CN115878907A - Social network forwarding behavior prediction method and device based on user dependency relationship - Google Patents
Social network forwarding behavior prediction method and device based on user dependency relationship Download PDFInfo
- Publication number
- CN115878907A CN115878907A CN202211697432.3A CN202211697432A CN115878907A CN 115878907 A CN115878907 A CN 115878907A CN 202211697432 A CN202211697432 A CN 202211697432A CN 115878907 A CN115878907 A CN 115878907A
- Authority
- CN
- China
- Prior art keywords
- user
- vector
- forwarding
- text
- text content
- 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.)
- Granted
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Machine Translation (AREA)
Abstract
The embodiment of the invention discloses a social network forwarding behavior prediction method and a device based on user dependency relationship, wherein the method comprises the following steps: determining a user group according to the social relationship of the users to obtain a first embedded vector of each user in the user group; calculating to obtain a second embedding vector of the user containing the context dependent vector according to the forwarding process and the first embedding vector of the user; performing cross-layer connection and normalization processing according to the second embedding vector of the user and a preset direction mask and based on a self-attention mechanism and according to the sequence of the text content forwarded by the user to obtain the forwarding user sequence embedding in the forwarding process; determining a forwarding text embedding vector of the text content according to the text content; according to the forward text embedding vector of the text content and the forward user sequence embedding, calculating to obtain an attention result of the forward user sequence embedding on the text content, and obtaining the prediction probability of the user forward text content according to the attention result.
Description
Technical Field
The embodiment of the invention relates to the technical field of social network application, in particular to a social network forwarding behavior prediction method and device based on user dependency.
Background
With the rapid development and popularization of the internet and web2.0 technologies, a large number of online social networks have appeared on the internet, and they have become the main platforms for people to share, spread and acquire information. The media can release content on the platform, and a large number of users accelerate the content transmission process through actions such as approval and forwarding.
The research on content diffusion in the prior art mainly aims at understanding the propagation mechanism of the content, and has important significance for commercial promotion, social hotspot tracking, sensitive content monitoring and the like. By analyzing the characteristics of content propagation, the propagation range of the content can be adjusted, the propagation user group is influenced, or the rapid propagation of the content is avoided.
At present, most predictions about content forwarding behaviors rely on an RNN (Recurrent Neural Network) model to extract observed features of a forwarding process, but the RNN model considers that next data of a sequence is influenced by a direct predecessor node, and the longer the distance is, the weaker the influence relationship is, and the processing mode is contrary to the propagation process of content. In addition, most existing models consider that the content propagation is only performed by depending on the influence relationship among users, and influence on the content propagation, such as the characteristics of the content, the friend relationship characteristics among the users, the attribute characteristics of the users, and the like, is ignored.
Disclosure of Invention
In view of the above, embodiments of the present invention are proposed to provide a social network forwarding behavior prediction method and apparatus based on user dependency relationships, which overcome or at least partially solve the above problems.
According to an aspect of the embodiments of the present invention, a social network forwarding behavior prediction method based on user dependency relationships is provided, which includes:
determining a user group according to the social relationship of the users to obtain a first embedded vector of each user in the user group;
calculating to obtain a second embedded vector of the user containing the context dependent vector according to the forwarding process and the first embedded vector of the user; performing cross-layer connection and normalization processing according to the second embedding vector of the user and a preset direction mask and based on a self-attention mechanism and according to the sequence of the text content forwarded by the user to obtain the forwarding user sequence embedding in the forwarding process;
determining a forwarding text embedding vector of the text content according to the text content;
and calculating an attention result of the forward user sequence embedding to the text content according to the forward text embedding vector and the forward user sequence embedding of the text content, and obtaining the prediction probability of the user forward text content according to the attention result.
Optionally, determining a user group according to the social relationship of the users, and obtaining the first embedded vector of each user in the user group further includes:
acquiring a user social relationship, determining a user group according to the user social relationship, and constructing a user social relationship network; the user social relationship network is constructed in a graph data structure and comprises a plurality of user nodes and neighbor nodes connected with the user nodes;
and aiming at the user group, inputting the vector of each user node into the graph neural network model to obtain a first embedded vector of each user in the user group.
Optionally, for the user group, inputting the vector of each user node into the graph neural network model, and obtaining the first embedded vector of each user in the user group further includes:
aiming at any user node in a user group, inputting a vector of the user node and a vector of a neighbor node of the user node into a graph neural network model, calculating the influence of each neighbor node on the user node, and splicing a plurality of obtained attention results of the user node according to the influence based on a multi-head attention mechanism; outputting an average value of the plurality of attention results as an embedding vector of the user node;
and splicing the preset random vector and the embedded vector of the user node to obtain a first embedded vector of the user.
Optionally, the calculating a second embedding vector of the user including the context-dependent vector according to the forwarding process and the first embedding vector of the user further includes:
calculating the influence of a previous forwarding user on a subsequent forwarding user according to the forwarding sequence of the forwarding process, and obtaining a context dependent vector of the subsequent forwarding user according to the weighted calculation of the influence;
and fusing the first embedding vector of the user and the context dependent vector of the user to obtain a second embedding vector of the user containing the context dependent vector.
Optionally, determining a forwarded text embedding vector of the text content further comprises, according to the text content:
determining a release time embedding vector, a text background, a content embedding vector and a text age embedding vector of the text content according to the text content;
and splicing the release time embedded vector, the text background and the content embedded vector of the text content and the text age embedded vector to obtain a forwarding text embedded vector of the text content.
Optionally, determining, according to the text content, a release time embedding vector, a text background, and a content embedding vector of the text content, and the text age embedding vector further includes:
determining the time characteristics of the release time of the text content according to the release time of the text content and a plurality of preset dimensions; the plurality of preset dimensions include daily, weekly, and/or holiday dimensions; the time characteristics are expressed by adopting one-hot coding;
determining time embedding vectors of a plurality of preset dimensions according to the time characteristics of the release time of the text contents of the plurality of preset dimensions;
splicing the time embedded vectors of a plurality of preset dimensions to obtain a release time embedded vector of the text content;
according to the text content, obtaining a corresponding text content embedding vector, and constructing a corresponding hot content embedding vector according to the hot content;
based on an attention mechanism, calculating the similarity between the hot content and the text content according to the text content embedding vector and the hot content embedding vector, and calculating to obtain a background feature vector of the text content according to the similarity;
calculating the weights of the background feature vector and the text content embedded vector, and calculating to obtain a text background and a text content embedded vector according to the weights; the weight is obtained based on a gating mechanism;
and determining the age characteristics of the text content based on the time attention mechanism, and obtaining a text age embedding vector according to the age characteristics.
Optionally, the obtaining an attention result of the text content by embedding the forwarding user sequence according to the forwarding text embedding vector of the text content and the embedding of the forwarding user sequence by calculation, and obtaining the prediction probability of the text content forwarded by the user according to the attention result further includes:
calculating the influence of the forward text embedding vector of the text content on each user for embedding the forward user sequence, and obtaining the attention result of the forward user sequence embedding on the text content according to the influence and the forward user sequence embedding;
and according to the attention result, obtaining the prediction probability of the text content forwarded to the next user by the user in the forwarding sequence.
