A data enhancement method for event information deduction

Abstract

The present invention relates to a data enhancement method for event information deduction, which belongs to the field of artificial intelligence. In order to solve the problem that available training data is scarce and difficult to obtain, the present invention uses an event graph to fuse the collected multi-source heterogeneous data, uses the multimodal embedding representation of the fusion network to uniformly characterize and describe the multi-source heterogeneous data, obtains the fused time series data, and then uses a generative adversarial network for training, so that the generation network G generates a latent vector Z based on random noise, and the recovery network R converts the latent vector Z back into new time series data. The present invention uses a generative adversarial network to expand the original data set to generate more diverse training samples, thereby improving the learning effect and generalization ability of the event deduction model, and effectively improving the ability to deduce small sample event information.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and in particular relates to a data enhancement method for event information deduction. Background Art

[0002] In intelligence processing, the demand for dynamic information analysis and processing of document and telegraphic data continues to increase. However, due to the special characteristics of certain fields, including sensitivity and confidentiality, available training data is scarce and difficult to obtain. This greatly limits the application of machine learning algorithms, especially deep learning models, in intelligence analysis. To address this challenge, this paper proposes a data enhancement method for event information deduction and conducts research on generative adversarial networks for time series data to achieve data enhancement, thereby effectively solving the small sample size problem and supporting the construction of more accurate and efficient information deduction and decision support systems. Summary of the Invention

[0003] (1) Technical issues to be resolved

[0004] The technical problem to be solved by the present invention is how to provide a data enhancement method for event information deduction to solve the problem that available training data is scarce and difficult to obtain.

[0005] (2) Technical solution

[0006] In order to solve the above technical problems, the present invention proposes a data enhancement method for event information deduction, which includes the following steps:

[0007] Step 1: Data Collection

[0008] Collect multi-dimensional intelligence data from different data sources, perform preliminary processing on the intelligence data to obtain multi-source heterogeneous data, and load it into a unified multimodal database;

[0009] Step 2: Data processing

[0010] The event graph is used to fuse the collected multi-source heterogeneous data. The multi-modal embedding representation of the fusion network is used to uniformly represent the multi-source heterogeneous data to obtain the fused time series data, which is then standardized and normalized.

[0011] Step 3: Model training

[0012] The generative adversarial network includes: embedding network E, recovery network R, generator network G, discriminator network D and supervisory network S. By training the generative adversarial network, the generator network G generates a latent vector Z based on random noise, and the recovery network R converts the latent vector Z back into new time series data.

[0013] Step 4: Data augmentation

[0014] After the model training is completed, random noise is provided to the generator network G to generate the latent vector Z, and the recovery network R is used to convert the latent vector Z back to the new time series data.

[0015] (3) Beneficial effects

[0016] The present invention proposes a data enhancement method for event information deduction. The present invention discloses a data enhancement method for event information deduction, which has the following main advantages:

[0017] (1) The present invention uses event graphs to fuse the collected multi-source heterogeneous data, and uses the multimodal embedding representation of the fusion network to uniformly characterize and describe the multi-source heterogeneous data, thereby achieving the unification of multi-source heterogeneous data;

[0018] (2) In order to accurately capture the temporal dependencies of events in intelligence, the LSTM structure is introduced into the traditional generative network to learn and reproduce the complex dynamic patterns in time series data, ensuring the coherence and logical consistency of the generated data points on the time axis.

[0019] (3) In order to extract multi-level intelligence information, a multi-scale feature extraction mechanism is adopted in the discriminant network to adapt to the data requirements of different granularities, so that the discriminant network can not only evaluate the authenticity of the overall sequence, but also perform fine-grained discrimination on local features to improve the discrimination accuracy.

[0020] (4) A new strategy combining adversarial loss and supervisory guidance is proposed, which enables the generative network to predict possible future development paths while imitating history, thereby enhancing the model's ability to deduce future events.

[0021] (5) The present invention uses a generative adversarial network to expand the original dataset to generate more diverse training samples, thereby improving the learning effect and generalization ability of the event deduction model, and effectively improving the ability to deduce small sample event information. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0023] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below with reference to the accompanying drawings and examples.

