A method for early warning of public opinion dissemination of digital content on-chain based on multimodal features

Abstract

The present invention discloses an on-chain digital content public opinion propagation early warning method based on multimodal features. The method starts with the calculation of the single-day public opinion propagation index and the generation of public opinion propagation early warning signals, and provides a data label production method. By utilizing multimodal time series data, through time series modeling, cross-modal attention, time series attention mechanism and feature fusion method in hyperbolic space, the influence of multimodal features on the on-chain digital content public opinion propagation is mined from the transaction mode, text mode and momentum mode, and the unique asset attributes of the on-chain digital content are fully considered to realize the early warning of the on-chain digital content public opinion propagation.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to an on-chain digital content public opinion propagation early warning method based on multimodal features. Background Art

[0002] Blockchain has the characteristics of decentralization, immutability, and anonymity. The more popular on-chain digital content of existing public blockchains usually exists in the form of contract code (such as ERC-721, ERC1155, etc.), and has the characteristics of both content attributes and financial attributes.

[0003] Users do not need to go through third-party review to publish digital content on the chain. Once the relevant topic spreads and ferments on the social platform, there may be a greater risk of public opinion, which will harm the network environment. This is called public opinion on the digital content on the chain. Therefore, the spread of relevant topics on social platforms will bring great challenges to blockchain supervision, and there is no public solution yet. In similar research fields, traditional public opinion analysis methods have been able to perform semantic understanding and analysis on the content of tweets on social platforms. For example, by learning the feature representation of sentences through deep learning methods, semantic features can be effectively extracted for downstream tasks, such as sentiment analysis and text classification; clustering methods based on topic models can filter out tweets containing risk information of digital content on the chain for subsequent analysis by regulators; however, the dimensions of traditional public opinion monitoring are mostly limited to the content of the tweets themselves, and lack attention to the time development dimension of public opinion dissemination.

[0004] In addition, the unique asset attributes of on-chain digital content have a significant impact on the dissemination of public opinion, which is also an influencing factor that needs to be considered. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides an on-chain digital content public opinion dissemination early warning method based on multimodal features. For the on-chain digital content to be regulated, the method takes its transaction mode, text mode, and momentum mode time series feature sequences as input and models them, and outputs whether there will be a public opinion dissemination early warning signal in the future.

[0006] The technical solution of the present invention is as follows: A method for early warning of public opinion dissemination of digital content on-chain based on multimodal features, comprising the following steps:

[0007] (1) The stage of defining public opinion propagation warning signals and producing data labels includes the following sub-steps:

[0008] (1.1) Collecting data sets: Using Twitter API, Weibo API or web page crawlers, collect tweet text data, tweet forwarding, commenting and liking data and transaction record data on the blockchain for topics related to digital content on the specified chain to obtain the required data sets; calculate the public opinion propagation degree I of all single tweets in the data setid ;

[0009] (1.2) Calculate the single - day public opinion dissemination index I of the digital content c on each chain on day d c，d : On the d - th day, the tweet set S of the digital content c on the chain c，d Among them, take the cardinality of the tweet set whose single - tweet public opinion dissemination degree I id is greater than the custom threshold N as the single - day public opinion dissemination index; the custom threshold N can be adjusted according to the actually monitored digital content c on the chain, and its expression is: I c，d = |{I id > N}|, id ∈ S c，d ;

[0010] (1.3) Generation of public opinion dissemination early - warning signals: For the single - day public opinion dissemination index sequence of the digital content c on the chain, first calculate the short - term moving average MA(N1) and the long - term moving average MA(N2) of this sequence, where N1 < N2. Then, take W (W >> N2) days as a cycle for sliding window, and perform outlier analysis on the public opinion dissemination index of this window through the Isolation Forest model to obtain the outlier set; Screen out the points in the outlier set whose single - day public opinion dissemination index is higher than MA(N2) as public opinion dissemination early - warning signals, that is, positive samples, and the rest are negative samples;

[0011] (1.4) Noise filtering and signal extension of public opinion dissemination early - warning signals: For the daily public opinion dissemination index sequence of the digital content c on the chain, mark the dates between two positive samples as positive samples, at most two; If the single - day public opinion dissemination index on the date after the positive sample is still higher than the moving average MA(N1), then mark it as a positive sample, at most three;

[0012] (2) Multi - modal data pre - processing stage, specifically including the following sub - steps:

[0013] (2.1) For the digital content c on the chain, set an observation window of D days for the early - warning model to predict whether a sudden public opinion dissemination early - warning signal will appear on day T; The early - warning model will start processing the multi - modal data of the previous D days at zero o'clock on day T;

[0014] (2.2) Construct the time - series features of the trading modality for the previous D days It includes the trading data of the digital content on the chain for a total of D days, and the daily trading data of the digital content on the chain includes 6 - dimensional features such as the earliest trading price, the latest trading price, the highest price, the lowest price, the buy trading volume, and the sell trading volume of the digital content c on the chain, forming a time - series representation

[0015] (2.3) Construct the time - series features of the text modality for the previous D days After the early warning model obtains a total of D days of tweets related to the digital content on the chain, it uses the emotional VAD dictionary to query the VAD score at the word level in each tweet, and adds the results of each word query in the three dimensions of alertness V, arousal A, and dominance D to obtain the sentence-level VAD score. At the same time, the length L of the tweet is recorded as the VAD-L feature of the tweet, a total of 4 dimensions; Calculate 3 aggregate representation data of all tweets on each day in the four dimensions of VAD-L features: average value, 25% quantile value, 75% quantile value, and finally splice to obtain the feature representation of text data Forming a time series representation

[0016] (2.4) Constructing the time series characteristics of the D-day before the momentum mode It contains the single-day public opinion dissemination index of the total digital content c on the chain on day D, recorded as Forming a time series representation

[0017]

[0018] (2.5) Performing data standardization processing on the extracted time series features of each modality in the time dimension;

[0019] (3) Time series data modeling stage, specifically: using gated recurrent units in hyperbolic space to perform time series encoding on the three modal data respectively, capturing the time series power law distribution and scale-free properties of the data;

[0020] (4) Cross-modal attention stage: Specifically, a cross-modal attention module CA is constructed to perform modal interaction calculation on the hidden layer outputs of two different modalities m1 and m2 for each on-chain digital content c to obtain the output

[0021]

[0022] (5) Temporal attention stage, which specifically includes the following sub-steps:

[0023] (5.1) When T day is taken as the warning target, the eigenvectors of mode m are concatenated according to the time dimension to obtain

[0024]

[0025] (5.2) Construct a temporal attention module TA, which assigns different weights to each on-chain digital content c according to the different temporal positions of each feature vector in the time series. With T day as the warning target, the modal features of the previous D day are input. Compute time series aggregate feature vector

[0026]

[0027] (6) Feature fusion warning stage, which specifically includes the following sub-steps:

[0028] (6.1) Select the interaction between momentum mode and trading mode to build the trading warning model E tx , output the prediction result

[0029] (6.2) Select the interaction between momentum mode and text mode to build the text warning model E text , output the prediction result

[0030] (6.3) Use the Focal loss function to train the text warning model E text and transaction warning model E tx , E text and E tx There is no shared relationship between the parameters of the two models, and it is determined whether there will be a public opinion propagation warning signal on day T;

[0031] (6.4) Use LightGBM learner to concatenate E text and E tx The knowledge representation output before entering Softmax is used to obtain the final judgment on whether there will be a public opinion propagation warning signal on day T.

[0032] Furthermore, the metadata of the public opinion dissemination degree id of a single tweet in the step (1.1) includes the number of reposts, the number of comments, and the number of likes; wherein the sum of the number of reposts, the number of comments, and the number of likes is equal to the public opinion dissemination degree of a single tweet.

[0033] Furthermore, the step (3.1) includes the following sub-steps, which are as follows:

[0034] (3.1.1) Using the Poincare disk model, the time series characteristics of each mode m with T day as the warning target are analyzed. To map the feature space, use Map the time series features from Euclidean space to hyperbolic space, where F is the mapping function;

[0035] (3.1.2) The gated recurrent unit in the hyperbolic space is represented as HGRU(·), for the hyperbolic space feature x at time t t ′, HGRU(·) is expressed as follows:

[0036]

[0037]

[0038]

[0039]

[0040] Among them, W, U, b are the learning weights, express Addition, express Multiplication;

[0041] (3.1.3) Calculate the temporal feature encoding under mode m:

[0042]

[0043] Furthermore, the construction of the attention module CA in step (4.1) specifically includes the following sub-steps:

[0044] (4.1.1) Obtaining the temporal feature encoding of two different modes m1 and m2 and Calculating cross-modal attention scores

[0045]

[0046] Among them, W 1 , W 2 is the learning weight;

[0047] (4.1.2) Calculating cross-modal attention weights

[0048]

[0049] (4.1.3) Calculate the eigenvector representation of two different modes m1 and m2 on day d

[0050]

[0051] Among them, modality m1 is processed by the cross-modal attention mechanism and integrates the information of modality m2; modality m2 remains unchanged.

[0052] Furthermore, the attention module TA construction of step (5.2) includes the following sub-steps:

[0053] (5.2.1) Calculate the daily attention score of modality m

[0054]

[0055] Among them, W last , W d ,u temporal is the learning weight;

[0056] (5.2.2) Calculate daily attention weight

[0057]

[0058] (5.2.3) Calculate the time series aggregate feature vector representation of mode m with T day as the warning target

[0059]

[0060] The beneficial effects of the present invention are as follows:

[0061] 1. Starting from the single-day public opinion dissemination index, anomalies are extracted from the time development dimension of public opinion dissemination based on the isolation forest algorithm and the moving average indicator. After signal denoising and extension, an on-chain digital content public opinion dissemination warning signal close to the real regulatory scenario is defined, providing a data labeling method for the training of the warning model.

[0062] 2. Adopting time series modeling, cross-modal attention, time series attention mechanism and feature fusion methods in hyperbolic space, we explored the impact of multimodal features on the public opinion dissemination of on-chain digital content from the transaction mode, text mode and momentum mode, and fully considered the unique asset attributes of on-chain digital content. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0064] Figure 1 This is a flow chart of the on-chain digital content public opinion propagation early warning method based on multimodal features of the present invention;

[0065] Figure 2 This is a module architecture diagram of the on-chain digital content public opinion dissemination early warning method based on multimodal features of the present invention. DETAILED DESCRIPTION

[0066] Here, exemplary embodiments are described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application.

[0067] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The singular forms of "a", "said" and "the" used in this application and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0068] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but the scope of the present invention is not limited in any way.

[0069] like Figure 1 As shown in the figure, the process of a method for early warning of public opinion propagation of digital content on the chain based on multimodal features includes six stages:

[0070] Phase 1: Definition of public opinion propagation warning signals and data labeling;

[0071] Phase 2: Multimodal data preprocessing phase;

[0072] Phase 3: Time series data modeling phase;

[0073] Stage 4: Cross-modal attention stage;

[0074] Stage 5: Temporal attention stage;

[0075] Stage 6: Feature fusion warning stage;

[0076] The first stage includes the following steps:

[0077] (1.1) Multimodal data collection of on-chain digital content: Use Twitter API, Weibo API or web page crawlers to collect tweet text data, tweet forwarding, commenting and liking data and blockchain transaction record data on topics related to designated on-chain digital content in recent years to obtain the required data set.

[0078] (1.2) Calculate the public opinion spread degree I of all single tweets in the data set id : The metadata of a single tweet id includes the number of reposts, comments, and likes. The sum of the number of reposts, comments, and likes is equal to the public opinion dissemination degree of a single tweet.

[0079] (1.3) Calculate the public opinion dissemination index I of each on-chain digital content c on day d c，d : On day d, the set of tweets S of the on-chain digital content c c，d In the above example, the number of tweets with a public opinion heat index greater than N will be used as the single-day public opinion dissemination index (N can be adjusted according to the actual monitored on-chain digital content c), and its calculation formula is L c，d =|{Iid >N}|, id ∈ S c，d 。

[0080] (1.4) Generation of public opinion dissemination early warning signals: For the daily public opinion dissemination index sequence of the digital content c on the chain, first calculate the short-term moving average MA(N1) and the long-term moving average MA(N2) of this sequence, where N1 < N2. Then, take W (W >> N2) days as a cycle for sliding window, and perform outlier analysis on the public opinion dissemination index of this window through the Isolation Forest model to obtain the outlier set. Further, screen out the points with a daily public opinion dissemination index higher than MA(N2) in the outlier set as public opinion dissemination early warning signals, that is, positive samples, and the rest are negative samples.

[0081] (1.5) Noise filtering and signal extension of public opinion dissemination early warning signals: For the daily public opinion dissemination index sequence of the digital content c on the chain, mark the dates existing between two positive samples as positive samples, at most two. If the daily public opinion dissemination index on the date after the positive sample is still higher than the moving average MA(N1), then mark it as a positive sample, at most three.

[0082] The second stage includes the following steps:

[0083] (2.1) For the digital content c on the chain, set an observation window of the early warning model for a total of D days to predict whether a sudden public opinion dissemination early warning signal will appear on the Tth day. The early warning model will start processing the multi-modal data of the previous D days at zero o'clock on the Tth day.

[0084] (2.2) Construct the time series features of the trading modality for the previous D days It contains a total of D days of on-chain digital content trading data, and the on-chain digital content trading data for each day includes 6-dimensional features such as the earliest trading price, the last trading price, the highest price, the lowest price, the buy trading volume, and the sell trading volume of the on-chain digital content c,

[0085] (2.3) Construct the time series features of the text modality for the previous D days After the early warning model obtains a total of D days of tweets related to the on-chain digital content, with the help of the VAD (Valence - Arousal - Dominance) dictionary, calculate the VAD scores of the words in each tweet, and add up the results in three dimensions of Valence (alertness), Arousal (arousal), and Dominance (degree of being dominated). At the same time, record the length L of the tweet, denoted as the VAD - L feature of this tweet. For the tweet text data of each day, calculate the average value, 25% quantile value, and 75% quantile value of the VAD - L feature of the tweets in each dimension, and finally splice them to obtain The VAD-L feature is a feature for each tweet, with a total of four dimensions: V, A, D, and L. Since it is necessary to form a daily representation, an aggregate representation (average, 25% quantile, 75% quantile) is required, and the final result is 12 dimensions.

[0086] (2.4) Constructing the time series characteristics of the D-day before the momentum mode It contains the single-day public opinion dissemination index of the total digital content c on the chain on day D, recorded as

[0087] (2.5) Perform data normalization processing on the time dimension for the extracted time series features of each modality.

[0088] The third stage includes the following steps:

[0089] (3.1) The gated recurrent unit in hyperbolic space is used to perform temporal encoding on the three modal data respectively, capturing the temporal power-law distribution and scale-free properties of the data.

[0090] The step (3.1) includes the following sub-steps, which are as follows:

[0091] (3.1.1) Using the Poincaré ball model, we can calculate the time series characteristics of each mode m with T day as the warning target. To map the feature space, use Map the time series features from Euclidean space to hyperbolic space, where F is the mapping function.

[0092] (3.1.2) The Hyperbolic Gated Recurrent Unit is denoted as HGRU(·). For the hyperbolic space feature x at time t, t ′, HGRU(·) is defined as follows:

[0093]

[0094]

[0095]

[0096]

[0097] Among them, W, U, b are the learning weights, express Addition, express Multiplication.

[0098] (3.1.3) Calculate the temporal feature encoding under mode m, where m is transaction mode, text mode, or momentum mode:

[0099]

[0100] The fourth stage includes the following steps:

[0101] (4.1) Construct a cross-modal attention module CA, and perform modal interaction calculation on each on-chain digital content c through the hidden layer outputs of two different modalities m1 and m2 to obtain the output like Figure 2 As shown, in the present invention, the momentum mode corresponds to the m2 mode, and the text mode or transaction mode corresponds to the m1 mode:

[0102]

[0103] Furthermore, the construction of the attention module CA in step (4.1) includes the following sub-steps:

[0104] (4.1.1) Obtaining the temporal feature encoding of two different modes m1 and m2 and Calculating cross-modal attention scores

[0105]

[0106] Among them, W 1 , W 2 is the learning weight.

[0107] (4.1.2) Calculating cross-modal attention weights

[0108]

[0109] (4.1.3) Calculate the eigenvector representation of two different modes m1 and m2 on day d

[0110]

[0111] Among them, modality m1 is processed by the cross-modal attention mechanism and incorporates the information of modality m2. Modality m2 remains unchanged.

[0112] The fifth stage includes the following steps:

[0113] (5.1) When T day is taken as the warning target, the eigenvectors of mode m are concatenated according to the time dimension to obtain

[0114]

[0115] (5.2) Construct a temporal attention module TA. For each digital content c on the chain, different weights are assigned according to the different temporal positions of each feature vector in the time series. With T day as the warning target, the modal features of the previous D day are input. Compute time series aggregate feature vector

[0116]

[0117] Furthermore, the construction of the attention module TA in step (5.2) includes the following sub-steps:

[0118] (5.2.1) Calculate the daily attention score of modality m

[0119]

[0120] Among them, W last , W d ,u tempral are the weights that can be learned.

[0121] (5.2.2) Calculate daily attention weight

[0122]

[0123] (5.2.3) Calculate the time series aggregate feature vector representation of mode m with T day as the warning target

[0124]

[0125] The stage six includes the following steps:

[0126] (6.1) Select the interaction between momentum mode and trading mode to build the trading warning model E tx .like Figure 2 As shown in the transaction warning model, in the cross-modal attention module, the momentum mode corresponds to the m2 mode, and the transaction mode corresponds to the m1 mode. It is the knowledge representation of the transaction warning model. Transaction warning model E tx The output prediction result is expressed as

[0127] (6.2) Select the interaction between momentum mode and text mode to build the text warning model E text .like Figure 2 As shown in the text warning model, in the cross-modal attention module, the momentum mode corresponds to the m2 mode, and the text mode corresponds to the m1 mode. is the knowledge representation of the text warning model. Text warning model E text The output prediction result is expressed as

[0128] (6.3) Use the Focal loss function to train E text and E tx Model, E text and E tx There is no shared relationship among the parameters of the model, which determines whether there will be a public opinion propagation warning signal on day T.

[0129] (6.4) Use LightGBM learner to concatenate E text and E tx The knowledge representation output by the model before entering Softmax is used to obtain the final judgment on whether there will be a public opinion propagation warning signal on day T. Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the contents disclosed herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common knowledge or customary technical means in the art that are not disclosed in the present application.

[0130] It should be understood that the present application is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

Claims

1. A method for early warning of public opinion dissemination of digital content on-chain based on multimodal features, characterized in that: Including the following steps: (1) The stage of defining early warning signals for public opinion dissemination and making data tags, specifically including the following sub-steps: (1.1) Collecting the dataset: Using the Twitter API, Weibo API, or web page crawler to collect tweet text data, tweet repost, comment, and like data, and blockchain transaction record data related to the specified on-chain digital content topics to obtain the required dataset; Calculate the public opinion spread degree I of all single tweets in the data set id ; (1.2) Calculate the public opinion dissemination index I of each on-chain digital content c on day d c，d : On day d, the set of tweets S of the on-chain digital content c c，d In the example, the public opinion propagation degree of a single tweet with a custom threshold I id The cardinality of the set of tweets greater than the custom threshold N is used as the single-day public opinion dissemination index; the custom threshold N can be adjusted according to the actual monitored on-chain digital content c, and its expression is: c, d=|{I id >N}, id∈S c，d ; (1.3) Generating early warning signals for public opinion dissemination: For the single-day public opinion dissemination index sequence of on-chain digital content c, first calculate the short-term moving average MA-N1 and the long-term moving average MA-N2 of this sequence, where N1 < N2. Then, use a sliding window with a period of W days, where W >> N2. Conduct outlier analysis on the public opinion dissemination index of this window through the Isolation Forest model to obtain the outlier set. Screen the points in the outlier set with a single-day public opinion dissemination index higher than MA-N2 as early warning signals for public opinion dissemination, that is, positive samples, and the rest are negative samples; (1.4) Noise filtering and signal extension of early warning signals for public opinion dissemination: For the daily public opinion dissemination index sequence of on-chain digital content c, mark the dates between two positive samples as positive samples, at most two; if the single-day public opinion dissemination index on the dates after the positive sample is still higher than the MA-N1 moving average, mark it as a positive sample, at most three; (2) The stage of multi-modal data preprocessing, specifically including the following sub-steps: (2.1) For on-chain digital content c, set the observation window of the early warning model to D days in total, and predict whether an early warning signal for sudden public opinion dissemination will appear on the Tth day; the early warning model will start processing the multi-modal data of the previous D days at zero o'clock on the Tth day; (2.2) Constructing the time series characteristics of the D-day before the transaction mode It contains the total on-chain digital content transaction data for D days and the daily on-chain digital content transaction data. Contains 6-dimensional features of the earliest transaction price, latest transaction price, highest price, lowest price, buy transaction volume, and sell transaction volume of the digital content c on the chain, forming a time series representation (2.3) Constructing the time series features of the D-day before the text modality After the early warning model obtains a total of D days of tweets related to the digital content on the chain, it uses the emotional VAD dictionary to query the VAD score at the word level in each tweet, and adds the results of each word query in the three dimensions of alertness V, arousal A, and dominance D to obtain the sentence-level VAD score. At the same time, the length L of the tweet is recorded as the VAD-L feature of the tweet, a total of 4 dimensions; Calculate 3 aggregate representation data of all tweets on each day in the four dimensions of VAD-L features: average value, 25% quantile value, 75% quantile value, and finally splice to obtain the feature representation of text data Forming a time series representation (2.4) Constructing the time series characteristics of the D-day before the momentum mode It contains the single-day public opinion dissemination index of the total digital content c on the chain on day D, recorded as Forming a time series representation (2.5) Perform data standardization processing on the extracted time series features of each modality in the time dimension; (3) The stage of time series data modeling, specifically: Use the gated recurrent unit in the hyperbolic space to perform time series encoding on the three modalities of data respectively to capture the time series power-law distribution and scale-free properties of the data; (4) Cross-modal attention stage: Specifically, a cross-modal attention module CA is constructed to perform modal interaction calculation on the hidden layer outputs of two different modalities m1 and m2 for each on-chain digital content c to obtain the output (5) The stage of time series attention, specifically including the following sub-steps: (5.1) When T day is taken as the warning target, the eigenvectors of mode m are concatenated according to the time dimension to obtain (5.2) Construct a temporal attention module TA. For each digital content c on the chain, different weights are assigned according to the different temporal positions of each feature vector in the time series. With T day as the warning target, the modal features of the previous D day are input. Compute time series aggregate feature vector (6) The stage of feature fusion and early warning, specifically including the following sub-steps: (6.1) Select the interaction between momentum mode and trading mode to build the trading warning model E tx , output the prediction result (6.2) Select the interaction between momentum mode and text mode to build the text warning model E text , output the prediction result (6.3) Use the Focal loss function to train the text warning model E text and transaction warning model E tx , E text and E tx There is no shared relationship between the parameters of the two models, and it is determined whether there will be a public opinion propagation warning signal on day T; (6.4) Use LightGBM learner to concatenate E text and E tx The knowledge representation output before entering Softmax is used to obtain the final judgment on whether there will be a public opinion propagation warning signal on day T.

2. According to claim 1, a method for early warning of public opinion dissemination of digital content on the chain based on multimodal features is characterized in that: The metadata of the public opinion dissemination degree id of a single tweet in step (1.1) includes the number of reposts, comments, and likes; the sum of the number of reposts, comments, and likes is equal to the public opinion dissemination degree of a single tweet.

3. According to claim 1, a method for early warning of public opinion dissemination of digital content on the chain based on multimodal features is characterized in that: Step (3.1) includes the following sub-steps, specifically as follows: (3.1.1) Using the Poincare disk model, the time series characteristics of each mode m with T day as the warning target are analyzed. To map the feature space, use Map time series features from Euclidean space to hyperbolic space; (3.1.2) The gated recurrent unit in the hyperbolic space is represented as HGRU(·), for the hyperbolic space feature x at time t t ′, HGRU(·) is expressed as follows: Among them, W, U, b are the learning weights, express Addition, express Multiplication; (3.1.3) Calculate the temporal feature encoding under mode m:

4. According to claim 1, a method for early warning of public opinion dissemination of digital content on the chain based on multimodal features is characterized in that: The construction of the attention module CA in step (4.1) specifically includes the following sub-steps: (4.1.1) Obtaining the temporal feature encoding of two different modes m1 and m2 and Calculating cross-modal attention scores Among them, W1 and W2 are learned weights; (4.1.2) Calculating cross-modal attention weights (4.1.3) Calculate the eigenvector representation of two different modes m1 and m2 on day d Among them, modality m1 is processed through the cross-modal attention mechanism and integrates the information of modality m2; modality m2 remains unchanged.

5. According to claim 1, a method for early warning of public opinion dissemination of digital content on the chain based on multimodal features is characterized in that: The construction of the attention module TA in step (5.2) includes the following sub-steps: (5.2.1) Calculate the daily attention score of modality m Among them, W last , W d ,u temporal is the learning weight; (5.2.2) Calculate daily attention weight (5.2.3) Calculate the time series aggregate feature vector representation of mode m with T day as the warning target

Images