Optionally, obtaining the predicted probability that the user forwards the text content to the next user in the forwarding order according to the attention result further includes:
and obtaining the prediction probability of the text content forwarded to the next user by the user in the forwarding sequence based on the attention result of the multi-layer perception computer.
Optionally, the method further comprises:
and predicting to obtain the target forwarding probability of the user group, and optimizing the prediction model according to the target forwarding probability and the sample forwarding probability in the training set.
According to another aspect of the embodiments of the present invention, there is provided a social network forwarding behavior prediction apparatus based on user dependency relationships, the apparatus including:
the first user vector module is suitable for determining a user group according to the social relationship of the users to obtain a first embedded vector of each user in the user group;
the second user vector module is suitable for calculating a second embedded vector of the user containing the context dependent vector according to the forwarding process and the first embedded vector of the user; performing cross-layer connection and normalization processing according to the second embedding vector of the user and a preset direction mask and based on a self-attention mechanism and according to the sequence of the text content forwarded by the user to obtain the forwarding user sequence embedding in the forwarding process;
the text vector module is suitable for determining a forwarding text embedded vector of the text content according to the text content;
and the prediction module is suitable for calculating an attention result of the forward user sequence embedding on the text content according to the forward text embedding vector and the forward user sequence embedding of the text content, and obtaining the prediction probability of the user forward text content according to the attention result.
According to still another aspect of an embodiment of the present invention, there is provided a computing device including: the system comprises a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface complete mutual communication through the communication bus;
the memory is used for storing at least one executable instruction, and the executable instruction enables the processor to execute the operation corresponding to the social network forwarding behavior prediction method based on the user dependency relationship.
According to a further aspect of the embodiments of the present invention, a computer storage medium is provided, where at least one executable instruction is stored in the storage medium, and the executable instruction causes a processor to perform an operation corresponding to the social network forwarding behavior prediction method based on user dependency relationship as described above.
According to the social network forwarding behavior prediction method and device based on the user dependency relationship, provided by the embodiment of the invention, the relevance of the behavior of forwarding text content in the user social network and various characteristics is analyzed, a prediction model for text content propagation based on the user social network is established, the propagation range of the text content can be predicted, the propagation of the text content is conveniently controlled, a platform is helped to recommend more appropriate text content, more accurate popularization is carried out, reasonable control over the content is realized, and the development of a text platform is promoted. Furthermore, the invention uses a single context feature extraction mode to extract the influence of the past forwarding user on the current user node, so as to obtain the context dependent vector. And the influence among the users is calculated by using a self-attention mechanism and a preset direction mask, so that the long-term dependence problem is relieved. The invention uses the Transformer as a sequence processing tool, is different from the existing deep learning model, the Transformer module completely abandons the structure of RNN and other sequence models, adopts an attention mechanism to construct the model, and any two nodes in the sequence can be directly associated, thereby avoiding the loss of characteristics. The invention greatly relieves the long-term dependence problem in the RNN model and avoids the problem that influence relation cannot be formed between non-adjacent nodes. Meanwhile, the invention determines a corresponding prediction model aiming at the characteristics of the forwarding process through the result of analyzing the forwarding behavior characteristics, thereby completing the prediction of the user forwarding behavior.
The foregoing description is only an overview of the technical solutions of the embodiments of the present invention, and the embodiments of the present invention can be implemented according to the content of the description in order to make the technical means of the embodiments of the present invention more clearly understood, and the detailed description of the embodiments of the present invention is provided below in order to make the foregoing and other objects, features, and advantages of the embodiments of the present invention more clearly understandable.
Drawings
Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the embodiments of the invention. Also, like reference numerals are used to refer to like parts throughout the drawings. In the drawings:
FIG. 1 illustrates a flow diagram of a social network forwarding behavior prediction method based on user dependencies, according to one embodiment of the invention;
FIG. 2 shows an overall architectural diagram of a predictive model;
FIG. 3 is a schematic structural diagram of a social network forwarding behavior prediction apparatus based on user dependency relationship according to an embodiment of the present invention;
FIG. 4 shows a block diagram of a computing device, according to one embodiment of the invention.
Detailed Description
Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.
Fig. 1 shows a flowchart of a social network forwarding behavior prediction method based on user dependency relationships according to an embodiment of the present invention, and as shown in fig. 1, the method includes the following steps:
step S101, determining a user group according to the social relationship of the users, and obtaining a first embedded vector of each user in the user group.
In the embodiment, a prediction model is adopted to predict the probability of forwarding the text content in the user group by the user based on the inter-user dependency relationship. The user dependency relationship comprises a user social relationship, a friend making condition of the user and the like, and the influence of the user social relationship on the forwarding behavior is analyzed based on the user social relationship. Specifically, the user social relationship may use a graph data structure manner of the friend relationship network in fig. 2 to construct a user social relationship network for each user included in the user social relationship, that is, a user group. In the graph data structure, each node represents a single user, and the connected edges represent the relationship characteristics between the users. Namely, a social relationship network of the user comprises a plurality of user nodes and is determined to be connected with the user nodes according to the connecting edgesOf the neighboring node. For the user relationship, the relationship features of the user in the friend network can be extracted through a GAT (Graph Attention Networks, graph neural network) model, and a first embedding vector of the user is obtained. The GAT is part of the predictive model in this embodiment, which is responsible for extracting the user's embedded vectors based on the user's social relationships. When the GAT is used for obtaining the first embedded vector of the user, sample data of a user social relationship network needs to be collected in advance, the GAT is trained according to the sample data, and finally the embedded vector of each user in a user group can be obtained according to the trained GAT. The sample data comprises a user social relationship network such as G (U, E) in a preset range and a content set S, wherein U is each user in the user social relationship network, E is a friend relationship (namely a dependency relationship) among different users, a single content S belongs to S, and a user U belongs to U at a time t 0 Send out, at a predetermined time range, e.g. Δ T obs Within a time range, the content s propagates through various online or offline approaches in the user social relationship network G, and the propagation process can be denoted as a time sequence process C = [ (u) 1 ,t 1 ),(u 2 ,t 2 ),…(u i ,t i )…(u j ,t j ),...(u n ,t n )]Wherein, when i < j, t i ＜t j And t is 1 ≥t 0 ，t n ＜t 0 +ΔT obs The method includes the steps that contents propagated in the user social relationship network by a plurality of users in a preset time range are obtained.
