Data processing method, electronic device and computer readable storage medium

Abstract

A data processing method comprises: establishing a data processing model based on a state switching Hawkes process, and processing data using the model; wherein a system state process is introduced into the state switching Hawkes process, and a parameter of the state switching Hawkes process changes with changes in the system state; an influence function of the state switching Hawkes process is a weighted superposition of a fixed number of basis functions, and a real-valued excitation in the influence function is mapped to a non-negative conditional intensity through a logical transformation function.

Description

Technical Field

[0001] The present invention relates to data processing technology, and more particularly to a data processing method, an electronic device and a computer-readable storage medium. Background Art

[0002] In some technologies, the conditional intensity function of the Hawkes process has three limitations: parameterization, linearization, and homogeneity, which results in limited expressiveness and a limited scope of application. Currently, there is no method that can solve these three limitations at the same time. Summary of the invention

[0003] The present application provides a data processing method, an electronic device and a computer-readable storage medium, which can enhance the expressive power of the Hawkes process model and expand its application scope.

[0004] In a first aspect, the present application provides a data processing method, comprising:

[0005] Establishing a data processing model based on the state switching Hawkes process, and processing data using the model;

[0006] Among them, the system state process is introduced into the state switching Hawkes process, and the parameters of the state switching Hawkes process change with the change of the system state; the influence function of the state switching Hawkes process is the weighted superposition of a fixed number of basis functions, and the real-valued excitation in the influence function is mapped to non-negative conditional intensity through a logical transformation function.

[0007] In a second aspect, the present application provides an electronic device including a memory and a processor, wherein the memory is used to store a program for performing data processing; the processor is used to read the program for performing data processing and execute the above-mentioned data processing method.

[0008] In a third aspect, the present application provides a computer-readable storage medium storing computer-executable instructions, and the computer-executable instructions are used to execute the above-mentioned data processing method.

[0009] In a fourth aspect, the present application provides a state switching Hawkes process, which refers to a Hawkes process that introduces a system state process and whose parameters change with the system state; the influence function of the state switching Hawkes process is a weighted superposition of a fixed number of basis functions, and the real-valued excitation in the influence function is mapped to non-negative conditional intensity through a logical transformation function.

[0010] In a fifth aspect, the present application also provides an inference algorithm for the above-mentioned state switching Hawkes process, wherein the inference algorithm includes two Bayesian inference algorithms proposed using Pólya-Gamma auxiliary latent variable technology or any one of them.

[0011] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by implementing the present application. Other advantages of the present application can be realized and obtained by the schemes described in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying drawings are used to provide an understanding of the technical solution of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the technical solution of the present application and do not constitute a limitation on the technical solution of the present application.

[0013] Figure 1 This is a flow chart of the data processing method in an embodiment of the present application;

[0014] Figure 2a The log-likelihood convergence curves of the four methods compared in the embodiments of the present application on the training data of the synthetic data;

[0015] Figure 2b The log-likelihood convergence curves of the four methods compared in the embodiments of the present application on the test data of the synthetic data;

[0016] Figure 3 This is a quantile graph of the four methods compared in the embodiments of the present application on the synthetic data;

[0017] Figure 4a This is a quantile diagram of the three methods compared in the embodiments of the present application on seismic data (SCE) of real data;

[0018] Figure 4b This is a quantile graph of the three methods compared in the embodiments of the present application on real high-frequency financial market transaction data (INTC). DETAILED DESCRIPTION

[0019] The present application describes multiple embodiments, but the description is exemplary rather than restrictive, and it is obvious to those skilled in the art that there may be more embodiments and implementations within the scope of the embodiments described in the present application. Although many possible feature combinations are shown in the drawings and discussed in the specific embodiments, many other combinations of the disclosed features are also possible. Unless specifically limited, any feature or element of any embodiment may be used in combination with any other feature or element in any other embodiment, or may replace any other feature or element in any other embodiment.

[0020] The present application includes and contemplates combinations of features and elements known to those of ordinary skill in the art. The embodiments, features and elements disclosed in the present application may also be combined with any conventional features or elements to form a unique invention scheme defined by the claims. Any features or elements of any embodiment may also be combined with features or elements from other invention schemes to form another unique invention scheme defined by the claims. Therefore, it should be understood that any feature shown and / or discussed in the present application may be implemented individually or in any appropriate combination. Therefore, except for the limitations made according to the attached claims and their equivalents, the embodiments are not subject to other restrictions. In addition, various modifications and changes may be made within the scope of protection of the attached claims.

