Social search hotspot propagation prediction method, system, program and equipment based on multi-feature sequence and storage medium

Abstract

The invention provides a social search hotspot propagation prediction method and system based on a multi-feature sequence, a program, equipment and a storage medium. According to the method, time information, event type information, user information and user attribute information in the event sequence are considered, representation learning is carried out on the social search hot event sequence by utilizing the Transform model, and the learned representation vector is used as the input of the Hawkes process, so that the mutual influence relationship between the events can be better captured. According to the method, the event sequence is predicted according to the intensity function of the historical event, the short-term prediction is to predict the occurrence time and the event type of the next event, and the long-term prediction is to predict the propagation size of the event in a given time interval, so that better sequence data fitting and sequence prediction effects are achieved.

Description

Technical Field

[0001] The present invention belongs to the field of social network security technology, and specifically relates to a method, system, program, device and storage medium for predicting the propagation of social search hot spots based on multiple feature sequences. Background Art

[0002] There are a lot of event sequence data in daily life [Zuo S, Jiang H, Li Z, et al. Transformer Hawkes Process [J]. 2020: 437-457.]. For example, in social media sites such as Twitter and Facebook, each user has a series of events, such as tweets and interactions, where people share experiences and respond to other users' information. Hundreds of millions of users generate a large number of tweets, which essentially form event sequences with different timestamps. Event sequence data shows dependence on its historical data, and users' past social behaviors (such as posting, commenting, sharing, etc.) will have an impact on current behaviors. In addition to sharing and forwarding on social media platforms, similar event sequence data can also be found in other (non-social media) environments, such as paper citations, blog spaces, and email forwarding. The core challenge of event sequence prediction is how to model the potential diffusion process of information, that is, to simulate its propagation trend.

[0003] A malicious incident on a social network can cause widespread public discussion and lead to public opinion storms. In order to avoid such problems, this study hopes to propose a model to simulate the trend of social search hotspot propagation, which can be used to predict the occurrence time and event type of the next social search hotspot event from the current observation time, as well as the propagation size of the social search hotspot event sequence within a period of time from the current observation time.

[0004] Social search hot spot propagation data is also a kind of time series data, so it is necessary to capture its temporal changes and dependencies. The Hawkes process can model such event sequence data [Hawkes AG. Spectra of some self-exciting and mutually exciting point processes [J]. Biometrika, 1971: 83-90.], but the process assumes that the occurrence of past events has a positive influence on the occurrence of current events, which simplifies the complex situation in reality, because the impact of past events on current events can be either enhancing or inhibiting. With the development of deep learning theory, Mei et al. proposed the Neural Hawkes model [Mei H, Eisner J. The Neural Hawkes Process: A Neurally Self-Modulating Multivariate Point Process [J]. Conference and Workshop on Neural Information Processing Systems, 2017: 6754-6764.], using LSTM to redesign the conditional intensity function. One of its significant defects is that it cannot capture the complex long-term dependencies between events. Inspired by the development of natural language processing, Simiao et al. [Zuo S, Jiang H, Li Z, et al. Transformer Hawkes Process [J]. 2020: 437-457.] proposed an attention-based Hawkes model, the Transformer Hawkes model, which can capture complex long-term and short-term dependencies in the sequence, but the model does not consider the users who forward the information and the user attribute information. Therefore, the present invention intends to draw on the Transformer Hawkes model, comprehensively consider the various features in the social search hot event sequence, including time features, event type features, user node features and user attribute features, and propose a social search hot propagation prediction model (Multi-Feature Sequence Public Opinion Propagation Prediction Model, referred to as MFSPOP) combined with a multi-feature sequence to improve the accuracy of social search hot data modeling and prediction. Summary of the invention

[0005] The object of the present invention is to provide a method, system, program, device and storage medium for predicting the propagation of social search hot spots based on multiple feature sequences.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] A method for predicting the spread of social search hot spots based on multiple feature sequences, comprising the following steps:

[0008] Step 1. Event Stream Analysis Probe

[0009] Define event streams as a sequence of discrete events that occur in continuous time. Each event is represented by a triple, including timestamp, event type, and user information. Analyzing and modeling event sequences can predict the type and time of the next event.