The main structure of GAT consists of a graph attention tier whose input is a set of node expression vectors h = [ h ] 1 ,h 2 ,h 3 ,...,h n ]I.e. a vector of individual user nodes, the output is also a set of node expression vectors h '= [ h' 1 ,h′ 2 , h′ 3 ,...,h′ n ]. Specifically, in this implementation, the input is a vector of the user node (such as various characteristics of the user node itself), and the input is an embedded vector of each user node in the user group. Aiming at the expression vector of each node in a certain user social relation network, the attention layer obtains the expression vectors of the neighbor nodes through the structural characteristics of the graph and counts the expression vectorsAnd calculating the attention coefficient. Influence alpha of neighbor node j of node i on node i ij The calculation formula is as follows:
wherein, W GAT In order to be a parameter after the training,
set of respective neighbor nodes, α, for node i ij The influence of the neighbor node j on the node i is strong or weak. In practical applications, F denotes a feed-forward neural network. LeakyReLU is the activation function.
For the middle layer of the GAT model, in order to make the result of the self-attention mechanism more stable, a multi-head attention mechanism can be adopted, that is, multiple sets of weights are adopted to learn the coefficients of the attention mechanism together, so that the condition that the fluctuation of a single coefficient is larger due to a data set or model initialization is avoided. Specifically, the following method can be adopted:
h i ＝[h ik ]
where σ () is the activation function, W k Is the coefficient of the kth attention mechanism, h ik For the attention result corresponding to the kth attention mechanism, [ 2 ]]For the splicing operation, [ h ] ik ]The attention results from multiple heads, e.g., k attention results, are stitched as the current result.
For the last layer of the GAT model, since it is directly used for output, the results of multiple attention mechanisms can be processed in an average manner to obtain a user embedded vector based on the user social relationship network, where the results of connecting the attention mechanisms are not used only in the last layer.
Further, in consideration of incompleteness of the social relationship network of the user, errors may be introduced in the graph embedding process, and the obtained user embedding vector is generally associated with the userIn order to eliminate the error, the present embodiment models the attribute that is not observed by the user by using another set of vectors, which represents the inherent attribute of the user that is irrelevant to the historical behavior, and is also not relevant to the social relationship network of the user. For each user U epsilon U, using a preset random vector v u ∈R 1×d The universality of the user embedded vector is increased. Wherein R is 1×d Presetting a random vector v for a real number space with 1 x d dimension u A situation setting may be implemented, which is not limited herein. Determining a first embedding vector of a user by a user embedding vector obtained by embedding the user social relationship network and a preset random vector together, and finally obtaining a first embedding vector e of each user in a user group u The first embedded vector of the user will be used for subsequent sequence processing etc., in particular, e u ＝[h u ,v u ]That is, the user presets a random vector v u User embedded vector h embedded into user social relationship network u Splicing to obtain a first embedded vector e of the user u And u represents the user.
Step S102, calculating to obtain a second embedded vector of the user containing the context dependent vector according to the forwarding process and the first embedded vector of the user; and performing cross-layer connection and normalization processing according to the second embedding vector of the user and a preset direction mask and based on a self-attention mechanism and according to the sequence of the text content forwarded by the user to obtain the forwarding user sequence embedding in the forwarding process.
The forwarding process is a series of processes in which text content is forwarded by a plurality of users, and includes, for example, forwarding the text content from user 1 to user 2, and forwarding the text content from user 2 to user 4 \8230, where the forwarding process involves a forwarding sequence among forwarding users, and according to the forwarding process, a mechanism of attention can be used to extract a dependency relationship among the users, and on the basis of retaining a first embedding vector of the user, a final embedding result of the user can be adjusted by changing a direction sequence of forwarding the text content.
Specifically, for the first embedded vector of each user, the influence of each previous forwarding user with a previous forwarding order on the current user (the subsequent forwarding user) is calculated by adopting an attention mechanism, and weighted summation is performed based on each influence to obtain the vector of the current forwarding process, wherein the vector represents the state of the content when the current user forwards, and integrates the information of each user in the historical forwarding process, so that the mutual influence among the users in the forwarding process is reflected. Specifically, the influence of the kth forwarding user on the jth forwarding user can be calculated by the following formula:
wherein the content of the first and second substances,
to predict the model parameters, adjustments may be made according to the training process,<>representing inner product operation, e k First embedded vector for kth user, e j For the first embedded vector of the jth user, k is less than j. e.g. of the type l The first embedded vector is the first user, and the value range of l is 1 to j-1. The attention mechanism adopted in the formula calculation has directionality, that is, only the users forwarded in the front have influence on the users forwarded in the back, and the other way, the attention mechanism does not have influence. Physically, only the previously forwarded user can have an effect on the later forwarded user, which is inconsistent with the way the natural language processing model processes because in text content, the later appearing text content may have a evidence of the earlier content. After calculation based on the attention mechanism, the attention weight may be normalized using a softmax function, or the like.
For the user node in the forwarding process, the context dependency vector is the first embedded vector of each user forwarded in advance and the weighted summation of the influence, as shown in the following formula:
d j a context dependent vector for the jth user, whichAnd calculating according to the first embedded vector of each user before j and the influence of each user on the jth user by weighted summation.
Because there is no dependency between each calculation result, the whole forwarding process can be calculated together in order to improve the calculation speed. Specifically, for a forwarding process with the length of l, a forwarding process matrix E epsilon R is constructed d×l Wherein d represents a user embedding dimension, each row in the matrix E represents a first embedding vector of a user at the position, and in order to ensure the directionality, a preset direction mask M E is introduced l×l When i is<j is, M i,j =0, otherwise M i,j = infinity, i.e. only the preceding forwarding user has an effect on the following forwarding user, and vice versa, the impact weight matrix of the attention mechanism is calculated as follows:
wherein A ∈ R l×l ，A ij Representing the influence of the ith user on the j users. In the calculation, if i>j, then M ij = infinity, i.e., after the softmax operation is performed, the entry becomes 0, thereby masking the influence of the post-forwarding user on the preceding forwarding user. Using matrix A, the computation of the context-dependent vector for each stage can be accelerated using matrix operations.
D＝AE T
Wherein, the ith row of the matrix D represents the context dependent vector of the ith user, the matrix A contains the influence among the users, A ij Representing the influence of the ith user on the j users. Each row in the matrix E represents a first embedded vector of a user at that position, and T is a transpose operation. The ith row in matrix D corresponds to D i 。
The features of the user include a first embedded vector e of the user i And its context dependency vector d i However, considering that the subsequent processing needs a single vector, the two can be fused by using methods such as connection, addition, multiplication, etc., in this embodiment, a gating mechanism is used to select the effective information in the twoAnd (4) merging the information, calculating a coefficient according to the values of the information and the coefficient, and merging the two vectors according to the coefficient in proportion to obtain a single vector of the user characteristics.
u i ＝ i ⊙ i +(1- i )⊙ i
Wherein the content of the first and second substances,
are all prediction model parameters, adjusted according to the training process, g i Characterize e i And d i In (e) i Degree of importance of 1- i Is shown as e i And d i In d i By the importance of g i Carrying out weighted average on the first embedding vector and the context dependent vector of the user, and fusing to obtain a second embedding vector u of the user containing the context dependent vector i . As an example, we can use Hadmard's function to represent the term-by-term multiplication of vectors.