[0024] This invention aims to address the existing problem of insufficient event deduction accuracy due to a lack of sufficient training samples. Specifically, to address the small sample size issue in event deduction scenarios, this invention uses a generative adversarial network to expand the original dataset to generate more diverse training samples, thereby improving the learning effect and generalization ability of the event deduction model.

[0025] The present invention proposes a data enhancement method for event information deduction, which is used to solve the task of small-sample event deduction and effectively improve the ability of small-sample event information deduction.

[0026] The present invention provides a data enhancement method for event information deduction. The overall process is as follows: Figure 1 As shown, in general, it includes four steps, namely data collection, data processing, model training and data augmentation.

[0027] Collect multi-dimensional intelligence data from different data sources, perform preliminary processing on the intelligence data to obtain multi-source heterogeneous data, and load it into a unified multimodal database.

[0028] Specifically, first, by integrating data sources (i.e., official document systems, open source data, and other data sources) and introducing high-concurrency real-time data streams, multi-dimensional intelligence is captured, and a temporary storage buffer is set up for the preliminary storage and preprocessing of large amounts of received raw data. The data in the buffer is preliminarily filtered and the format is standardized. ETL tools are used to extract and transform data from different sources to obtain multi-source heterogeneous data, and then load it into a unified multimodal database.

[0029] The event graph is used to fuse the collected multi-source heterogeneous data, and the multi-modal embedding representation of the fusion network is used to uniformly characterize the multi-source heterogeneous data to obtain the fused time series data, and then the data is standardized and normalized.

[0030] S21. First, construct an event graph. The event nodes in the event graph are composed of multi-source heterogeneous data (images, text, etc.). That is, the event graph is a heterogeneous network constructed from multimodal data. The DeepWalk algorithm is used to learn the embedding of the event nodes in the event graph (each node represents an event). A truncated random walk algorithm is used to generate a sequence of event nodes (such as [A, B, C]). The co-occurrence relationship of the event nodes in the sequence is used to represent their local neighborhood structure (for example, node B is a neighbor of A, and C is a neighbor of B). The Word2Vec word embedding model is then applied to the resulting sequence to obtain a network representation of the event nodes.

[0031] For example, suppose there is an action-related event, the event node includes:

[0032] 1. Intelligence collection (data types: text, images): For example, obtaining deployment information through satellite imagery and intelligence reports.

[0033] 2. Action decision (data type: timestamp): For example, the time point when deciding whether to initiate an action.

[0034] 3. Invested resources (Data type: structured data): For example, the number of personnel and equipment invested.

[0035] Random walk sequence: [Intelligence gathering, resource investment, action decision], nodes randomly walk to form a sequence. Event nodes form an event graph, which is defined based on all event nodes. Multimodal data is equivalent to the values ​​in the event nodes.

[0036] S22. For a heterogeneous network constructed for multimodal data, an embedding representation is first constructed for each modality; text data can be output as a vector representation using the TF-IDF method, and image data can be processed by a convolutional network and then obtained as a vector representation through a fully connected layer. The embedding representations of different modalities are then mapped to an embedding space of the same dimension. Each iteration maximizes the similarity between the embedding representations of linked nodes and minimizes the similarity of unlinked nodes. By continuously adjusting the parameters in the vector representation generation process, a multimodal embedding representation of the fusion network is finally obtained.

[0037] The fusion network comes after the heterogeneous network. It refers to the multimodal embedding representation generated by the heterogeneous network (text data outputs vector representation through the TF-IDF method, and image data is processed by the convolutional network and then obtained by the fully connected layer). After that, it passes through the fusion network to map the embedding representations of different modalities to the embedding space of the same dimension, and finally obtain the multimodal embedding representation of the fusion network.

[0038] S23, the multimodal embedding representation of the fusion network is arranged in chronological order, and the fused time series data X={x1,x2,...,x T}, where each x t is the multimodal embedding representation of the fusion network at time t. These data are standardized and normalized so that all features have the same numerical range [-1, 1]. This helps the model learn the relative relationship between data and generate standardized time series data.

[0039] The generative adversarial network includes: embedding network E, recovery network R, generator network G, discriminator network D and supervisory network S. By training the generative adversarial network, the generator network G generates a latent vector Z based on random noise, and the recovery network R converts the latent vector Z back into new time series data.

