A modeling method of a statistical mixture model in a big data distributed scene

Abstract

The application relates to the computer technical field and discloses a modeling method of a statistical mixed model in a big data distributed scene. The method comprises the following steps: distributing and storing data shards; initializing model parameters; iteratively performing an expectation step and a maximization step; the expectation step is scheduled to a GPU node to perform parallel calculation on posterior probability; the maximization step is scheduled to a CPU node to aggregate statistics and update parameters, perform component merging / deletion, and perform convergence judgment; meanwhile, a memory reuse mechanism based on reference counting and scope analysis is adopted to reduce redundant data transmission. Through heterogeneous task scheduling and memory collaborative optimization, the training speed, resource utilization rate, and model self-adaptation capability are improved.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and specifically relates to a modeling method for statistical hybrid models in a big data distributed scenario. Background Technology

[0002] With the deep penetration of big data technology into fields such as financial risk control, smart healthcare, bioinformatics, and the Internet of Things, statistical mixture model-based analytical methods are widely used due to their powerful modeling capabilities for complex data distributions. Mixture models approximate the multimodal characteristics of real data by combining multiple probability distribution components, and the training process typically employs the Expectation-Maximization (EM) algorithm for iterative optimization. While the EM algorithm theoretically exhibits good convergence, it faces significant challenges in computational efficiency and resource adaptation during large-scale distributed deployments. Particularly in heterogeneous hardware environments, the dynamic nature of model training and the static scheduling of hardware characteristics present problems, hindering the efficient implementation of mixture models in cloud-edge collaborative scenarios.

[0003] The EM iteration process of hybrid models inherently exhibits computational heterogeneity: the E-step involves calculating the posterior probabilities of a large number of samples for each component, resulting in high-density matrix operations with high parallelism, making it suitable for accelerators such as GPUs; while the M-step includes logic-intensive operations such as model parameter updates, component merging or splitting decisions, sparse data processing, and convergence checks, relying on complex control flow and conditional branches, making it more suitable for general-purpose computing architectures like CPUs and GPUs. However, current mainstream distributed computing frameworks generally adopt coarse-grained task partitioning strategies, treating the entire EM iteration as a single computing unit for scheduling. This fails to perceive the fine-grained computational differences within the hybrid model, leading to idle GPU resources in the M-step phase and CPUs becoming a bottleneck in the E-step phase, resulting in a significant imbalance in hardware utilization.

[0004] Existing technologies face the following challenges in addressing the aforementioned issues: they lack dynamic awareness of runtime characteristics such as the hybrid model structure and data sparsity, making it impossible to accurately identify the computational load boundaries between the E-step and M-step before task execution; the task scheduling mechanism is not deeply coupled with underlying hardware characteristics, hindering intelligent offloading of subtasks and optimization of execution paths across devices; and frequent cross-device data transfers are not effectively pipelined, causing data movement overhead between GPUs and CPUs to become a performance bottleneck, particularly in memory-constrained edge device scenarios. These problems make it difficult to adaptively optimize the same hybrid model algorithm across clusters with different hardware configurations, severely limiting its deployment capabilities across all scenarios, from high-performance cloud servers to resource-constrained edge nodes. Therefore, a novel modeling method is urgently needed that can perceive the computational characteristics of hybrid models, support collaborative scheduling of heterogeneous hardware, and optimize data flow. Summary of the Invention

[0005] This invention provides a modeling method for statistical hybrid models in big data distributed scenarios. It aims to address the technical problems of low computational efficiency, insufficient hardware resource utilization, and slow model convergence speed caused by the failure of existing distributed computing frameworks to perform fine-grained resource scheduling and memory co-optimization for the heterogeneous computational characteristics of the expectation step and maximization step in the expectation-maximization algorithm when processing statistical hybrid models. This invention achieves efficient co-execution of the dense matrix operations of the expectation step and the sparse logic processing of the maximization step by constructing a task decomposition and dynamic scheduling mechanism oriented towards heterogeneous computing units, combined with a memory management strategy based on data locality and lifecycle.

[0006] This invention provides a modeling method for statistical hybrid models in a big data distributed scenario, comprising:

[0007] The input raw observation dataset is divided into multiple data fragments according to a preset fragmentation rule, and the data fragments are distributed and stored on multiple working nodes of a distributed computing cluster.

[0008] Initialize the parameter set of the mixture model, which includes the weight coefficients, mean vector and covariance matrix of each component;

[0009] The iterative optimization process is performed until the preset convergence condition is met; the iterative optimization process includes alternating between the expected step and the maximization step.

[0010] During the expected step, the expected step task is scheduled to a computing node equipped with a CPU graphics processor, and the CPU graphics processor is used to compute the posterior probability of each data slice with respect to each mixture component in parallel.

[0011] During the maximization step, the maximization step task is scheduled to a computing node equipped with a central processing unit. The central processing unit aggregates the posterior probability statistics from all data shards and performs component parameter updates, sparse data filtering, component merging or splitting determination, and convergence evaluation.

[0012] Between the expectation step and the maximization step, a memory reuse mechanism based on reference counting and scope analysis is adopted to manage the lifecycle of intermediate calculation results, avoiding redundant data copying and frequent data migration between host devices.

[0013] Preferably, the input raw observation dataset is divided into multiple data partitions according to a preset partitioning rule, including:

[0014] Based on the primary key hash value or timestamp range of the observed data, the data is evenly distributed to a preset number of data fragments;

[0015] The size of each data shard is no greater than the maximum memory capacity of a single worker node, and the number of data shards matches the number of CPU graphics processors in the cluster.

[0016] Preferably, the parameter set for initializing the hybrid model includes:

[0017] Randomly sample several subsets from the original observation dataset and calculate their global mean and covariance as the initial mean vector and covariance matrix;

[0018] Set the initial weighting coefficients of each component to be equal, with the value being the reciprocal of the number of components.

[0019] Preferably, when executing the desired step, scheduling the desired step task to a computing node equipped with a CPU graphics processor includes:

[0020] Construct a desired step computation task graph that defines the computational dependencies between each data slice and each blend component;

[0021] The computational subtasks in the task graph are assigned to idle GPU graphics processors in the cluster;

[0022] On the GPU graphics processor, the corresponding data slices and current model parameters are loaded, matrix multiplication, exponential function and normalization operations are performed to generate the posterior probability matrix of each observation sample with respect to each mixture component;

[0023] The posterior probability matrix is ​​stored in the video memory of the CPU graphics processor in column-major format.

[0024] Preferably, the posterior probability is specifically expressed as follows:

[0025] ;

[0026] For the sample about the first The posterior probability of each component. For the edge likelihood of the sample, For the first Likelihood values ​​for a mixture of components These are the corresponding weighting coefficients.

[0027] Preferably, during the maximization step, scheduling the maximization step task to a computing node equipped with a central processing unit includes:

[0028] An aggregation process is initiated on the central processing unit. This aggregation process retrieves a statistical summary of the posterior probability matrix from the video memory of each CPU's graphics processor via remote direct memory access technology. The statistical summary includes the total responsibility weight, weighted observation sum, and weighted outer product sum for each component.

[0029] Based on the statistical summary, calculate the new mean vector, covariance matrix, and weight coefficients;

[0030] Perform a positive definiteness check on the newly calculated covariance matrix; if it does not meet the requirement, apply a diagonal loading correction.

[0031] A threshold is used to determine the responsibility weight of each component. If it is less than the preset minimum responsibility threshold, the component is marked as pending deletion.

[0032] Calculate the distance between the mean vectors of any two components. If the distance is less than the preset merging distance threshold and the difference in the Frobenius norm of the covariance structure is less than the preset similarity threshold, then perform the component merging operation.

[0033] Based on the log-likelihood difference between the old and new parameter sets, determine whether the convergence condition is met.

[0034] Preferably, the component merging operation includes:

[0035] The weight coefficients of the two components to be merged are added together to obtain the weight of the new component.

[0036] The sum of the weighted mean vectors of the two components is divided by the total weight to obtain the mean vector of the new component.

[0037] After merging the weighted covariance matrices, weighted outer product terms, and cross terms of the two components, the covariance matrix of the new component is recalculated.

[0038] Preferably, a memory reuse mechanism based on reference counting and scope analysis is adopted, including:

[0039] Before the expected step begins, a fixed-size buffer is pre-allocated in the GPU graphics processor memory for each data slice;

[0040] After the expectation step calculation is completed, the posterior probability matrix is ​​retained until the aggregation process of the maximization step completes the reading of it;

[0041] After the parameter update is completed in the maximization step, immediately release all intermediate data buffers related to the previous iteration;

[0042] For model parameters, a single authoritative copy is maintained in the central processing unit's memory. Before each desired step begins, it is copied to the video memory of each CPU's graphics processor via asynchronous data transfer. The transfer process overlaps with the computation process to hide communication latency.

[0043] Preferably, the preset convergence conditions include:

[0044] In three consecutive iterations, the absolute value of the increment of the log-likelihood of the statistical mixture model is less than the preset likelihood convergence threshold; or the total number of iterations reaches the preset maximum number of iterations limit.

[0045] Preferred, the first The likelihood values ​​for each mixture component are given by the multivariate Gaussian probability density function:

[0046] ;

[0047] For the first The first segment The feature vector of each sample and The first The current mean vector and covariance matrix of each component. Given the feature dimension, the edge likelihood of the sample is:

[0048] ;

[0049] for The total number.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] 1. This invention achieves precise adaptation to heterogeneous computing load during hybrid model training by mapping the expectation step and maximization step of the expectation maximization algorithm to the GPU graphics processor and the central processing unit for execution, respectively.

[0052] 2. The dense matrix operations of the expected step achieve high throughput acceleration on the GPU graphics processor, while the sparse logic, component management and convergence judgment in the maximization step are efficiently processed serially on the central processing unit, avoiding the performance bottleneck caused by performing unsuitable tasks on a single type of hardware.

[0053] 3. The memory collaborative management mechanism designed in this invention precisely controls the lifecycle of intermediate data through reference counting and scope analysis, reducing redundant data transmission between the host and the device and lowering communication overhead.

[0054] 4. The dynamic merging and deletion mechanism of components enhances the model's ability to adapt to the inherent structure of the data, avoiding overfitting and underfitting.

[0055] 5. This invention improves the training speed and resource utilization of large-scale hybrid models in distributed heterogeneous environments while ensuring model accuracy. Attached Figure Description

[0056] Figure 1This is a schematic diagram of the overall technical solution architecture of the present invention;

[0057] Figure 2 This is a flowchart illustrating the logical flow of the parallel posterior probability calculation and memory buffer management expected to be performed on a GPU graphics processor in this invention.

[0058] Figure 3 This is a logical flow diagram of the statistical aggregation, parameter update, and dynamic component management that maximizes the steps on the central processing unit in this invention. Detailed Implementation

[0059] refer to Figures 1 to 3 This invention provides a modeling method for statistical hybrid models in a big data distributed scenario. It addresses the heterogeneous computational characteristics of the expectation step and maximization step in the expectation-maximization algorithm by constructing a fine-grained task scheduling mechanism for GPUs and CPUs, and combining it with a memory collaborative management strategy based on reference counting and scope analysis to improve the efficiency, resource utilization, and convergence speed of large-scale hybrid model training. The specific implementation of this method will be described in detail below.

[0060] The method first performs a data sharding operation. The input raw observation dataset is divided into multiple data shards according to a preset sharding rule, and these data shards are distributed and stored across multiple worker nodes in a distributed computing cluster. The sharding rule is based on the primary key hash value or timestamp range of the observation data to ensure that the data is evenly distributed across the nodes. The size of each data shard is strictly controlled, not exceeding the memory capacity limit of a single worker node, to avoid task failure due to memory overflow. Simultaneously, the number of data shards is matched to the number of CPUs and GPUs in the cluster, thus providing sufficient hardware resource alignment for the parallel computation of subsequent desired steps. After sharding, each data shard is persistently stored in the local storage medium of the corresponding worker node, and a global index mapping table is established for quickly locating and loading data during iteration.

[0061] The model initialization process is performed. The parameter set of the mixture model is initialized, including the weight coefficients, mean vector, and covariance matrix of each component. The initialization process involves randomly sampling several subsets from the original observation dataset and calculating their global mean and covariance as the initial mean vector and covariance matrix. Specifically, the arithmetic mean of the feature dimensions of all observed samples is calculated to form the initial mean vector; the average of the outer product matrices after sample centering is calculated to obtain the initial covariance matrix. The initial weight coefficients of each mixture component are set to equal values, which are equal to the reciprocal of the number of mixture components. This model initialization strategy ensures good numerical stability of the model in the early stages of training and avoids convergence difficulties caused by excessive initial parameter bias. The initialized parameter set is stored in the main memory of the central processing unit as a single authoritative copy of the model for subsequent iterations.

[0062] After data partitioning and model initialization are completed, the iterative optimization process begins. This process alternates between expectation steps and maximization steps until a preset convergence condition is met. The convergence condition includes: in three consecutive iterations, the absolute value of the increment of the model's log-likelihood is less than a preset likelihood convergence threshold; or the total number of iterations reaches a preset maximum iteration limit. The iteration terminates when either condition is met.

[0063] During the expected step, the expected step tasks are scheduled to compute nodes equipped with CPU graphics processors. Specifically, an expected step computation task graph is constructed, which defines the computational dependencies between each data shard and each mixture component. The task graph is represented as a directed acyclic graph, where nodes represent computational subtasks and edges represent data dependencies. The task scheduling unit traverses this task graph, dynamically allocating computational subtasks to currently idle GPU graphics processors in the cluster. Each GPU graphics processor receives one or more data shards and their corresponding mixture component parameters, loading them into its video memory. On the CPU graphics processor, intensive matrix operations are performed to compute the posterior probability of each observation sample with respect to each mixture component. For the first... The first data shard in the data shard For the nth observation sample, calculate its nth... Likelihood values ​​for each mixed component The likelihood value is given by the multivariate Gaussian probability density function:

[0064] ;

[0065] For the first The first segment The feature vector of each sample and The first The current mean vector and covariance matrix of each component. For the feature dimension, the likelihood values ​​of all components are correlated with their corresponding weight coefficients. Multiply and then sum to obtain the marginal likelihood of the sample. :

[0066] ;

[0067] for The total number;

[0068] Then, by dividing the weighted likelihood value of each component by the marginal likelihood, we obtain the sample's likelihood with respect to the first component. The posterior probability of each component :

[0069] ;z

[0070] The above calculations are implemented through highly parallelized matrix multiplication, exponential functions, and normalization operations, fully utilizing the single instruction multiple data (SID) architecture of the GPU. The generated posterior probability matrix is ​​stored in the GPU's memory in column-major format, with each column corresponding to a mixture component and each row corresponding to an observed sample. This column-major storage format facilitates efficient aggregation of subsequent statistics.

[0071] After the expected step is completed, the process immediately proceeds to the maximization step. The maximization step task is scheduled to a computing node equipped with a central processing unit (CPU). An aggregation process is initiated on the CPU, which retrieves a statistical summary of the posterior probability matrix from the video memory of each CPU's graphics processing unit (GPU) via remote direct memory access (RDA). This statistical summary includes the total responsibility weight for each component, the weighted sum of observations, and the weighted sum of outer products, mathematically expressed as follows:

[0072] ;

[0073] For the first The total responsibility weight of each component For weighted observation sum, This is a weighted outer product sum. Based on the statistical summary, new model parameters are calculated. The new weight coefficients are... , The total number of samples; the new mean vector The new covariance matrix is .

[0074] After parameter updates, component management logic is executed. The newly calculated covariance matrix undergoes positive definiteness verification. If its minimum eigenvalue is less than a preset positive definiteness threshold, a diagonal loading correction is applied, adding a small constant to the diagonal of the covariance matrix to ensure its invertibility. A threshold judgment is performed on the responsibility weight of each component. If the responsibility weight is less than a preset minimum responsibility threshold, the component is marked as pending deletion and removed from the parameter set before the next iteration. The distance between the mean vectors of any two components is calculated. If the Euclidean distance is less than a preset merging distance threshold and the difference in the Frobenius norm of their covariance structures is less than a preset similarity threshold, a component merging operation is performed. The merging operation includes: adding the weight coefficients of the two components to be merged as the weight of the new component; dividing the sum of the weighted mean vectors of the two components by the total weight to obtain the mean vector of the new component; merging the weighted covariance matrices, weighted outer product terms, and cross terms of the two components, and then recalculating the covariance matrix of the new component. Dynamic component management mechanisms enable models to adapt to the inherent structure of the data, avoiding overfitting or underfitting.

[0075] In the final stage of the maximization step, a convergence evaluation is performed. The log-likelihood value of the model for the current iteration is calculated. :

[0076] ;

[0077] Compare the increment with the log-likelihood value from the previous iteration. If the absolute value of the increment is less than the preset likelihood convergence threshold, record the convergence state. If the condition is met for three consecutive iterations, or if the number of iterations reaches the maximum limit, terminate the iteration.

[0078] Throughout the iteration process, the memory co-management module runs continuously. Before the expectation step begins, a fixed-size buffer is pre-allocated in the GPU's video memory for each data slice to store observation data, model parameters, and the posterior probability matrix. After the expectation step calculation is completed, the posterior probability matrix is ​​retained until the aggregation process of the maximization step finishes reading it. The retention mechanism is implemented through reference counting: whenever the aggregation process begins reading a block of posterior probability data, the reference count is incremented by one; after reading, the reference count is decremented by one. When the reference count reaches zero, the data block is marked as free. After the parameter update is completed in the maximization step, all intermediate data buffers related to the previous iteration are immediately released, including the old posterior probability matrix, statistical summary, and temporary calculation results. For model parameters, a single authoritative copy is maintained in the CPU's memory. Before each expectation step begins, these parameters are copied to the video memory of each CPU's GPU via asynchronous data transfer. This asynchronous data transfer process overlaps with other computational preparation operations of the GPU to hide communication latency and improve overall throughput.

[0079] The execution of the above method relies on a distributed computing cluster containing multiple worker nodes, each equipped with at least one central processing unit (CPU) and one GPU (graphics processing unit), connected via a high-speed interconnect network. The cluster deploys a task scheduler, a memory manager, and a communication coordinator. The task scheduler is responsible for parsing the computation graph of the desired step and the maximization step, and allocating subtasks to appropriate hardware resources. The memory manager maintains a global memory view, tracking the lifecycle, reference count, and physical location of each data block. The communication coordinator manages remote direct memory access operations, ensuring efficient retrieval of statistical summaries. All components work together to ensure the scalability and robustness of the method on large-scale datasets.

[0080] In summary, this embodiment achieves efficient heterogeneous acceleration of the training process for statistical mixture models by scheduling the expectation step to the GPU for intensive matrix operations and the maximization step to the CPU for sparse logic and component management, supplemented by memory reuse and asynchronous parameter synchronization mechanisms. This method improves training speed and hardware resource utilization while maintaining model accuracy, making it suitable for scenarios involving massive amounts of high-dimensional observation data.

[0081] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0082] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A modeling method for statistical hybrid models in a big data distributed scenario, characterized in that, include: The input raw observation dataset is divided into multiple data fragments according to a preset fragmentation rule, and the data fragments are distributed and stored on multiple working nodes of a distributed computing cluster. Initialize the parameter set of the mixture model, which includes the weight coefficients, mean vector and covariance matrix of each component; The iterative optimization process is performed until the preset convergence condition is met; the iterative optimization process includes alternating between the expected step and the maximization step. During the expected step, the expected step task is scheduled to a computing node equipped with a CPU graphics processor, and the CPU graphics processor is used to compute the posterior probability of each data slice with respect to each mixture component in parallel. During the maximization step, the maximization step task is scheduled to a computing node equipped with a central processing unit. The central processing unit aggregates the posterior probability statistics from all data shards and performs component parameter updates, sparse data filtering, component merging or splitting determination, and convergence evaluation. Between the expectation step and the maximization step, a memory reuse mechanism based on reference counting and scope analysis is adopted to manage the lifecycle of intermediate calculation results, avoiding redundant data copying and frequent data migration between host devices.

2. The modeling method for statistical hybrid models in a big data distributed scenario according to claim 1, characterized in that, The input raw observation dataset is divided into multiple data partitions according to a preset partitioning rule, including: Based on the primary key hash value or timestamp range of the observed data, the data is evenly distributed to a preset number of data fragments; The size of each data shard is no greater than the maximum memory capacity of a single worker node, and the number of data shards matches the number of CPU graphics processors in the cluster.

3. The modeling method for statistical hybrid models in a big data distributed scenario according to claim 2, characterized in that, The parameter set for initializing the hybrid model includes: Randomly sample several subsets from the original observation dataset and calculate their global mean and covariance as the initial mean vector and covariance matrix; Set the initial weighting coefficients of each component to be equal, with the value being the reciprocal of the number of components.

4. The modeling method for statistical hybrid models in a big data distributed scenario according to claim 3, characterized in that, During the execution of the desired step, scheduling the desired step task to a computing node equipped with a GPU graphics processor includes: Construct a desired step computation task graph that defines the computational dependencies between each data slice and each blend component; The computational subtasks in the task graph are assigned to idle CPU graphics processors in the cluster; On the GPU graphics processor, the corresponding data slices and current model parameters are loaded, matrix multiplication, exponential function and normalization operations are performed to generate the posterior probability matrix of each observation sample with respect to each mixture component; The posterior probability matrix is ​​stored in the video memory of the CPU graphics processor in column-major format.

5. The modeling method for statistical hybrid models in a big data distributed scenario according to claim 4, characterized in that, The specific expression of the posterior probability is as follows: ; For the sample about the first The posterior probability of each component. For the edge likelihood of the sample, For the first Likelihood values ​​for a mixture of components These are the corresponding weighting coefficients.

6. The modeling method for statistical hybrid models in a big data distributed scenario according to claim 5, characterized in that, During the maximization step, scheduling the maximization step task to a computing node equipped with a central processing unit includes: An aggregation process is initiated on the central processing unit. This aggregation process retrieves a statistical summary of the posterior probability matrix from the video memory of each CPU's graphics processor via remote direct memory access technology. The statistical summary includes the total responsibility weight, weighted observation sum, and weighted outer product sum for each component. Based on the statistical summary, calculate the new mean vector, covariance matrix, and weight coefficients; Perform a positive definiteness check on the newly calculated covariance matrix; if it does not meet the requirement, apply a diagonal loading correction. A threshold is used to determine the responsibility weight of each component. If it is less than the preset minimum responsibility threshold, the component is marked as pending deletion. Calculate the distance between the mean vectors of any two components. If the distance is less than the preset merging distance threshold and the difference in the Frobenius norm of the covariance structure is less than the preset similarity threshold, then perform the component merging operation. Based on the log-likelihood difference between the old and new parameter sets, determine whether the convergence condition is met.

7. The modeling method for statistical hybrid models in a big data distributed scenario according to claim 6, characterized in that, Performing component merging operations includes: The weight coefficients of the two components to be merged are added together to obtain the weight of the new component. The sum of the weighted mean vectors of the two components is divided by the total weight to obtain the mean vector of the new component. After merging the weighted covariance matrices, weighted outer product terms, and cross terms of the two components, the covariance matrix of the new component is recalculated.

8. The modeling method for statistical hybrid models in a big data distributed scenario according to claim 7, characterized in that, A memory reuse mechanism based on reference counting and scope analysis is adopted, including: Before the expected step begins, a fixed-size buffer is pre-allocated in the GPU graphics processor memory for each data slice; After the expectation step calculation is completed, the posterior probability matrix is ​​retained until the aggregation process of the maximization step completes the reading of it; After the parameter update is completed in the maximization step, immediately release all intermediate data buffers related to the previous iteration; For model parameters, a single authoritative copy is maintained in the central processing unit's memory. Before each desired step begins, it is copied to the video memory of each CPU's graphics processor via asynchronous data transfer. The transfer process overlaps with the computation process to hide communication latency.

9. The modeling method for statistical hybrid models in a big data distributed scenario according to claim 8, characterized in that, The preset convergence conditions include: In three consecutive iterations, the absolute value of the increment of the log-likelihood of the statistical mixture model is less than the preset likelihood convergence threshold; or the total number of iterations reaches the preset maximum number of iterations limit.

10. The modeling method for statistical hybrid models in a big data distributed scenario according to claim 9, characterized in that, No. The likelihood values ​​for each mixture component are given by the multivariate Gaussian probability density function: ; For the first The first segment The feature vector of each sample and The first The current mean vector and covariance matrix of each component. Given the feature dimension, the edge likelihood of the sample is: ; for The total number.

Images