The forwarding sequence embedding process mainly uses a Transformer module. Since the embodiment does not predict the forwarding sequence but predicts whether a single user forwards, only the Transformer encoder part is used here, and the attention mechanism in the encoder can effectively extract the association relationship between user nodes. In each layer of the transform, the input data contains for each user a second embedded vector u of users of the context dependent vector i Arranging according to the order of forwarding text content by each user to obtain
Wherein l represents the sequence length, d up Dimension set according to implementation. In order to better reflect the influence of the position characteristics of each user, a preset position code (a fixed position code can be adopted) is added to each position to add the position characteristics, and the final result is mainly completed through a self-attention mechanism, which is shown in the following formula:
wherein SA denotes a self-attention mechanism, f x Denotes the softmax function, W Q 、W K 、W V All are prediction model parameters, which can be adjusted according to the training process, and M is a preset direction mask. To SA (C) of the output o ) Performing cross-layer connection and normalization processing, wherein the following formula is shown:
e c ＝BatchNorm(SA(C o )+C o )
wherein BatchNorm is a normalization function. A cross-layer connection structure such as ResNet can be adopted, so that the problem of gradient disappearance in the training process is avoided. In order to achieve better processing effect, the above operations can be repeated for multiple times to obtain the final forwarding user sequence embedding e in the forwarding process c ＝{e c1 ,e c2 ,...,e cn }. Wherein e is c1 ,e c2 ,...,e cn I.e. the user sequence of n users arranged in the order in which the n users forward the text content. The forwarding user sequence embedding comprises a plurality of user dependency relationships such as user social relationships, user forwarding relationships and the like, and the user dependency relationships have great influence on social network forwarding of the user.
And step S103, determining a forwarding text embedding vector of the text content according to the text content.
The forwarding behavior characteristics of the text content can be summarized into various categories, such as publishing source, publishing time, text content and the like. The processing of various features specifically includes feature extraction of text content, feature extraction of release time, feature extraction of a heat decay pattern, and the like.
The release time characteristics of the text content are a limited set, and a network relation structure does not exist, so that the processing mode of the text content is completely different from the text characteristics. And combining daily work and rest time arrangement, selecting to model the release time according to periodic change of the work and rest time, for example, according to the current working mode, selecting a plurality of preset dimensions to determine the time characteristics in the text, for example, selecting to embed the time characteristics according to dimensions such as daily dimension, weekly dimension, holiday dimension and the like.
The time-of-day characteristic is a time range in which the content is delivered on the current day, the time-of-week characteristic is a time range in which the content is delivered on the current week, and the holiday time characteristic is whether the content is a holiday or not. Determining the distribution time of the content based on the above time characteristics uses three one-hot unique encodings, e.g.
To represent e pd As a characteristic of time of day, e pw As a feature of the week time e ph The time characteristics of the holidays are shown. The above time characteristics are represented by one-hot codes, each one-hot code belongs to a real number space with different preset dimensions, and the specific preset dimensions are set according to implementation conditions and are not limited here. After the unique hot code of each time feature is obtained, the unique hot code is converted into a corresponding time embedding vector through a linear layer, as follows:
h pd ＝W pd e pd
h pw ＝W pw e pw
h ph ＝W ph e ph
wherein e is pd ,e pw ,e ph A unique hot code representing the time of day characteristic, the time of week characteristic and the holiday time characteristic respectively,
corresponding to the three, the time embedded vectors are respectively the time embedded vectors,
to predict the model parameters, adjustments may be made according to a training process. d is a radical of d,w,h Representing the dimension, T, of each temporal embedding vector d ,T w ,T h And representing the dimensions, and setting specific numerical values of the dimensions according to implementation conditions. According to the three time embedding vectors, representing the release time characteristics together to obtain the release time embedding vector：
h pt ＝[h pd ,h pw ,h ph ]
Wherein [ 2 ]]Represents vector stitching, h pt I.e. the release time embedding vector.
According to the text content, the text content issued by the user can be converted by a natural language processing model to obtain the text content characteristics, such as a large-scale pre-training model BERT, to obtain the corresponding text content embedding vector v p 。
Considering that the forwarding of the text content is influenced by the contents such as the hot topic when the text is published, the text forwarding can be predicted by combining the contents of the hot topic. Specifically, a group of hot content embedding vectors V of hot topics and other contents (hereinafter referred to as hot contents) are embedded and constructed according to the hot topic contents h ∈R 50×d The dimension of R is for illustration, and may be specifically set according to the implementation. Although the hot content may represent public opinion and background information of the current text content, the hot content is not combined with the specific text content, and the forwarding condition of the specific text content cannot be determined, and the hot content is numerous, and the hot content having an influence on the current text content needs to be obtained through screening as the background information of the current text content. In order to filter out hot contents irrelevant to the current text contents in the hot contents, an attention mechanism can be adopted to filter data of each hot topic, and the similarity between the hot contents and the currently processed text contents is calculated. Specifically, the calculation is performed according to the following formula:
v pt ＝f p (W p v p +b p )
V ht ＝f h (W h V h +b h )
wherein, W p ,W h ,b p ,b h V can be adjusted according to the training process in order to predict the model parameters pt 、V ht According to respective function f p 、f h Embedding vectors v into text content p Embedded vector V of hot content h Is calculated to obtain N h I.e. number of hot spot contents, V hti Denotes V ht I.e. the embedding result, beta, of the ith hot content i Represents the importance degree of the ith hot content to the text content, beta i The similarity between the text content and the hot content is calculated. After the embedded vector of each hot content and the importance of the hot content to the text content are obtained, the background feature of the text content can be obtained through weighted summation.
Wherein v is bg Background feature vectors representing the current text content, in order to fuse features within the text with the background features, a gating cell may be used to fuse the two. And calculating the output of a gating mechanism through the background feature vector and the text content embedding vector, namely outputting the proportion of the background feature vector in a final result by using a gating unit.
h hg ＝[v bg ,v p ]
α＝sigmoid(f(W hg h hg +b hg ))
Wherein h is hg Input representing gating information, W hg 、b hg For predicting model parameters, f is an activation function, and alpha represents the weight occupied by the background feature vector. And according to the background feature weight, performing weighted summation on the background feature vector and the text content embedded vector, and calculating to obtain a text background and a content embedded vector containing the background feature vector.
e bg ＝α*v bg +(1-α)*v p
Considering that the text content has a phenomenon of heat fading, that is, the heat declines with time, the self-attention mechanism cannot completely extract the relevant features of the heat fading, so that the text content needs to be modeled by using the time attention mechanism to extract the heat fading features, and the heat of the text content can be defined as the 'age features' of the text content, which fade with time.
The age characteristic of the text content is similar to that of the publishing time, but the age characteristic has no periodicity and no special value, so that the text content only needs to be embedded by using the age characteristic, and the periodicity, the special value and the like are not required to be considered, and the text content can be obtained by the following formula:
h dt ＝E dt e dt
wherein e is dt ∈R D Is a one-hot code characteristic of age, R D Is the number of age ranges, and D is the dimension set according to the implementation.
Adjustments are made according to the training process to predict model parameters. />
Is a text age embedding vector.