[0021] In addition, when describing representative embodiments, the specification may have presented the method and / or process as a specific sequence of steps. However, to the extent that the method or process does not rely on the specific order of the steps described herein, the method or process should not be limited to the steps of the specific order described. As will be understood by those of ordinary skill in the art, other sequences of steps are also possible. Therefore, the specific sequence of the steps set forth in the specification should not be interpreted as a limitation to the claims. In addition, the claims for the method and / or process should not be limited to the steps of performing them in the order written, and those skilled in the art can easily understand that these sequences can be changed and still remain within the spirit and scope of the embodiments of the present application.

[0022] The combination of “# and numbers” that appears after the formula in this article indicates the number of the corresponding formula.

[0023] In many fields (such as high-frequency financial trading, neural pulse sequences, seismology, etc.), data processing is required based on Hawkes process modeling. The application scope of the Hawkes process model depends on its expressive power. The way the Hawkes process model is established also affects the inference method of the model.

[0024] Hawkes process is a kind of random point process that can describe self-excitation and mutual excitation phenomena. It has been applied in a wide range of fields including high-frequency financial trading, neural pulse sequences, seismology, etc. The event dynamics of the point process is described by the conditional intensity function. An M-dimensional multivariate Hawkes process contains M sequences of random event timestamps. Where N i is the number of random events in the i-th dimension, and the conditional intensity function of the M-th dimension is:

[0025]

[0026] in, is the historical information before time t, μ i is the baseline intensity of dimension i, fij (·) is the influence function of j dimension on i dimension.

[0027] In some techniques, the expressive power of the conditional intensity function of the Hawkes process is limited due to the following three limitations: (1) Parameterization. In some techniques, the influence function in the Hawkes process is assumed to be a parameterized function, such as an exponential decay function, which makes it impossible to fit the complex influence functions in practical applications; (2) Linearization. In some techniques, the conditional intensity function in the Hawkes process is assumed to be a linear superposition of influence functions, which makes it impossible to express the incentive effect but not the inhibitory effect; (3) Homogeneity. In some techniques, the baseline intensity and influence function in the Hawkes process do not change with time, which makes it impossible to describe the dynamics of events that change with the system state.

[0028] There are the following solutions for the above three limitations:

[0029] (1) In order to solve the parameterization restriction problem, some works have adopted the non-parametric method of the frequentist school; some works have adopted the Bayesian non-parametric method.

[0030] (2) The classic solution to the linearization limitation problem is to map the real-valued excitation to a non-negative conditional intensity through a nonlinear transformation function, so that the influence function can express both excitation and inhibition. For example, exponential transformation, logical transformation and other functions are used.

[0031] (3) To address the problem of homogeneity restriction, discrete system state variables can be introduced into the model. The baseline strength and influence function of the Hawkes process are different under different system states. For example, some works use an open-loop system state process, which develops independently and is not affected by the Hawkes process; some works use a closed-loop system state process, which interacts with the Hawkes process.

[0032] However, the above solutions only solve one or two of the limitations among parameterization, linearization, and homogeneity, and there is no method that can solve these three limitations at the same time.

[0033] In response to the above problems, Figure 1 As shown, the embodiment of the present application provides a data processing method, including:

[0034] 101 Establish a data processing model based on the state switching Hawkes process;

[0035] 102 Use models to process data;

[0036] Among them, the state switching Hawkes process introduces the system state process, and the parameters of the state switching Hawkes process change with the change of the system state (that is, the state switching Hawkes process refers to a Hawkes process that introduces the system state process and the parameters change with the change of the system state); the influence function of the state switching Hawkes process is the weighted superposition of a fixed number of basis functions, and the real-valued excitation in the influence function is mapped to non-negative conditional intensity through a logical transformation function.

[0037] The data processing method proposed in the embodiment of the present disclosure utilizes the state switching Hawkes process to establish a model, which assumes that the influence function is a weighted superposition of a fixed number of basis functions, so that the influence function becomes flexible; it maps the real-valued excitation to a non-negative conditional intensity through a logical transformation function, so that the influence function can express both excitation and inhibition; it introduces a system state process, and assumes that the parameters of the Hawkes process change with the change of the system state, so that the baseline intensity and influence function of the Hawkes process change from time-invariant to time-varying. The data processing method in the embodiment of the present application can simultaneously solve the three limitations mentioned above.