[0010] Step 2. Event sequence encoding

[0011] The input sequence is processed by a multi-head self-attention mechanism and a fully connected layer to generate an encoder output containing global semantic information. The decoder decodes the encoder output and previously generated symbols into a new output sequence and extracts the hidden state of each event, including the long-term and short-term dependencies of the events. The multi-layer mechanism is used to sequentially implement the correlation coding of the sequence and the weighted coding of the input sequence.

[0012] Step 3. Event Representation Learning

[0013] After the initial encoding and multi-head attention mechanism, the result is input into the feedforward neural network to further generate the hidden layer representation of the input event sequence, completing the generation process of the representation learning vector of the event sequence;

[0014] Step 4. Prediction of social search hot spots

[0015] The obtained historical event sequence is used as the input of this process to describe the event sequence that changes over time. The dynamic changes of the time point process are described by a continuous conditional intensity function to capture the mutual influence relationship between events and predict the occurrence of future events.

[0016] Furthermore, the step 1 comprises the following steps:

[0017] Step 1.1. Each event (t j ,k j ,u j ) is a triplet, indicating that user u j At time t j Retweeted k j Type of information; represents an event sequence, where k j represents the jth event in the sequence, t j Indicates the timestamp of the event; u j Indicates the user corresponding to the event;

[0018] Step 1.2. In the time interval [t A ,tE ], there is a sequence of retweets (t A ,k1,u A ),(t B ,k1,u B ),(t F ,k2,u F ),(t C ,k1,u C ),…,(t H ,k2,u H ),(t E ,k1,u E );

[0019] Step 1.3. Given a social search hot event sequence Make short-term and long-term forecasts. Short-term forecasts refer to predicting when the next event will occur. With event type Long-term forecasting refers to the p -t i , predict the spread of social search hot events

[0020] Furthermore, the step 2 comprises the following steps:

[0021] Step 2.1. Assume that the social search hotspot propagation event sequence with a given event length L where t j Indicates the event (t j ,k j ,u j ) occurs; given the following time code definitions:

[0022]

[0023] Step 2.2. Given For each event type, an embedding matrix U∈R is learned M*K , M represents the embedding dimension, K represents the number of event types, and the jth column of U represents event type k j M-dimensional embedding of;

[0024] Step 2.3. After obtaining the event time code, event type code, node information code, and user attribute code, the event sequence The embedding of is X=(E+Z+H+G) T .

[0025] Furthermore, the step 3 comprises the following steps:

[0026] Step 3.1. Use the self-attention module to obtain the long-term and short-term dependencies between sequences; calculate the attention output S =

[0027]

[0028] Step 3.2. After the social search hotspot propagation event sequence S passes through the initial encoding and multi-head self-attention mechanism, the attention output S out Generate a hidden representation of the input event sequence S through a feedforward neural network and h(t j )=H(j,:).

[0029] Furthermore, step 4 includes the following steps:

[0030] Step 4.1. For each k∈{1,2,…,K}, set λ k (t|H t ) is defined as the conditional intensity function of k-type events, and the conditional intensity function of the entire social search hotspot event sequence is

[0031] Step 4.2. The specific intensity of each event type is adopted

[0032] Step 4.3. The probability of the next social search hot event occurring at any time is expressed as

[0033] Step 4.4. The timestamp prediction of the next event and the event type prediction of the next event are expressed as and

[0034] Step 4.5. Social search hot event sequence at t p -t i Estimated spread size during this time The prediction is expressed as

[0035] A multi-constraint guided Chinese rumor adversarial sample generation system, including an event flow analysis device, an event sequence encoding device, an event representation learning device, and a social search hotspot prediction device;

[0036] The event stream analysis device captures the event stream occurring in the social network in a continuous time, including timestamps, event types and user information;

[0037] The event sequence coding device codes and combines the time code, event type code, node information code, and user attribute code information into an event sequence code;

[0038] The event representation learning device captures and forms long-term and short-term dependency information between events, and completes the generation process of the representation learning vector of the event sequence;

[0039] The social search hotspot prediction device captures the mutual influence relationship between historical events and predicts the occurrence of future events.

[0040] A computer device / equipment / system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a method for predicting the propagation of social search hot spots based on multiple feature sequences.

[0041] A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: when the computer program / instruction is executed by a processor, the steps of the above-mentioned social search hotspot propagation prediction method based on multiple feature sequences are implemented.

[0042] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of a method for predicting the propagation of social search hot spots based on multiple feature sequences.

[0043] The beneficial effects of the present invention are:

[0044] The present invention can effectively predict the spread of social search hot spots in social networks and can flexibly simulate the development trend of events by using data fitting. The present invention completes the representation process through multi-feature sequences and representation vectors, so that it can better capture the mutual influence relationship between events and complete the prediction of event sequences based on the intensity function of historical events. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of the event sequence consisting of tweets forwarded by different users;

[0046] Figure 2 A structural block diagram of the system described in the present invention;

[0047] Figure 3 Transformer model architecture diagram;

[0048] Figure 4 The structure diagram of the social search hotspot propagation prediction model combining multiple feature sequences;

[0049] Figure 5 Schematic diagram of the information cascade diffusion process;

[0050] Figure 6 Self-Correcting and Hawkes intensity function changes;

[0051] Figure 7 Comparison of log-likelihood values ​​between MFSPOP and NHP;

[0052] Figure 8 Comparison of prediction accuracy of event types;

[0053] Fig. 9 Comparison of log-likelihood values ​​between MFSPOP and MFSPOP+RNN. DETAILED DESCRIPTION

[0054] The present invention is further described below in conjunction with the accompanying drawings.

[0055] Figure 1 Schematic diagram of the event sequence of retweets by different users, including:

[0056] Basic information such as event flow, dissemination trend representation, social search hot spot forwarding increment, dissemination size and event sequence representation.

[0057] Figure 2 The structure block diagram of the system of the present invention is shown, comprising:

[0058] Event stream analysis device: captures the event stream occurring in social networks in continuous time, including timestamps, event types, and user information.

[0059] Event sequence coding device: completes the coding and synthesizes the time coding, event type coding, node information coding, user attribute coding and other information into event sequence coding.

[0060] Event representation learning device: captures and forms long-term and short-term dependency information between events, and completes the generation process of representation learning vectors of event sequences.

[0061] Social search hotspot prediction device: captures the mutual influence between historical events and predicts the occurrence of future events.

[0062] Figure 3 This is the architecture diagram of the Transformer model, including:

[0063] Multi-Head Attention, Residual Connection and Layer Normalization (Add&Norm) and Feed Forward Network. The most important part is the Multi-Head Attention, which is composed of multiple self-attention mechanisms. The self-attention mechanism is used to calculate the "degree of correlation" between input sequences. In essence, it encodes by associating input sequences with each other. Its core is three vectors: query vector, key vector and value vector, which calculate the attention score by inner product multiplication and weighted sum. Add in Add&Norm refers to a residual connection, where the input is residually connected to its output, specifically expressed as X+MultiHeadAttention(X). Norm refers to layer normalization, which converts the input of each layer into the same mean variance. Feed Forward Network is a two-layer fully connected layer that provides nonlinear transformation based on position information and feature information, enhancing the expressive power of the model. Muti-head attention, Add&Norm, and Feed Forward Network construct an Encoder block, and multiple Encoder blocks are superimposed to form an encoder.

[0064] Figure 4 The architecture of the social search hotspot propagation prediction model combining multiple feature sequences is given, specifically:

[0065] The first part is to perform representation learning on social search hot event sequence data based on the Transformer model. It can encode the input event sequence and capture the intrinsic relationship between the input sequences through the self-attention layer. The second part is to predict the occurrence of future events based on the Hawkes process. Hawkes can model social search hot events and predict the occurrence of future events by analyzing the propagation trends of past events.

[0066] Figure 5 A schematic diagram of the information cascade diffusion process is given, specifically:

[0067] In the real world, it can be found that the social behavior of users is easily influenced by their neighbor nodes. Therefore, when modeling event sequence data, the structural information of the network and the attribute information of users should also be taken into account. As shown in the figure, user A at t A A message is posted at a certain moment, and then user B sees it and forwards it. User C and user D forward the message forwarded by B, and user E forwards the message of user D. In this way, a tree structure is constructed by users A, B, C, D and E.

[0068] In order to verify the effectiveness of the present invention, synthetic data and real data are used for experiments, and each data set is divided into a training set, a validation set, and a test set. The validation set is used to adjust hyperparameters, and the test set is used to measure model performance. The data sets cover multiple fields such as social media, programming question-answering networks, medical care, and finance. The present invention is aimed at the study of social search hot spots on social networks, and four representative data sets (Retweets, MemeTrack, StackOverflow, and Electrical Medical Records) are used for experimental analysis. Using these data sets for event sequence modeling and prediction can better evaluate the performance and applicability of the model.

[0069] The present invention designs three experiments to verify the modeling effect and prediction ability of MFSPOP proposed in the present invention on event sequences, mainly including:

[0070] 1. Event sequence fitting experiment. (1) The real potential strength of the known synthetic dataset is compared with the strength predicted by the model to measure the modeling ability of MFSPOP for social search hot event data; (2) Among the four datasets, only Retweets includes user information and user attribute information of the event. Therefore, this experiment only compares the log-likelihood value of MFSPOP and NHP on this dataset. The larger the log-likelihood value, the better the data modeling ability of the model is and the better it can handle real-world event sequence data; (3) Experiments are conducted on the remaining three datasets, and compared with RMTPP, NHP, HP and SAHP to verify the effectiveness of MFSPOP in capturing the propagation trend of event sequences.

[0071] We extracted 100 continuous events from the data sets generated by the Hawkes process and the Self-Correcting process, and compared their true conditional strengths with the conditional strengths learned by the model to observe whether MFSPOP can adaptively capture the changes in different time series data. Figure 6 As shown in the figure, the dotted line represents the conditional intensity learned by the model, and the blue curve represents the true conditional intensity. The experimental results show that MFSPOP can accurately capture the general shape of the true intensity function and the changing trend of each point, so it can flexibly simulate the development trend of events and has a strong modeling ability for social search hot events.

[0072] The larger the log-likelihood value during the model training process, the better the data fitting effect. The Retweets dataset contains the time information, event type information, user information, and attribute information of the user. This experiment compares the log-likelihood values ​​of the results of running MFSPOP and NHP on the Retweets dataset. The horizontal axis in the figure represents the epoch change, from 1 to 50, and the vertical axis represents the log-likelihood value of each training. NE represents node encoding information, AE represents user attribute encoding information. The original NHP model does not encode the node information and user attribute information in the event sequence. NHP+NE means that the user information of the event is taken into account, and NHP+NE+AE means that the user information and user attribute information are taken into account. Figure 7 As shown in the figure, during the entire training process, MFSPOP outperforms NHP on this dataset. The reason is that the Transformer model used by MFSPOP is more capable of simulating the dynamics of information dissemination due to the complexity of social media data. Table 1 summarizes the log-likelihood value results of each model on the MemeTrack, StackOverflow, and MIMIC-II datasets. It can be seen that MFSPOP can simulate the dissemination trend of different sequence data very well and outperform other methods in all datasets.

[0073] Table 1

[0074]

[0075] 2. Event prediction experiment. The purpose is to verify the model's prediction effect on the time and type of event occurrence and the spread size. The evaluation indicators used are the prediction accuracy of event type, the root mean square error of event occurrence time, and the mean square logarithmic error and mean absolute percentage error of spread size.

[0076] Figure 8 Table 2 summarizes the accuracy (%) comparison of event type prediction on the datasets Retweets, MIMIC-II and StackOverFlow, as well as the RMSE comparison results of event time prediction between the RMTPP, NHP, SAHP models and the MFSPOP model proposed in this study.

[0077] Table 2

[0078]

[0079] like Figure 8As shown in the figure, the prediction accuracy of MFSPOP on the Retweets dataset reached 60.95%, which is 10.7% higher than that of NHP. The experimental results show that the Transformer model used in MFSPOP can better capture short-term and long-term dependencies than the LSTM used in NHP. In Table 2, the RMSE of MFSPOP on the dataset MIMIC-11 is more than doubled compared to RMTPP. The above results prove that MFSPOP has better event type and time prediction effects than other methods. In Table 3, the performance comparison of MFSPOP and NHP propagation size prediction is summarized. It can be observed that as the observation time increases, the model has more information and data to learn, and its prediction performance is better.

[0080] Table 3

[0081]

[0082] 3. Ablation experiment. (1) Check the model variants by using time encoding, user encoding and user attribute encoding to verify the importance of time features, node features and attribute features to modeling; (2) Verify the low sensitivity of the model to parameters by conducting experiments on different numbers of parameters. (3) Experimentally verify the data fitting effect and event type prediction ability of MFSPOP+RNN by adding a layer of RNN to the Transformer model.

[0083] Considering the variants of MFSPOP that are checked by time coding, node information coding, and attribute information coding, there are three types: without considering time coding (Atten), considering time coding (Atten+TE), and considering time coding and node information coding (Atten+TE+NE). MFSPOP considers all features. Table 4 shows that on the Retweets dataset, the log-likelihood value of the model MFSPOP after model training dropped from -1.94 to -6.46 after ignoring time coding, the log-likelihood value dropped to -3.47 after ignoring node coding, and the log-likelihood value dropped to -2.74 after ignoring attribute coding. It can be seen that the richer the information we consider, the better the model fits the event sequence and the better it can simulate the propagation trend of the event sequence.

[0084] Table 4

[0085]

[0086] The present invention also tests the sensitivity of the model to the number of parameters on the Retweets dataset. Table 5 shows that low sensitivity to the number of parameters means that relatively fewer parameters can be used to train the model. A small number of parameters means that the training time required is relatively short, so the training speed is faster.

[0087] Table 5

[0088]

[0089] Fig. 9 The comparison of the log-likelihood values ​​of MFSPOP and MFSPOP+RNN is described. The larger the log-likelihood value, the better the data fitting effect. As can be seen from the figure, in the initial stage of training, the data fitting effect of MFSPOP+RNN is better than that of MFSPOP, but as the number of training times increases, the log-likelihood value of MFSPOP is greater than that of MFSPOP+RNN, indicating that stacking more network structures in the neural network does not necessarily improve the effect of the model.

[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the spread of social search hot spots based on multiple feature sequences, characterized by: The following steps are involved: Step 1: Event flow analysis detection; Define an event stream, a sequence of discrete events that occur in continuous time, where each event is represented by a triple, including a timestamp, event type, and user information; Analyzing and modeling event sequences can predict the type and time of the next event; Step 2: Event sequence encoding; The input sequence is processed by a multi-head self-attention mechanism and a fully connected layer to generate an encoder output containing global semantic information. The decoder decodes the encoder output and previously generated symbols into a new output sequence and extracts the hidden state of each event, including the long-term and short-term dependencies of the events. The multi-layer mechanism is used to sequentially implement the correlation coding of the sequence and the weighted coding of the input sequence. Step 3: Event representation learning; After the initial encoding and multi-head attention mechanism, the result is input into the feedforward neural network to further generate the hidden layer representation of the input event sequence, completing the generation process of the representation learning vector of the event sequence; Step 4: Prediction of social search hot spots; The obtained historical event sequence is used as the input of this process to describe the event sequence that changes over time. The dynamic changes of the time point process are described by a continuous conditional intensity function to capture the mutual influence relationship between events and predict the occurrence of future events.

2. The method for predicting the spread of social search hot spots based on multiple feature sequences according to claim 1, characterized in that: The step 1 comprises the following steps: Step 1.1: Each event (t j ,k j ,u j ) is a triplet, indicating that user u j At time t j Retweeted k j Type of information; represents an event sequence, where k j represents the jth event in the sequence, t j Indicates the timestamp of the event; u j Indicates the user corresponding to the event; Step 1.2: In the time interval [t A ,t E ], there is a sequence of retweets (t A ,k1,u A ),(t B ,k1,u B ),(t F ,k2,u F ),(t C ,k1,u C ),…,(t H ,k2,u H ),(t E ,k1,u E ); Step 1.3: Given a social search hot event sequence Make short-term and long-term forecasts. Short-term forecasts refer to predicting when the next event will occur. With event type Long-term forecasting refers to the p -t i , predict the spread of social search hot events 3. The method for predicting the spread of social search hot spots based on multiple feature sequences according to claim 1, characterized in that: The step 2 comprises the following steps: Step 2.1: Assume that the social search hotspot propagation event sequence with a given event length L where t j Represents an event (t j ,k j ,u j ) occurs; given the following time code definitions: Step 2.2: Given For each event type, an embedding matrix U∈R is learned M*K , M represents the embedding dimension, K represents the number of event types, and the jth column of U represents event type k j M-dimensional embedding of; Step 2.3: After obtaining the event time code, event type code, node information code, and user attribute code, the event sequence The embedding of is X=(E+Z+H+G) T .

4. The method for predicting the spread of social search hot spots based on multiple feature sequences according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1: Use the self-attention module to obtain the long-term and short-term dependencies between sequences; calculate the attention output Step 3.2: After the social search hotspot propagation event sequence S passes through the initial encoding and multi-head self-attention mechanism, the attention output S out Generate a hidden representation of the input event sequence S through a feedforward neural network and h9t j )=H(j,:).

5. The method for predicting the spread of social search hot spots based on multiple feature sequences according to claim 1, characterized in that: The step 4 comprises the following steps: Step 4.1: For each k∈{1,2,…,K}, set λ k (t|H t ) is defined as the conditional intensity function of k-type events, and the conditional intensity function of the entire social search hotspot event sequence is: Step 4.2: The specific intensity of each event type is determined using: Step 4.3: The probability of the next social search hot event occurring at any time is expressed as: Step 4.4: The timestamp prediction of the next event and the event type prediction of the next event are expressed as: Step 4.5: Social search hot event sequence at t p -t i Estimated spread size during this time The prediction is expressed as:

6. A multi-constraint guided Chinese rumor adversarial sample generation system according to any one of claims 1-5, characterized in that: It includes an event flow analysis device, an event sequence encoding device, an event representation learning device and a social search hot spot prediction device; The event stream analysis device captures the event stream occurring in the social network in a continuous time, including timestamps, event types and user information; The event sequence coding device codes and combines the time code, event type code, node information code, and user attribute code information into an event sequence code; The event representation learning device captures and forms long-term and short-term dependency information between events, and completes the generation process of the representation learning vector of the event sequence; The social search hotspot prediction device captures the mutual influence relationship between historical events and predicts the occurrence of future events.

7. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

9. A computer program product comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images