Splicing the text age embedded vector with the text content release time embedded vector and the text background and content embedded vector to obtain e age ＝[h dt ,h pt,bg ]And obtaining the forwarding text embedded vector containing the text age, the release time, the text background characteristics and the text content characteristics.
The characteristics of the user dependency relationship and the characteristics of the forwarded text content play an important role in the forwarding behavior prediction, and the content information can have more accurate expression capability by combining the characteristics of the user dependency relationship and the characteristics of the forwarded text content, so that the prediction accuracy is improved.
Here, steps S101 to S102 are related calculation of the embedded vector of the user, step S103 is related calculation of the embedded vector of the text content, and the execution order of the two may be prior to any one of the calculations, and the execution order of steps S101 to S102 and step S103 is not limited herein.
And step S104, calculating an attention result of the text content embedded by the forwarding user sequence according to the forwarding text embedding vector and the forwarding user sequence embedding of the text content, and obtaining the prediction probability of the text content forwarded by the user according to the attention result.
Calculating the influence on the embedding of the forwarding user sequence by using the forwarding text embedding vector, and calculating the attention mechanism coefficient as follows:
wherein, w f ∈R d To predict the model parameters, adjustments may be made according to the training process, e age Playing a role of gate control information, controlling the feature strength of the text age, the release time, the text background feature, the text content feature and the like to enter the subsequent processing, e ci Refers to the embedding of the ith user in the embedding of the forwarded user sequence,
is the result of the attention of the ith user. As an example, we can use Hadmard's function to represent the term-by-term multiplication of vectors. Calculating aiming at the embedding of the forwarding user sequence containing n users to obtain the attention result of the whole forwarding user sequence embedding on the forwarding text content, wherein the attention result is a cascade embedding vector:
from the above attention results, for any time t i Cascading attention results e cas Get user u i For the next user u i+1 Forwarding text content s i The prediction probability of (2) is shown by the following formula:
among them, MLP (Multilayer perceivron),
represents each of the user groupsAnd the probability that each user forwards the text content for the next forwarding user, wherein the user group is the user group specified according to the social relationship of the users.
Optionally, this embodiment further includes the following steps:
and step S105, predicting the target forwarding probability of the user group, and optimizing the prediction model according to the target forwarding probability and the sample forwarding probability in the training set.
In the embodiment, when the probability that the user forwards the text content to the next user is predicted, the prediction model is trained based on the pre-collected sample data. According to the training set in the collected sample data, the sample forwarding probabilities of all user groups in the training set can be obtained, and each prediction model parameter of the prediction model in the training process can be adjusted according to the probability of the sample forwarding users. Calculating the forwarding probability of the whole user group according to the probability of the predicted user forwarding the text content to the next user, as follows:
the objective of optimizing the prediction model is to make L reach the maximum value as much as possible, during optimization, through an Adam optimizer and a back propagation algorithm, for example, and taking the result of L as a reference, each parameter in the prediction model is optimized, and finally, the parameters of the prediction model obtained through training are parameters of each prediction model when the prediction model is used for prediction.
The schematic architecture diagram of the prediction model is shown in fig. 2, the input information includes a friend relationship network of a user, a forwarding process of text content, release content (i.e., text content) and hot content, and the prediction model includes different stages such as representation learning, user dependence, timing dependence, external dependence and prediction. In the expression learning stage, the expression learning is carried out on a friend relationship network (namely a user social network), for example, a first embedded vector of a user is obtained by utilizing GAT extraction; for the forwarding process, obtaining a forwarding sequence (for example, a forwarding process matrix formed by a text forwarding sequence by a user), embedding time attenuation and the like in the forwarding process, and the like; for post content and hot content, text embedding (i.e., text content embedding vectors) and the like can be extracted. In the user-dependent stage, the influence of each user in the forwarding process is weighted and summed through an attention mechanism to form a context-dependent vector (i.e. for user sequence embedding in fig. 2), and the context-dependent vector is fused with the first embedding vector of the user (i.e. the context information in fig. 2) to form a second embedding vector of the user. And the second embedded vector of the user is used as the input of a time sequence dependency stage, and the time sequence dependency characteristics in the sequence are extracted through the processing of a Transformer encoder to obtain the embedding of the forwarding user sequence in the forwarding process, wherein the embedding of the forwarding user sequence comprises the addition of position coding, and the embedding of the forwarding user sequence is output by the Transformer encoder through multi-head attention, weighted summation, normalization, feed-forward network and the like. In the external dependency and prediction stage, feature vectors composed of the features of the hot content and the release content, time information and the like, namely forwarding text embedding vectors, are subjected to weighted summation on the embedding influence of the obtained forwarding user sequences to form, for example, a cascade embedding vector, namely an attention result (namely, external dependency attention in fig. 2), and finally the probability of forwarding text content of each user is output through an MLP model according to the attention result.
According to the social network forwarding behavior prediction method based on the user dependency relationship, provided by the embodiment of the invention, the relevance of the behavior of forwarding the text content in the user social network and various characteristics is analyzed, the prediction model of text content propagation based on the user social network is established, the propagation range of the text content can be predicted, the propagation of the text content is conveniently controlled, a more appropriate text content is recommended by a help platform, more accurate popularization is carried out, reasonable control on the content is realized, and the development of a text platform is promoted. Furthermore, the invention uses a single context feature extraction mode to extract the influence of the past forwarding user on the current user node, so as to obtain the context dependent vector. And the influence among the users is calculated by using a self-attention mechanism and a preset direction mask, so that the long-term dependence problem is relieved. The invention uses the Transformer as a sequence processing tool, is different from the existing deep learning model, the Transformer module completely abandons the structure of RNN and other sequence models, and adopts an attention mechanism to construct the model, any two nodes in the sequence can be directly associated, and the loss of characteristics is avoided. The invention greatly relieves the long-term dependence problem in the RNN model and avoids the problem that influence relation cannot be formed between non-adjacent nodes. Meanwhile, the invention determines a corresponding prediction model aiming at the characteristics of the forwarding process through the result of the characteristic analysis of the forwarding behavior, thereby completing the prediction of the user forwarding behavior.
Fig. 3 is a schematic structural diagram illustrating a social network forwarding behavior prediction apparatus based on user dependency according to an embodiment of the present invention. As shown in fig. 3, the apparatus includes:
the first user vector module 310 is adapted to determine a user group according to the social relationship of the users, and obtain a first embedded vector of each user in the user group;
a second user vector module 320, adapted to calculate a second embedded vector of the user including the context dependent vector according to the forwarding process and the first embedded vector of the user; according to the second embedding vector of the user and a preset direction mask, performing cross-layer connection and normalization processing according to the sequence of the text contents forwarded by the user on the basis of a self-attention mechanism to obtain the forwarding user sequence embedding in the forwarding process;
a text vector module 330 adapted to determine a forward text embedding vector of the text content according to the text content;
the prediction module 340 is adapted to calculate an attention result of the forward user sequence embedding on the text content according to the forward text embedding vector of the text content and the forward user sequence embedding, and obtain a prediction probability of the user forward text content according to the attention result.