[0038] For the proposed model, the embodiment of the present invention proposes two Bayesian inference algorithms; both algorithms use the Pólya-Gamma auxiliary latent variable technology to obtain a closed-form iterative formula, thereby making the inference simple and fast. The method of the embodiment of the present application improves the data processing speed based on the Hawkes process model.

[0039] Embodiment 1

[0040] The data processing method of the present application is described below through Example 1. This example will be described from two perspectives: model building and inference method.

[0041] Model building

[0042] In the state switching Hawkes process of this embodiment, we introduce a state process z(t) to express the state of the system. The value space of z(t) is a discrete finite state space K represents the number of states. An M-dimensional state switching Hawkes process contains M sequences of event timestamps and their corresponding system states. The observation period is [0, T]. Here, we assume that z(t) is a Markov state process, which interacts with the Hawkes process in the following way: the parameters of the Hawkes process depend on the current system state z(t); the switching of the system state z(t) only occurs when a new event occurs; the state transition matrix of z(t) depends on the type of the new event. Given a set of state transition matrices Φ = {Φ1, ..., Φ M}, where Φ iis the K×K state transition matrix corresponding to dimension i. At , the transition probability of z(t) is:

[0043]

[0044] Here, we assume that the state process z(t) is left continuous, yes The right limit of i (k, k′) is Φ i The corresponding elements of , k, k′∈1,...,K.

[0045] At the same time, we established a nonlinear Hawkes process, whose influence function can express both excitation and inhibition and depends on the current state of the system. The conditional intensity function of the i-th dimension of the nonlinear Hawkes process is:

[0046]

[0047]

[0048] Among them, h i (t, z(t)) is a state-dependent real-valued stimulus that can be transformed by the logic function σ(·) and the strength upper bound (The upper bound of the intensity It can be inferred from formulas 10d and 13c in the text that the scaling of ) can be used to obtain a non-negative conditional intensity function. and are the baseline incentive and influence functions in the z(t) state, both of which can be negative. To meet the needs of flexibility, we assume It is a mixture of B basis functions:

[0049]

[0050] in, are predefined B basis functions, and the B basis functions can be selected according to requirements, and this embodiment does not impose any limitation on this. is the state-dependent mixing weight, which represents the influence of dimension j on dimension i through the bth basis function in the z(t) state. The excitation of dimension i can also be written in vector form:

[0051]

[0052] in, The cumulative effect of dimension j on time t through the bth basis function can be pre-calculated. F(t)=[1,F 11(t), ..., F MB (t)] T Combining formulas (2), (3), and (4), we get the state switching Hawkes process. The state switching Hawkes process can be considered as a closed-loop interactive system. The parameters it contains are the event type-dependent state transfer matrix, the strength upper bound, and the state-dependent activation weight, which can be summarized as

[0053] After obtaining the above-mentioned state switching Hawkes process, we can establish a data processing model based on the state switching Hawkes process and use the model to process data.

[0054] Inference method

[0055] This embodiment uses the Pólya-Gamma auxiliary latent variable technology to propose two Bayesian inference algorithms. Both algorithms have closed-form iterative formulas, which can make the inference process simple and fast.

[0056] Given a set of observation data D, the likelihood function of the state switching Hawkes process can be expressed as:

[0057] In order to learn the parameters of the state switching Hawkes process using Bayesian inference, we use a conjugate Dirichlet prior on each row of the state transfer matrix, a non-normal prior on the intensity upper bound, and a symmetric Gaussian prior on the activation weights, respectively:

[0058]

[0059]

[0060]

[0061] in, is Φ i The kth row of the hyperparameter is α = [α1, ..., α K ] T and the covariance matrix According to the prior distribution in formula (6) and the likelihood function in formula (5), we can derive the joint distribution. The posterior distribution of the state transfer matrix is ​​easy to calculate because the Dirichlet prior is conjugate to the likelihood function (classification distribution) of the state process. However, the likelihood function of the point process is not conjugate to its corresponding prior distribution, which makes the inference problem complicated. Many works have proposed to use Laplace approximation or expectation propagation methods to approximate the posterior of the point process parameters, but in this embodiment, we use the Pólya-Gamma auxiliary latent variable technique to augment the likelihood function so that the augmented likelihood function is conjugate to the Gaussian prior. Based on the augmented likelihood function, we construct an efficient Gibbs sampler that can accurately characterize the posterior distribution of the point process parameters; and, in order to further improve efficiency, we also construct a mean field variational inference algorithm that can approximately characterize the posterior distribution of the point process parameters.