[0040] S31. First, perform the pre-training phase:

[0041] S311, pre-training embedding network E and recovery network R. Use the processed standardized time series data X as the data set to train the embedding network E and the recovery network R, so that the recovery network R can accurately extract the latent vector output by the embedding network E. Reconstruct the original time series The goal of this stage is to minimize the reconstruction loss .

[0042] S312, pre-training generation network G and supervision network S. Random noise z attempts to generate hidden vectors through the generation network G , and the supervisory network S strives to predict whether these latent vectors come from the embedding network E. The discriminant network D is not involved at this time. The goal is to let the generating network G learn to generate latent vectors close to the output of the embedding network E. The loss function of the pre-training is .

[0043] The LSTM structure is introduced into the generative network G to learn and reproduce the complex dynamic patterns in time series data, ensuring the coherence and logical consistency of the generated data points on the time axis.

[0044] S32. Secondly, conduct adversarial training:

[0045] S321. Update the generative network G and the supervisory network S. Fix all other components and update only the generative network G and the supervisory network S. This allows G to generate more realistic latent vectors, and S to better predict the authenticity of these latent vectors. During this process, G receives not only feedback from the discriminative network D, namely the adversarial loss, but also additional guidance from S, namely the supervisory loss. The input of the discriminative network D is the latent vector from the embedding network E or the generative network G (either a real sample from E or a sample generated by G), and the output is the probability of the sample being authentic. The overall loss function is:

[0046]

[0047] Where γ is a hyperparameter used to adjust the adversarial loss and monitoring losses The balance between adversarial loss and supervisory loss is as follows:

[0048]

[0049] in, is a random noise distribution, D is the discriminant network, represents the expectation, and G(z) represents the time series data generated by the generator network G based on the input noise z. is the latent vector of the real time series data after being transformed by the embedding network, and S is the supervision network.

[0050] S322: Update the discriminant network D. Fix the generative network G and other components and update only the discriminant network D to enable it to more effectively distinguish true and false time series. This step is achieved by minimizing the discriminant loss.

[0051] A multi-scale feature extraction mechanism is adopted in the discriminant network D to adapt to data requirements of different granularities, so that the discriminant network D can not only evaluate the authenticity of the overall sequence, but also perform fine-grained discrimination of local features to improve discrimination accuracy.

[0052] The discriminant loss of the discriminant network D typically aims to maximize the accuracy of judging real and generated samples (a smaller loss indicates a better training of the discriminant network, resulting in higher accuracy). The adversarial loss of the generative network G, on the other hand, aims to trick the discriminant network into believing the generated samples are real. Therefore, in step S322, the discriminant loss refers to the loss function of the discriminant network, while the adversarial loss is part of the loss of the generative network. Together, they constitute the adversarial training process.

[0053] S323, update the embedding network E and the recovery network R. Adjust the embedding network E and the recovery network R again. By using both the true latent vector E(X) and the generated latent vector G(z), the recovery network R is forced to maintain high-precision reconstruction capabilities when facing both the true and generated latent vectors. The loss function at this stage is:

[0054]

[0055] Where β is a hyperparameter used to adjust the balance between reconstruction loss and supervision loss. The reconstruction loss and supervision loss are as follows:

[0056]

[0057] in, Denotes E(X) and G(z), where E(X) is the latent vector of the real time series data after being transformed by the embedding network, and G(z) represents the latent vector generated by the generator network based on the input noise z.

[0058] S33, then enter the alternating optimization stage:

[0059] In each iteration, the parameters of each network are updated in the order of S321-S323 to ensure that each network is fully trained without overfitting. This alternating update strategy helps stabilize the entire training process. According to the experiment, the model's learning rate, batch size and other hyperparameters are adjusted to obtain better generation results. When the model reaches the predetermined stopping condition (such as the maximum number of iterations T max, the discriminant network cannot distinguish the authenticity of the generated data), the training process ends.

[0060] After the model training is completed, random noise is provided to the generator network G to generate the latent vector Z. The recovery network R is used to convert the latent vector Z back into new time series data. These newly generated time series data should be as close as possible to the actual observed data distribution, but not simply copied and pasted. The specific operations are as follows:

[0061] S41. Prepare random noise z: Draw a set of random numbers from a predefined noise distribution (such as standard normal distribution or uniform distribution) as input.

[0062] S42. Generate latent vector: Input random noise z into the generative network G to obtain latent vector Z=G(z).

[0063] S43. Convert to time series: Use the recovery network R to convert the latent vector Z back to the time series form X ^ =R(Z). At this time, X ^ This is the newly generated time series data. To ensure the quality of the generated data, multiple sets of samples are typically generated and those closest to the true distribution are selected. Furthermore, evaluation metrics such as KL divergence and Wasserstein distance can be used to quantify the similarity between the generated data and the true data.

[0064] S44. Post-processing: The generated time series data is often standardized or normalized, so destandardization / denormalization is required to restore its original numerical range.

[0065] The augmented data generated after post-processing can be used as input for subsequent event information deduction models, thereby improving the training quality of the event information deduction models.

[0066] The present invention discloses a data enhancement method for event information deduction, which has the following main advantages:

[0067] (1) The present invention uses event graphs to fuse the collected multi-source heterogeneous data, and uses the multimodal embedding representation of the fusion network to uniformly characterize and describe the multi-source heterogeneous data, thereby achieving the unification of multi-source heterogeneous data;

[0068] (2) In order to accurately capture the temporal dependencies of events in intelligence, the LSTM structure is introduced into the traditional generative network to learn and reproduce the complex dynamic patterns in time series data, ensuring the coherence and logical consistency of the generated data points on the time axis.

[0069] (3) In order to extract multi-level intelligence information, a multi-scale feature extraction mechanism is adopted in the discriminant network to adapt to the data requirements of different granularities, so that the discriminant network can not only evaluate the authenticity of the overall sequence, but also perform fine-grained discrimination on local features to improve the discrimination accuracy.

[0070] (4) A new strategy combining adversarial loss and supervisory guidance is proposed, which enables the generative network to predict possible future development paths while imitating history, thereby enhancing the model's ability to deduce future events.

[0071] (5) The present invention uses a generative adversarial network to expand the original dataset to generate more diverse training samples, thereby improving the learning effect and generalization ability of the event deduction model, and effectively improving the ability to deduce small sample event information.

[0072] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A data enhancement method for event information deduction, characterized in that: The method comprises the following steps: Step 1: Data Collection Collect multi-dimensional intelligence data from different data sources, perform preliminary processing on the intelligence data to obtain multi-source heterogeneous data, and load it into a unified multimodal database; the multi-source heterogeneous data includes: image data and text data; Step 2: Data processing The event graph is used to fuse the collected multi-source heterogeneous data. The multi-modal embedding representation of the fusion network is used to uniformly represent the multi-source heterogeneous data to obtain the fused time series data, which is then standardized and normalized. Step 3: Model training The generative adversarial network includes: embedding network E, recovery network R, generator network G, discriminator network D and supervisory network S. By training the generative adversarial network, the generator network G generates a latent vector Z based on random noise, and the recovery network R converts the latent vector Z back into new time series data. Step 4: Data augmentation After the model training is completed, random noise is provided to the generator network G to generate the latent vector Z, and the recovery network R is used to convert the latent vector Z back to the new time series data; in, The second step includes: S21. First, an event graph is constructed. The event nodes in the event graph are composed of multi-source heterogeneous data. The DeepWalk algorithm is used to embed the event nodes in the event graph, and a truncated random walk algorithm is used to generate an event node sequence. The co-occurrence relationship of the event nodes in the sequence is used to represent its local neighborhood structure. The Word2Vec word embedding model is then applied to the obtained sequence to obtain a network representation of the event nodes. S22. For heterogeneous networks built for multimodal data, an embedding representation is first constructed for each modality. These embedding representations of different modalities are then mapped to an embedding space of the same dimension. Each iteration maximizes the similarity between the embedding representations of linked nodes and minimizes the similarity between unlinked nodes. By continuously adjusting the parameters in the vector representation generation process, a multimodal embedding representation of the fused network is ultimately obtained. S23. The multimodal embedding representations of the fused network are arranged in chronological order to obtain the fused time series data X={x1,x2,...,xt,...,xT}, where each xt is the multimodal embedding representation of the fused network at time t; these data are standardized and normalized so that all features have the same numerical range [-1,1] to generate standardized time series data.

2. The data enhancement method for event information deduction according to claim 1, characterized in that: The first step includes: integrating data sources and introducing high-concurrency real-time data streams to capture multi-dimensional intelligence, establishing a temporary storage buffer for preliminary storage and preprocessing of large amounts of received raw data, performing preliminary filtering and format standardization on the data in the buffer, and using ETL tools to extract and transform data from different sources to obtain multi-source heterogeneous data, and loading the data into a unified multimodal database.

3. The data enhancement method for event information deduction according to claim 1, characterized in that: Text data is represented by vectors using the TF-IDF method.

4. The data enhancement method for event information deduction according to claim 1, characterized in that: The image data is processed by the convolutional network and then passed through the fully connected layer to obtain a vector representation.

5. The data enhancement method for event information deduction according to claim 1, characterized in that: The step three includes: S31. First, perform the pre-training phase: S311, pre-training embedding network E and recovery network R; using the processed standardized time series data X as the data set to train the embedding network E and the recovery network R, so that the recovery network R can accurately extract the latent vector output by the embedding network E Reconstruct the original time series ; The goal of this stage is to minimize the reconstruction loss ; S312, pre-training generation network G and supervision network S; random noise z attempts to generate latent vectors through the generation network G , and the supervisory network S strives to predict whether these latent vectors come from the embedding network E; the discriminant network D is not involved at this time, and the purpose is to let the generating network G learn to generate latent vectors close to the output of the embedding network E; the loss function of the pre-training is ; S32. Secondly, conduct adversarial training: S321, update the generative network G and the supervisory network S; fix the other parts and only update the generative network G and the supervisory network S, so that G can generate more realistic latent vectors and S can better predict the authenticity of these latent vectors; in this process, G not only receives feedback from the discriminative network D, that is, adversarial loss, but also receives additional guidance from S, that is, supervisory loss; the input of the discriminative network D is the latent vector from the embedding network E or the generative network G, and the output is the probability of sample authenticity. The total loss function is: Where γ is a hyperparameter used to adjust the adversarial loss and monitoring losses the balance between; S322, update the discriminant network D; fix the generative network G and other components, and only update the discriminant network D so that it can more effectively distinguish true and false time series; this step is achieved by minimizing the discriminant loss; S323, update the embedding network E and the recovery network R; adjust the embedding network E and the recovery network R again, and force the recovery network R to maintain high-precision reconstruction capability when facing the real or generated latent vector by using the real latent vector E(X) and the generated latent vector G(z) at the same time; the loss function of this stage is: Where β is a hyperparameter used to adjust the balance between reconstruction loss and supervision loss. The reconstruction loss and supervision loss are as follows: in, Represents E(X) or G(z), where E(X) is the latent vector of the real time series data after being transformed by the embedding network, and G(z) represents the latent vector generated by the generator network based on the input noise z; S33, then enter the alternating optimization stage: In each iteration, the parameters of each network are updated in sequence according to the above S321-S323 to ensure that each network can be fully trained without overfitting; when the model reaches the predetermined stopping condition, the training process ends.

6. The data enhancement method for event information deduction according to claim 5, characterized in that: In S312, an LSTM structure is introduced into the generative network G to learn and reproduce complex dynamic patterns in time series data, ensuring the coherence and logical consistency of the generated data points on the time axis.

7. The data enhancement method for event information deduction according to claim 5, characterized in that: In S321, the adversarial loss and the supervisory loss are specifically as follows: in, is the random noise distribution, D is the discriminant network, Represents the expectation, and G(z) represents the time series data generated by the generator network G based on the input noise z; is the latent vector of the real time series data after being transformed by the embedding network, and S is the supervision network.

8. The data enhancement method for event information deduction according to claim 5, characterized in that: In S322, a multi-scale feature extraction mechanism is adopted in the discriminant network D to adapt to data requirements of different granularities, so that the discriminant network D can not only evaluate the authenticity of the overall sequence, but also perform fine-grained discrimination on local features.

9. The data enhancement method for event information deduction according to claim 5, characterized in that: The step 4 specifically includes: S41. Prepare random noise z: extract a set of random numbers from a predefined noise distribution as input; S42. Generate latent vector: Input random noise z into the generative network G to obtain latent vector Z=G(z); S43. Convert to time series: Use the recovery network R to convert the latent vector Z back to the time series form X ^ =R(Z); at this time, X ^ That is the newly generated time series data; S44. Post-processing: The generated time series data is de-standardized and denormalized to restore its original value range.

Images