Optionally, the first user vector module 310 is further adapted to:
acquiring a user social relationship, determining a user group according to the user social relationship, and constructing a user social relationship network; the user social relation network is constructed in a graph data structure and comprises a plurality of user nodes and neighbor nodes connected with the user nodes;
and aiming at the user group, inputting the vector of each user node into the graph neural network model to obtain a first embedded vector of each user in the user group.
Optionally, the first user vector module 310 is further adapted to:
aiming at any user node in a user group, inputting a vector of the user node and a vector of a neighbor node of the user node into a graph neural network model, calculating the influence of each neighbor node on the user node, and splicing a plurality of obtained attention results of the user node according to the influence based on a multi-head attention mechanism; outputting an average value of the plurality of attention results as an embedding vector of the user node;
and splicing the preset random vector and the embedded vector of the user node to obtain a first embedded vector of the user.
Optionally, the second user vector module 320 is further adapted to:
calculating the influence of a previous forwarding user on a subsequent forwarding user according to the forwarding sequence of the forwarding process, and obtaining a context dependent vector of the subsequent forwarding user according to the weighted calculation of the influence;
and fusing the first embedding vector of the user and the context dependent vector of the user to obtain a second embedding vector of the user containing the context dependent vector.
Optionally, the text vector module 330 is further adapted to:
determining a release time embedding vector, a text background, a content embedding vector and a text age embedding vector of the text content according to the text content;
and splicing the release time embedded vector, the text background and the content embedded vector of the text content and the text age embedded vector to obtain a forwarding text embedded vector of the text content.
Optionally, the text vector module 330 is further adapted to:
determining the time characteristics of the release time of the text content according to the release time of the text content and a plurality of preset dimensions; the plurality of preset dimensions include daily, weekly, and/or holiday dimensions; the time characteristics are expressed by adopting one-hot coding;
determining time embedding vectors of a plurality of preset dimensions according to the time characteristics of the release time of the text contents of the plurality of preset dimensions;
splicing the time embedded vectors of a plurality of preset dimensions to obtain a release time embedded vector of the text content;
according to the text content, obtaining a corresponding text content embedding vector, and constructing a corresponding hot content embedding vector according to the hot content;
based on an attention mechanism, calculating the similarity between the hot content and the text content according to the text content embedding vector and the hot content embedding vector, and calculating to obtain a background feature vector of the text content according to the similarity;
calculating the weights of the background feature vector and the text content embedded vector, and calculating to obtain a text background and a content embedded vector according to the weights; the weight is obtained based on a gating mechanism;
and determining the age characteristics of the text content based on the time attention mechanism, and obtaining a text age embedding vector according to the age characteristics.
Optionally, the prediction module 340 is further adapted to:
calculating the influence on each user embedded in the forwarding user sequence according to the forwarding text embedding vector of the text content, and obtaining the attention result of the text content embedded in the forwarding user sequence according to the influence and the embedding of the forwarding user sequence;
and according to the attention result, obtaining the prediction probability of the text content forwarded to the next user by the user in the forwarding sequence.
Optionally, the prediction module 340 is further adapted to:
and obtaining the prediction probability of the text content forwarded to the next user by the user in the forwarding sequence based on the attention result of the multi-layer perception computer.
Optionally, the apparatus further comprises:
and the optimization module 350 is adapted to predict the target forwarding probability of the obtained user group and optimize the prediction model according to the target forwarding probability and the sample forwarding probability in the training set.
The descriptions of the modules refer to the corresponding descriptions in the method embodiments, and are not repeated herein.
The embodiment of the invention also provides a nonvolatile computer storage medium, wherein the computer storage medium stores at least one executable instruction, and the executable instruction can execute the social network forwarding behavior prediction method based on the user dependency relationship in any method embodiment.
Fig. 4 is a schematic structural diagram of a computing device according to an embodiment of the present invention, and a specific embodiment of the present invention does not limit a specific implementation of the computing device.
As shown in fig. 4, the computing device may include: a processor (processor) 402, a Communications Interface 404, a memory 406, and a Communications bus 408.
Wherein:
the processor 402, communication interface 404, and memory 406 communicate with each other via a communication bus 408.
A communication interface 404 for communicating with network elements of other devices, such as clients or other servers.
The processor 402 is configured to execute the program 410, and may specifically execute relevant steps in the foregoing social network forwarding behavior prediction method embodiment based on the user dependency relationship.
In particular, program 410 may include program code comprising computer operating instructions.
The processor 402 may be a central processing unit CPU, or an Application Specific Integrated Circuit ASIC (Application Specific Integrated Circuit), or one or more Integrated circuits configured to implement an embodiment of the present invention. The computing device includes one or more processors, which may be the same type of processor, such as one or more CPUs; or may be different types of processors such as one or more CPUs and one or more ASICs.
And a memory 406 for storing a program 410. Memory 406 may comprise high-speed RAM memory, and may also include non-volatile memory (non-volatile memory), such as at least one disk memory.
The program 410 may be specifically configured to enable the processor 402 to execute the social network forwarding behavior prediction method based on the user dependency relationship in any of the method embodiments described above. For specific implementation of each step in the program 410, reference may be made to corresponding steps and corresponding descriptions in units in the foregoing social network forwarding behavior prediction embodiment based on the user dependency relationship, which are not described herein again. It can be clearly understood by those skilled in the art that, for convenience and brevity of description, the specific working processes of the above-described devices and modules may refer to the corresponding process descriptions in the foregoing method embodiments, and are not described herein again.
The algorithms or displays presented herein are not inherently related to any particular computer, virtual system, or other apparatus. Various general purpose systems may also be used with the teachings herein. The required structure for constructing such a system is apparent from the description above. In addition, embodiments of the present invention are not directed to any particular programming language. It is appreciated that a variety of programming languages may be used to implement the teachings of embodiments of the present invention as described herein, and any descriptions of specific languages are provided above to disclose preferred embodiments of the invention.
In the description provided herein, numerous specific details are set forth. It is understood, however, that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.
Similarly, it should be appreciated that in the foregoing description of exemplary embodiments of the invention, various features of the embodiments of the invention are sometimes grouped together in a single embodiment, figure, or description thereof for the purpose of streamlining the disclosure and aiding in the understanding of one or more of the various inventive aspects. However, the disclosed method should not be construed to reflect the intent: that is, the claimed embodiments of the invention require more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of this invention.
Those skilled in the art will appreciate that the modules in the device in an embodiment may be adaptively changed and disposed in one or more devices different from the embodiment. The modules or units or components of the embodiments may be combined into one module or unit or component, and furthermore they may be divided into a plurality of sub-modules or sub-units or sub-components. All of the features disclosed in this specification (including any accompanying claims, abstract and drawings), and all of the processes or elements of any method or apparatus so disclosed, may be combined in any combination, except combinations where at least some of such features and/or processes or elements are mutually exclusive. Each feature disclosed in this specification (including any accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise.