[0062] In order to transform the non-conjugate likelihood function into the conjugate likelihood function, we augment the original likelihood function with two auxiliary latent variables:

[0063] The first is the Pólya-Gamma latent variable, whose core formula is:

[0064]

[0065] Among them, PG(ω|1,0) is the Pólya-Gamma distribution, g(ω,x)=x / 2-x 2 ω / 2-log2. After substituting formula (7) into formula (5), The product term is transformed into Gaussian form.

[0066] The second is a labeled Poisson latent process, whose core formula is:

[0067]

[0068] in, is a stochastic realization of a labeled Poisson process, For i The probability measure of is, and the corresponding intensity function is Timestamp Subject to a strength of Homogeneous Poisson process, Pólya-Gamma latent variable For each timestamp The labels are independent and identically distributed. Substituting formula (8) into formula (5), Items are converted to Gaussian form.

[0069] After the above two transformations, we get the augmented form of the likelihood function:

[0070]

[0071] in, It can be seen In formula (9), both appear in Gaussian form, so the augmented form of the likelihood function (Formula 9) is conjugate to the Gaussian prior distribution.

[0072] According to the prior distribution in formula (6) and the augmented likelihood function in formula (9), we can derive the augmented joint distribution Based on the augmented joint distribution, we derive a Gibbs sampler:

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] Wherein, formula (10a) is the posterior distribution of the Pólya-Gamma variable. In its derivation process, we use the skewed Pólya-Gamma distribution Formula (10b) is the posterior intensity of the labeled Poisson process; Formula (10c) is the posterior distribution of the state transfer matrix, is the number of switching from state k to state k′ in dimension i; formula (10d) is the posterior distribution of the upper bound of the strength, R i =|Π i | for Π i The number of points on the graph; Formula (10e) is the posterior distribution of the incentive weight, we define is the observation timestamp and corresponding hidden label of state k in dimension i, is the hidden timestamp and corresponding hidden label under state k on dimension i, the covariance matrix in, is a diagonal matrix, The first N i，k The elements are After R i，k The elements are is the prior covariance matrix (i.e. the hyperparameters above ), in The first N i，kThe first element is 0.5, and the last R i，k The element is -0.5. According to formula (10), we perform iterative sampling and finally obtain a series of samples to characterize the posterior distribution of the parameters.

[0079] In order to further improve the efficiency, we derive a mean field variational inference algorithm based on the augmented joint distribution. Since the Dirichlet prior is conjugate to the likelihood function of the state process, the posterior distribution of the state transfer matrix can be directly written analytically:

[0080]

[0081] in, is the number of switches from state k to state k′ in dimension i. In the mean field variational inference algorithm, we only need to approximate the parameter posterior of the point process. Here, it is assumed that the parameter variational distribution of the point process dimension i is:

[0082]

[0083] By using the calculus of variations, it can be shown that the optimal distribution of each factor that minimizes the Kullback-Leibler divergence is:

[0084]

[0085]

[0086] Substituting the augmented joint distribution into formula (12), we can obtain the optimal distribution of each factor:

[0087]

[0088]

[0089]

[0090]

[0091] Where, formula (13a) is the optimal variational distribution of the Pólya-Gamma variable. We use the skewed Pólya-Gamma distribution Formula (13b) is the optimal variational intensity of the labeled Poisson process, Formula (13c) is the optimal variational distribution of the upper strength bound, Where ψ(·) is the digamma function; formula (13d) is the optimal variational distribution of the excitation weight, in, According to formula (11), the posterior distribution of the state transfer matrix is ​​directly calculated, and then ω is iteratively updated through formula (13): i , Π i , w i Finally, we get a set of variational distributions to approximate the posterior distribution of parameters.

[0092] In this embodiment, two Bayesian inference algorithms are proposed. Both algorithms use the Pólya-Gamma auxiliary hidden variable technology to obtain a closed-form iterative formula, thereby making the inference simple and fast, and can improve the data processing speed of the above-mentioned state switching Hawkes process model.

