combination of a graph sequence attention network and a decoder for iterative training
In the step S5, graph sequence data of the academic quotation is input, and popularity of the academic quotation at different moments is predicted by using a macro dynamic popularity prediction model obtained by training and based on the combination of a graph sequence attention network and a decoder, so that the academic quotation propagation popularity prediction values at different moments are obtained.
Example 3:
the embodiment is different from the embodiment 1 in that a twitter propagation popularity prediction method based on bidirectional social influence learning is provided;
step S1, collecting twitter data and constructing twitter propagation popularity observation graph sequence data; in this embodiment, a total of 20000 pieces of twitter data are collected. The observation time is set to 1 hour, the Tutt propagation popularity is predicted every 0.5 hour, and the model learning rate is set to 1e -5 。
In the step S2, the process is carried out,
in step S4, the graph sequence data of 1 hour observation time is input into a macroscopic dynamic popularity prediction model based on the combination of a graph sequence attention network and a decoder for iterative training.
In step S5, graph sequence data of the twitter propagation is input, and the popularity of the twitter propagation at different moments is predicted by using a macro dynamic popularity prediction model obtained by training and based on the combination of a graph sequence attention network and a decoder, so that the twitter propagation popularity prediction values at different moments are obtained.
The above description is only for the preferred embodiments of the present invention, but the protection scope of the present invention is not limited thereto, and any person skilled in the art can substitute or change the technical solution of the present invention and the inventive concept within the scope of the present invention, which is disclosed by the present invention, and the equivalent or change thereof belongs to the protection scope of the present invention.
Claims (10)
1. A propagation popularity prediction method for bidirectional social influence learning is characterized by comprising the following steps:
s1, collecting information propagation network data, and constructing graph sequence data of information propagation, wherein the graph sequence data comprises a graph adjacency matrix, a node representation feature and time sequence representation feature data;
s2, constructing a model based on a graph sequence attention network, wherein the model comprises a bidirectional local graph attention mechanism based on a node level and a graph deformation coding and decoding module based on a bidirectional graph sequence attention mechanism based on a graph level;
s3, constructing a macroscopic dynamic popularity prediction model based on the combination of the graph sequence attention network and a decoder, wherein the macroscopic dynamic popularity prediction model comprises the graph sequence attention network and a time sequence perception attention decoder;
s4, inputting the graph sequence data into a macroscopic dynamic popularity prediction model based on the combination of a graph sequence attention network and a decoder for iterative training;
and S5, inputting graph sequence data of information propagation, and predicting the popularity of the information propagation at different moments by using a model obtained by training.
2. The method for predicting the propagation popularity of the bidirectional social influence learning according to claim 1, wherein in step S1, information propagation network data is collected, information is propagated in a social network to form a propagation network, and a post m issued by a user u is forwarded by a user v at a time t to form a propagation quadruple of data (v, u, t, m);
at time i, the publisher and all forwarders of a post m form a propagation subgraph G i (ii) a Modeling a propagation network as a sequence of graphs, the propagation network G at all times i Construct a graph sequence data, denoted as { G } 0 ，G 1 ，...，G T ，G T+1 ，G T+2 A construction, where T is the observation time, and the node features and the subgraph structure of the corresponding subgraph snapshot at time i are denoted X, respectively i And A i ，
N i Is subgraph G of time i i F is the initial characteristic dimension of the node, and the time sequence characteristic of the time i is marked as tp i ，
The time sequence characteristics, the node characteristics and the sub-graph structure at all the time respectively form time sequence representation characteristics, node representation characteristics and graph adjacency matrix data of graph sequence data.
3. The method for predicting the popularity of propagation of bidirectional social influence learning according to claim 1, wherein in step S2, the attention based on the bidirectional local graph is specifically as follows:
features of a node in a propagation subgraph
Not only has a relation with the current neighbor, but also is influenced by the subgraphs of the forward and backward moments, and the calculation formula is as follows:
wherein f is combine Is a method of combining the current node characteristics and the previous and subsequent time node characteristics,
representing the characteristics of the propagation subgraph of node v at time t.
4. The propagation popularity prediction method for bidirectional social influence learning according to claim 3, wherein social influences of neighbor nodes of a node in the attention of the bidirectional local graph on the node are specifically as follows:
initial feature x of node v v Is converted into conversion characteristics through a neural network layer
The calculation formula is as follows:
wherein, W l Are parameters of the neural network layer and,
f represents x v L layers of (1) initialize the feature dimension, F l Is a characteristic of layer conversion
Is the activation function;
transformation characteristics of node v after transformation through neural network layer
And translation characteristics of neighbor node u of node v
And carrying out aggregation, wherein u belongs to N (v), and N (v) is a neighbor node set of the node v, and the method comprises the following steps:
wherein f is prop For calculating attention values, e.g.
Is the attention mechanism sharing neural network parameters, e uv The attention value of the node v for the neighbor node u of the node v,
calculating the attention value of the node v to the node u;
the weight calculation mode of the neighbor node u of the node v acting on the node v is as follows:
wherein alpha is uv Is the attention coefficient of the neighbor node u of the node v to the node v after normalization, the node k is one of the neighbor node set of the node v, e kv The attention value of a neighbor node k of the node v to the node v;
compute node v aggregates updated transformation features
The calculation method is as follows:
where σ is an activation function
5. The method of claim 4, wherein the characteristics of a node in the propagation subgraph are also influenced by the subgraphs at forward and backward time, and the sub-network G at the previous time t-1 is the sub-network G t-1 Social impact of intermediate node u on node v
The calculation method is as follows:
wherein u ∈ V t-1 ，V t-1 F is a neighbor node set of a node v in a sub-network at the previous time t-1, a calculation method of node aggregation coefficients at the current time and the forward time is set as an Euclidean space of node characteristics,
n is the dimension of the feature, sub-network G of the later time t +1 t+1 Middle jointSocial impact of point u on point v
The calculation method is as follows:
wherein u ∈ V t+1 ，V t+1 A neighbor node set of a node v in the sub-network at the later moment t + 1; considering the forward and backward characteristics of a node, the updated node characteristics are as follows:
6. the method for predicting the propagation popularity of the bidirectional social influence learning as claimed in claim 1, wherein in the step S2, the attention mechanism based on the digraph sequence is specifically as follows:
the subgraph and a non-adjacent subgraph at a certain time of a non-neighbor generate a long dependency relationship, a multi-head self-attention mechanism is adopted by the model, and bidirectional long dependency of the subgraph on a time sequence is learned; spatio-temporal features Q incorporating propagation structure and timing features t The calculation formula of (c) is:
wherein,
is the feature of the sub-graph node feature at the time t after average pooling, tp t Is a time-sequential representation of time tCharacteristic of, X t For the sub-graph node feature at time t, sumPool is the sum pooling operation, W t Is a Linear neural network parameter;
the space-time characteristics are learned through a multi-head self-attention mechanism, so that the graph sequence attention network model can deeply fuse the propagation structure and the time sequence characteristics, and the multi-head self-attention mechanism formula is as follows:
wherein,
is the ith attention head, has the same K, V and Q and comes from the same space-time characteristic input,
is a scaling factor; operating a plurality of multi-head self-attention machines to learn the sequence characteristics of the graph respectively, finally connecting the output characteristics of the c multi-head self-attention machines together and outputting a group of characteristic vectors
F o Is the dimension of the group feature vector H, c is the number of subgraphs; each sub-graph corresponds to a feature vector H, calculated by the formula, where W attn Is a multi-head attention mechanism weight parameter:
the multi-head self-attention mechanism can learn a group of influence weights of the bipartite graph level of the subgraph and subgraphs at other moments, and the characteristics of the subgraphs at different moments are fused.
7. The method for predicting the popularity of propagation of bidirectional social influence learning according to claim 1, wherein in step S2, the model based on the graph sequence attention network includes a graph deformation encoding module and a decoding module;
the graph deformation module comprises a bidirectional graph local attention layer, a graph pooling layer and a bidirectional graph sequence attention layer, and the deformation coding module comprises an Add + Norm layer and a Feed Forward layer; the Add + Norm layer and the Feed Forward layer are the same as a transform Encoder Block; the Add + Norm layer is subjected to batch normalization processing after a residual connecting network, so that the generalization capability of the model can be enhanced, and the neural network learning characterization capability is prevented from declining along with the number of layers;
after the computation of the graph deformation coding module is completed, a group of graph sequence characteristics H are output gtb Features H of map sequence gtb Performing dynamic Pooling (Adaptive Pooling), which is a common operation in deep learning, and fusing features at different moments to output a feature H representing the sequence of the whole graph gtb Based on which a regression model is trained.
8. The method for predicting the popularity of propagation of bi-directional social influence learning according to claim 1, wherein in step S3, the macro dynamic popularity prediction model based on the combination of the graph sequence attention network and the decoder comprises a graph sequence attention network module and a decoder module;
in the structure of a Decoder, a time sequence perception information is added to a graph sequence Attention network to enhance the time-dependent learning of information propagation by a model, the basic structure of a time sequence perception Attention Decoder TTA (Temporal Aware attachment) Decoder Block is close to a Decoder in a Transformer, a time sequence perception part is added on the basis, the time sequence characteristics of a current time subgraph are obtained by representing and learning the position information and the current time information, and the learning characteristics are closely associated with the time sequence by learning through a self-Attention mechanism of the Decoder.
9. The method for predicting the popularity of propagation of bidirectional social influence learning according to claim 1, wherein the step S4 comprises the steps of:
s4.1, a Graph transformation module (Graph transform Block) receives Graph sequence data comprising a Graph adjacency matrix, a node representation characteristic and a time sequence representation characteristic as input;
s4.2, the local attention layer of the bipartite graph of the graph deformation module pools all node features in a propagation subgraph and outputs a feature vector corresponding to each subgraph
Learning node characteristics including structural information of mutual influence in the forward and backward directions;
s4.3, running a plurality of multi-head self-attention mechanisms on the bidirectional graph sequence attention layer of the graph deformation module to learn graph sequence characteristics respectively, and for the characteristic vector of each sub-graph
The self-attention mechanism can learn a group of influence weights of bipartite graph levels of the subgraph and the subgraphs at other moments, finally, output features of the c multi-head self-attention mechanism are connected together, and a group of feature vectors are output
Fo is the dimension of the output feature of the multi-head self-attention mechanism, and the set feature vector H comprises the feature vector H corresponding to the cascade subgraph at each time i i Dynamically pooling the group feature vector H, and outputting a vector representing the features of the whole image series;
s4.4, the time sequence perception attention decoder takes the output vector of the image deformation module as V, K and Q, and the output of the time i-1 (Power embedding) at the time on the time sequence perception attention decoder
The Position Encoding (Position Encoding) and the time sequence representation feature are added, and the output of the time sequence perception attention decoder time i is
S4.5, in the model training stage, shifting (Shift) is carried out on the propagation popularity, the propagation popularity whole sequence is shifted to the left by one bit, the true value of the propagation popularity at the previous moment i-1 becomes the input of the time sequence perception attention decoder of the model at the moment i, and the TeacherForcing training is carried out.
10. The method as claimed in claim 1, wherein the predicted value of the propagation popularity at the current time is output in step S5 by using the predicted value of the propagation popularity at the previous time as an input to a time-series perceptual attention decoder of the model at the current time.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211196722.XA CN115618124A (en) | 2022-09-28 | 2022-09-28 | Propagation popularity prediction method for bidirectional social influence learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211196722.XA CN115618124A (en) | 2022-09-28 | 2022-09-28 | Propagation popularity prediction method for bidirectional social influence learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115618124A true CN115618124A (en) | 2023-01-17 |
Family
ID=84859991
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211196722.XA Pending CN115618124A (en) | 2022-09-28 | 2022-09-28 | Propagation popularity prediction method for bidirectional social influence learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115618124A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117610717A (en) * | 2023-11-13 | 2024-02-27 | 重庆大学 | Information popularity prediction method based on double-variation cascade self-encoder |
-
2022
- 2022-09-28 CN CN202211196722.XA patent/CN115618124A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117610717A (en) * | 2023-11-13 | 2024-02-27 | 重庆大学 | Information popularity prediction method based on double-variation cascade self-encoder |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Chen et al. | TrajVAE: A Variational AutoEncoder model for trajectory generation | |
| CN111310063B (en) | Neural network-based article recommendation method for memory perception gated factorization machine | |
| CN112488055B (en) | Video question-answering method based on progressive graph attention network | |
| CN109117943B (en) | Method for enhancing network representation learning by utilizing multi-attribute information | |
| Sheng et al. | Enhanced graph neural network for session-based recommendation | |
| CN114971784B (en) | Session recommendation method and system based on graph neural network by fusing self-attention mechanism | |
| CN113590976A (en) | Recommendation method of space self-adaptive graph convolution network | |
| CN115470406A (en) | Graph neural network session recommendation method based on dual-channel information fusion | |
| CN113704438A (en) | Conversation recommendation method of abnormal picture based on layered attention mechanism | |
| CN115618124A (en) | Propagation popularity prediction method for bidirectional social influence learning | |
| CN114637911A (en) | Next interest point recommendation method of attention fusion perception network | |
| CN115017405A (en) | Graph neural network travel package recommendation method based on multitask self-coding | |
| CN114925270A (en) | Session recommendation method and model | |
| CN117633371B (en) | Recommendation method, device and readable storage medium based on multi-attention mechanism | |
| CN114676332A (en) | Network API recommendation method facing developers | |
| CN117994011A (en) | E-commerce dynamic perception data recommendation method based on memory updating and neighbor transfer | |
| CN114820139A (en) | Multi-user recommendation system based on knowledge graph path reasoning | |
| CN114528490A (en) | Self-supervision sequence recommendation method based on long-term and short-term interests of user | |
| CN112559904B (en) | Conversational social recommendation method based on door mechanism and multi-modal graph network | |
| Zheng et al. | Tourism demand forecasting using complex network theory | |
| CN117390267A (en) | Knowledge graph-based personalized multitask enhanced recommendation model | |
| CN117237140A (en) | Social network influence maximization method fusing graph convolution neural network and transducer | |
| CN118396685B (en) | Advertisement marketing recommendation method based on deep reinforcement learning | |
| CN115660147B (en) | Information propagation prediction method and system based on influence modeling between propagation paths and in propagation paths | |
| Shou et al. | Research on Knowledge Concept Recommendation Algorithm with Spatial-Temporal Information |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |