Call center operation risk prediction method based on federated learning

Abstract

The application relates to the technical field of federated learning, and particularly discloses a call center operation risk prediction method based on federated learning, wherein a center end initializes a global risk identification base parameter and distributes the same to a participation end selected through stratified sampling; after the participation end receives the base parameter, the participation end performs inner loop updating by using a local first data set, and constructs an elastic mask matrix according to parameter difference and local language dispersion to reserve key responses of general risk knowledge, so as to obtain a local risk identification parameter; the participation end evaluates performance by using a local second data set, obtains a meta-gradient by backtracking, generates an adaptive weight according to a data size and a prediction confidence, and then uploads the adaptive weight; the center end performs spectral clustering and clustering on the meta-gradient, performs weighted fusion according to the clusters, and then updates the global base parameter; the application solves the gradient conflict and catastrophic forgetting problem caused by the distribution difference of cross-region data, improves the risk prediction accuracy and federated aggregation robustness in a non-independent and identically distributed scene while ensuring data privacy.

Description

Technical Field

[0001] This invention relates to the field of federated learning technology, and more specifically to a method for predicting operational risks in call centers based on federated learning. Background Technology

[0002] With the online and intelligent development of financial services, call centers of banks, insurance companies, and other institutions generate massive amounts of customer service call data daily, containing rich operational risk information, including escalation of customer complaints, identification of fraudulent calls, and warnings of operational errors. Traditional centralized machine learning methods require the raw call data from various branches or outsourced call centers to be uniformly aggregated to a central server for model training. Federated learning, as an emerging distributed learning paradigm, can achieve joint modeling through gradient aggregation without the data leaving the local machine, providing a feasible technical path for predicting operational risks in call centers.

[0003] Existing technologies lack a way to avoid gradient conflicts and catastrophic forgetting caused by the forced uniform aggregation of traditional federated averaging algorithms, given that the data on the participating side exhibits significant non-independent identical distribution characteristics (i.e., there is a natural isolation between dialect systems, business types, and risk distributions in different regions). This would allow for the simultaneous preservation of the ability to represent general risk knowledge across regions and the ability to identify personalized risk characteristics in each region. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting call center operational risks based on federated learning, in order to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: A federated learning-based approach to predicting call center operational risks includes the following steps: S1: The central end initializes a global risk identification base parameter, which is used to characterize general knowledge of cross-regional call center operational risks; S2: The central terminal selects several participating terminals from multiple participating terminals to form a task set, and distributes the current global risk identification base parameters to each selected participating terminal respectively; S3: After each selected participant receives the global risk identification base parameters, it uses the local first dataset to perform an inner loop update. During the update process, it introduces elastic regularization constraints based on the difference between the local data distribution and the historical global distribution to retain key responses of general risk knowledge, thereby obtaining local risk identification parameters that are adapted to local dialect characteristics and business risk distribution. S4: Each selected participant uses its local second dataset to evaluate the performance of local risk identification parameters, calculates the meta-gradient of the global risk identification basis parameters based on the evaluation loss, and generates adaptive weights based on the size of the second dataset and the label confidence. The meta-gradient and adaptive weights are then uploaded to the central terminal. S5: The central terminal receives the meta-gradients and adaptive weights uploaded by all selected participating terminals, performs directional similarity clustering on the meta-gradients, performs weighted fusion according to the adaptive weights by cluster, and updates the global risk identification basis parameters based on the fusion results, so that the updated global risk identification basis parameters can be quickly adapted locally by each participating terminal in the next iteration to more accurately predict the operational risks of different regions.

[0006] As a further aspect of the present invention: S1 specifically includes: The central terminal collects historical call text data from each participating terminal and extracts location-independent risk semantic units from the historical call text data. Risk semantic units include monetary figures, emotional intensity values, and high-frequency fraud keywords. The central end performs distribution statistics on risk semantic units and constructs an initial general knowledge vector based on the statistical results. The initial general knowledge vector is used to describe risk patterns that co-occur across regions. The central terminal sets a learnable parameter matrix based on the initial general knowledge vector, and uses the parameter matrix as the basis parameter for global risk identification.

[0007] As a further aspect of the present invention: S2 specifically includes: The central terminal receives the local data risk entropy value uploaded by each participating terminal. The local data risk entropy value is calculated based on the frequency of occurrence and the degree of disorder in the distribution of risk keywords in the historical call text of each participating terminal. The central end sorts the participating ends according to the risk entropy value of local data, divides them into several gradient layers according to the entropy value from high to low, and randomly selects one participating end from each gradient layer as the selected participating end. The central terminal will distribute the current global risk identification base parameters to each selected participating terminal.

[0008] As a further aspect of the present invention: S3 specifically includes: After each selected participant receives the global risk identification basis parameters, it randomly selects a support subset from the local first dataset, and performs the first round of gradient descent on the global risk identification basis parameters based on the support subset to obtain the first intermediate parameters. Each selected participant calculates the parameter difference value between the first intermediate parameter and the global risk identification base parameter of the previous round, and constructs an elastic mask matrix based on the size of each dimension of the parameter difference value. Each selected participant extracts a query subset from the local first dataset, and performs a second round of gradient descent on the first intermediate parameter based on the query subset and using the elastic mask matrix as the weight of the regularization term, to obtain local risk identification parameters that are adapted to local dialect features and business risk distribution.

[0009] As a further aspect of the present invention: the construction of an elastic mask matrix based on the magnitude of each dimension of the parameter difference values ​​specifically includes: Each selected participant calculates the dimension-wise difference between the first intermediate parameter and the global risk identification baseline parameter of the previous round, and obtains the parameter difference vector by converting the dimension-wise difference into absolute value. Each selected participant obtains the dispersion of dialect features in the first local dataset, calculates an elastic adjustment coefficient based on the dispersion of dialect features, and multiplies each element in the parameter difference vector with the elastic adjustment coefficient to obtain a scaled difference vector. Each selected participant inputs a scaled difference vector into a monotonically increasing saturation function. The output value of the saturation function is between 0 and 1. The output values ​​of the saturation function are arranged according to the original dimensions to form an elastic mask matrix.

[0010] As a further aspect of the present invention: S4 specifically includes: Each selected participant performs forward computation on the second dataset to obtain the cross-entropy between the predicted risk value and the true risk label, and uses the cross-entropy as the evaluation loss. Each selected participant differentiates the local risk identification parameters based on the assessed loss to obtain the first gradient, and then uses the chain rule to backtrack the first gradient to the global risk identification basis parameters to obtain the meta-gradient. The number of samples in the second dataset is used as the base size for each selected participant. At the same time, the average probability entropy of the predicted risk value on the second dataset is calculated as the inverse of the confidence level. The base size is divided by the inverse of the confidence level and the logarithm is taken to obtain the adaptive weight. Each selected participant multiplies the meta-gradient by adaptive weights and encapsulates it into an upload data packet, which is then sent to the central endpoint.

[0011] As a further aspect of the present invention: the specific process of obtaining the meta-gradient is as follows: Each selected participant will evaluate the loss by calculating the partial derivative of the local risk identification parameters to obtain the first gradient, and at the same time record the gradient propagation path of the global risk identification basis parameters on which the local risk identification parameters depend during the update process. Each selected participant applies a path decay coefficient to the first gradient based on the gradient propagation path. The path decay coefficient is negatively correlated with the step size and number of iterations of the inner loop update. The first gradient after applying the decay is used as the intermediate backtracking gradient. Each selected participant propagates the intermediate backtracking gradient backward along the gradient propagation path layer by layer, stopping when it reaches the global risk identification basis parameters, and taking the gradient value at the stopping position as the meta-gradient.

[0012] As a further aspect of the present invention: S5 specifically includes: The central end receives the meta-gradients and adaptive weights uploaded by all selected participating ends, calculates the cosine similarity between every two meta-gradients, and constructs a similarity matrix based on the cosine similarity. The central end inputs the similarity matrix into the spectral clustering algorithm, and divides the meta-gradient into several gradient clusters according to the preset number of clusters, so that the meta-gradient directions within the same cluster are similar and the meta-gradient directions between different clusters are different. For each gradient cluster, the central end performs a weighted sum of all the meta-gradients within the cluster according to the corresponding adaptive weights to obtain the cluster representative gradient of the corresponding cluster. Then, the cluster representative gradients are superimposed with equal weights to obtain the global update direction. The central end multiplies the global update direction by the preset learning rate and adds it to the current global risk identification basis parameters to obtain the updated global risk identification basis parameters.

[0013] As a further aspect of the present invention: the central end inputs the similarity matrix into a spectral clustering algorithm, and divides the meta-gradient into several gradient clusters according to a preset number of clusters, so that the meta-gradient directions within the same cluster are similar and the meta-gradient directions between different clusters are different, specifically including: The center performs row normalization on the similarity matrix to obtain a normalized similarity matrix. Based on the normalized similarity matrix, a diagonal matrix is ​​calculated. The diagonal elements of the diagonal matrix are the sum of the elements of each row of the normalized similarity matrix. The center subtracts the normalized similarity matrix from the diagonal matrix to obtain the Laplacian matrix. The eigenvectors corresponding to the first few smallest eigenvalues ​​of the Laplacian matrix are calculated, and the eigenvectors are arranged in columns to form an eigenvector matrix. The central end treats each row of the feature vector matrix as a low-dimensional embedding representation of a meta-gradient. Based on the Euclidean distance between the low-dimensional embedding representations, a greedy iterative strategy is used to divide the meta-gradient into a preset number of gradient clusters, so that the meta-gradients within the same cluster are adjacent in the low-dimensional embedding space, while the meta-gradients between different clusters are far apart.

[0014] The beneficial effects of this invention are: (1) This invention effectively solves the model bias problem caused by the differences in cross-regional data distribution by introducing an elastic mask matrix and a meta-gradient backtracking mechanism. When each participating end updates locally, it constructs an elastic mask matrix based on the parameter difference value and dialect dispersion, and applies strong regularization constraints to the key responses of general risk knowledge, avoiding catastrophic forgetting caused by forced uniform aggregation. At the same time, the central end performs spectral clustering and fusion of meta-gradients, so that the gradients of participating ends in the same dialect area or similar business models are aggregated first. The generated global risk identification base parameters can be accurately adapted to local dialect features and complex financial product risks with a small amount of local updates in the next iteration, improving the prediction accuracy of the model in non-independent and identically distributed scenarios.

[0015] (2) This invention reduces communication overhead and the risk of malicious gradient pollution through stratified sampling and adaptive weight generation mechanisms. The central end performs stratified random sampling on each participating end based on the local data risk entropy value, ensuring that representative samples of high, medium, and low risk regions are covered in each iteration, avoiding model update deviations caused by improper selection of participating ends. Each participating end generates adaptive weights based on the size of the second dataset and the prediction confidence. Participating ends with larger data volumes and more certain predictions receive higher weights, effectively suppressing the interference of low-quality data or malicious annotations on global updates. Cluster-weighted fusion after spectral clustering further amplifies the consensus information within clusters in the same gradient direction, making the global update direction more robust and improving the robustness and convergence efficiency of the federated learning system. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a call center operational risk prediction method based on federated learning, comprising the following steps: S1: The central end initializes a global risk identification base parameter, which is used to characterize general knowledge of cross-regional call center operational risks; S2: The central terminal selects several participating terminals from multiple participating terminals to form a task set, and distributes the current global risk identification base parameters to each selected participating terminal respectively; S3: After each selected participant receives the global risk identification base parameters, it uses the local first dataset to perform an inner loop update. During the update process, it introduces elastic regularization constraints based on the difference between the local data distribution and the historical global distribution to retain key responses of general risk knowledge, thereby obtaining local risk identification parameters that are adapted to local dialect characteristics and business risk distribution. S4: Each selected participant uses its local second dataset to evaluate the performance of local risk identification parameters, calculates the meta-gradient of the global risk identification basis parameters based on the evaluation loss, and generates adaptive weights based on the size of the second dataset and the label confidence. The meta-gradient and adaptive weights are then uploaded to the central terminal. S5: The central terminal receives the meta-gradients and adaptive weights uploaded by all selected participating terminals, performs directional similarity clustering on the meta-gradients, performs weighted fusion according to the adaptive weights by cluster, and updates the global risk identification basis parameters based on the fusion results, so that the updated global risk identification basis parameters can be quickly adapted locally by each participating terminal in the next iteration to more accurately predict the operational risks of different regions.

[0020] In S1, the central terminal initializes a global risk identification base parameter. This global risk identification base parameter is used to characterize general knowledge of cross-regional call center operational risks, specifically including: The first step is for the central terminal to collect historical call text data from each participating terminal. Specifically, the central terminal establishes a connection with the local database of each participating terminal through a secure interface to obtain the transcribed text of all customer service call recordings within a preset time period (e.g., the past 12 months). During the collection process, the central terminal performs anonymization processing on the received raw text data, removing personal privacy information such as customer names, ID numbers, and contact information. Simultaneously, it performs text cleaning, removing blank characters, meaningless interjections, and garbled symbols caused by transcription errors, resulting in standardized historical call text data.

[0021] The second step involves the central endpoint extracting location-independent risk semantic units from the historical call text data. This is achieved by first accessing a pre-defined risk keyword lexicon, which contains at least three categories of words: the first category is monetary values, achieved by matching all consecutive numerical combinations in the text using regular expressions and labeling their values; the second category is emotional intensity values, achieved by loading a pre-trained Chinese sentiment dictionary and scoring each call text for emotional intensity, using the absolute difference between positive and negative sentiment as the emotional intensity value of that sentence; the third category is high-frequency fraud keywords, achieved by statistically analyzing the frequency of words in all historical call texts and selecting words whose frequency is significantly higher than normal samples in confirmed fraud samples, thus forming a set of high-frequency fraud keywords. The central endpoint then extracts these three types of features from each call text to form the initial set of risk semantic units.

[0022] The third step involves the central unit performing distribution statistics on the risk semantic units and constructing an initial general knowledge vector based on the statistical results. The specific calculation method is as follows: For monetary values, the mean and variance are calculated across all samples, and the normalized mean and variance are concatenated into two dimensions; for emotional intensity values, the average emotional intensity value across all samples and the proportion of samples with emotional intensity values ​​exceeding a preset threshold are calculated, and these two values ​​are used as two dimensions; for high-frequency fraud keywords, the total number of times each keyword appears in all samples is counted, forming a word frequency vector with a dimension equal to the number of high-frequency fraud keywords. The central unit sequentially concatenates all the statistically obtained dimensions to obtain a fixed-length numerical vector, which is the initial general knowledge vector. The value of each dimension reflects the overall distribution of a certain risk feature across geographical regions.

[0023] Fourth, the central endpoint sets a learnable parameter matrix based on the initial general knowledge vector, using this parameter matrix as the basis parameters for global risk identification. Specifically, the central endpoint creates an empty matrix with the same dimensions as the initial general knowledge vector, and sequentially fills each value of the initial general knowledge vector into the corresponding position of the empty matrix, serving as its initial value. During subsequent federated learning iterations, this parameter matrix will be continuously adjusted based on the meta-gradients uploaded by each participating endpoint, thereby gradually optimizing its adaptability to regionally specific risks while retaining cross-regional general risk knowledge.

[0024] In S2, the central terminal selects several participating terminals from multiple participating terminals to form a task set, and distributes the current global risk identification base parameters to each selected participating terminal, specifically including: The first step involves the central terminal receiving the local data risk entropy values ​​uploaded by each participating terminal. Specifically, each participating terminal pre-calculates the risk entropy value of its historical call text data locally and uploads this value to the central terminal via an encrypted channel. The calculation method for the local data risk entropy value is as follows: First, the participating terminal counts the occurrence frequency of each risk keyword in a preset risk keyword library within its local historical call texts. This risk keyword library is consistent with that of the central terminal and contains at least 100 words related to fraud, complaints, and operational errors. Then, it calculates the proportion of each risk keyword's occurrence frequency to the total occurrence frequency of all risk keywords, recording this proportion as the first probability. Next, it multiplies each first probability by the logarithm of the first probability to base 2, sums the products, and takes the negative of the sum to obtain the preliminary entropy value. Finally, the participating terminal counts the total number of all historical call texts and multiplies the preliminary entropy value by the natural logarithm of the total number of texts to obtain the final local data risk entropy value. The larger this value, the more chaotic the distribution of risk keywords and the higher the uncertainty in the participating terminal's local data.

[0025] The second step is for the central end to sort each participating end according to the local data risk entropy value, divide it into several gradient layers according to the entropy value from high to low, and randomly select one participating end from each gradient layer as the selected participating end. The specific implementation is as follows: The central end sorts all participating ends in descending order of their uploaded local data risk entropy values ​​to obtain a sorted list; the central end presets a gradient layer number, which is set to 3 layers; the central end calculates the total length of the sorted list, divides the total length by the number of gradient layers to obtain the baseline number of each layer, and divides the sorted list into the first gradient layer, the second gradient layer, and the third gradient layer from top to bottom according to the baseline number. The first gradient layer contains several participating ends with the highest entropy values, and the third gradient layer contains several participating ends with the lowest entropy values. If the total length is not divisible, priority is given to ensuring that the first and second gradient layers have complete baseline numbers, and the remaining participating ends are assigned to the third gradient layer; for each gradient layer, the central end uses a random number generation algorithm to extract one participating end from that layer, finally obtaining 3 selected participating ends, which represent the data distribution characteristics of the high entropy value region, the medium entropy value region, and the low entropy value region, respectively.

[0026] Third, the central terminal distributes the current global risk identification basis parameters to each selected participant. Specifically, the central terminal reads the global risk identification basis parameters stored in the current iteration round; these parameters are a multi-dimensional matrix. The central terminal establishes a secure communication link with the three selected participants and sends a complete copy of the global risk identification basis parameters to each participant. Each selected participant, upon receiving the parameter matrix, stores it in its local memory for subsequent inner loop update steps. Through this stratified random sampling method, the central terminal can obtain broadly representative gradient information in each iteration with minimal communication overhead, avoiding model update bias caused by improper participant selection.

[0027] In S3, each selected participant receives the global risk identification base parameters and performs an inner loop update using its local first dataset. During the update process, elastic regularization constraints are introduced based on the difference between the local data distribution and the historical global distribution to retain key responses that preserve general risk knowledge. This results in local risk identification parameters adapted to local dialect characteristics and business risk distribution, specifically including: In the first step, after each selected participant receives the global risk identification basis parameters, it randomly selects a support subset from the local first dataset, and performs a first round of gradient descent on the global risk identification basis parameters based on the support subset to obtain the first intermediate parameters. The specific implementation is as follows: The selected participant divides the locally stored historical call text data into a first dataset and a second dataset according to a preset ratio. The first dataset is used for inner loop updates, and the second dataset is used for performance evaluation. The selected participant randomly selects a preset number of samples from the first dataset, which is set to 32, to form a support subset. The selected participant uses the global risk identification basis parameters as initial values, inputs each sample in the support subset into a preset risk prediction network to obtain a predicted risk value, and calculates the cross-entropy loss value between the predicted risk value and the sample's true risk label. The selected participant differentiates the global risk identification basis parameters based on the cross-entropy loss value to obtain a gradient value, multiplies the gradient value by a preset inner loop learning rate, which is set to 0.01, and then subtracts the product result from the global risk identification basis parameters to obtain the updated parameter matrix. This parameter matrix is ​​recorded as the first intermediate parameter.

[0028] The second step involves each selected participant calculating the parameter difference between the first intermediate parameter and the baseline parameters of the previous round of global risk identification, and constructing an elastic mask matrix based on the magnitude of each dimension of the parameter difference. The specific implementation of this step will be described in detail in subsequent sub-steps.

[0029] The third step involves each selected participant extracting a query subset from the local first dataset, and performing a second round of gradient descent on the first intermediate parameter based on the query subset and using the elastic mask matrix as the regularization term weight, to obtain local risk identification parameters that are adapted to local dialect features and business risk distribution. The specific implementation is as follows: The selected participant randomly selects a preset number of samples from the first dataset, which is set to 32, to form a query subset, ensuring that the query subset and the support subset do not overlap; the selected participant uses the first intermediate parameter as the initial value, inputs each sample in the query subset into the risk prediction network to obtain the predicted risk value, and calculates the cross-entropy loss value between the predicted risk value and the sample's true risk label; simultaneously, the selected participant calculates the squared difference between the first intermediate parameter and the global risk identification basis parameter, multiplies the squared difference by the value of the corresponding dimension in the elastic mask matrix to obtain the regularization term; the selected participant adds the cross-entropy loss value to the regularization term to obtain the comprehensive loss value, where the weight coefficient of the regularization term is set to 0.001; the selected participant differentiates the first intermediate parameter based on the comprehensive loss value to obtain the second gradient, multiplies the second gradient by the inner loop learning rate, and then subtracts the product result from the first intermediate parameter to obtain the updated parameter matrix, which is used as the local risk identification parameter.

[0030] The step of constructing an elastic mask matrix based on the magnitude of each dimension of the parameter difference values ​​specifically includes the following sub-steps: Each selected participant calculates the dimension-wise difference between the first intermediate parameter and the previous round's global risk identification baseline parameters, and then converts the absolute value of the dimension-wise difference to obtain a parameter difference vector. Specifically, the selected participant reads each value in the first intermediate parameter matrix and each value at the corresponding position in the global risk identification baseline parameter matrix received from the central terminal in the previous round. The two values ​​are subtracted to obtain the difference. The absolute value of this difference is taken, and all absolute values ​​are arranged in their original dimensional order to form a one-dimensional vector. This vector is the parameter difference vector, and its length is equal to the total number of dimensions of the global risk identification baseline parameters.

[0031] Each selected participant acquires the dispersion of dialect features in the first local dataset, calculates an elastic adjustment coefficient based on the dispersion of the dialect features, and multiplies each element of the parameter difference vector by the elastic adjustment coefficient to obtain a scaled difference vector. Specifically, the selected participant statistically analyzes all call text data in the first dataset, extracts the dialect features of each text, and loads a preset dialect recognition lexicon containing typical words from dialects such as Cantonese, Minnan, and Sichuanese. The participant calculates the proportion of dialect words appearing in each text to the total number of words, and calculates the variance of this proportion for all texts; the variance obtained is the dispersion of the dialect features. The selected participant adds a very small constant of 0.0001 to the dispersion, takes the reciprocal, and multiplies the reciprocal by a preset scaling factor set to 0.1 to obtain the elastic adjustment coefficient. The selected participant multiplies each element of the parameter difference vector by the elastic adjustment coefficient to obtain a new vector, which is the scaled difference vector.

[0032] Each selected participant inputs the scaled difference vector into a monotonically increasing saturation function. The output value of the saturation function is between 0 and 1. The output values ​​of the saturation function are arranged according to their original dimensions to form the elastic mask matrix. Specifically, the selected participant uses the Sigmoid function as the saturation function. For each element in the scaled difference vector, the exponent of the negative number of that element with the natural constant e as the base is calculated. The result of adding 1 to the exponent and taking the reciprocal yields the mask value corresponding to that element. This mask value is always between 0 and 1; when the element value is large, the mask value approaches 1, and when the element value is small, the mask value approaches 0. The selected participant rearranges the mask values ​​corresponding to all elements according to their original dimensional order into a matrix. The shape of this matrix is ​​exactly the same as the global risk identification basis parameters. This matrix is ​​used as the elastic mask matrix. In the second round of gradient descent, the dimensions with larger values ​​in this matrix will impose a strong regularization constraint on the update of the first intermediate parameter, thereby preserving those parameter dimensions that are crucial to general risk knowledge.

[0033] In S4, each selected participant uses its local second dataset to evaluate the performance of its local risk identification parameters. Based on the evaluation loss, it calculates the meta-gradient with respect to the global risk identification basis parameters. Simultaneously, it generates adaptive weights based on the size of the second dataset and the label confidence level. The meta-gradient and adaptive weights are then uploaded to the central endpoint. Specifically, this includes: The first step involves each selected participant performing forward computation on the second dataset using the local risk identification parameters to obtain the cross-entropy between the predicted risk value and the actual risk label. This cross-entropy is then used as the evaluation loss. Specifically, the selected participant reads a pre-divided second dataset from local storage. This dataset contains several call text samples, each labeled with an actual operational risk label. The label value is either 0 or 1, where 0 indicates no risk and 1 indicates risk. The selected participant uses the local risk identification parameters as weights for the current risk prediction network. Each sample in the second dataset is sequentially input into the network, which outputs a predicted risk value between 0 and 1. For each sample, the selected participant calculates the natural logarithm of the actual risk label multiplied by the predicted risk value, then adds 1 and subtracts the natural logarithm of the actual risk label multiplied by 1 and the predicted risk value. The result is then inversely represented to obtain the cross-entropy loss value for that sample. The selected participant sums the cross-entropy loss values ​​of all samples and divides the sum by the total number of samples in the second dataset to obtain the average cross-entropy loss value. This average cross-entropy loss value is then used as the evaluation loss.

[0034] The second step involves each selected participant differentiating the local risk identification parameters based on the assessed loss to obtain a first gradient. Then, using the chain rule, this first gradient is backtracked to the global risk identification basis parameters to obtain the meta-gradient. The specific implementation of this step will be detailed in subsequent sub-steps.

[0035] The third step involves counting the number of samples in the second dataset for each selected participant as the scale base, and simultaneously calculating the average probability entropy of the predicted risk value on the second dataset as the inverse of the confidence level. The scale base is then divided by the inverse of the confidence level, and the logarithm is taken to obtain the adaptive weight. The specific implementation is as follows: The selected participant counts the total number of samples in the second dataset and records this value as the scale base. For each sample in the second dataset, the selected participant takes the predicted risk value of the sample as the first probability, subtracts the predicted risk value from 1 to obtain the second probability, calculates the first probability multiplied by the logarithm of the first probability to the base 2, adds the second probability multiplied by the logarithm of the first probability to the base 2, and takes the opposite of the sum to obtain the predicted entropy value of the sample. The selected participant adds the predicted entropy values ​​of all samples and divides them by the total number of samples to obtain the average predicted entropy value. The larger the average predicted entropy value, the more uncertain the model is about the prediction result, and it is used as the reciprocal of the confidence level. The selected participant divides the scale base by the reciprocal of the confidence level to obtain an intermediate ratio, and then calculates the logarithm of this intermediate ratio to the base of the natural constant e. The calculation result is used as the adaptive weight. This weight value increases with the increase of the number of samples and decreases with the increase of prediction uncertainty.

[0036] Fourth, each selected participant multiplies the meta-gradient by the adaptive weight and encapsulates it into an upload data packet, which is then sent to the central terminal. Specifically, the selected participant multiplies each value in the meta-gradient matrix calculated in the second step by the adaptive weight calculated in the third step to obtain a weighted meta-gradient matrix. The selected participant combines the weighted meta-gradient matrix with a preset header, which includes at least the participant's identifier and the current iteration round number. The selected participant then sends the combined data to the central terminal via an encrypted communication protocol for aggregation processing in subsequent steps.

[0037] The specific process of differentiating the local risk identification parameters based on the evaluation loss and backtracking to obtain the meta-gradient includes the following sub-steps: Step 4-2-1: Each selected participant calculates the partial derivative of the evaluation loss with respect to the local risk identification parameters to obtain the first gradient. Simultaneously, it records the gradient propagation path of the global risk identification basis parameters upon which the local risk identification parameters depend during the update process. Specifically, the selected participant uses automatic differentiation to backpropagate from the evaluation loss, calculating the partial derivative of the evaluation loss with respect to each value in the local risk identification parameter matrix. All partial derivatives are arranged according to their original dimensions to obtain the first gradient matrix. Simultaneously, the selected participant traces back the generation process of the local risk identification parameters, starting from the global risk identification basis parameters, obtaining the first intermediate parameter through the first round of gradient descent on the support subset, and then obtaining the local risk identification parameters through the second round of gradient descent on the query subset. All computational operations involved in this process and their dependencies are recorded, forming a computational graph path. This path contains each layer of mapping from the global risk identification basis parameters to the local risk identification parameters.

[0038] Step 4-2-2: Each selected participant applies a path decay coefficient to the first gradient according to the gradient propagation path. The path decay coefficient is negatively correlated with the step size and iteration number of the inner loop update. The first gradient after decay is used as the intermediate backtracking gradient. Specifically, the selected participant counts the total number of iterations of the first and second rounds of gradient descent during the inner loop update process, which is set to 2. The value of the inner loop learning rate is read and set to 0.01. The path decay coefficient is calculated by dividing 1 by the 1 in parentheses and adding the product of the inner loop learning rate and the total number of iterations. The selected participant multiplies each value in the first gradient matrix by the path decay coefficient to obtain the decayed gradient matrix, which is recorded as the intermediate backtracking gradient.

[0039] Step 4-2-3: Each selected participant propagates the intermediate backtracking gradient backward layer by layer along the gradient propagation path, stopping when it reaches the global risk identification basis parameter. The gradient value at the stopping position is taken as the meta-gradient. Specifically, the selected participant starts with the intermediate backtracking gradient and, following the computational graph path recorded in step 4-2-1, propagates the intermediate backtracking gradient backward to each intermediate variable, starting from the local risk identification parameter and sequentially performing the second and first rounds of gradient descent operations. During propagation, each gradient descent operation requires a corresponding transformation of the gradient based on the support subset or query subset data used in that operation. When the propagation path backtracks to the node containing the global risk identification basis parameter, propagation stops, and the accumulated gradient value at that node is the final backtracking result. The selected participant arranges all gradient values ​​at this node according to their original dimensions to obtain a matrix with the same shape as the global risk identification basis parameter, and uses this matrix as the meta-gradient. This meta-gradient reflects the degree of indirect influence of the evaluation loss on the local second dataset on the global risk identification basis parameter through the inner loop update process.

[0040] In S5, the central terminal receives the meta-gradients and adaptive weights uploaded by all selected participating terminals. It performs directional similarity clustering on the meta-gradients, then performs weighted fusion based on the adaptive weights within each cluster. The global risk identification basis parameters are updated based on the fusion results. This allows the updated global risk identification basis parameters to be quickly adapted locally by each participating terminal in the next iteration, enabling more accurate prediction of operational risks in different regions. Specifically, this includes: The first step involves the central endpoint receiving the meta-gradients and adaptive weights uploaded by all selected participants, calculating the cosine similarity between every two meta-gradients, and constructing a similarity matrix based on the cosine similarity. Specifically, at the end of each iteration, the central endpoint waits for all selected participants to return their uploaded data packets, from which it parses the meta-gradient matrix and adaptive weight scalar values ​​corresponding to each participant. Assuming there are M selected participants in this round, the central endpoint obtains M meta-gradient matrices. The dimension of each matrix is ​​exactly the same as the global risk identification basis parameters, denoted as a D-dimensional vector (in practice, the matrix is ​​flattened into a one-dimensional vector). The central endpoint flattens each meta-gradient matrix into a numerical vector of length D, denoted as... ,in From 1 to M. For any two distinct participants... and The cosine similarity between their meta-gradient vectors is calculated as follows: First, calculate the dot product of the two vectors (multiply the corresponding dimensions and sum them). Then, calculate the magnitudes of the two vectors (sum of the squares of each dimension and take the square root). Finally, divide the dot product by the product of the two magnitudes. The result is the cosine similarity, which ranges from -1 to 1. The closer the value is to 1, the more consistent the gradient directions are. This calculation process can be expressed mathematically as follows:

[0041] in, and They represent the first The and the first The meta-gradient vector of each participating end, This represents the dot product of the two. Representing vectors The modulus of the matrix is ​​denoted as . The cosine similarity calculation results between all pairs of elements are filled into an M x M symmetric matrix at the center end. Line number The column element is the first one. The and the first The cosine similarity of the individual gradients is calculated, with the diagonal elements set to 1. This matrix is ​​the similarity matrix, denoted as S.

[0042] The second step involves the central terminal inputting the similarity matrix into a spectral clustering algorithm. Based on a preset number of clusters, the meta-gradients are divided into several gradient clusters, ensuring that the meta-gradient directions within the same cluster are similar, while the meta-gradient directions between different clusters are different. The specific implementation of this step will be detailed in subsequent sub-steps.

[0043] The third step involves the central end performing a weighted sum of all meta-gradients within each gradient cluster, based on their corresponding adaptive weights, to obtain the cluster-representative gradient. Then, all cluster-representative gradients are weighted and summed to obtain the global update direction. Specifically, after spectral clustering, M meta-gradients are divided into K clusters (K is the preset number of clusters, for example, 3), and each cluster contains several participating ends. For the k-th cluster, let its set of participating end indices be... Then calculate the cluster representative gradient of the cluster. For: the meta-gradient vector of each participating end within the cluster Multiply by its corresponding adaptive weight Then sum them up, and divide by the sum of all adaptive weights within the cluster to obtain the weighted average gradient. That is... The global update direction u is equal to the sum of the products of the gradient vectors of each element within the cluster and their corresponding adaptive weights, divided by the sum of the adaptive weights within the cluster. After obtaining K representative gradients for each cluster, the central end sums them with equal weights, i.e., calculates the average of all representative gradients for each cluster, resulting in the global update direction u, which is equal to the sum of all representative gradients for each cluster divided by K. This global update direction integrates representative information from clusters with different gradient directions, avoiding the dominance of a single cluster in the update.

[0044] The fourth step involves the central endpoint multiplying the global update direction by a preset learning rate and then adding it to the current global risk identification basis parameters to obtain the updated global risk identification basis parameters. Specifically, the central endpoint reads the preset global learning rate η, which is set to 0.1. Each value in the global update direction vector u is multiplied by η to obtain the update vector. Then, the current global risk identification basis parameter vector (i.e., the parameter values ​​after the previous iteration) is added to the corresponding dimension of the update vector to obtain a new parameter vector. This vector is then rearranged into a matrix according to its original dimensions, which is the updated global risk identification basis parameter, used for distribution in the next iteration.

[0045] The central end inputs the similarity matrix into the spectral clustering algorithm, and divides the meta-gradient into several gradient clusters according to the preset number of clusters. Specifically, it includes the following sub-steps: Step 5-2-1: The central end performs row normalization on the similarity matrix to obtain a normalized similarity matrix. Based on the normalized similarity matrix, a diagonal matrix is ​​calculated, where the diagonal elements are the sum of the elements in each row of the normalized similarity matrix. Specifically, the central end performs row normalization on the similarity matrix S, that is, for each row, each element of that row is divided by the sum of all elements in that row, making the sum of each row equal to 1. Let the first element of S be... Line number Column elements are Then the normalized elements are Divide by the sum of all elements in that row to obtain the normalized similarity matrix. Then, compute a diagonal matrix D, whose diagonal elements... It equals the sum of all elements in the i-th row of the normalized similarity matrix, i.e. For the first The result of adding the elements in each row and column.

[0046] Step 5-2-2: At the center end, subtract the normalized similarity matrix from the diagonal matrix to obtain the Laplacian matrix. Calculate the eigenvectors corresponding to the first few smallest eigenvalues ​​of the Laplacian matrix, and arrange the eigenvectors column-wise to form an eigenvector matrix. Specifically, this is implemented by constructing the Laplacian matrix. Where D is a diagonal matrix, To normalize the similarity matrix, L is an M x M symmetric matrix. Next, the center needs to calculate the eigenvalues ​​and eigenvectors of L. Since M is usually small (e.g., dozens of participating ends), a numerical method can be used. The center calls an eigenvalue decomposition algorithm to obtain all the eigenvalues ​​of L, arranged in ascending order, and simultaneously obtains the corresponding eigenvectors. The eigenvectors corresponding to the first K smallest eigenvalues ​​(excluding 0) are selected, where K is the preset number of clusters. Let these K eigenvectors be... , , ..., Each of them is a column vector of length M. Arranging them column-wise side-by-side forms an eigenvector matrix V with M rows and K columns. This eigenvalue decomposition process can be represented by solving the following equation: Where L is the Laplace matrix, It is the k-th smallest eigenvalue. These are the corresponding eigenvectors. The eigenvector matrix V is the first eigenvector of the eigenvector matrix. The element in the kth column is the element in the kth row. The component of the gradient of each element on the k-th eigenvector.

[0047] Step 5-2-3: The central end treats each row of the feature vector matrix as a low-dimensional embedding representation of a meta-gradient. Based on the Euclidean distance between the low-dimensional embedding representations, a greedy iterative strategy is used to divide the meta-gradients into a predetermined number of gradient clusters, ensuring that meta-gradients within the same cluster are adjacent in the low-dimensional embedding space, while meta-gradients between different clusters are far apart. Specifically, each row of the feature vector matrix V corresponds to the coordinates of a participating end's meta-gradient in K-dimensional space, i.e., the... row vector Indicates the first Low-dimensional embedding of individual gradients. Now, these M points need to be divided into K clusters. Iterative clustering is performed using the K-means algorithm: First, K row vectors are randomly selected as initial cluster centers; then, for each row vector, the Euclidean distance to each cluster center is calculated. The Euclidean distance is calculated as follows for two K-dimensional vectors... and The process involves calculating the square root of the sum of squares of the differences in each dimension; assigning each point to the nearest cluster center; then recalculating the center of each cluster by taking the mean of all points within the cluster; repeating this assignment and update process until the cluster centers no longer change or the maximum number of iterations (e.g., 100) is reached. Ultimately, each meta-gradient is assigned to a cluster, with points within the same cluster being closer in low-dimensional space and points in different clusters being farther apart. This cluster partitioning reflects the similarity grouping along the meta-gradient direction, providing a basis for subsequent cluster-weighted fusion.

[0048] The working principle of this invention is as follows: The central terminal initializes a global risk identification base parameter to characterize general knowledge of cross-regional call center operational risks; the central terminal performs stratified random sampling based on the risk entropy values ​​of local data uploaded by each participating terminal, selects several participating terminals to form a task set, and distributes the current global risk identification base parameter to each selected participating terminal; after receiving the base parameter, each selected participating terminal performs an inner loop update using the local first dataset, and during the update process, based on the difference between the local data distribution and the historical global distribution, constructs an elastic mask matrix by calculating the parameter difference vector and combining it with the dispersion of local dialect features, and uses this matrix as the weight of the regularization term to perform a second round of gradient descent, obtaining local risk identification parameters adapted to local dialect features and business risk distribution; each selected participating terminal uses the local first dataset... The two datasets are used to evaluate the performance of the local risk identification parameters. The first gradient is calculated based on the evaluation loss and backtracked along the gradient propagation path to the global risk identification basis parameters to obtain the meta-gradient. At the same time, adaptive weights are generated based on the size of the second dataset and the average probability entropy of the predicted risk values. The meta-gradient is multiplied by the adaptive weights and then uploaded to the central end. After receiving all meta-gradients and adaptive weights, the central end calculates the cosine similarity between each pair to construct a similarity matrix. Spectral clustering is performed on this matrix to divide the meta-gradients into several gradient clusters. The gradients of each cluster are weighted and fused according to the adaptive weights to obtain the representative gradients of each cluster. Then, the representative gradients of all clusters are superimposed with equal weights to obtain the global update direction. Finally, the global update direction is multiplied by the learning rate and added to the current global risk identification basis parameters to obtain the updated basis parameters for the next iteration.

[0049] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for predicting operational risks in call centers based on federated learning, characterized in that, Includes the following steps: S1: The central end initializes a global risk identification base parameter, which is used to characterize general knowledge of cross-regional call center operational risks; S2: The central terminal selects several participating terminals from multiple participating terminals to form a task set, and distributes the current global risk identification base parameters to each selected participating terminal respectively; S3: After each selected participant receives the global risk identification base parameters, it uses the local first dataset to perform an inner loop update. During the update process, it introduces elastic regularization constraints based on the difference between the local data distribution and the historical global distribution to retain key responses of general risk knowledge, thereby obtaining local risk identification parameters that are adapted to local dialect characteristics and business risk distribution. S4: Each selected participant uses its local second dataset to evaluate the performance of local risk identification parameters, calculates the meta-gradient of the global risk identification basis parameters based on the evaluation loss, and generates adaptive weights based on the size of the second dataset and the label confidence. The meta-gradient and adaptive weights are then uploaded to the central terminal. S5: The central terminal receives the meta-gradients and adaptive weights uploaded by all selected participating terminals, performs directional similarity clustering on the meta-gradients, performs weighted fusion according to the adaptive weights by cluster, and updates the global risk identification basis parameters based on the fusion results, so that the updated global risk identification basis parameters can be quickly adapted locally by each participating terminal in the next iteration to more accurately predict the operational risks of different regions.

2. The method for predicting call center operational risks based on federated learning according to claim 1, characterized in that, S1 specifically includes: The central terminal collects historical call text data from each participating terminal and extracts location-independent risk semantic units from the historical call text data. Risk semantic units include monetary figures, emotional intensity values, and high-frequency fraud keywords. The central end performs distribution statistics on risk semantic units and constructs an initial general knowledge vector based on the statistical results. The initial general knowledge vector is used to describe risk patterns that co-occur across regions. The central terminal sets a learnable parameter matrix based on the initial general knowledge vector, and uses the parameter matrix as the basis parameter for global risk identification.

3. The method for predicting call center operational risks based on federated learning according to claim 1, characterized in that, S2 specifically includes: The central terminal receives the local data risk entropy value uploaded by each participating terminal. The local data risk entropy value is calculated based on the frequency of occurrence and the degree of disorder in the distribution of risk keywords in the historical call text of each participating terminal. The central end sorts the participating ends according to the risk entropy value of local data, divides them into several gradient layers according to the entropy value from high to low, and randomly selects one participating end from each gradient layer as the selected participating end. The central terminal will distribute the current global risk identification base parameters to each selected participating terminal.

4. The method for predicting call center operational risks based on federated learning according to claim 1, characterized in that, S3 specifically includes: After each selected participant receives the global risk identification basis parameters, it randomly selects a support subset from the local first dataset, and performs the first round of gradient descent on the global risk identification basis parameters based on the support subset to obtain the first intermediate parameters. Each selected participant calculates the parameter difference value between the first intermediate parameter and the global risk identification base parameter of the previous round, and constructs an elastic mask matrix based on the size of each dimension of the parameter difference value. Each selected participant extracts a query subset from the local first dataset, and performs a second round of gradient descent on the first intermediate parameter based on the query subset and using the elastic mask matrix as the weight of the regularization term, to obtain local risk identification parameters that are adapted to local dialect features and business risk distribution.

5. The call center operational risk prediction method based on federated learning according to claim 4, characterized in that, The construction of an elastic mask matrix based on the magnitude of each dimension of the parameter difference values ​​specifically includes: Each selected participant calculates the dimension-wise difference between the first intermediate parameter and the global risk identification baseline parameter of the previous round, and obtains the parameter difference vector by converting the dimension-wise difference into absolute value. Each selected participant obtains the dispersion of dialect features in the first local dataset, calculates an elastic adjustment coefficient based on the dispersion of dialect features, and multiplies each element in the parameter difference vector with the elastic adjustment coefficient to obtain a scaled difference vector. Each selected participant inputs a scaled difference vector into a monotonically increasing saturation function. The output value of the saturation function is between 0 and 1. The output values ​​of the saturation function are arranged according to the original dimensions to form an elastic mask matrix.

6. The method for predicting call center operational risks based on federated learning according to claim 1, characterized in that, S4 specifically includes: Each selected participant performs forward computation on the second dataset to obtain the cross-entropy between the predicted risk value and the true risk label, and uses the cross-entropy as the evaluation loss. Each selected participant differentiates the local risk identification parameters based on the assessed loss to obtain the first gradient, and then uses the chain rule to backtrack the first gradient to the global risk identification basis parameters to obtain the meta-gradient. The number of samples in the second dataset is used as the base size for each selected participant. At the same time, the average probability entropy of the predicted risk value on the second dataset is calculated as the inverse of the confidence level. The base size is divided by the inverse of the confidence level and the logarithm is taken to obtain the adaptive weight. Each selected participant multiplies the meta-gradient by adaptive weights and encapsulates it into an upload data packet, which is then sent to the central endpoint.

7. The call center operational risk prediction method based on federated learning according to claim 6, characterized in that, The specific process for obtaining the meta-gradient is as follows: Each selected participant will evaluate the loss by calculating the partial derivative of the local risk identification parameters to obtain the first gradient, and at the same time record the gradient propagation path of the global risk identification basis parameters on which the local risk identification parameters depend during the update process. Each selected participant applies a path decay coefficient to the first gradient based on the gradient propagation path. The path decay coefficient is negatively correlated with the step size and number of iterations of the inner loop update. The first gradient after applying the decay is used as the intermediate backtracking gradient. Each selected participant propagates the intermediate backtracking gradient backward along the gradient propagation path layer by layer, stopping when it reaches the global risk identification basis parameters, and taking the gradient value at the stopping position as the meta-gradient.

8. The method for predicting call center operational risks based on federated learning according to claim 1, characterized in that, S5 specifically includes: The central end receives the meta-gradients and adaptive weights uploaded by all selected participating ends, calculates the cosine similarity between every two meta-gradients, and constructs a similarity matrix based on the cosine similarity. The central end inputs the similarity matrix into the spectral clustering algorithm, and divides the meta-gradient into several gradient clusters according to the preset number of clusters, so that the meta-gradient directions within the same cluster are similar and the meta-gradient directions between different clusters are different. For each gradient cluster, the central end performs a weighted sum of all the meta-gradients within the cluster according to the corresponding adaptive weights to obtain the cluster representative gradient of the corresponding cluster. Then, the cluster representative gradients are superimposed with equal weights to obtain the global update direction. The central end multiplies the global update direction by the preset learning rate and adds it to the current global risk identification basis parameters to obtain the updated global risk identification basis parameters.

9. The method for predicting call center operational risks based on federated learning according to claim 8, characterized in that, The central end inputs the similarity matrix into the spectral clustering algorithm, and divides the meta-gradient into several gradient clusters according to the preset number of clusters, so that the meta-gradient directions within the same cluster are similar and the meta-gradient directions between different clusters are different, specifically including: The center performs row normalization on the similarity matrix to obtain a normalized similarity matrix. Based on the normalized similarity matrix, a diagonal matrix is ​​calculated. The diagonal elements of the diagonal matrix are the sum of the elements of each row of the normalized similarity matrix. The center subtracts the normalized similarity matrix from the diagonal matrix to obtain the Laplacian matrix. The eigenvectors corresponding to the first few smallest eigenvalues ​​of the Laplacian matrix are calculated, and the eigenvectors are arranged in columns to form an eigenvector matrix. The central end treats each row of the feature vector matrix as a low-dimensional embedding representation of a meta-gradient. Based on the Euclidean distance between the low-dimensional embedding representations, a greedy iterative strategy is used to divide the meta-gradient into a preset number of gradient clusters, so that the meta-gradients within the same cluster are adjacent in the low-dimensional embedding space, while the meta-gradients between different clusters are far apart.

Images