[0093] For the state switching Hawkes process (FS-Hawkes) in the embodiment of the present disclosure, the synthetic data and real data are compared with the statistical Hawkes process and the deep Hawkes process in some technologies. The results of the comparison are analyzed as follows:

[0094] In the experiment of synthetic data, a two-state two-dimensional self-excited, mutually inhibited multivariate Hawkes process was first synthesized; and two sets of data were generated using it, one set for training data and the other for testing data. The objects for comparison are: the Gibbs sampler (Gibbs) and the mean-field variational inference algorithm (Mean-filed) proposed in the embodiments of the present disclosure, the state-dependent Hawkes process (SD-Hawkes) proposed in the document entitled State dependent Hawkes processes and their application to limit order book modelling (Document 1), and the nonlinear Hawkes process (NL-Hawkes) proposed in the document entitled Efficient inference of flexible interaction in spiking-neuronnetworks (Document 2). Figure 2a The following table shows the performance of the four methods on the training dataset. Figure 2a The horizontal axis represents the number of iterations, and the vertical axis represents the training log-likelihood; Figure 2b The performance of the four methods on the test dataset is shown in Figure 2b The horizontal axis represents the number of iterations, and the vertical axis represents the test log-likelihood; Figure 3 The horizontal and vertical axes in represent quantiles.

[0095] For the fitting ability, the comparison results are shown in the log-likelihood convergence curve in Figure 2 and Figure 3 Quantile plots of the four methods on synthetic data. Figure 2aIn the curve, at the position where the horizontal axis is 100, the order from bottom to top along the vertical axis is: the SD-Hawkes training data curve of document 1, the NL-Hawkes training data curve of document 2, the Gibbs training data curve of the embodiment of the present application; and the Mean-filed training data curve of the embodiment of the present application. Figure 2b In the curve, the order from bottom to top along the vertical axis is: the SD-Hawkes test data curve of document 1, the NL-Hawkes test data curve of document 2, the Gibbs test data curve of the embodiment of the present application; and the Mean-filed test data curve of the embodiment of the present application. Figure 3 The middle dotted line is a diagonal line. The curve located farther above the dotted line is the SD-Hawkes curve of document 1, the curve located farther below the dotted line is the NL-Hawkes curve of document 2, and the two curves close to the dotted line are the Gibbs curve and the Mean-filed curve of the embodiment of the present application.

[0096] from Figure 2a and Figure 2b As can be seen from the figure, Gibbs and Mean-filed converge to higher log-likelihoods than the baseline model on both training and test data. Figure 3 On the quantile graph, the two methods proposed in the embodiments of the present disclosure are also closer to the diagonal line than the baseline model, which indicates that the model and the two inference algorithms proposed in the embodiments of the present disclosure have better fitting performance.

[0097] In the experiment of real data, a set of earthquake data (SCE) and a set of high-frequency financial market transaction data (INTC) were used for the experiment, and each set of data sets was divided into training data and test data. For the purpose of saving time, the objects for comparison are: the mean field variational inference algorithm proposed in the embodiment of the present disclosure, the Transformer Hawkes process (TR-Hawkes) proposed in the document entitled Transformer Hawkes Process (Document 3), and the update process (TM-MRP) of triangular mapping modeling proposed in the document entitled Fast and Flexible Temporal Point Processes with Triangular Maps (Document 4). The horizontal axis and the vertical axis in Figure 4 both represent quantiles.

[0098] For the fitting ability, the comparison results are shown in Table 1 and Figure 4a and Figure 4b . Figure 4a and Figure 4b These are the quantile diagrams of the three methods on real data, where the dotted line represents the diagonal line. Figure 4aThe middle is a data graph of SCE. Taking the position of 0.6 on the horizontal axis as an example, the three curves from bottom to top along the vertical axis are: TR-Hawkes, FS-Hawkes of the embodiment of the present application, and TM-MRP; Figure 4b The data graph of INTC is shown in the figure. At the position of 0.6 on the horizontal axis, the three curves from bottom to top along the vertical axis are: TR-Hawkes, TM-MRP and FS-Hawkes of the embodiment of the present application.

[0099] Table 1 Log-likelihood of three methods on real data

[0100]

[0101] It can be seen from Table 1 and Figure 4 that the FS-Hawkes model proposed in the embodiment of the present disclosure has better fitting performance than the baseline model. The reason is that the FS-Hawkes model proposed in the embodiment of the present disclosure has parameters that can change with the system state, which plays a vital role in the application of system state driving event dynamics.

[0102] The disclosed embodiment also proposes an electronic device, including a memory and a processor, wherein the memory is used to store a program for data processing; the processor is used to read the program for data processing and execute the data processing method in any of the above embodiments.

[0103] The embodiments of the present disclosure further provide a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the data processing method in any of the above embodiments.