Furthermore, those skilled in the art will appreciate that while some embodiments herein include some features included in other embodiments, rather than other features, combinations of features of different embodiments are meant to be within the scope of the invention and form different embodiments. For example, in the following claims, any of the claimed embodiments may be used in any combination.
Various component embodiments of the invention may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. Those skilled in the art will appreciate that a microprocessor or Digital Signal Processor (DSP) may be used in practice to implement some or all of the functionality of some or all of the components according to embodiments of the present invention. Embodiments of the invention may also be implemented as apparatus or device programs (e.g., computer programs and computer program products) for performing a portion or all of the methods described herein. Such programs implementing embodiments of the present invention may be stored on computer-readable media or may be in the form of one or more signals. Such a signal may be downloaded from an internet website, or provided on a carrier signal, or provided in any other form.
It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and that those skilled in the art will be able to design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. Embodiments of the invention may be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by one and the same item of hardware. The usage of the words first, second and third, etcetera do not indicate any ordering. These words may be interpreted as names. The steps in the above embodiments should not be construed as limiting the order of execution unless specified otherwise.
Claims (12)
1. A social network forwarding behavior prediction method based on user dependency is characterized by comprising the following steps:
determining a user group according to the social relationship of the users to obtain a first embedded vector of each user in the user group;
calculating to obtain a second embedded vector of the user containing the context dependent vector according to the forwarding process and the first embedded vector of the user; according to the second embedded vector of the user and a preset direction mask, performing cross-layer connection and normalization processing according to the sequence of the text contents forwarded by the user on the basis of a self-attention mechanism to obtain the forwarding user sequence embedding in the forwarding process;
determining a forwarding text embedding vector of the text content according to the text content;
and calculating an attention result of the forward user sequence embedding on the text content according to the forward text embedding vector of the text content and the forward user sequence embedding, and obtaining the prediction probability of the user forward text content according to the attention result.
2. The method of claim 1, wherein determining a user population according to user social relationships, and obtaining the first embedded vector for each user in the user population further comprises:
acquiring a user social relationship, determining a user group according to the user social relationship, and constructing a user social relationship network; the user social relationship network is constructed in a graph data structure and comprises a plurality of user nodes and neighbor nodes connected with the user nodes;
and aiming at the user group, inputting the vector of each user node into the graph neural network model to obtain a first embedded vector of each user in the user group.
3. The method of claim 2, wherein the inputting the vector of each user node into the neural network model for the user population to obtain the first embedded vector of each user in the user population further comprises:
aiming at any user node in the user group, inputting the vector of the user node and the vector of the neighbor node of the user node into a graph neural network model, calculating the influence of each neighbor node on the user node, and splicing a plurality of obtained attention results of the user node according to the influence based on a multi-head attention mechanism; outputting an average of a plurality of attention results as an embedded vector for the user node;
and splicing a preset random vector with the embedded vector of the user node to obtain a first embedded vector of the user.
4. The method of claim 1, wherein calculating a second embedding vector for a user comprising a context dependent vector according to the forwarding process and the first embedding vector for the user further comprises:
calculating the influence of a previous forwarding user on a subsequent forwarding user according to the forwarding sequence in the forwarding process, and obtaining a context dependence vector of the subsequent forwarding user according to the influence weighting calculation;
and fusing the first embedded vector of the user and the context dependent vector of the user to obtain a second embedded vector of the user containing the context dependent vector.
5. The method of claim 1, wherein determining a forward text embedding vector for the text content based on the text content further comprises:
determining a release time embedding vector, a text background and a content embedding vector of the text content and a text age embedding vector according to the text content;
and splicing the release time embedded vector, the text background and the content embedded vector of the text content and the text age embedded vector to obtain a forwarding text embedded vector of the text content.
6. The method of claim 5, wherein determining from the text content a publication time embedding vector, a text background and a content embedding vector for the text content, and wherein determining the text age embedding vector further comprises:
determining the time characteristics of the release time of the text content according to a plurality of preset dimensions according to the release time of the text content; the plurality of preset dimensions include daily, weekly, and/or holiday dimensions; the time characteristics are represented by one-hot coding;
determining time embedding vectors of a plurality of preset dimensions according to the time characteristics of the release time of the text contents of the plurality of preset dimensions;
splicing the time embedded vectors of the plurality of preset dimensions to obtain the release time embedded vector of the text content;
according to the text content, obtaining a corresponding text content embedding vector, and constructing a corresponding hot content embedding vector according to the hot content;
based on an attention mechanism, calculating the similarity between the hot content and the text content according to the text content embedding vector and the hot content embedding vector, and calculating to obtain a background feature vector of the text content according to the similarity;
calculating the weights of the background feature vector and the text content embedded vector, and calculating to obtain a text background and a text content embedded vector according to the weights; the weights are derived based on a gating mechanism;
and determining the age characteristics of the text content based on the time attention mechanism, and obtaining a text age embedding vector according to the age characteristics.
7. The method of claim 1, wherein the calculating an attention result of the forward user sequence embedding to the text content according to the forward text embedding vector of the text content and the forward user sequence embedding, and obtaining a predicted probability of the user forward text content according to the attention result further comprises:
calculating influence on each user embedded in the forwarding user sequence according to the forwarding text embedding vector of the text content, and obtaining an attention result of the forwarding user sequence embedding on the text content according to the influence and the forwarding user sequence embedding;
and according to the attention result, obtaining the prediction probability of the text content forwarded to the next user by the user in the forwarding sequence.
8. The method of claim 7, wherein obtaining a predicted probability of forwarding text content from a user to a next user in a forwarding order based on the attention result further comprises:
and calculating the attention result based on the multilayer perception computer to obtain the prediction probability of the text content forwarded to the next user by the user in the forwarding sequence.
9. The method according to any one of claims 1-8, further comprising:
and predicting to obtain the target forwarding probability of the user group, and optimizing the prediction model according to the target forwarding probability and the sample forwarding probability in the training set.
10. An apparatus for predicting social network forwarding behavior based on user dependency, the apparatus comprising:
the first user vector module is suitable for determining a user group according to the social relationship of users to obtain a first embedded vector of each user in the user group;
the second user vector module is suitable for calculating a second embedded vector of the user containing the context dependent vector according to the forwarding process and the first embedded vector of the user; performing cross-layer connection and normalization processing according to the second embedded vector of the user and a preset direction mask and based on a self-attention mechanism and according to the sequence of the text contents forwarded by the user to obtain the forwarding user sequence embedding in the forwarding process;
the text vector module is suitable for determining a forwarding text embedding vector of the text content according to the text content;
and the prediction module is suitable for calculating an attention result of the text content embedded by the forwarding user sequence according to the forwarding text embedded vector of the text content and the embedding of the forwarding user sequence, and obtaining the prediction probability of the text content forwarded by the user according to the attention result.
11. A computing device, comprising: the system comprises a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface complete mutual communication through the communication bus;
the memory is used for storing at least one executable instruction, and the executable instruction causes the processor to execute the operation corresponding to the social network forwarding behavior prediction method based on the user dependency relationship in any one of claims 1-9.
12. A computer storage medium, wherein at least one executable instruction is stored in the storage medium, and the executable instruction causes a processor to perform an operation corresponding to the social network forwarding behavior prediction method based on user dependency relationship according to any one of claims 1 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211697432.3A CN115878907B (en) | 2022-12-28 | 2022-12-28 | Social network forwarding behavior prediction method and device based on user dependency relationship |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211697432.3A CN115878907B (en) | 2022-12-28 | 2022-12-28 | Social network forwarding behavior prediction method and device based on user dependency relationship |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115878907A true CN115878907A (en) | 2023-03-31 |
| CN115878907B CN115878907B (en) | 2023-08-11 |
Family
ID=85755723
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211697432.3A Active CN115878907B (en) | 2022-12-28 | 2022-12-28 | Social network forwarding behavior prediction method and device based on user dependency relationship |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115878907B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116091260A (en) * | 2023-04-07 | 2023-05-09 | 吕梁学院 | Cross-domain entity identity association method and system based on Hub-node |
| CN116842398A (en) * | 2023-06-27 | 2023-10-03 | 哈尔滨工业大学 | Topic perception information forwarding prediction method and system based on shielding self-attention network |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180365562A1 (en) * | 2017-06-20 | 2018-12-20 | Battelle Memorial Institute | Prediction of social media postings as trusted news or as types of suspicious news |
| CN112418525A (en) * | 2020-11-24 | 2021-02-26 | 重庆邮电大学 | Method and device for predicting social topic group behaviors and computer storage medium |
| CN113139134A (en) * | 2021-04-26 | 2021-07-20 | 中国人民解放军国防科技大学 | Method and device for predicting popularity of user-generated content in social network |
| CN113807616A (en) * | 2021-10-22 | 2021-12-17 | 重庆理工大学 | Information diffusion prediction system based on space-time attention and heterogeneous graph convolution network |
| CN113901298A (en) * | 2021-10-29 | 2022-01-07 | 杭州师范大学钱江学院 | Social network forwarding behavior prediction method based on microblog text embedded learning |
| CN115168743A (en) * | 2022-07-27 | 2022-10-11 | 天津大学 | Crowd-sourcing body forwarding behavior prediction system based on heterogeneous graph neural network |
| CN115293437A (en) * | 2022-08-15 | 2022-11-04 | 重庆邮电大学 | Relationship prediction method based on social network platform dynamic interaction |
| CN115470991A (en) * | 2022-09-15 | 2022-12-13 | 重庆邮电大学 | Network rumor propagation prediction method based on user short-time emotion and evolutionary game |
-
2022
- 2022-12-28 CN CN202211697432.3A patent/CN115878907B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180365562A1 (en) * | 2017-06-20 | 2018-12-20 | Battelle Memorial Institute | Prediction of social media postings as trusted news or as types of suspicious news |
| CN112418525A (en) * | 2020-11-24 | 2021-02-26 | 重庆邮电大学 | Method and device for predicting social topic group behaviors and computer storage medium |
| CN113139134A (en) * | 2021-04-26 | 2021-07-20 | 中国人民解放军国防科技大学 | Method and device for predicting popularity of user-generated content in social network |
| CN113807616A (en) * | 2021-10-22 | 2021-12-17 | 重庆理工大学 | Information diffusion prediction system based on space-time attention and heterogeneous graph convolution network |
| CN113901298A (en) * | 2021-10-29 | 2022-01-07 | 杭州师范大学钱江学院 | Social network forwarding behavior prediction method based on microblog text embedded learning |
| CN115168743A (en) * | 2022-07-27 | 2022-10-11 | 天津大学 | Crowd-sourcing body forwarding behavior prediction system based on heterogeneous graph neural network |
| CN115293437A (en) * | 2022-08-15 | 2022-11-04 | 重庆邮电大学 | Relationship prediction method based on social network platform dynamic interaction |
| CN115470991A (en) * | 2022-09-15 | 2022-12-13 | 重庆邮电大学 | Network rumor propagation prediction method based on user short-time emotion and evolutionary game |
Non-Patent Citations (1)
| Title |
|---|
| 张志扬 等: "基于分层注意力的信息级联预测模型", 计算机科学, vol. 47, no. 06, pages 201 - 209 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116091260A (en) * | 2023-04-07 | 2023-05-09 | 吕梁学院 | Cross-domain entity identity association method and system based on Hub-node |
| CN116091260B (en) * | 2023-04-07 | 2023-07-25 | 吕梁学院 | Cross-domain entity identity association method and system based on Hub-node |
| CN116842398A (en) * | 2023-06-27 | 2023-10-03 | 哈尔滨工业大学 | Topic perception information forwarding prediction method and system based on shielding self-attention network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115878907B (en) | 2023-08-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112150210B (en) | Improved neural network recommendation method and system based on GGNN (global warming network) | |
| US20210381712A1 (en) | Determining demand curves from comfort curves | |
| CN115878907A (en) | Social network forwarding behavior prediction method and device based on user dependency relationship | |
| WO2021047593A1 (en) | Method for training recommendation model, and method and apparatus for predicting selection probability | |
| Nasiri et al. | A simulation optimisation approach for real-time scheduling in an open shop environment using a composite dispatching rule | |
| Chen et al. | Modeling and optimization of complex building energy systems with deep neural networks | |
| CN108021983A (en) | Neural framework search | |
| CN112541124B (en) | Method, apparatus, device, medium and program product for generating a multitasking model | |
| CN108876044B (en) | Online content popularity prediction method based on knowledge-enhanced neural network | |
| CN111723910A (en) | Method and device for constructing multi-task learning model, electronic equipment and storage medium | |
| Costa et al. | Exploring genetic programming and boosting techniques to model software reliability | |
| CN115018193A (en) | Time series wind energy data prediction method based on LSTM-GA model | |
| CN115221396A (en) | Information recommendation method and device based on artificial intelligence and electronic equipment | |
| CN116684330A (en) | Traffic prediction method, device, equipment and storage medium based on artificial intelligence | |
| JP6219509B2 (en) | Assigning and examining synaptic delays dynamically | |
| US20200167605A1 (en) | Computer architecture for artificial image generation | |
| CN115953215B (en) | Search type recommendation method based on time and graph structure | |
| Chen et al. | Design of manufacturing systems by a hybrid approach with neural network metamodelling and stochastic local search | |
| CN114338416A (en) | Space-time multi-index prediction method and device and storage medium | |
| CN115409217A (en) | Multitask predictive maintenance method based on multi-expert hybrid network | |
| Zhou et al. | A Novel Ensemble Strategy Combining Gradient Boosted Decision Trees and Factorization Machine Based Neural Network for Clicks Prediction | |
| HIRANO et al. | Data-driven Agent Design for Artificial Market Simulation | |
| CN114548382A (en) | Migration training method, device, equipment, storage medium and program product | |
| CN118043802A (en) | Recommendation model training method and device | |
| Su et al. | Neural network based fusion of global and local information in predicting time series |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |