Financial service volume sequence value determination method and apparatus, and electronic device

Abstract

The present application relates to the field of financial technology, and discloses a financial service volume sequence value determination method and apparatus, and an electronic device. The method comprises: decomposing service volume driving sequences into different frequency domain spaces to obtain intrinsic mode functions; in a combined attention layer of each group of encoders, performing encoding operations on the intrinsic mode functions, using LSTM units to combine encoding hidden state information of a previous moment with the intrinsic mode functions, and outputting the hidden state of the current moment; updating the hidden state of the current moment on the basis of mode parameters of the current moment, to obtain a hidden state matrix of the current moment; in a temporal attention layer of each decoder, on the basis of a context vector and a decoder hidden state of the previous moment, calculating an attention weight by means of a combined attention mechanism to obtain a weighted sum of the encoding hidden states of all time steps; and on the basis of the weighted sum of the encoding hidden states output by the last decoder, determining a service volume sequence value required by a financial service at the next time point.

Description

Method and device for determining financial service quantity sequence value, and electronic equipment

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on December 18, 2023, with application number 2023117474377 and application name “Method and device for determining the sequence value of financial service quantity, and electronic device”, all contents of which are incorporated by reference into this application. Technical Field

[0002] The present application relates to the field of financial technology or other related fields, and specifically, to a method and device for determining the sequence value of financial service volume, and an electronic device. Background Art

[0003] Financial institutions' customer service centers require a rational staffing structure. This ensures key queue management metrics are achieved while reducing staffing costs. However, service volume distribution is unstable and often fluctuates due to numerous factors. Service volume series are characterized by high noise, non-stationarity, multi-scale, and long memory, and exhibit complex relationships with other factors. Therefore, they can be considered a time series problem with multiple input features.

[0004] In related technologies, when processing time series problems with multiple feature inputs, the feature sequence modeling method is used, which makes it difficult to model the short-term / long-term dependencies within the time series and the correlations between sequences; the prediction error for non-stationary time series is large, and the results of the time series prediction model based on machine learning methods have a lag phenomenon. The commonly used deep learning methods are unable to describe the dynamic interdependence between sequences when processing multivariate time series problems.

[0005] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0006] The embodiments of the present application provide a method and device for determining the sequence value of financial service volume, and an electronic device, so as to at least solve the technical problem in the related art that the multi-feature time series indicating the arrangement of financial institution service volume cannot accurately obtain the mutual dependence between the sequences, resulting in poor arrangement effect.

[0007] According to one aspect of an embodiment of the present application, a method for determining a financial service volume sequence value is provided, comprising: receiving N service volume driven sequences associated with a target financial service, decomposing each of the service volume driven sequences into different frequency domain spaces using a multi-scale decomposition algorithm, obtaining a corresponding intrinsic mode function, and using the intrinsic mode function as input information for a first encoder in a codec set, wherein the service volume driven sequence is a sequence of multiple time points involving the target financial service in a historical process, the intrinsic mode function includes different fluctuation feature information in the service volume driven sequence, the codec set includes M groups of encoders and decoders, N is a positive integer greater than 1, and M is a positive integer; in the combined attention layer of each group of encoders, encoding the input intrinsic mode function, and using an LSTM unit to combine the encoded hidden state information of the previous moment with the input intrinsic mode function, Output the hidden state at the current moment; update the hidden state at the current moment according to the modal parameters at the current moment, and aggregate to obtain the hidden state matrix at the current moment, wherein the hidden state matrix at the current moment contains the importance of sequences at different time points and the time dependency within the sequence; in the time attention layer of the decoder, based on the input context vector and the decoder hidden state at the previous moment, calculate the attention weight by combining the attention mechanism to obtain the weighted sum of the encoded hidden states of all time steps, wherein the context vector is obtained by concatenating the weighted sum of the encoded hidden states at the historical time points with the weighted sum of the new hidden states, and the new hidden state weighted sum is obtained through the time dependency in the hidden state matrix and the sequence fluctuation characteristic information; based on the output amount indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group, determine the service volume sequence value required for the target financial business at the next time point.

[0008] In some embodiments, a multi-scale decomposition algorithm is used to decompose each of the service volume drive sequences into different frequency domain spaces to obtain the corresponding intrinsic mode function, including: step S1, for each of the service volume drive sequences, determining the maximum and minimum values ​​in the service volume drive sequence; step S2, using a preset interpolation polynomial method to construct the upper and lower envelopes of the service volume drive sequence according to the maximum and minimum values; step S3, calculating the local mean of the service volume drive sequence according to the upper and lower envelopes, and subtracting the local mean from the service volume drive sequence to obtain a new drive sequence; step S4, using the new drive sequence The service volume drive sequence is replaced by a new sequence, and steps S1 to S3 are repeatedly performed until the variance between the latest new drive sequence and the previous new drive sequence is less than a preset variance threshold, and the latest new drive sequence is defined as the intrinsic mode function of the service volume drive sequence; step S5, the intrinsic mode function is subtracted and separated from the service volume drive sequence to obtain a difference sequence; step S6, the service volume drive sequence is replaced by the difference sequence, and steps S1 to S5 are repeatedly performed until a termination condition is met, wherein the termination condition includes: the difference sequence satisfies monotonicity, or the difference sequence has only one zero point.

[0009] In some embodiments, an LSTM unit is used to combine the encoded hidden state information of the previous moment with the input intrinsic modal function to output the hidden state of the current moment, including: extracting the hidden state feature column vector from the intrinsic modal function using the LSTM unit, wherein the hidden state column vector contains forward and reverse information of the sequence; processing the hidden state feature column vector using the additive attention mechanism to obtain the importance of the sequence at different time points in the intrinsic modal function; constructing an initial hidden state matrix based on multiple hidden state feature column vectors, wherein the row vectors in the initial hidden state matrix represent the state vectors of the same input feature under T time steps, and the column vectors represent the state vectors of all input features under a single time step; performing enhancement processing on the row vectors of the initial hidden state matrix using the convolution kernel of a convolutional neural network, and using a combined attention mechanism to obtain the mutual dependencies within the initial hidden state matrix across time steps to obtain the time dependencies within the sequence, wherein the combined attention mechanism refers to a mechanism obtained by combining the additive attention mechanism with the temporal pattern attention mechanism.

[0010] In some embodiments, the steps of updating the hidden state at the current moment according to the modal parameters at the current moment and aggregating to obtain the hidden state matrix at the current moment include: determining the time pattern matrix based on the time dependency contained in the hidden state matrix at the current moment; extracting the row vectors in the time pattern matrix; using the inter-vector similarity measurement function to multiply the row vectors in the time pattern matrix and the corresponding weight coefficients to obtain the initial context vector; performing full connection processing on the initial context vector and the corresponding first change coefficient matrix, the hidden state at the current moment and the corresponding second change coefficient matrix, and aggregating to obtain the hidden state matrix at the current moment.

[0011] In some embodiments, the modal parameters at the current moment are pre-generated. When generating the modal parameters at the current moment, it includes: using a feedforward network containing a hidden layer to combine the encoder hidden state and cell state at the previous moment, calculating the distribution of attention, and obtaining an attention score; using a softmax function to normalize the attention score into a weight value; multiplying the weight value with the observed value of the intrinsic modal function at the current moment to obtain the modal parameters at the current moment.

[0012] In some embodiments, for the input context vector, it includes: integrating the time dependency in the hidden state matrix at the current moment and the fluctuation feature information in the intrinsic mode function, generating the new hidden state weighted sum through the time pattern attention mechanism; splicing the new hidden state weighted sum and the weighted sum of the encoded hidden states at the historical time points to obtain the spliced ​​context vector.

[0013] In some embodiments, the step of determining the service volume sequence value required for the target financial business at the next time point based on the output indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group includes: inputting N service volume driving sequences associated with the target financial business into an autoregressive model, and having the autoregressive model perform autoregressive processing on the N service volume driving sequences to obtain a service regression output; and accumulating the output indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group and the service regression output to obtain the service volume sequence value required for the target financial business at the next time point.

[0014] According to another aspect of an embodiment of the present application, a device for determining a sequence value of a financial service volume is provided, including: a sequence decomposition unit for receiving N service volume driven sequences associated with a target financial business, and using a multi-scale decomposition algorithm to decompose each of the service volume driven sequences into different frequency domain spaces to obtain a corresponding intrinsic mode function, and using the intrinsic mode function as input information of a first encoder in a codec set, wherein the service volume driven sequence is a sequence of multiple time points involving the target financial business in a historical process, and the intrinsic mode function contains different fluctuation feature information in the service volume driven sequence, and the codec set contains M groups of encoders and decoders, N is a positive integer greater than 1, and M is a positive integer; an encoding unit for encoding the input intrinsic mode function in the combined attention layer of each group of encoders, and using an LSTM unit to combine the encoded hidden state information of the previous moment with the input intrinsic mode function to output the current moment. Hidden state; a hidden aggregation unit, used to update the hidden state at the current moment according to the modal parameters at the current moment, and aggregate to obtain the hidden state matrix at the current moment, wherein the hidden state matrix at the current moment contains the importance of sequences at different time points and the time dependency within the sequence; a combined attention calculation unit, used to calculate the attention weight through a combined attention mechanism based on the input context vector and the decoder hidden state at the previous moment in the time attention layer of the decoder, to obtain the weighted sum of the encoded hidden states of all time steps, wherein the context vector is obtained by concatenating the weighted sum of the encoded hidden states at the historical time points with the weighted sum of the new hidden states, and the new hidden state weighted sum is obtained through the time dependency in the hidden state matrix and the sequence fluctuation characteristic information; a service volume determination unit, used to determine the service volume sequence value required for the target financial business at the next time point based on the output volume indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group.

[0015] In some embodiments, the sequence decomposition unit includes: a first determination module for executing step S1, for each of the service volume driven sequences, determining the maximum and minimum values ​​in the service volume driven sequence; a first construction module for executing step S2, for constructing the upper and lower envelopes of the service volume driven sequence according to the maximum and minimum values ​​using a preset interpolation polynomial method; a first calculation module for executing step S3, for calculating the local mean of the service volume driven sequence according to the upper and lower envelopes, and subtracting the local mean from the service volume driven sequence to obtain a new driving sequence; a first repetition module for executing step S4, for replacing the service volume driven sequence with the new driving sequence. A driving sequence is obtained, and steps S1 to S3 are repeatedly executed until the variance between the latest new driving sequence and the previous new driving sequence is less than a preset variance threshold, and the latest new driving sequence is defined as the intrinsic mode function of the service volume driving sequence; a first subtraction module is used to execute step S5, to subtract and separate the intrinsic mode function from the service volume driving sequence to obtain a differential sequence; a second repetition module is used to execute step S6, to replace the service volume driving sequence with the differential sequence, and to repeat steps S1 to S5 until a termination condition is met, wherein the termination condition includes: the differential sequence satisfies monotonicity, or the differential sequence has only one zero point.

[0016] In some embodiments, the encoding unit includes: a first extraction module, used to use an LSTM unit to extract the hidden state feature column vector in the intrinsic mode function, wherein the hidden state column vector contains forward and reverse information of the sequence; a first processing unit, used to use an additive attention mechanism to process the hidden state feature column vector to obtain the importance of sequences at different time points in the intrinsic mode function; a second construction module, used to construct an initial hidden state matrix based on multiple hidden state feature column vectors, wherein the row vector in the initial hidden state matrix represents the state vector of the same input feature under T time steps, and the column vector represents the state vector of all input features under a single time step; a second processing module, used to perform enhancement processing on the row vector of the initial hidden state matrix through the convolution kernel of the convolutional neural network, and use a combined attention mechanism to obtain the mutual dependence within the initial hidden state matrix across time steps to obtain the time dependence within the sequence, wherein the combined attention mechanism refers to a mechanism obtained by combining the additive attention mechanism and the temporal pattern attention mechanism.

[0017] In some embodiments, the hidden aggregation unit includes: a second determination module, used to determine the time pattern matrix based on the time dependency contained in the hidden state matrix at the current moment; a second extraction module, used to extract the row vectors in the time pattern matrix; a first multiplication module, used to use the inter-vector similarity measurement function to multiply the row vectors in the time pattern matrix and the corresponding weight coefficients to obtain the initial context vector; a third processing module, used to perform full connection processing on the initial context vector and the corresponding first change coefficient matrix, the hidden state at the current moment and the corresponding second change coefficient matrix, and aggregate to obtain the hidden state matrix at the current moment.

[0018] In some embodiments, the modal parameters at the current moment are pre-generated. When generating the modal parameters at the current moment, the device for determining the sequence value of the financial service volume includes: an attention calculation unit, which is used to use a feedforward network containing a hidden layer to combine the encoder hidden state and cell state at the previous moment, calculate the distribution of attention, and obtain an attention score; a normalization unit, which is used to use a softmax function to normalize the attention score into a weight value; and a second multiplication module, which is used to multiply the weight value with the intrinsic modal function observation value at the current moment to obtain the modal parameter at the current moment.

[0019] In some embodiments, for the input context vector, the device for determining the sequence value of the financial service volume includes: a generation unit, which is used to integrate the time dependency in the hidden state matrix at the current moment and the fluctuation characteristic information in the intrinsic mode function, and generate the new hidden state weighted sum through the time pattern attention mechanism; a splicing unit, which is used to splice the new hidden state weighted sum and the weighted sum of the encoded hidden states at the historical time points to obtain the spliced ​​context vector.

[0020] In some embodiments, the service volume determination unit includes: an autoregressive module, which is used to input N service volume driving sequences associated with the target financial business into an autoregressive model, and the autoregressive model performs autoregressive processing on the N service volume driving sequences to obtain a service regression output; an accumulation module, which is used to accumulate the output indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group with the service regression output to obtain the service volume sequence value required for the target financial business at the next time point.

[0021] According to another aspect of an embodiment of the present application, a computer-readable storage medium is also provided, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute any of the above-mentioned methods for determining the sequence value of the financial service volume.

[0022] According to another aspect of an embodiment of the present application, an electronic device is also provided, comprising one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement any one of the above-mentioned methods for determining the sequence value of the financial service volume.

[0023] In the present disclosure, N service volume driven sequences of the associated target financial business are received, and a multi-scale decomposition algorithm is used to decompose each service volume driven sequence into different frequency domain spaces to obtain a corresponding intrinsic mode function, and the intrinsic mode function is used as the input information of the first encoder in the codec set, wherein the service volume driven sequence is a sequence of multiple time points involving the target financial business in the historical process, and the intrinsic mode function contains different fluctuation feature information in the service volume driven sequence. The codec set includes M groups of encoders and decoders, and in the combined attention layer of each group of encoders, the input intrinsic mode function is encoded, and the LSTM unit is used to combine the encoded hidden state information of the previous moment with the input intrinsic mode function to output the hidden state of the current moment; the current moment is updated according to the modal parameters of the current moment. The hidden states of the previous moment are aggregated to obtain the hidden state matrix of the current moment, wherein the hidden state matrix of the current moment contains the importance of sequences at different time points and the time dependency within the sequence; in the time attention layer of the decoder, based on the input context vector and the decoder hidden state of the previous moment, the attention weight is calculated by the combined attention mechanism to obtain the weighted sum of the encoded hidden states of all time steps, wherein the context vector is obtained by concatenating the weighted sum of the encoded hidden states of the historical time points with the weighted sum of the new hidden states, and the new hidden state weighted sum is obtained through the time dependency in the hidden state matrix and the sequence fluctuation feature information; based on the output amount indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group, the service volume sequence value required for the target financial business at the next time point is determined.

[0024] In the present disclosure, the ability of different attention mechanisms to extract features can be combined, and the attention mechanisms can be combined with each other and added to the encoder-decoder to more accurately capture the mutual correlation between different input sequences and the hidden states within the sequences, and more effectively extract the short-term and long-term dependency information of the sequences. Finally, the service volume sequence value required for the next time point is obtained through the dependency relationship and sequence fluctuation characteristics, thereby enhancing the ability to arrange financial services, thereby solving the technical problem in the related art that the multi-feature time series indicating the arrangement of the service volume of financial institutions cannot accurately obtain the mutual dependency relationship between the sequences, resulting in poor arrangement effect.

[0025] In this disclosure, it is proposed to construct a combined attention mechanism model for feature extraction of multiple input features. By combining it with the encoder-decoder structure, the short-term / long-term dependencies and mutual dependencies between multiple driving sequences and target sequences as well as within the sequences are analyzed to improve the accuracy of model prediction.

[0026] In this disclosure, a combined prediction framework of multi-scale decomposition and attention mechanism is proposed, which combines the multi-scale decomposition method with the long-short-term memory network and the attention mechanism. By inputting the set of intrinsic modal functions obtained through modal decomposition into the attention mechanism, the adaptive mechanism of the attention model selects information useful for prediction, and can learn relevant information across the frequency domain, ultimately achieving the purpose of denoising, effectively suppressing the noise in the data and realizing high-precision prediction of service volume. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0028] FIG1 is a flowchart of an optional method for determining a sequence value of a financial service volume according to an embodiment of the present application;

[0029] FIG2 is a schematic diagram of an optional multi-feature time service volume sequence output based on a multi-scale decomposition method and an attention mechanism according to an embodiment of the present application;

[0030] FIG3 is a schematic diagram of an operation of an optional combined attention mechanism according to an embodiment of the present application;

[0031] FIG4 is a schematic diagram of an optional device for determining a sequence value of a financial service volume according to an embodiment of the present application;

[0032] FIG5 is a hardware structure block diagram of an electronic device (or mobile device) for determining a method for a sequence value of a financial service volume according to an embodiment of the present application. DETAILED DESCRIPTION

[0033] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0034] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0035] To facilitate those skilled in the art to understand the present application, some of the terms or nouns involved in the embodiments of the present application are explained below:

[0036] LSTM, or Long Short-Term Memory, is a variant of recurrent neural networks (RNNs) designed to process time series data and other data with long-term dependencies. It includes input, forget, and output gates, along with cell states, enabling efficient memorization and forgetting of information.

[0037] The Encoder-Decoder architecture is a neural network structure used to process sequence data. It consists of two parts: an encoder and a decoder. The encoder maps the input sequence data to an intermediate representation of a fixed dimension, and the decoder maps the intermediate representation to the output sequence data. This can effectively process input and output sequence data of varying lengths.

[0038] Convolutional neural network, or CNN for short, is a deep learning neural network used to process and analyze data with a grid structure.

[0039] CAM, or Class Activation Mapping, is a technique used to visualize the attention mechanism in deep learning models. It helps us understand the areas that the model focuses on when classifying inputs, improving the model's interpretability and visualization capabilities. In this application, CAM is combined with a convolutional neural network (CNN) to visualize the distribution of the model's attention on the input.

[0040] Empirical Mode Decomposition (EMD) is a method for decomposing nonlinear and nonstationary signals. EMD decomposes complex signals into multiple essential mode functions (IMFs), each representing distinct characteristics and frequency components of the signal.

[0041] Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMD) is an improvement and extension of empirical mode decomposition (EMD). Similar to EMD, CEEMD decomposes complex signals into multiple essential mode functions (IMFs), each representing distinct signal characteristics and frequency components. Compared to standard EMD, CEEMD better handles noise in signals, improving the accuracy and stability of signal decomposition.

[0042] The EMD-Attention architecture is an attention mechanism based on empirical mode decomposition (EMD). EMD is a signal processing technique used to decompose a signal into multiple intrinsic mode functions (IMFs), each representing a specific frequency component of the signal. The EMD-Attention architecture uses this decomposition method to decompose the input signal into multiple IMFs. These IMFs are then weighted using an attention mechanism to extract important information and perform feature fusion.

[0043] It should be noted that the method for determining the financial service volume sequence value and the device thereof in the present disclosure can be used in the field of financial technology to output high-precision financial service volume sequence values, and can also be used in any field other than the field of financial technology to output high-precision financial service volume sequence values. The present disclosure does not limit the application field of the method for determining the financial service volume sequence value and the device thereof.

[0044] It should be noted that the relevant information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of each region, and provide corresponding operation portals for users to choose to authorize or refuse. For example, an interface is set up between this system and relevant users or institutions. Before obtaining relevant information, it is necessary to send an acquisition request to the aforementioned user or institution through the interface, and obtain relevant information after receiving the consent information fed back by the aforementioned user or institution.

[0045] The following embodiments of this application can be applied to various systems / applications / devices that output financial service volume sequence values. This application provides a service volume sequence value output strategy based on a multi-scale decomposition method and an attention mechanism. On the basis of enhancing the feature extraction capability of a single attention mechanism, it can flexibly increase the weight of important features of sequence information of different lengths. The CAM-RNN method is constructed by combining the Encoder-Decoder structure, in which the combination of the encoder-decoder network and different attention mechanisms has a strong ability to mine fluctuation features from short sequences. The use of the temporal mode attention mechanism can combine long-term dependencies with short-term fluctuation information by adjusting the receptive field to process sequences of different lengths. At the same time, the EMD-Attention structure is also constructed in this application. The multi-scale decomposition method, empirical mode decomposition (EMD), is used to pre-process the sequence to separate high-frequency information and low-frequency information in the sequence, which plays the role of sequence smoothing to obtain trend information for subsequent model input. The attention mechanism is used to adaptively identify high-frequency noise in the sequence and assign low weights, thereby achieving the purpose of denoising and realizing high-precision service volume prediction.

[0046] To address the problem of multi-feature time series used in prior art to rank financial institutions' service volumes, which cannot accurately capture the interdependencies between sequences, resulting in poor ranking results, this application combines the feature extraction capabilities of different attention mechanisms and incorporates them into the encoder-decoder. This allows for more accurate capture of the correlations between different input sequences and between hidden states within sequences, more effectively extracting short-term and long-term dependency information from sequences, and thus leveraging this rich information to predict future fluctuations in service volume. Finally, the multi-scale decomposition method is combined with a long-short-term memory network and an attention mechanism to achieve high-precision predictions by effectively suppressing noise in the data.

[0047] The present application is described in detail below with reference to various embodiments.

[0048] Example 1

[0049] According to an embodiment of the present application, an embodiment of a method for determining a sequence value of a financial service volume is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0050] FIG1 is a flowchart of an optional method for determining a sequence value of a financial service volume according to an embodiment of the present application. As shown in FIG1 , the method includes the following steps:

[0051] Step S101: Receive N service volume driven sequences associated with the target financial business, use a multi-scale decomposition algorithm to decompose each service volume driven sequence into different frequency domain spaces, obtain the corresponding intrinsic mode function, and use the intrinsic mode function as the input information of the first encoder in the codec set, wherein the service volume driven sequence is a sequence of multiple time points involving the target financial business in the historical process, the intrinsic mode function contains different fluctuation characteristic information in the service volume driven sequence, and the codec set contains M groups of encoders and decoders, N is a positive integer greater than 1, and M is a positive integer.

[0052] In step S102, in the combined attention layer of each group of encoders, the input intrinsic modal function is encoded, and the LSTM unit is used to combine the encoded hidden state information of the previous moment with the input intrinsic modal function to output the hidden state of the current moment.

[0053] Step S103 , updating the hidden state at the current moment according to the modal parameters at the current moment, and aggregating to obtain the hidden state matrix at the current moment, wherein the hidden state matrix at the current moment includes the importance of sequences at different time points and the time dependency within the sequence.

[0054] Step S104: In the temporal attention layer of the decoder, based on the input context vector and the decoder hidden state at the previous moment, the attention weight is calculated by the combined attention mechanism to obtain the weighted sum of the encoded hidden states of all time steps, wherein the context vector is obtained by concatenating the weighted sum of the encoded hidden states at the historical time points with the weighted sum of the new hidden states, and the new hidden state weighted sum is obtained through the time dependency in the hidden state matrix and the sequence fluctuation feature information.

[0055] Step S105 : determining a service amount sequence value required for the target financial service at the next time point based on an output amount indicated by a weighted sum of the encoded hidden states output by the decoders of the Mth group.

[0056] Through the above steps, N service volume driven sequences associated with the target financial business can be received, and each service volume driven sequence can be decomposed into different frequency domain spaces by using a multi-scale decomposition algorithm to obtain a corresponding intrinsic mode function, and the intrinsic mode function is used as the input information of the first encoder in the codec set, wherein the service volume driven sequence is a sequence of multiple time points involving the target financial business in the historical process, and the intrinsic mode function contains different fluctuation feature information in the service volume driven sequence. The codec set contains M groups of encoders and decoders, N is a positive integer greater than 1, and M is a positive integer; in the combined attention layer of each group of encoders, the input intrinsic mode function is encoded, and the LSTM unit is used to combine the encoded hidden state information of the previous moment with the input intrinsic mode function to output the hidden state of the current moment; according to the current moment The modal parameters of the modal parameters are used to update the hidden state at the current moment, and the hidden state matrix at the current moment is aggregated, where the hidden state matrix at the current moment contains the importance of sequences at different time points and the time dependency within the sequence; in the time attention layer of the decoder, based on the input context vector and the decoder hidden state at the previous moment, the attention weight is calculated by the combined attention mechanism to obtain the weighted sum of the encoded hidden states of all time steps, where the context vector is obtained by concatenating the weighted sum of the encoded hidden states at the historical time points with the weighted sum of the new hidden states, and the new hidden state weighted sum is obtained through the time dependency in the hidden state matrix and the sequence fluctuation feature information; based on the output amount indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group, the service volume sequence value required for the target financial business at the next time point is determined. In this embodiment, the ability of different attention mechanisms to extract features can be combined, and the attention mechanisms can be combined with each other and added to the encoder-decoder to more accurately capture the mutual correlation between different input sequences and the hidden states within the sequences, and more effectively extract the short-term and long-term dependency information of the sequences. Finally, the service volume sequence value required for the next time point is obtained through the dependency relationship and sequence fluctuation characteristics, thereby enhancing the ability to arrange financial services, thereby solving the technical problem in related technologies that the multi-feature time series indicating the arrangement of financial institutions' service volumes cannot accurately obtain the mutual dependency relationship between sequences, resulting in poor arrangement effects.

[0057] The embodiments of the present application are described in detail below in combination with the above steps.

[0058] Step S101: receive N service volume drive sequences associated with the target financial business, use a multi-scale decomposition algorithm to decompose each service volume drive sequence into different frequency domain spaces, obtain the corresponding intrinsic mode function, and use the intrinsic mode function as the input information of the first encoder in the codec set.

[0059] It should be noted that in this embodiment, the encoder-decoder structure is configured with M groups of encoders and decoders. The output value of each group of decoders is used as the input of the next encoder, and the structure is continuously optimized. Finally, the output amount indicated by the weighted sum of the encoded hidden states output by the Mth group of decoders determines the service amount sequence value required for the target financial business at the next time point.

[0060] It should be noted that the service volume driven sequence in this embodiment includes multiple historical time points and corresponding historical service volume observation values. The sequence smoothing and decomposition module is used to decompose the observed service volume driven sequence using EMD or CEEMD in the multi-scale decomposition algorithm to decompose the service volume historical observation values ​​and other related sequences. Based on the stationary characteristics of the decomposed sequence data, the sequence is normalized to the maximum and minimum values.

[0061] In some embodiments, a multi-scale decomposition algorithm is used to decompose each service volume drive sequence into different frequency domain spaces to obtain the corresponding intrinsic mode function, including: step S1, for each service volume drive sequence, determining the maximum and minimum values ​​in the service volume drive sequence; step S2, using a preset interpolation polynomial method to construct the upper and lower envelopes of the service volume drive sequence according to the maximum and minimum values; step S3, calculating the local mean of the service volume drive sequence according to the upper and lower envelopes, and subtracting the local mean from the service volume drive sequence to obtain a new drive sequence; step S4, using the new drive sequence The service volume drive sequence is replaced by a new sequence, and steps S1 to S3 are repeatedly performed until the variance between the latest new drive sequence and the previous new drive sequence is less than a preset variance threshold, and the latest new drive sequence is defined as the intrinsic mode function of the service volume drive sequence; step S5, the intrinsic mode function is subtracted from the service volume drive sequence to obtain a difference sequence; step S6, the service volume drive sequence is replaced by the difference sequence, and steps S1 to S5 are repeatedly performed until a termination condition is met, wherein the termination condition includes: the difference sequence satisfies monotonicity, or the difference sequence has only one zero point.

[0062] The preset interpolation polynomial method may be a cubic spline interpolation method or a piecewise cubic Hermite interpolation polynomial method, or a Lagrange interpolation method.

[0063] When executing the above step S2 and using the preset interpolation polynomial method to construct the upper and lower envelopes of the service volume driven sequence according to the maximum and minimum values, the interpolation nodes can be determined first, and then the upper and lower envelopes of the service volume driven sequence can be constructed using the selected interpolation polynomial method.

[0064] In this embodiment, a data dimensionality reduction and feature extraction module can be used to combine the encoder in the Encoder-Decoder structure with the combined attention mechanism (CAM) to screen valid sequence information and perform feature processing and feature representation on the sequence information to solve the problem of mismatch between the input sequence and the output sequence length.

[0065] By combining the attention mechanism, we can effectively utilize the parallel processing of different attention convolution structures, the mapping operation of arbitrary receptive fields, and the ability of the attention mechanism to identify important information to capture the long-term dependencies within the sequence and the dynamic correlations between sequences.

[0066] In step S102, in the combined attention layer of each group of encoders, the input intrinsic modal function is encoded, and the LSTM unit is used to combine the encoded hidden state information of the previous moment with the input intrinsic modal function to output the hidden state of the current moment.

[0067] In some embodiments, step S102 includes: using an LSTM unit to extract a hidden state feature column vector in the intrinsic mode function, wherein the hidden state column vector contains forward and reverse information of the sequence; using an additive attention mechanism to process the hidden state feature column vector to obtain the importance of sequences at different time points in the intrinsic mode function; constructing an initial hidden state matrix based on multiple hidden state feature column vectors, wherein the row vector in the initial hidden state matrix represents the state vector of the same input feature under T time steps, and the column vector represents the state vector of all input features under a single time step; performing enhancement processing on the row vector of the initial hidden state matrix through the convolution kernel of the convolutional neural network, and using a combined attention mechanism to obtain the mutual dependencies within the initial hidden state matrix across time steps to obtain the time dependency within the sequence, wherein the combined attention mechanism refers to a mechanism obtained by combining the additive attention mechanism and the temporal mode attention mechanism.

[0068] Step S103 , updating the hidden state at the current moment according to the modal parameters at the current moment, and aggregating to obtain the hidden state matrix at the current moment, wherein the hidden state matrix at the current moment includes the importance of sequences at different time points and the time dependency within the sequence.

[0069] In some embodiments, step S103 includes: determining a time pattern matrix based on the time dependency contained in the hidden state matrix at the current moment; extracting row vectors in the time pattern matrix; using a vector similarity measurement function to multiply the row vectors in the time pattern matrix and the corresponding weight coefficients to obtain an initial context vector; performing full connection processing on the initial context vector and the corresponding first change coefficient matrix, the hidden state at the current moment, and the corresponding second change coefficient matrix, and aggregating to obtain the hidden state matrix at the current moment.

[0070] In some embodiments, the modal parameters at the current moment are pre-generated. When generating the modal parameters at the current moment, it includes: using a feedforward network containing a hidden layer to combine the encoder hidden state and cell state at the previous moment, calculating the distribution of attention, and obtaining an attention score; using a softmax function to normalize the attention score into a weight value; multiplying the weight value by the observed value of the intrinsic modal function at the current moment to obtain the modal parameters at the current moment.

[0071] Step S104: In the temporal attention layer of the decoder, based on the input context vector and the decoder hidden state at the previous moment, the attention weight is calculated by the combined attention mechanism to obtain the weighted sum of the encoded hidden states of all time steps, wherein the context vector is obtained by concatenating the weighted sum of the encoded hidden states at the historical time points with the weighted sum of the new hidden states, and the new hidden state weighted sum is obtained through the time dependency in the hidden state matrix and the sequence fluctuation feature information.

[0072] In some embodiments, for the input context vector, it includes: integrating the time dependency in the hidden state matrix at the current moment and the fluctuation characteristic information in the intrinsic mode function, generating a new hidden state weighted sum through the time pattern attention mechanism; splicing the new hidden state weighted sum and the weighted sum of the encoded hidden states at the historical time points to obtain the spliced ​​context vector.

[0073] Step S105 : determining a service amount sequence value required for the target financial service at the next time point based on an output amount indicated by a weighted sum of the encoded hidden states output by the decoders of the Mth group.

[0074] In some embodiments, the step of determining the service volume sequence value required for the target financial business at the next time point based on the output indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group includes: inputting N service volume driving sequences associated with the target financial business into an autoregressive model, and using the autoregressive model to perform autoregressive processing on the N service volume driving sequences to obtain a service regression output; and accumulating the output indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group and the service regression output to obtain the service volume sequence value required for the target financial business at the next time point.

[0075] The following describes in detail another optional specific implementation.

[0076] Figure 2 is a schematic diagram of an optional multi-feature time service volume sequence output based on a multi-scale decomposition method and an attention mechanism according to an embodiment of the present application. The following is a schematic explanation in conjunction with the content of Figure 2.

[0077] This application discloses a service volume prediction method based on a multi-scale decomposition method and an attention mechanism, which mainly includes the following process steps:

[0078] S1. The service volume is unevenly distributed, with high noise, non-stationary, multi-scale and obvious trend characteristics. At the same time, the service volume has a complex correlation with other influencing sequences. The historical service volume information of the target financial business is used as a multi-feature driving sequence. For example, taking stock information as an example, the stock service volume is used as the target sequence. The multi-scale decomposition algorithm is used to convert the input service volume driving sequence X=(x 1 ,x 2 ,…,x n ) · =(x1,x2,…,x T ) is decomposed into different frequency domain spaces to obtain the corresponding intrinsic mode functions IMFs i (i=1,2,…,N) is used as encoder input, and the decomposed intrinsic mode functions IMFs i Contains different fluctuation characteristic information in the original sequence.

[0079] In this embodiment, a sequence smoothing and decomposition module is used to decompose the observed driving sequence using EMD and CEEMD in the multi-scale decomposition algorithm to decompose the historical service volume observations and other related sequences (as shown in FIG2 , EMD and CEEMD are performed on the service volume driving sequence to obtain the intrinsic mode function). Based on the stationary nature of the decomposed sequence data, the sequence is normalized to its maximum and minimum values. The specific steps are as follows:

[0080] (1) Determine the original service volume driving sequence x k (t)∈R T The maximum and minimum values ​​of ;

[0081] (2) Use cubic spline interpolation or piecewise cubic Hermite interpolation polynomial method to construct x according to the maximum and minimum values k (t) upper and lower envelopes;

[0082] (3) Calculate x based on the upper and lower envelopes k The local mean m of (t) 1j (t), using x k (t) minus the local mean m 1j (t) Get the new driving sequence h 1j (t);

[0083] (4). Use the new drive sequence h 1j (t) replaces the original service quantity driving sequence x k (t), and repeat the above steps (1)-(3) until h 1j(t) and h 1(j-1) Until the variance of (t) is lower than the set threshold and satisfies the two conditions of IMF. Then the first IMF component IMFs1(t) is h 1j (t), denoted as IMFs1(t)=h 1j (t);

[0084] (5) Through IMFs1(t) and the original service quantity driving sequence x k (t) is separated by subtraction to obtain the differential sequence r1(t)=x k (t)-IMFs1(t) is used to filter out the high-frequency components in the original sequence;

[0085] (6) Replace the original sequence x with the differential sequence r1(t) k (t), and repeat the above steps (1)-(5) until the termination condition is met (usually making the last residual sequence meet monotonicity, or have only one zero point).

[0086] This embodiment can transform the service volume sequence from the time domain to the frequency domain through modal decomposition to explain the different characteristic information of the sequence. The inherent mode function obtained by decomposition reflects the different fluctuation information in the service volume, has advantages such as stability, and reduces the difficulty of prediction.

[0087] Using the data dimensionality reduction and feature extraction modules, the encoder in the Encoder-Decoder structure is combined with the combined attention mechanism (CAM) to filter effective sequence information, and perform feature processing and feature representation on the sequence information to solve the problem of mismatch between the input sequence and the output sequence length.

[0088] S2. In the combined attention layer, the input intrinsic mode functions (IMFs) are first i (i=1,2,…,N) performs encoding operation. Use LSTM unit to encode the hidden state information h of the previous moment t-1 ∈R m IMFs with input sequence information t Combine and output the hidden state h at the current moment t .

[0089] Figure 3 is a schematic diagram of the operation of an optional combined attention mechanism according to an embodiment of the present application. The operation content of the combined attention mechanism is schematically explained below in conjunction with Figure 3. The driving sequence in Figure 3 is the service volume driving sequence mentioned above. After the driving sequence is spectrally decomposed using the EMD method (Figure 3 uses the spectral domain to illustrate), the intrinsic mode function is obtained. After that, the additive attention mechanism is processed and input into the combined attention layer of the decoder for processing.

[0090] By combining the attention mechanism, we can effectively utilize the parallel processing of different attention convolution structures, the mapping operation of arbitrary receptive fields, and the ability of the attention mechanism to identify important information to capture the long-term dependencies within the sequence and the dynamic correlations between sequences. The specific steps are as follows:

[0091] 1). The input sequence obtained by multi-scale decomposition is IMFs = (IMFs1, IMFs2,…, IMFs N ) · ∈R N×T ;

[0092] 2) Obtain the hidden state feature column vector h of dimension m through the LSTM-based encoder i (i=1,2,...,T), this vector contains the forward and reverse direction information of the sequence, and obtains the importance information of the input sequence through the additive attention mechanism;

[0093] 3) Use several output vectors to form an initial hidden state matrix H = [h1,h2,…,h T ]∈R m×T , where the row vector represents the state vector of the same input feature at T time steps, and the column vector represents the state vector of all input features at a single time step;

[0094] 4) Based on the CNN structure’s ability to extract features from the data matrix, the CNN convolution kernel is applied to the row vectors of the initial hidden state matrix H to enhance the model’s learning ability. The combined attention mechanism formed by the additive attention mechanism and the temporal pattern attention mechanism is used to obtain the temporal dependency within the hidden state matrix across time steps. Assuming there are k one-dimensional convolution kernels C i ∈R 1×T , where the time window size T is the maximum length of the model attention, that is, the size of the convolution kernel. Then the hidden state matrix H = (h1,h2,…,h T ) Perform convolution operation on the whole to obtain the time pattern matrix H of the model C ∈R m×k ,in Represents the i-th row vector of H and the j-th convolution kernel C j The convolution value of . Formally, the convolution operation is defined by the following formula:

[0095] 5) By including the cross-time step hidden state related dependency information H C and the hidden state h at the current moment T Perform multiplication attention mechanism (dot product model), calculate and h TThe correlation between the time pattern attention weight vector is normalized using the sigmoid activation function, and the final intermediate vector v t H C The weighted sum of the row vectors of , in order to achieve the purpose of updating the hidden state at the current moment;

[0096] 6) Hidden state obtained by updating It contains dependency information with a time step size of at least T, effectively solving the problem of information loss caused by long-distance transmission. At the same time, the updated hidden state matrix is ​​used as the input of the decoding stage. Because it contains sufficient long-term information, combined with the weighted input sequence encoding information and the dependency information across time steps, it can effectively improve the accuracy of prediction.

[0097] Among them, the time pattern matrix H in this embodiment is C and the hidden state h at the current moment T The formula for the multiplicative attention mechanism is as follows:

[0098] Among them: H i C ∈R k H C Row vector of R; f1:R k ×R m →R is the measurement function of the similarity between vectors; W a ∈R k×m is the transformation coefficient matrix; δ i for The weight coefficient of

[0099] 7) Multiply the vector by the weight and then add them together to get the context vector v T ;

[0100] 8). The context vector v T and the hidden state h at the current moment T Perform full connection to obtain the updated hidden state The input for the decoding stage is as follows.

[0101] Where: W h ∈R m×m , W v ∈R m×T is the transformation coefficient matrix.

[0102] This embodiment is based on a combination of multiple attention mechanisms, which can adaptively select corresponding external influencing factors and noise information to achieve the purpose of sequence smoothing and noise suppression, while integrating the short-term and long-term time information input by the encoder across time steps.

[0103] S3. Use a feedforward network (fully connected) with one hidden layer to calculate the distribution of attention, and transform the encoder hidden state h at time t-1 t-1 and cell states t-1 Combined, generate the alignment score. Then use the softmax function to normalize the attention score into a weight value, and multiply the weight value by the intrinsic mode function observation value IMFs at time t t Get the modal parameters at the current moment

[0104] S4, according to the weighted t moment Update the hidden state at the corresponding moment Aggregate to get the hidden state matrix H C ∈R m×T , which contains the importance information of different sequences and the time dependency within the sequences;

[0105] S5. Use different attention mechanisms to perform weighted updates on the feature information across time, and finally input it into the combined attention layer to obtain a new hidden state matrix Better extract the complex, nonlinear relationships and interdependencies between different variables in high-dimensional space, and improve the model's ability to model long-sequence dependencies;

[0106] S6. In the Temporal Pattern Attention layer, first the input context vector c j (j=1,2,...,T-1) based on the previous decoder hidden state d t-1 ∈R p , calculate the attention weight by combining the attention mechanism Represent the information as the weighted sum c of the encoded hidden states at all time steps t ;

[0107] S7, integrate the time dependency of the hidden state matrix and the input information, and obtain the new hidden state weighted sum c′ through the time pattern attention mechanism t . The concatenated context vector [c t ,c′ t ] Perform a full connection operation to obtain a new context vector Used as input to the decoder LSTM unit;

[0108] S8. Add the output of the last decoder LSTM unit to the output of the linear unit AR model to obtain This is the service volume prediction result of the model.

[0109] Through the above embodiments, the technical effects that can be achieved are as follows:

[0110] 1. By introducing a combined attention mechanism into the encoder-decoder structure, the internal structure and dependencies of the input data can be better expressed, thereby improving the accuracy of both short-term and long-term service volume sequence value prediction.

[0111] 2. By combining the multi-scale decomposition algorithm with the attention mechanism, high-frequency noise in the service volume sequence can be suppressed while ensuring the integrity of the sequence information. At the same time, the attention mechanism can be used to quantify the importance of influencing factors for the prediction of service volume sequence values, further improving the overall prediction accuracy of the model.

[0112] The following describes it in detail with reference to another embodiment.

[0113] Example 2

[0114] The device for determining the sequence value of a financial service volume provided in this embodiment includes multiple implementation units, each implementation unit corresponding to each implementation step in the above-mentioned embodiment 1.

[0115] Figure 4 is a schematic diagram of an optional device for determining the sequence value of financial service volume according to an embodiment of the present application. As shown in Figure 4, the device for determining the sequence value of financial service volume may include: a sequence decomposition unit 41, an encoding unit 42, a hidden aggregation unit 43, a combined attention calculation unit 44, and a service volume determination unit 45.

[0116] The sequence decomposition unit 41 is configured to receive N service-volume-driven sequences associated with a target financial business, decompose each service-volume-driven sequence into different frequency domain spaces using a multi-scale decomposition algorithm, obtain a corresponding intrinsic mode function, and use the intrinsic mode function as input information for a first encoder in a codec set. The service-volume-driven sequence is a sequence of multiple time points associated with the target financial business in a historical process, and the intrinsic mode function includes different fluctuation characteristic information in the service-volume-driven sequence. The codec set includes M sets of encoders and decoders, where N is a positive integer greater than 1 and M is a positive integer.

[0117] An encoding unit 42 is configured to encode the input intrinsic mode function in the combined attention layer of each set of encoders, and combine the encoded hidden state information at the previous moment with the input intrinsic mode function using an LSTM unit to output the hidden state at the current moment;

[0118] A hidden aggregation unit 43 is configured to update the hidden state at the current moment according to the modal parameters at the current moment, and aggregate to obtain a hidden state matrix at the current moment, wherein the hidden state matrix at the current moment includes the importance of sequences at different time points and the time dependency within the sequence;

[0119] A combined attention calculation unit 44 is configured to calculate an attention weight in the decoder's temporal attention layer based on an input context vector and the decoder's hidden state at the previous moment, using a combined attention mechanism to obtain a weighted sum of the encoding hidden states at all time steps, wherein the context vector is obtained by concatenating the weighted sum of the encoding hidden states at the historical time points with the weighted sum of the new hidden states, and the new hidden state weighted sum is obtained by using the temporal dependency in the hidden state matrix and sequence fluctuation feature information;

[0120] The service amount determination unit 45 is configured to determine a service amount sequence value required for the target financial service at a next time point based on an output amount indicated by a weighted sum of the coded hidden states output by the decoders of the Mth group.

[0121] The above-mentioned device for determining the value of the financial service volume sequence can receive N service volume driving sequences associated with the target financial business through the sequence decomposition unit 41, and use the multi-scale decomposition algorithm to decompose each service volume driving sequence into different frequency domain spaces to obtain the corresponding intrinsic mode function, and use the intrinsic mode function as the input information of the first encoder in the codec set, wherein the service volume driving sequence is a sequence of multiple time points involving the target financial business in the historical process, and the intrinsic mode function contains different fluctuation feature information in the service volume driving sequence. The codec set contains M groups of encoders and decoders, and the encoding unit 42 performs encoding operation on the input intrinsic mode function in the combined attention layer of each group of encoders, and uses the LSTM unit to combine the encoded hidden state information of the previous moment with the input intrinsic mode function to output the hidden state of the current moment; through the hidden aggregation unit 43, the hidden state is obtained according to the mode of the current moment. The state parameters are used to update the hidden state at the current moment, and the hidden state matrix at the current moment is obtained by aggregation, wherein the hidden state matrix at the current moment contains the importance of sequences at different time points and the time dependency within the sequence; the combined attention calculation unit 44 calculates the attention weight in the time attention layer of the decoder based on the input context vector and the decoder hidden state at the previous moment, and obtains the weighted sum of the encoded hidden states of all time steps, wherein the context vector is obtained by concatenating the weighted sum of the encoded hidden states at the historical time points with the weighted sum of the new hidden states, and the new hidden state weighted sum is obtained through the time dependency in the hidden state matrix and the sequence fluctuation characteristic information; the service volume sequence value required for the target financial business at the next time point is determined by the service volume determination unit 45 based on the output indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group. In this embodiment, the ability of different attention mechanisms to extract features can be combined, and the attention mechanisms can be combined with each other and added to the encoder-decoder to more accurately capture the mutual correlation between different input sequences and the hidden states within the sequences, and more effectively extract the short-term and long-term dependency information of the sequences. Finally, the service volume sequence value required for the next time point is obtained through the dependency relationship and sequence fluctuation characteristics, thereby enhancing the ability to arrange financial services, thereby solving the technical problem in related technologies that the multi-feature time series indicating the arrangement of financial institutions' service volumes cannot accurately obtain the mutual dependency relationship between sequences, resulting in poor arrangement effects.

[0122] In some embodiments, the sequence decomposition unit includes: a first determination module for executing step S1, for each service volume driven sequence, determining the maximum value and the minimum value in the service volume driven sequence; a first construction module for executing step S4, for constructing the upper and lower envelopes of the service volume driven sequence according to the maximum value and the minimum value using a preset interpolation polynomial method; a first calculation module for executing step S3, for calculating the local mean of the service volume driven sequence according to the upper and lower envelopes, and subtracting the local mean from the service volume driven sequence to obtain a new driving sequence; a first repetition module for executing step S4, for replacing the service volume driven sequence with the new driving sequence The method further comprises the steps of: first, performing step S5, performing subtraction and separation of the intrinsic mode function from the service volume drive sequence, and repeatedly executing steps S1 to S3 until the variance between the latest new drive sequence and the previous new drive sequence is less than a preset variance threshold, and defining the latest new drive sequence as the intrinsic mode function of the service volume drive sequence; and second, performing step S6, replacing the service volume drive sequence with the differential sequence, and repeatedly executing steps S1 to S5 until a termination condition is satisfied, wherein the termination condition includes: the differential sequence satisfies monotonicity, or the differential sequence has only one zero point.

[0123] In some embodiments, the encoding unit includes: a first extraction module, used to use an LSTM unit to extract the hidden state feature column vector in the intrinsic mode function, wherein the hidden state column vector contains forward and reverse information of the sequence; a first processing unit, used to use an additive attention mechanism to process the hidden state feature column vector to obtain the importance of sequences at different time points in the intrinsic mode function; a second construction module, used to construct an initial hidden state matrix based on multiple hidden state feature column vectors, wherein the row vector in the initial hidden state matrix represents the state vector of the same input feature under T time steps, and the column vector represents the state vector of all input features under a single time step; a second processing module, used to perform enhancement processing on the row vector of the initial hidden state matrix through the convolution kernel of the convolutional neural network, and use a combined attention mechanism to obtain the mutual dependence within the initial hidden state matrix across time steps to obtain the time dependence within the sequence, wherein the combined attention mechanism refers to a mechanism obtained by combining the additive attention mechanism and the temporal pattern attention mechanism.

[0124] In some embodiments, the hidden aggregation unit includes: a second determination module, used to determine the time pattern matrix based on the time dependency contained in the hidden state matrix at the current moment; a second extraction module, used to extract the row vectors in the time pattern matrix; a first multiplication module, used to use the inter-vector similarity measurement function to multiply the row vectors in the time pattern matrix and the corresponding weight coefficients to obtain the initial context vector; a third processing module, used to perform full connection processing on the initial context vector and the corresponding first change coefficient matrix, the hidden state at the current moment and the corresponding second change coefficient matrix, and aggregate to obtain the hidden state matrix at the current moment.

[0125] In some embodiments, the modal parameters at the current moment are pre-generated. When generating the modal parameters at the current moment, the device for determining the sequence value of the financial service volume includes: an attention calculation unit, which is used to use a feedforward network containing a hidden layer to combine the encoder hidden state and cell state at the previous moment, calculate the distribution of attention, and obtain an attention score; a normalization unit, which is used to normalize the attention score into a weight value using a softmax function; and a second multiplication module, which is used to multiply the weight value with the observed value of the intrinsic modal function at the current moment to obtain the modal parameter at the current moment.

[0126] In some embodiments, for the input context vector, the device for determining the sequence value of the financial service volume includes: a generation unit, which is used to integrate the time dependency in the hidden state matrix at the current moment and the fluctuation characteristic information in the intrinsic mode function, and generate a new hidden state weighted sum through the time pattern attention mechanism; a splicing unit, which is used to splice the new hidden state weighted sum and the weighted sum of the encoded hidden states at historical time points to obtain a spliced ​​context vector.

[0127] In some embodiments, the service volume determination unit includes: an autoregressive module, which is used to input N service volume driving sequences associated with the target financial business into the autoregressive model, and the autoregressive model performs autoregressive processing on the N service volume driving sequences to obtain a service regression output; an accumulation module, which is used to accumulate the output indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group and the service regression output to obtain the service volume sequence value required for the target financial business at the next time point.

[0128] The above-mentioned device for determining the sequence value of financial service volume may also include a processor and a memory. The above-mentioned sequence decomposition unit 41, encoding unit 42, hidden aggregation unit 43, combined attention calculation unit 44, service volume determination unit 45, etc. are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize the corresponding functions.

[0129] The processor includes a kernel, which retrieves the corresponding program unit from the memory. One or more kernels can be set, and the value of the financial service volume sequence can be output by adjusting the kernel parameters.

[0130] The above-mentioned memory may include non-permanent memory in a computer-readable medium, random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0131] The present application also provides a computer program product, which, when executed on a data processing device, is suitable for executing a program initialized with the following method steps: receiving N service volume driven sequences associated with the target financial business, using a multi-scale decomposition algorithm to decompose each service volume driven sequence into different frequency domain spaces, obtaining a corresponding intrinsic mode function, and using the intrinsic mode function as the input information of the first encoder in the codec set, wherein the service volume driven sequence is a sequence of multiple time points involving the target financial business in the historical process, the intrinsic mode function contains different fluctuation characteristic information in the service volume driven sequence, the codec set contains M groups of encoders and decoders, N is a positive integer greater than 1, and M is a positive integer; in the combined attention layer of each group of encoders, the input intrinsic mode function is encoded, and the LSTM unit is used to combine the encoded hidden state information of the previous moment with the input intrinsic mode function. , output the hidden state at the current moment; update the hidden state at the current moment according to the modal parameters at the current moment, and aggregate to obtain the hidden state matrix at the current moment, wherein the hidden state matrix at the current moment contains the importance of sequences at different time points and the time dependency within the sequence; in the time attention layer of the decoder, based on the input context vector and the decoder hidden state at the previous moment, the attention weight is calculated by the combined attention mechanism to obtain the weighted sum of the encoded hidden states of all time steps, wherein the context vector is obtained by concatenating the weighted sum of the encoded hidden states at the historical time points with the weighted sum of the new hidden states, and the new hidden state weighted sum is obtained through the time dependency in the hidden state matrix and the sequence fluctuation feature information; based on the output quantity indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group, determine the service volume sequence value required for the target financial business at the next time point.

[0132] According to another aspect of an embodiment of the present application, a computer-readable storage medium is also provided, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the method for determining the financial service quantity sequence value of any one of the above-mentioned embodiments.

[0133] According to another aspect of an embodiment of the present application, an electronic device is also provided, comprising one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors implement the method for determining the financial service volume sequence value of any one of the above-mentioned embodiments.

[0134] FIG5 is a block diagram of the hardware structure of an electronic device (or mobile device) for a method for determining a sequence value of a financial service volume according to an embodiment of the present application. As shown in FIG5 , the electronic device may include one or more (illustrated in FIG5 using 502a, 502b, ..., 502n) processors 502 (the processor 502 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), and a memory 504 for storing data. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a keyboard, a power supply, and / or a camera. It will be understood by those skilled in the art that the structure shown in FIG5 is merely illustrative and does not limit the structure of the electronic device described above. For example, the electronic device may also include more or fewer components than those shown in FIG5 , or have a configuration different from that shown in FIG5 .

[0135] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0136] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0137] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only exemplary. For example, the division of the units can be a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0138] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0139] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0140] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0141] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for determining a sequence value of a financial service volume, comprising: Receive N service volume drive sequences associated with the target financial business, use a multi-scale decomposition algorithm to decompose each of the service volume drive sequences into different frequency domain spaces, obtain a corresponding intrinsic mode function, and use the intrinsic mode function as input information of the first encoder in the codec set, wherein the service volume drive sequence is a sequence of multiple time points involving the target financial business in a historical process, the intrinsic mode function includes different fluctuation feature information in the service volume drive sequence, and the codec set includes M groups of encoders and decoders, N is a positive integer greater than 1, and M is a positive integer; In the combined attention layer of each group of encoders, the input intrinsic modal function is encoded, and the encoded hidden state information at the previous moment is combined with the input intrinsic modal function using an LSTM unit to output the hidden state at the current moment; Update the hidden state at the current moment according to the modal parameters at the current moment, and aggregate to obtain the hidden state matrix at the current moment, wherein the hidden state matrix at the current moment includes the importance of sequences at different time points and the time dependency within the sequence; In the time attention layer of the decoder, based on the input context vector and the decoder hidden state at the previous moment, the attention weight is calculated by the combined attention mechanism to obtain the weighted sum of the encoding hidden state of all time steps, wherein the context vector is obtained by concatenating the weighted sum of the encoding hidden state at the historical time point with the weighted sum of the new hidden state, and the new hidden state weighted sum is obtained by the time dependency in the hidden state matrix and the sequence fluctuation feature information; Based on the output amount indicated by the weighted sum of the encoded hidden states output by the decoders of the Mth group, a service amount sequence value required for the target financial service at the next time point is determined.

2. The method for determining the sequence value of financial service volume according to claim 1, wherein: The step of using a multi-scale decomposition algorithm to decompose each of the service volume drive sequences into different frequency domain spaces to obtain the corresponding intrinsic mode function includes: Step S1, for each of the service volume driving sequences, determining the maximum value and the minimum value in the service volume driving sequence; Step S2, using a preset interpolation polynomial method to construct upper and lower envelopes of the service volume driven sequence according to the maximum and minimum values; Step S3, calculating the local mean of the service volume driving sequence according to the upper and lower envelopes, and subtracting the local mean from the service volume driving sequence to obtain a new driving sequence; Step S4, using the new drive sequence to replace the service volume drive sequence, and repeatedly performing steps S1 to S3 until the variance between the latest new drive sequence and the previous new drive sequence is less than a preset variance threshold, and defining the latest new drive sequence as the intrinsic mode function of the service volume drive sequence; Step S5, subtracting and separating the intrinsic mode function from the service volume driving sequence to obtain a difference sequence; Step S6, using the differential sequence to replace the service volume driving sequence, and repeating steps S1 to S5 until a termination condition is met, wherein the termination condition includes: the differential sequence satisfies monotonicity, or the differential sequence has only one zero point.

3. The method for determining the sequence value of financial service volume according to claim 1, wherein: The step of combining the encoded hidden state information at the previous moment with the input intrinsic modal function using an LSTM unit to output the hidden state at the current moment includes: Using an LSTM unit to extract a hidden state feature column vector in the intrinsic mode function, wherein the hidden state column vector contains forward and reverse sequence information; Using an additive attention mechanism to process the hidden state feature column vector, to obtain the importance of different time point sequences in the intrinsic mode function; Constructing an initial hidden state matrix based on the plurality of hidden state feature column vectors, wherein the row vectors in the initial hidden state matrix represent the state vectors of the same input feature at T time steps, and the column vectors represent the state vectors of all input features at a single time step; The row vectors of the initial hidden state matrix are enhanced by the convolution kernel of a convolutional neural network, and the mutual dependencies within the initial hidden state matrix are obtained across time steps using a combined attention mechanism to obtain the temporal dependencies within the sequence, wherein the combined attention mechanism refers to a mechanism obtained by combining an additive attention mechanism with a temporal pattern attention mechanism.

4. The method for determining the sequence value of financial service volume according to claim 1, wherein: The step of updating the hidden state at the current moment according to the modal parameters at the current moment and aggregating to obtain the hidden state matrix at the current moment includes: Determine a temporal pattern matrix based on the temporal dependency relationship contained in the hidden state matrix at the current moment; Extracting row vectors from the time pattern matrix; Using a vector similarity measurement function, multiplying the row vectors in the temporal pattern matrix and the corresponding weight coefficients to obtain an initial context vector; The initial context vector and the corresponding first change coefficient matrix, the hidden state at the current moment and the corresponding second change coefficient matrix are fully connected to obtain the hidden state matrix at the current moment.

5. The method for determining the sequence value of financial service volume according to claim 1, wherein: The modal parameters at the current moment are generated in advance. When generating the modal parameters at the current moment, the following steps are included: A feedforward network with one hidden layer is used to combine the encoder hidden state and cell state at the previous moment, calculate the distribution of attention, and obtain the attention score; Use the softmax function to normalize the attention score into a weight value; The weight value is multiplied by the intrinsic modal function observation value at the current moment to obtain the modal parameter at the current moment.

6. The method for determining the sequence value of financial service volume according to claim 1, wherein: For the input context vector, it includes: Combining the time dependency in the hidden state matrix at the current moment and the fluctuation characteristic information in the intrinsic mode function, generating the new hidden state weighted sum through the time mode attention mechanism; The new hidden state weighted sum and the weighted sum of the encoded hidden states at the historical time point are concatenated to obtain the concatenated context vector.

7. The method for determining the sequence value of financial service volume according to claim 1, wherein: The step of determining the service amount sequence value required for the target financial service at the next time point based on the output amount indicated by the weighted sum of the encoded hidden states output by the decoders of the Mth group comprises: Inputting N service volume driving sequences associated with the target financial business into an autoregressive model, and using the autoregressive model to perform autoregressive processing on the N service volume driving sequences to obtain a service regression output; The output quantity indicated by the weighted sum of the encoded hidden states output by the decoder of the Mth group is accumulated with the service regression output quantity to obtain a service quantity sequence value required for the target financial service at the next time point.

8. A device for determining a sequence value of a financial service volume, comprising: A sequence decomposition unit is used to receive N service volume driving sequences associated with the target financial business, decompose each of the service volume driving sequences into different frequency domain spaces by using a multi-scale decomposition algorithm, obtain a corresponding intrinsic mode function, and use the intrinsic mode function as input information of a first encoder in a codec set, wherein the service volume driving sequence is a sequence of multiple time points involving the target financial business in a historical process, the intrinsic mode function includes different fluctuation feature information in the service volume driving sequence, and the codec set includes M groups of encoders and decoders, N is a positive integer greater than 1, and M is a positive integer; An encoding unit, used for encoding the input intrinsic modal function in the combined attention layer of each group of encoders, and combining the encoded hidden state information of the previous moment with the input intrinsic modal function by using an LSTM unit, and outputting the hidden state of the current moment; A hidden aggregation unit, used to update the hidden state at the current moment according to the modal parameters at the current moment, and aggregate to obtain a hidden state matrix at the current moment, wherein the hidden state matrix at the current moment includes the importance of sequences at different time points and the time dependency within the sequence; A combined attention calculation unit is used to calculate the attention weight through a combined attention mechanism in the time attention layer of the decoder based on the input context vector and the decoder hidden state at the previous moment, so as to obtain the weighted sum of the encoding hidden state of all time steps, wherein the context vector is obtained by concatenating the weighted sum of the encoding hidden state at the historical time point with the weighted sum of the new hidden state, and the new hidden state weighted sum is obtained by the time dependency relationship in the hidden state matrix and the sequence fluctuation feature information; The service volume determination unit is used to determine the service volume sequence value required for the target financial service at the next time point based on the output volume indicated by the weighted sum of the coded hidden states output by the decoders of the Mth group.

9. A computer-readable storage medium comprising a stored computer program, wherein: When the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the method for determining the financial service volume sequence value described in any one of claims 1 to 7.

10. An electronic device comprising one or more processors and a memory, wherein the memory is used to store one or more programs, wherein: When the one or more programs are executed by the one or more processors, the one or more processors implement the method for determining the financial service volume sequence value as described in any one of claims 1 to 7.

Images