[0104] It will be appreciated by those skilled in the art that all or some of the steps, systems, and functional modules / units in the methods disclosed above may be implemented as software, firmware, hardware, and appropriate combinations thereof. In hardware implementations, the division between the functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed by several physical components in cooperation. Some or all components may be implemented as software executed by a processor, such as a digital signal processor or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or temporary medium). As known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, it is well known to those of ordinary skill in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

Claims

1. A data processing method, characterized in that: include: Establishing a data processing model based on the state switching Hawkes process, and processing data using the model; The state switching Hawkes process introduces a system state process, and the parameters of the state switching Hawkes process change with the change of the system state; the influence function of the state switching Hawkes process is a weighted superposition of a fixed number of basis functions, and the real-valued excitation in the influence function is mapped to a non-negative conditional intensity through a logical transformation function; the data processed by the model includes neural pulse sequence data; The parameters of the Hawkes process depend on the current system state z(t); when a new event occurs, the system state z(t) switches; the state transition matrix Φ of z(t) depends on the type of the new event; Wherein, the state transfer matrix Φ={Φ1,…,Φ M }, where Φ i is the K×K state transfer matrix corresponding to dimension i; In timestamp At , the transition probability of z(t) is: Among them, the state process z(t) is left continuous, yes The right limit of i (k,k ′ ) is Φ i The corresponding element of k,k ′ ∈1,…,K; The conditional intensity function of the i-th dimension of the state switching Hawkes process is: Among them, h i (t,z(t)) is a state-dependent real-valued stimulus, σ(·) represents a logic function, represents the upper bound of intensity; and are the reference excitation and influence functions in the z(t) state; Said It is a mixture of B basis functions: in, are the predefined B basis functions, is the state-dependent mixing weight, which represents the influence of dimension j on dimension i through the bth basis function in state z(t); The parameters of the state switching Hawkes process are the event type-dependent state transfer matrix, the strength upper bound, and the state-dependent activation weight, which can be summarized as in, is the state-dependent activation weight.

2. The data processing method according to claim 1, characterized in that: For a given observation parameter, the likelihood function of the state switching Hawkes process is: Among them, the parameters of the state switching Hawkes process are learned using the Bayesian inference method, and the following conjugate Dirichlet prior is used on each row of the state transfer matrix, the following non-normal prior is used on the intensity upper bound, and the following symmetric Gaussian prior is used on the activation weight: in, is Φ i The kth row of K ] T and the covariance matrix 3. The data processing method according to claim 2, characterized in that: The following two auxiliary latent variables are augmented into the original likelihood function: The first is the Pólya-Gamma latent variable, whose core formula is: Among them, PG(ω|1,0) is the Pólya-Gamma distribution, g(ω,x)=x / 2-x 2 ω / 2-log2; The second is a labeled Poisson latent process, whose core formula is: in, is a stochastic realization of a labeled Poisson process, For i The probability measure of is, and the corresponding intensity function is Timestamp Subject to a strength of Homogeneous Poisson process, Pólya-Gamma latent variable For each timestamp Independent and identically distributed labels.

4. The data processing method according to claim 3, characterized in that: The following Gibbs sampler is used: Wherein, formula (10a) is the posterior distribution of the Pólya-Gamma variable; formula (10b) is the posterior intensity of the labeled Poisson process; formula (10c) is the posterior distribution of the state transfer matrix, is from state k to state k in dimension i ′ The number of switching times; Formula (10d) is the posterior distribution of the upper bound of the intensity, R i =|Π i | for Π i The number of points on the graph; Formula (10e) is the posterior distribution of the incentive weight, defined as is the observation timestamp and corresponding hidden label of state k in dimension i, is the hidden timestamp and corresponding hidden label under state k on dimension i, the covariance matrix in, is a diagonal matrix, The first N i,k The elements are After R i,k The elements are is the prior covariance matrix, in, The first N i,k The first element is 0.5, and the last R i,k The element is -0.

5.

5. The data processing method according to claim 2, characterized in that: The posterior distribution of the state transfer matrix is: in, is from state k to state k in dimension i ′ The number of switching times.

6. The data processing method according to claim 5, characterized in that: The method further includes: iteratively updating ω by the following formula i ,Π i , w i The variational distribution of : Wherein, formula (13a) is the optimal variational distribution of the Pólya-Gamma variable, which uses the skewed Pólya-Gamma distribution Formula (13b) is the optimal variational intensity of the labeled Poisson process, Formula (13c) is the optimal variational distribution of the upper strength bound, Where ψ(·) is the digamma function; formula (13d) is the optimal variational distribution of the excitation weight, in, 7. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store a program for data processing; the processor is used to read the program for data processing and execute the data processing method according to any one of claims 1 to 6.

8. A computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the data processing method according to any one of claims 1 to 6.

Images