Timing data processing method and device, and electronic equipment

Abstract

The application discloses a time series data processing method and device and electronic equipment, and relates to the field of artificial intelligence and financial technology. The method comprises the following steps: acquiring first time series data, behavior statistical sequence, second time series data and fluctuation statistical sequence, wherein the first time series data is a time series data sequence about a first object; the second time series data is a time series data sequence about a second object; determining a behavior feature vector of the first object according to the first time series data and the behavior statistical sequence; determining a fluctuation feature vector of the second object according to the second time series data and the fluctuation statistical sequence; determining a matching score between the first object and the second object according to the behavior feature vector and the fluctuation feature vector, and outputting the matching score to a display device or generating a control signal according to the matching score. The application solves the problem that the prior art cannot simultaneously process dynamic behavior features and fluctuation mode features of different objects, thereby reducing the accuracy of the correlation matching degree evaluation between different objects.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and financial technology, and more specifically, to a time-series data processing method, apparatus, and electronic device. Background Technology

[0002] In computer data processing, it is often necessary to evaluate the degree of correlation between different objects. For example, in a business matching scenario, the data processing system needs to calculate the matching score between the first object and the second object based on the behavioral data of the first object and the change data of the second object. Traditional data processing methods typically perform statistical analysis or feature extraction on the raw data of the first and second objects separately, and then calculate the similarity between the statistical measures as the matching basis.

[0003] Taking business data processing in the fintech field as an example, data processing systems need to assess the matching degree between users and products based on user transaction behavior data and product volatility data. Existing data processing methods, when dealing with this scenario, typically only perform statistical analysis on the user's static attributes or extract features from the product's return data in a single dimension. User transaction behavior data contains time-series variation information, while product volatility data contains variation information in terms of amplitude and frequency; these two types of data differ in their data structures. Existing methods struggle to simultaneously utilize the time-series behavioral information of the first object and the volatility variation information of the second object for correlation assessment, resulting in low accuracy in assessing the matching degree between the first and second objects.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a time-series data processing method, apparatus, and electronic device to at least solve the technical problem that the prior art cannot simultaneously process the dynamic behavior characteristics and fluctuation pattern characteristics of different objects, resulting in low accuracy in evaluating the correlation matching degree between different objects.

[0006] According to one aspect of the embodiments of this application, a time-series data processing method is provided, comprising: acquiring first time-series data, a behavioral statistics sequence, second time-series data, and a fluctuation statistics sequence, wherein the first time-series data is a time-series data sequence concerning a first object, and the behavioral statistics sequence includes at least the occurrence frequency of various business behaviors of the first object; the second time-series data is a time-series data sequence concerning a second object, and the fluctuation statistics sequence includes at least the fluctuation amplitude and extreme value decrease amplitude of the second time-series data, and the first object and the second object are different objects with business interactions; determining a behavioral feature vector of the first object based on the first time-series data and the behavioral statistics sequence, wherein the behavioral feature vector is used to characterize the time-series behavioral pattern and behavioral statistical characteristics of the first object; determining a fluctuation feature vector of the second object based on the second time-series data and the fluctuation statistics sequence, wherein the fluctuation feature vector is used to characterize the frequency domain characteristics and data fluctuation patterns of the second object over time; determining a matching score between the first object and the second object based on the behavioral feature vector and the fluctuation feature vector, and outputting the matching score to a display device or generating a control signal based on the matching score.

[0007] According to another aspect of the embodiments of this application, a time-series data processing apparatus is also provided, comprising: an acquisition unit, configured to acquire first time-series data, a behavioral statistics sequence, second time-series data, and a fluctuation statistics sequence, wherein the first time-series data is a time-series data sequence concerning a first object, and the behavioral statistics sequence includes at least the occurrence frequency of various business behaviors of the first object; the second time-series data is a time-series data sequence concerning a second object, and the fluctuation statistics sequence includes at least the fluctuation amplitude and extreme value decrease amplitude of the second time-series data, wherein the first object and the second object are different objects with business dealings; a first feature vector determination unit, configured to determine a behavioral feature vector of the first object based on the first time-series data and the behavioral statistics sequence, wherein the behavioral feature vector is used to characterize the time-series behavioral pattern and behavioral statistics characteristics of the first object; a second feature vector determination unit, configured to determine a fluctuation feature vector of the second object based on the second time-series data and the fluctuation statistics sequence, wherein the fluctuation feature vector is used to characterize the frequency domain characteristics and data fluctuation patterns of the second object over time; and a feature vector processing unit, configured to determine a matching score between the first object and the second object based on the behavioral feature vector and the fluctuation feature vector, and output the matching score to a display device or generate a control signal based on the matching score.

[0008] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described timing data processing method.

[0009] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when one or more programs are executed by one or more processors, the one or more processors cause the one or more processors to perform the timing data processing method described above.

[0010] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which implement the above-described timing data processing method when executed by a processor.

[0011] In this embodiment, the data processing system first acquires first time-series data, a behavioral statistical sequence, second time-series data, and a fluctuation statistical sequence. The first time-series data is a time-series data sequence about a first object, reflecting the numerical trajectory of the first object over time. The behavioral statistical sequence includes at least the frequency of occurrence of various business behaviors of the first object, supplementing the behavioral characteristics of the first object at the statistical level. The second time-series data is a time-series data sequence about a second object, reflecting the original numerical process of the second object over time. The fluctuation statistical sequence includes at least the fluctuation amplitude and extreme value decrease amplitude of the second time-series data, characterizing the intensity of fluctuations and extreme decrease events of the second object in the time domain. The data processing system determines the behavioral feature vector of the first object based on the first time-series data and the behavioral statistical sequence. The behavioral feature vector characterizes the time-series behavioral pattern and behavioral statistical characteristics of the first object, thereby integrating the dynamic behavioral change pattern of the first object in the time series with the static behavioral frequency information into a unified vector representation, which helps to reduce the information loss caused by relying on only a single dimension to describe the first object in the prior art. Furthermore, the data processing system determines the fluctuation feature vector of the second object based on the second time-series data and the fluctuation statistical sequence. The fluctuation feature vector is used to characterize the frequency domain characteristics and data fluctuation patterns of the second object over time. This integrates the energy distribution information of the second object at different frequency components with the fluctuation amplitude and extreme value decrease amplitude information in the time domain into a unified vector representation, which helps to avoid the defect of relying solely on the original numerical sequence and ignoring the fluctuation pattern characteristics.

[0012] Furthermore, the data processing system determines the matching score between the first object and the second object based on the behavioral feature vector and the fluctuation feature vector. Since the behavioral feature vector simultaneously contains the temporal behavioral pattern and behavioral statistical characteristics of the first object, and the fluctuation feature vector simultaneously contains the frequency domain characteristics and data fluctuation patterns of the second object, the matching score determined based on these two vectors can simultaneously reflect the correlation between the behavioral patterns of the first object and the fluctuation patterns of the second object. The data processing system outputs the matching score to a display device or generates a control signal based on the matching score, enabling the matching result to be output or used for subsequent processing. This embodiment of the application, by acquiring first object data containing dynamic behavioral information and statistical information, and second object data containing frequency domain fluctuation information and temporal fluctuation patterns, and extracting the corresponding behavioral feature vector and fluctuation feature vector respectively, and then determining the matching score based on the behavioral feature vector and fluctuation feature vector, facilitates the joint processing of the dynamic behavioral characteristics and fluctuation pattern characteristics of different objects. This improves the accuracy of the evaluation of the correlation matching degree between the first object and the second object, thereby solving the technical problem that the prior art cannot simultaneously process the dynamic behavioral characteristics and fluctuation pattern characteristics of different objects, resulting in low accuracy in evaluating the correlation matching degree between different objects. Attached Figure Description

[0013] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0014] Figure 1 This is a schematic diagram of an optional time-series data processing method according to an embodiment of this application;

[0015] Figure 2 This is a schematic diagram of an optional timing data processing apparatus according to an embodiment of this application. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] According to an embodiment of this application, a method embodiment for time-series data processing is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0019] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding access points are provided for users to choose to authorize or refuse. For example, interfaces are set up between this system and relevant users or organizations, providing users with corresponding access points to choose to agree to or refuse automated decision-making results; if the user chooses to refuse, the process proceeds to the expert decision-making stage.

[0020] According to the embodiments of this application, a data processing system can be used as the execution subject of the time-series data processing method of this application embodiment. The system can be a software system or an embedded system combining software and hardware. Of course, the method execution subject in the embodiments of this application can also be other forms of execution subject, such as devices, equipment, etc. It should be known by those skilled in the art that this application does not particularly limit the specific form of the method execution subject.

[0021] Figure 1 This is a schematic diagram of a timing data processing method according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0022] Step S101: Obtain first time series data, behavioral statistics sequence, second time series data, and fluctuation statistics sequence. The first time series data is a time series data sequence about a first object, and the behavioral statistics sequence includes at least the frequency of occurrence of various business behaviors of the first object. The second time series data is a time series data sequence about a second object, and the fluctuation statistics sequence includes at least the fluctuation amplitude and extreme value decrease amplitude of the second time series data. The first object and the second object are different objects with business dealings.

[0023] For example, the first time-series data can refer to a numerical sequence about a first object collected in chronological order. The behavioral statistical sequence can refer to a set of values ​​obtained by statistically analyzing various actions of the first object in business activities, and this set of values ​​includes at least the frequency of occurrence of various business behaviors. The second time-series data can refer to a numerical sequence about a second object collected in chronological order. The fluctuation statistical sequence can refer to statistics obtained by statistically analyzing the second time-series data, including at least the fluctuation amplitude and the extreme value decrease amplitude. The first object and the second object are different objects with business interactions. The first object can be an entity with time-series behavioral data, and the second object can be an entity with fluctuation change data. For example, the first object can be a service requester, and the second object can be a service provider. In a financial business scenario, the first object can be a fund customer, and the second object can be a fund product; or, in a branch management scenario, the first object can be a bank branch, and the second object can be a certain business of that branch. By acquiring the four types of data mentioned above—first time-series data, behavioral statistical sequences, second time-series data, and fluctuation statistical sequences—the data processing system can simultaneously provide a data foundation for subsequent feature extraction and matching degree calculation from two dimensions: the behavioral patterns of the first object and the fluctuation patterns of the second object. This is beneficial for comprehensively utilizing the temporal behavioral patterns of the first object and the data fluctuation patterns of the second object, thereby improving the evaluation accuracy of the matching score between the first and second objects.

[0024] In some embodiments, the first time-series data acquired by the data processing system is a numerical change sequence of a first object within a preset time window, and the behavioral statistics sequence is a behavioral statistics sequence of the first object, which includes the frequency of behavior occurrence and the duration of the state. The second time-series data acquired by the data processing system is a numerical change sequence of a candidate second object, and the fluctuation statistics sequence is a fluctuation statistics sequence of the candidate second object, which includes statistics on the local fluctuation amplitude and extreme value decrease amplitude of the numerical change sequence. The above acquisition method is beneficial for extracting dynamic behavioral patterns from the real behavioral data of the first object and extracting change features from the numerical fluctuations of the candidate second object, which is beneficial for subsequent matching degree calculation to simultaneously utilize the behavioral patterns of the first object and the change patterns of the second object.

[0025] In other embodiments, the first time-series data acquired by the data processing system is a numerical sequence of a first object within a preset time window, and the behavioral statistics sequence is the frequency of the first object's behavior types and the duration of behavior intervals. The second time-series data acquired is a service node's traffic volume sequence, and the fluctuation statistics sequence is the statistics of the fluctuation amplitude and the statistics of the extreme value decrease amplitude of the traffic volume sequence. The traffic volume sequence can refer to the sequence of business processing quantities recorded by the service node at fixed time intervals within the preset time window. This acquisition method facilitates correlation analysis from two dimensions: the behavior pattern of the first object and the traffic volume fluctuation pattern of the service node, which is beneficial for providing data basis for resource scheduling or service strategy adjustment.

[0026] Step S102: Determine the behavioral feature vector of the first object based on the first time series data and the behavioral statistical sequence, wherein the behavioral feature vector is used to characterize the time series behavioral pattern and behavioral statistical features of the first object.

[0027] For example, the behavioral feature vector can be represented as a multi-dimensional numerical vector, which is used to quantitatively describe the behavioral change patterns of the first object over time and its distribution characteristics at the behavioral statistical level. The temporal behavioral pattern of the first object can refer to the dynamic patterns such as behavioral trends, periodicity, or abrupt changes exhibited by the first object in a continuous time series. Behavioral statistical features can refer to the static features obtained after statistical processing such as counting, frequency, or proportion of various behaviors of the first object in business activities. The data processing system determines the behavioral feature vector of the first object based on the first time series data and the behavioral statistical sequence, which can integrate the dynamic time series information and static statistical information of the first object into a unified vector representation. This is beneficial for a more comprehensive characterization of the behavioral attributes of the first object and for more accurate matching calculations with the second object.

[0028] In some embodiments, the data processing system can divide the first time series data into multiple subsequences according to a preset time window, extract the statistical feature values ​​of each subsequence, and concatenate the extracted multiple statistical feature values ​​with the occurrence frequency in the behavioral statistical sequence to form a behavioral feature vector. The embodiments of this application are beneficial to reducing the dimensionality of the first time series data, while retaining the key statistical information in the time series data, thereby improving the computational efficiency of the behavioral feature vector.

[0029] In other embodiments, the data processing system uses a deep learning network to jointly encode the first time-series data and the behavioral statistical sequence. The system uses the numerical value at each time point in the first time-series data and the frequency information in the corresponding behavioral statistical sequence as multidimensional inputs for that time point. After layer-by-layer transformation by the deep learning network, a behavioral feature vector is output. This embodiment of the application is beneficial for uncovering the deep nonlinear correlation between the first time-series data and the behavioral statistical sequence, thereby generating a more representative behavioral feature vector.

[0030] Step S103: Determine the fluctuation feature vector of the second object based on the second time series data and the fluctuation statistical sequence, wherein the fluctuation feature vector is used to characterize the frequency domain characteristics and data fluctuation patterns of the second object over time.

[0031] For example, the fluctuation feature vector can be represented as a multi-dimensional numerical vector, which is used to quantitatively describe the frequency domain energy distribution characteristics of the second object in the time dimension and the fluctuation change pattern of the data sequence. The frequency domain characteristics of the second object changing over time can refer to the spectral energy, dominant frequency component, or frequency band distribution extracted after transforming the second time series data from the time domain to the frequency domain. The data fluctuation pattern can refer to the dynamic characteristics of the second time series data on the time axis, such as the amplitude of change, rate of change, or extreme value distribution. The data processing system determines the fluctuation feature vector of the second object based on the second time series data and the fluctuation statistical sequence, which can integrate the time domain fluctuation information and frequency domain energy information of the second object into a unified vector representation. This is beneficial for capturing the dynamic change pattern of the second object from multiple perspectives and provides a more comprehensive feature basis for subsequent matching calculations with the first object.

[0032] In some embodiments, the data processing system can perform a Fourier transform on the second time-series data, extract the energy values ​​of multiple frequency bands in the spectrum as frequency domain features, and use the fluctuation amplitude and extreme value decrease amplitude in the fluctuation statistical sequence as supplementary features. All features, including the frequency domain features and supplementary features, are then concatenated to form a fluctuation feature vector. This embodiment of the application is advantageous in combining the periodic fluctuation information and statistical fluctuation information of the second time-series data, thereby improving the ability of the fluctuation feature vector to characterize the changing patterns of the second object.

[0033] In other embodiments, the data processing system can perform wavelet transform on the second time-series data, extract wavelet coefficient energy from different scales as frequency domain features, and use the fluctuation amplitude and extreme value descent amplitude in the fluctuation statistical sequence as time domain statistical features. The frequency domain features and time domain statistical features are then weighted and fused to generate a fluctuation feature vector. This embodiment of the application is advantageous for simultaneously capturing the fluctuation characteristics of the second time-series data in both low-frequency trends and high-frequency details, thereby enhancing the expressive diversity of the fluctuation feature vector.

[0034] Step S104: Determine the matching score between the first object and the second object based on the behavior feature vector and the fluctuation feature vector, and output the matching score to the display device or generate a control signal based on the matching score.

[0035] For example, the matching score can be represented as a numerical value or vector, which is used to quantify the degree of association or fit between the first object and the second object. The data processing system determines the matching score based on the behavioral feature vector and the fluctuation feature vector, enabling a quantitative comparison of the behavioral pattern of the first object and the fluctuation pattern of the second object, thereby generating a quantitative index reflecting the degree of matching between the two. The data processing system outputs the matching score to a display device, allowing operators to visually view the matching results; alternatively, the data processing system generates a control signal based on the matching score, which can trigger subsequent business processing operations. The data processing system of this application embodiment facilitates the transformation of the complex association between the first object and the second object into a quantifiable numerical index.

[0036] In some embodiments, the data processing system can calculate the vector distance between the behavioral feature vector and the fluctuation feature vector, and use the reciprocal or negative correlation transformation value of this vector distance as the matching score. The vector distance can include at least one of Euclidean distance, cosine distance, or Mahalanobis distance. This approach facilitates the rapid calculation of feature similarity between the first object and the second object, thereby improving the efficiency of matching score calculation.

[0037] In other embodiments, the data processing system inputs behavioral feature vectors and fluctuation feature vectors into a pre-trained matching scoring model, and outputs a matching score through forward computation of the matching scoring model. The matching scoring model can be built based on any machine learning model, such as neural networks, support vector machines, or logistic regression. This approach is beneficial for learning the non-linear mapping relationship between behavioral feature vectors and fluctuation feature vectors, thereby improving the prediction accuracy of the matching score.

[0038] In some optional embodiments, determining the behavioral feature vector of the first object based on the first time-series data and the behavioral statistical sequence includes: the data processing system can input the first time-series data and the behavioral statistical sequence into a bidirectional long short-term memory network, extract the temporal dependency features between the first time-series data and the behavioral statistical sequence through the bidirectional long short-term memory network, and determine the hidden state vector of the temporal dependency features at each time step, wherein the hidden state vector is used to characterize the cumulative temporal features from the starting time step to the current time step; obtain the attention weight of each time step, wherein the attention weight is used to characterize the degree of contribution of the hidden state vector at the corresponding time step to the behavioral feature vector; and perform a weighted summation of the hidden state vector at each time step according to the attention weight of each time step to obtain the behavioral feature vector.

[0039] For example, a bidirectional long short-term memory network can be represented as a recurrent neural network structure composed of a forward long short-term memory network and a backward long short-term memory network, used to capture the contextual dependencies in time-series data from two directions. Since there is a temporal correlation between the first time-series data and the behavioral statistical sequence, processing from only one direction easily loses the impact of future information on the current time step. Therefore, a bidirectional structure can utilize both past and future contextual information simultaneously. The hidden state vector can represent the encoding result output by the bidirectional long short-term memory network at each time step. This encoding result can fuse historical information from the starting time step to the current time step and future information from the current time step to the ending time step, thereby providing a feature representation containing global temporal dependencies for each time step. However, the contribution of the hidden state vectors at different time steps to characterizing the overall behavioral pattern of the first object is not the same; for example, certain time points where significant trading activities occur are more representative than ordinary trading days. Therefore, this application introduces an attention mechanism, where each time step corresponds to an attention weight, and this weight value is used to measure the importance of the hidden state vector at that time step in the final generation of the behavioral feature vector. The data processing system first extracts temporal dependency features using a bidirectional long short-term memory network and outputs the hidden state vectors for each time step. Then, it calculates the attention weight for each time step and finally weights and sums the hidden state vectors of each time step according to their corresponding attention weights. This process helps obtain behavioral feature vectors that combine temporal dependencies with key event information. By linking the bidirectional long short-term memory network with the attention mechanism, the system can automatically focus on time steps that contribute significantly to the expression of behavioral patterns and suppress interference from irrelevant or noisy time steps, thereby giving the generated behavioral feature vectors stronger representational capabilities.

[0040] For example, the data processing system uses a customer channel to encode the first time-series data and behavioral statistical sequences. The input to the customer channel can be expressed as formula (1):

[0041] Formula (1);

[0042] in, Indicates the first time series data. This indicates the total number of time points within the preset time window, for example... This is beneficial for capturing the long-term behavioral dependencies of the first object. This represents a sequence of behavioral statistics, including the frequency of the first object's behavior. Duration of state and the magnitude of extreme value decrease The extreme value decrease range can represent the maximum decrease in the first time series data from the peak to the subsequent trough.

[0043] The network structure of the customer channel can be represented as follows:

[0044] Formula (2);

[0045] in, Indicates the first time series data. Represents a behavioral statistical sequence. This represents a vector concatenation operation. This represents a multilayer perceptron. This represents a bidirectional long short-term memory network. This represents the attention mechanism. The data processing system will process the first time-series data. Inputting the data into a bidirectional long short-term memory network, the hidden state vector at each time step is extracted; the behavioral statistics sequence is then processed. The input is fed into a multilayer perceptron, which maps the behavioral statistical features into feature vectors of the same dimension as the hidden state vectors; the output of the bidirectional long short-term memory network is concatenated with the output of the multilayer perceptron. This represents a vector concatenation operation; then, an attention mechanism is used to perform a weighted summation of the concatenated features to obtain the behavioral feature vector. The dimension of the behavioral feature vector can be set to 128. Through the aforementioned customer channels, the data processing system can simultaneously utilize the long-term dependency information of the first time series data and the static statistical characteristics of the behavioral statistical sequence, which is beneficial for generating a more comprehensive behavioral feature vector.

[0046] In some embodiments, when acquiring the attention weights for each time step, the data processing system inputs the hidden state vector of each time step into a single-layer perceptron, calculates the unnormalized attention score for each time step using the single-layer perceptron, and then performs a normalized exponential function operation on the unnormalized attention scores of multiple time steps to obtain the attention weight for each time step. This approach helps to constrain the attention weights to between zero and one, and ensures that the sum of the attention weights for multiple time steps is one, thereby improving the interpretability and stability of the attention weights.

[0047] In other embodiments, before the data processing system performs a weighted summation of the hidden state vectors at each time step based on the attention weights at each time step, it can first apply a temperature coefficient scaling process to the attention weights across multiple time steps. This involves dividing the attention weights at each time step by the temperature coefficient before performing the normalized exponential function operation. This approach helps adjust the smoothness of the attention weight distribution, thereby controlling the degree to which the behavioral feature vector focuses on features at different time steps.

[0048] In some optional embodiments, determining the fluctuation feature vector of the second object based on the second time-series data and the fluctuation statistical sequence includes: the data processing system can convert the second time-series data and the fluctuation statistical sequence into a time-series feature sequence; inputting the time-series feature sequence into a convolutional neural network, wherein the convolutional neural network extracts local features of the time-series feature sequence at different window lengths through multiple convolutional kernels, wherein the local features include the upward trend, downward trend, or fluctuation amplitude of the time-series feature sequence at consecutive time points; combining all local features into a convolutional feature sequence; inputting the convolutional feature sequence into a gated recurrent unit in time step order, and outputting the fluctuation feature vector after processing the convolutional feature sequence by the gated recurrent unit.

[0049] For example, a temporal feature sequence can be represented as a matrix or vector sequence, which can be formed by concatenating the values ​​at each time point in the second time-series data with the corresponding statistics in the fluctuation statistics sequence. Since the second time-series data reflects absolute numerical changes, while the fluctuation statistics sequence can provide supplementary information on the intensity of local fluctuations, aligning and concatenating the two at time points allows the input at each time step to simultaneously contain the original value and statistical features, which is beneficial for providing more diverse input representations for subsequent local pattern extraction. A convolutional neural network can be represented as a neural network structure composed of multiple convolutional layers, each containing at least one convolutional kernel. The convolutional kernel is used to slide across the temporal feature sequence and perform convolution operations to extract local patterns. Local features can refer to the pattern representation extracted by the convolutional kernel within a local window of the temporal feature sequence. For example, an upward trend can represent a pattern of gradually increasing values ​​within the window, a downward trend can represent a pattern of gradually decreasing values ​​within the window, and fluctuation amplitude can represent the difference between the maximum and minimum values ​​within the window. Since fluctuation patterns at different time scales have different meanings in characterizing the second object—for example, short-term fluctuations reflect noise or high-frequency trading, while long-term trends reflect fundamental changes—multiple convolutional kernels with different window lengths can be used in parallel to extract local features, simultaneously capturing various fluctuation patterns from short-term to long-term. The total number of time points in the first time series data within the preset time window can be set to greater than or equal to 360, which is beneficial for capturing the long-term behavioral dependencies of the first object.

[0050] Optionally, combining the local features extracted by all convolutional kernels along the feature dimension to form a convolutional feature sequence is beneficial for preserving temporal order, but the representation at each time step already incorporates multi-scale local contextual information. A gated recurrent unit (ROU) can represent a recurrent neural network structure with reset and update gates, used for progressively encoding sequential data. The data processing system inputs the convolutional feature sequence into the gated ROU in time-step order. The gated ROU, through its reset and update gate mechanisms, can selectively forget or pass historical information, thereby encoding long-term temporal dependencies in the convolutional feature sequence. This approach facilitates the use of convolutional neural networks to capture local fluctuation patterns in second-series data at different time scales, and the use of gated ROUs to capture long-term temporal dependencies between these fluctuation patterns, thus generating more expressive fluctuation feature vectors.

[0051] In some embodiments, before inputting the temporal feature sequence into the convolutional neural network, the data processing system can perform a padding operation on the temporal feature sequence to ensure that the output convolutional feature sequence of the convolutional neural network maintains the same number of time steps as the input temporal feature sequence. This approach helps maintain temporal alignment and ensures that the gated recurrent unit can correctly correspond to the local features of each time step when processing in time-step order.

[0052] In other embodiments, when the data processing system combines all local features into a convolutional feature sequence, it concatenates multiple local features extracted by multiple convolutional kernels at each time step along that time step dimension to generate a multidimensional convolutional feature vector corresponding to each time step. Then, it arranges the multidimensional convolutional feature vectors from multiple time steps in chronological order to form a convolutional feature sequence. This approach is beneficial for fusing local feature information under different window lengths, thereby enriching the expressive dimensions of the convolutional feature sequence.

[0053] In some optional embodiments, the output of a fluctuation feature vector after processing the convolutional feature sequence by a gated recurrent unit includes: the data processing system can obtain the feature value corresponding to the (i+1)th time step in the convolutional feature sequence as the input information of the (i+1)th time step, where i is an integer greater than 1; obtain the hidden state vector of the i-th time step output by the gated recurrent unit, where the hidden state vector of the i-th time step is used to characterize the feature representation obtained by the gated recurrent unit after encoding the input information from the first time step to the i-th time step; and determine the output value of the reset gate and the update gate based on the input information of the (i+1)th time step and the hidden state vector of the i-th time step. The output values ​​are as follows: the output value of the reset gate controls the proportion of information ignored in the hidden state vector of the i-th time step; the output value of the update gate controls the data fusion ratio between the input information of the (i+1)-th time step and the hidden state vector of the i-th time step; based on the output value of the reset gate and the input information of the (i+1)-th time step, candidate hidden state vectors are determined, where the candidate hidden state vectors represent the current time step feature representation of the candidate; based on the output value of the update gate, the hidden state vector of the i-th time step, and the candidate hidden state vectors, the hidden state vector of the (i+1)-th time step is determined; and the hidden state vector of the last time step is used as the fluctuation feature vector.

[0054] For example, a reset gate can represent a gating structure within a gated recurrent unit. The output value of this gating structure controls the proportion of information ignored in the hidden state vector of the previous time step; the closer the output value is to one, the more historical information is retained, and the closer the output value is to zero, the more historical information is ignored. An update gate can represent another gating structure within the gated recurrent unit. The output value of this gating structure controls the data fusion ratio between the input information of the current time step and the hidden state vector of the previous time step; the closer the output value is to one, the greater the contribution of the current input information, and the closer the output value is to zero, the greater the contribution of the historical hidden state. A candidate hidden state vector can represent a temporary feature representation generated by the gated recurrent unit at the current time step, which fuses the historical information processed by the reset gate with the current input information.

[0055] Optionally, the gated recurrent unit's processing of the convolutional feature sequence follows a time-step progression principle. Starting from the first time step, it sequentially updates the hidden state vector of the current time step using the hidden state vector output from the previous time step. When processing the (i+1)th time step, the data processing system first obtains the feature value corresponding to the (i+1)th time step in the convolutional feature sequence as the input information for the (i+1)th time step, and simultaneously obtains the hidden state vector of the i-th time step already calculated by the gated recurrent unit at the i-th time step. The hidden state vector of the i-th time step has already encoded the sequence information from the first to the i-th time step into a fixed-dimensional feature representation through forward computation, while the input information of the (i+1)th time step provides the local features of the (i+1)th time step. To adaptively determine how much information is retained in the hidden state vector of the i-th time step and how much of the input information of the (i+1)th time step is fused, the data processing system calculates the output values ​​of the reset gate and update gate based on the input information of the (i+1)th time step and the hidden state vector of the i-th time step, respectively. The output of the reset gate is used to filter which parts of the hidden state vector at time step i should be ignored, resulting in reset historical information. The output of the update gate is used to balance the weights of the hidden state vector at time step i and the candidate hidden state vectors in the final output. After obtaining the output of the reset gate, the data processing system multiplies the output of the reset gate element-wise with the hidden state vector at time step i to obtain the reset historical information vector. This reset historical information vector is then concatenated with the input information at time step i+1, and a candidate hidden state vector is generated through a nonlinear transformation. The candidate hidden state vector represents a temporary representation that depends only on the input information at time step i+1 and the reset historical information. Subsequently, based on the output of the update gate, the data processing system performs a weighted fusion between the hidden state vector at time step i and the candidate hidden state vectors to obtain the hidden state vector at time step i+1. The weighted fusion calculation method can be as follows: multiply the output value of the update gate by the input information or candidate hidden state vector of the (i+1)th time step, multiply the output value of the update gate by one minus the output value of the update gate by the hidden state vector of the ith time step, and then add the two together. Through the collaborative calculation of the reset gate and the update gate, the gated recurrent unit can flexibly adjust the retention ratio between long-term dependency information and short-term input information. After recursively processing from the first time step to the last time step in the above manner, the data processing system outputs the hidden state vector of the last time step as the fluctuation feature vector. Since the hidden state vector of the last time step has already encoded the temporal dependency relationship of the entire convolutional feature sequence into a fixed-dimensional feature representation through the gating mechanism, using the hidden state vector of the last time step as the fluctuation feature vector is beneficial to ensure that the fluctuation feature vector contains the long-term dependency information in the second time series data.

[0056] In some embodiments, when determining the output value of the reset gate, the data processing system concatenates the input information of the (i+1)th time step with the hidden state vector of the ith time step and inputs the result into the first fully connected layer. Then, a first activation function constrains the output value to between zero and one. When determining the output value of the update gate, the system concatenates the input information of the (i+1)th time step with the hidden state vector of the ith time step and inputs the result into the second fully connected layer. Then, a second activation function constrains the output value to between zero and one. This approach facilitates the independent learning of the parameters of the reset and update gates, thereby improving the selective modeling capability of the gated recurrent unit for temporal features.

[0057] In other embodiments, when the data processing system determines the candidate hidden state vector based on the output value of the reset gate and the input information at the (i+1)th time step, it first multiplies the output value of the reset gate element-wise with the hidden state vector at the ith time step to obtain the reset historical information vector. Then, it concatenates the reset historical information vector with the input information at the (i+1)th time step, inputs the concatenation result to the third fully connected layer, and processes it through the hyperbolic tangent activation function to obtain the candidate hidden state vector. This approach allows for flexible control of the retention ratio of historical information when generating the candidate hidden state vector, thereby enhancing the gated recurrent unit's ability to capture important patterns in the time series.

[0058] For example, the data processing system uses a fund channel to encode the second time series data and volatility statistics sequence. The input to the fund channel can be represented as a multi-scale feature matrix:

[0059] Formula (3);

[0060] Among them, the first column The first column represents the second time-series data, i.e., the numerical change sequence of the second object. The remaining columns represent different statistics in the fluctuation statistics sequence, such as fluctuation amplitude statistics, extreme value decline amplitude statistics, and other quantitative indicators reflecting the fluctuation characteristics of the second object, such as the degree of correlation with the market benchmark. Each row of the above input matrix corresponds to a multidimensional feature of a time step.

[0061] The network structure of the fund channel can be represented as:

[0062] Formula (4);

[0063] in, Indicates the second time series data. This represents the feature matrix corresponding to the fluctuation statistics sequence. This represents a one-dimensional convolutional neural network. Indicates a gated loop unit. This represents feature concatenation. The data processing system can output a volatility feature vector, which can be set to 128 dimensions. The aforementioned fund channel is beneficial for simultaneously capturing local volatility patterns and multidimensional volatility statistical features of the second time series data, thereby enhancing the expressive power of the volatility feature vector.

[0064] In some optional embodiments, determining the matching score between the first object and the second object based on the behavioral feature vector and the fluctuation feature vector includes: the data processing system can calculate the first derivative value of the first time series data at each time point to obtain a first derivative sequence; calculate the first derivative value of the second time series data at each time point to obtain a second derivative sequence; obtain the dynamic time warping distance between the first derivative sequence and the second derivative sequence, wherein the dynamic time warping distance is the minimum cumulative sum of the distances between corresponding points of the first derivative sequence and the second derivative sequence under the alignment path that satisfies the boundary conditions, continuity and monotonicity constraints; determine the sequence matching degree based on the dynamic time warping distance, wherein the sequence matching degree is inversely proportional to the dynamic time warping distance; generate matching degree feature data based on the sequence matching degree, wherein the matching degree feature data is used to characterize the degree of matching between the first object and the second object; and determine the matching score based on the behavioral feature vector, the fluctuation feature vector and the matching degree feature data.

[0065] For example, the first derivative value can represent the rate of change of the first or second time series data at each time point, i.e., the difference between adjacent time points. The first derivative sequence can represent a sequence formed by arranging the first derivative values ​​of all time points of the first time series data in chronological order. The dynamic time warping distance can represent the minimum cumulative sum of distances between all corresponding points of two sequences when aligning their time axes, under the constraints of boundary conditions (e.g., start-point alignment, end-point alignment), continuity (e.g., the alignment path moves one step at a time), and monotonicity (i.e., time does not regress). The sequence matching degree can represent a numerical value that is inversely proportional to the dynamic time warping distance; that is, the smaller the dynamic time warping distance, the greater the sequence matching degree, and vice versa. Matching degree feature data can represent feature values ​​or feature vectors generated based on the sequence matching degree, used to quantify the degree of matching between the first and second objects.

[0066] Optionally, the first and second time-series data have different numerical ranges and absolute value levels. Directly calculating the distance between the original data sequences is easily affected by absolute value shifts. The data processing system calculates the first derivative of the first time-series data at each time point to obtain the first derivative sequence, and calculates the first derivative of the second time-series data at each time point to obtain the second derivative sequence. This helps reduce the interference from the absolute value shifts between the first and second time-series data, allowing subsequent comparisons to focus on the trend itself. The first and second time-series data may exhibit local stretching or compression misalignment on the time axis. For example, the behavioral changes of the first object and the fluctuation changes of the second object may have a phase difference in time. Directly comparing the first and second derivative sequences point-by-point would introduce alignment errors. The data processing system obtains the dynamic time-warped distance between the first and second derivative sequences. This allows for the calculation of the minimum cumulative sum of distances between corresponding points in the two derivative sequences through the optimal alignment path, under the conditions of satisfying boundary conditions, continuity, and monotonicity constraints. This enables non-linear time axis alignment and reduces alignment problems caused by phase differences. The data processing system determines the sequence matching degree based on the dynamic time warping distance. The sequence matching degree is inversely proportional to the dynamic time warping distance, which is beneficial because a smaller dynamic time warping distance indicates a more synchronized trend between the first and second derivative sequences, resulting in a higher sequence matching degree; conversely, a larger dynamic time warping distance indicates a less synchronized trend, resulting in a lower sequence matching degree. The data processing system generates matching degree feature data based on the sequence matching degree, which is beneficial for converting the sequence matching degree into feature values ​​or feature vectors that can be processed by subsequent neural networks. The data processing system determines the matching score based on the behavioral feature vector, fluctuation feature vector, and matching degree feature data. This allows for the comprehensive utilization of the temporal behavioral features of the first object, the fluctuation features of the second object, and the degree of matching of the changing trends between the first and second objects, thereby improving the accuracy of the matching score in characterizing the dynamic relationship between the first and second objects.

[0067] In some embodiments, when obtaining the dynamic time-warped distance between the first derivative sequence and the second derivative sequence, the data processing system can first normalize the first and second derivative sequences, mapping their numerical ranges to the same interval, and then calculate the dynamic time-warped distance between the normalized first and second derivative sequences. This approach helps reduce the impact of differences in dimensions or amplitudes between the first and second derivative sequences on the calculation of the dynamic time-warped distance, thereby improving the accuracy of sequence matching evaluation.

[0068] In other embodiments, when determining the sequence matching degree based on the dynamic time warp distance, the data processing system can set the sequence matching degree as the reciprocal of the dynamic time warp distance, and truncate the sequence matching degree to a preset maximum value when the dynamic time warp distance is less than a preset threshold. This approach helps avoid the problem of the sequence matching degree becoming too large and unstable due to the small dynamic time warp distance, thereby improving the numerical stability and processability of the sequence matching degree.

[0069] For example, the matching degree feature data between the first object and the second object can be calculated using the following improved dynamic time warping algorithm:

[0070] Formula (5);

[0071] in, The input data representing the first object includes the first time-series data and the behavioral statistics sequence. The input data for the second object includes the second time series data and the fluctuation statistics sequence. Indicates the first time series data in the 1st... The first derivative values ​​at each time point Indicates the second time series data in the first... The first derivative values ​​at each time point. This represents the set of aligned paths that satisfy boundary conditions, continuity, and monotonicity constraints. This represents the minimum cumulative distance among all possible alignment paths. The output value is... This is the matching feature data, used to characterize the degree of synchronization between the first object and the second object in terms of their changing trends.

[0072] In some optional embodiments, determining the matching score based on behavioral feature vectors, fluctuation feature vectors, and matching degree feature data includes: the data processing system can fuse the behavioral feature vectors, fluctuation feature vectors, and matching degree feature data to obtain a fused feature tensor; inputting the fused feature tensor into a gated cross-network, wherein the gated cross-network includes a first branch and a second branch, the first branch generates a gate signal through a gated activation function, and the second branch generates residual features through a linear transformation, wherein the gate signal is used to control the retention ratio of different feature dimensions in the fused feature tensor, and the residual features are feature vectors obtained after linear mapping of the fused feature tensor, used to retain the original information of the fused feature tensor; and determining the matching score based on the gate signal and the residual features.

[0073] For example, the fusion feature tensor can represent a multidimensional array obtained by concatenating or combining behavioral feature vectors, fluctuation feature vectors, and matching degree feature data along the feature dimensions. A gated cross-network can represent a neural network structure containing two parallel branches, where the first branch generates a gating signal and the second branch generates residual features. The gating signal can represent a vector, where each element takes a value between zero and one, used to control the proportion of the corresponding feature dimension retained in the fusion feature tensor. The residual features can represent feature vectors obtained by linearly transforming the fusion feature tensor, used to retain the original information in the fusion feature tensor.

[0074] Optionally, the behavioral feature vector, fluctuation feature vector, and matching degree feature data come from different feature extraction channels. Directly using these three features for matching score prediction presents two problems: firstly, the importance of different features varies, and simple concatenation can introduce redundancy or noise; secondly, deep networks are prone to losing original information in the input during layer-by-layer transformation. The data processing system first fuses the behavioral feature vector, fluctuation feature vector, and matching degree feature data to obtain a fused feature tensor. This fused feature tensor organizes the features from the three sources into a unified multidimensional array, which is beneficial for providing a unified input for subsequent adaptive filtering. The data processing system inputs the fused feature tensor into a gated cross-network. In the first branch of the gated cross-network, a gating signal is generated through a gating activation function. Each element of the gating signal independently controls the retention ratio of the corresponding feature dimension in the fused feature tensor, which is beneficial for adaptively filtering out feature dimensions that contribute significantly to matching score prediction while suppressing feature dimensions that contribute less or contain noise. In the second branch of the gated cross-network, residual features are generated through linear transformation. These residual features are feature vectors obtained by linearly mapping the fused feature tensor, which helps preserve the original information in the fused feature tensor and avoids information loss due to over-filtering by the gating signal. The data processing system determines the matching score based on the gating signal and the residual features. For example, the gating signal is multiplied element-wise with the fused feature tensor to obtain the gated feature vector, then the gated feature vector is added element-wise with the residual features to obtain the fused output vector. Finally, the fused output vector is mapped to the matching score. Through this parallel input of the fused feature tensor to the gating and residual branches, the gating signal can adaptively filter important feature dimensions, while the residual branch preserves the original information, which helps alleviate the information decay problem in deep networks and improves the prediction accuracy of the matching score.

[0075] In some embodiments, when the data processing system generates a gating signal through the first branch, it inputs the fused feature tensor into the first fully connected layer, and then constrains the output value of the first fully connected layer to between zero and one using a gating activation function to obtain the gating signal. The gating activation function can be a normalized exponential function or a logistic function. This approach allows each element of the gating signal to independently control the retention ratio of the corresponding feature dimension, thereby improving the fine-grained filtering capability of the gating cross-network for the fused feature tensor.

[0076] In other embodiments, when determining the matching score based on the gating signal and residual features, the data processing system multiplies the gating signal element-wise with the fused feature tensor to obtain a gated feature vector. Then, it adds the gated feature vector element-wise with the residual features to obtain a fused output vector. This fused output vector is then input to the output layer to obtain the matching score. This approach facilitates combining the gated features with the original residual features, thereby preserving important features while avoiding the gradient vanishing problem and enhancing the training stability of the gated cross-network.

[0077] For example, data processing systems use gated cross-networks to process behavioral feature vectors. , fluctuation eigenvector Matching feature data The process of fusion can be represented as follows:

[0078] Formula (6);

[0079] in, This indicates that behavioral feature vectors, matching degree feature data, and fluctuation feature vectors are concatenated or combined along the feature dimension. Represents the gated activation function, for example function, and This represents the learnable weight matrix. Represents the linear rectification activation function. This represents the concatenation operation. Through the above calculations, the data processing system obtains the fused feature tensor. .

[0080] After obtaining the fused feature tensor, the data processing system can determine the matching score and the classification label of the first object.

[0081] The matching score can be calculated as follows:

[0082] Formula (7);

[0083] in, The function maps the fused feature tensor to a value between 0 and 1, which is the matching score between the first and second objects.

[0084] The calculation method for the category label of the first object can be expressed as follows:

[0085] Formula (8);

[0086] in, The function maps the fused feature tensor to a probability distribution across multiple categories, and selects the category with the highest probability as the classification label for the first object. This classification label can be used to represent the category to which the first object belongs, which is beneficial for grouping the first object into different categories.

[0087] In this embodiment, the data processing system can employ a dual-channel deep neural network to extract features and calculate matching degrees for a first object and a second object. A dual-channel deep neural network can refer to a parallel deep learning topology consisting of two independent feature extraction channels, each processing heterogeneous input data types, and achieving cross-channel feature interaction through a fusion layer. The mathematical formalization of a dual-channel deep neural network can be expressed as:

[0088] Formula (9);

[0089] in, The input data for the first object includes the first time series data and the behavioral statistics sequence. The input data for the second object includes the second time series data and the fluctuation statistics sequence. Represents a behavioral feature vector; Represents the eigenvector of fluctuation; This represents the matching degree feature data; This represents the feature tensor after fusion.

[0090] In some optional embodiments, after determining the matching score between the first object and the second object based on the behavioral feature vector and the fluctuation feature vector, the method further includes: the data processing system can obtain the matching score between the first object and each second object for multiple second objects; sort the second objects from high to low according to the matching score, and select the top K second objects to generate a sorted list, where K is a preset positive integer.

[0091] For example, the sorted list can represent a sequence formed by arranging multiple second objects in descending order of matching scores, where the first K positions of the sequence correspond to the K second objects with the highest matching scores. K can represent a preset positive integer, such as 5, 10, or 20. For each of the multiple second objects, the data processing system obtains the matching score between the first object and that second object, then sorts the multiple second objects in descending order of matching scores, and selects the second objects in the first K positions after sorting to generate the sorted list. This method facilitates the rapid selection of the top K second objects with the highest matching degree from a large number of second objects, thereby providing a refined candidate set for subsequent recommendations or decisions and reducing the number of candidate objects that the data processing system needs to process.

[0092] In some embodiments, after the data processing system sorts the second objects from highest to lowest according to their matching scores, if multiple second objects have the same matching score, it performs a secondary sorting according to the preset priority of the second objects or according to the update time of the corresponding time-series data, and then selects the top K second objects to generate a sorted list. This method is beneficial for further distinguishing the order of second objects when matching scores are the same, thereby improving the certainty and stability of the sorting results.

[0093] In other embodiments, after generating a sorted list by selecting the top K second objects, the data processing system also calculates the difference between the matching score of each second object in the sorted list and its corresponding historical matching score. When the difference exceeds a preset change threshold, the second object is marked as a candidate object with a significant change in matching degree in the sorted list. This method helps to highlight second objects with large changes in matching degree, thereby providing operators with richer decision-making reference information.

[0094] In some other embodiments, after generating the sorted list, the data processing system also compares the matching degree feature data with a preset threshold; when the matching degree feature data is lower than the preset threshold, a prompt message is generated, which is used to indicate the adjustment of the association between the first object and the corresponding second object.

[0095] In one optional implementation, the data processing system performs the time-series data processing method according to the following process:

[0096] Step one is data preprocessing, which involves the data processing system performing sliding window standardization on the first time series data. For example, for each time point in the first time series data, data within a preset window length before and after that time point are taken, and the mean and standard deviation within that preset window are calculated. Then, the original value at that time point is subtracted from the mean and divided by the standard deviation to obtain the standardized value. Sliding window standardization helps reduce systematic fluctuations in the first time series data, allowing subsequent processing to focus on the dynamic behavioral characteristics of the first object itself.

[0097] Step two involves parallel feature extraction, where the data processing system executes two feature extraction channels in parallel: in the first channel, the behavioral feature vector of the first object is determined based on the first time-series data and the behavioral statistical sequence; in the second channel, the fluctuation feature vector of the second object is determined based on the second time-series data and the fluctuation statistical sequence. Parallel execution improves the processing efficiency of the data processing system and shortens the calculation time for the matching score.

[0098] Step three involves calculating the dynamic fit, whereby the data processing system calculates the dynamic fit between the first and second objects. The dynamic fit can be achieved using an improved dynamic time warping algorithm. This algorithm calculates the dynamic time warping distance between the first derivative sequences of the first and second time series data and generates matching feature data based on this distance. This matching feature data acts as a bridge, quantifying the microscopic synchronicity of the first and second objects in their changing trends.

[0099] Step four involves fusion layer processing and dual-objective decision-making. The data processing system inputs behavioral feature vectors, fluctuation feature vectors, and matching degree feature data into the fusion layer to obtain a fusion feature tensor. Based on this tensor, the system executes a dual-objective decision: a first objective and a second objective. The first objective generates cluster labels for first objects. The system can perform cluster analysis on the fusion feature tensors of multiple first objects or different time periods for the same first object. For example, using the K-means clustering algorithm, the system divides the fusion feature tensor into a predetermined number of categories and generates a corresponding cluster label for each first object. These cluster labels represent the group type to which the first object belongs, such as category one, category two, category three, or category four, facilitating group segmentation and differentiated processing of the first object. The second objective generates a ranked list of second objects. Based on the matching score between the first object and each second object, the system sorts the second objects from highest to lowest score, selecting the top K second objects to generate a ranked list for recommendation or filtering.

[0100] In some optional embodiments, the first time series data is the total user return sequence of the target user within a preset time window, and the behavioral statistics sequence is the transaction behavior statistics sequence of the target user, which includes the frequency of transactions and the duration of holding positions of the target user; the second time series data is the fund return sequence of the fund to be recommended, and the volatility statistics sequence is the fund volatility statistics sequence of the fund to be recommended, which includes the local volatility amplitude statistics and the extreme value decline amplitude statistics of the fund return sequence.

[0101] For example, a target user can represent an object for which fund recommendation services need to be provided. A preset time window can represent a pre-defined time length, such as 360 days or 180 days. The user's total return sequence can represent a sequence formed by arranging daily or monthly returns within the preset time window. The transaction behavior statistics sequence can represent a set of values ​​obtained by statistically analyzing various transaction actions of the target user within the preset time window. This set of values ​​includes transaction frequency and holding duration, where transaction frequency can represent the total number of times the target user performs fund buy or sell operations within the preset time window, and holding duration can represent the average or total number of days the target user holds a particular fund. Funds to be recommended can represent candidate fund products that can be recommended to the target user. The fund return sequence can represent a sequence formed by arranging the funds to be recommended within the preset time window according to their daily or monthly returns. A fund volatility statistical sequence can represent the set of statistics obtained after volatility analysis of a fund return sequence. This set of statistics includes local volatility statistics and extreme value decline statistics. Local volatility statistics represent the range of fluctuations in the fund return sequence across multiple consecutive time points, while extreme value decline statistics represent the maximum decline in the fund return sequence from its peak to its subsequent trough. The data processing system uses this data as input, which helps to uncover dynamic risk preferences from the target users' actual trading behavior and extract risk characteristics from the fund's historical return volatility. This provides a data foundation adapted to the fund recommendation scenario for subsequent matching degree calculations.

[0102] In some embodiments, the data processing system acquires a daily cumulative numerical sequence of the numerical changes of the first object within a preset time window, and the behavioral statistics sequence includes the frequency of behavior occurrence and the average duration of the state. The data processing system uses the numerical change sequence of the second object as second time-series data, and uses the local fluctuation amplitude statistics and extreme value decrease amplitude statistics of the second object as fluctuation statistics sequences. The above acquisition method is beneficial for matching and analyzing the long-term behavioral data of the first object and the fluctuation characteristics of the second object, and is suitable for business scenarios that require correlation assessment based on historical behavior and object fluctuation patterns.

[0103] Optionally, the statistical sequence of target user's trading behavior acquired by the data processing system also includes the trading interval duration, which represents the time interval between two adjacent trading actions. The data processing system incorporates trading frequency, holding duration, and trading interval duration into the statistical sequence of behavior. This approach helps to more comprehensively characterize the trading behavior patterns of target users, thereby improving the ability of the behavioral feature vector to distinguish between user types such as market timing or long-term holding.

[0104] In other embodiments, the numerical change sequence of the first object acquired by the data processing system within a preset time window is a sequence after sliding window standardization, and the behavioral statistics sequence includes the frequency of occurrence of different behavioral types. The numerical change sequence of the second object acquired by the data processing system is a smoothed sequence, and the local fluctuation amplitude statistics in the fluctuation statistics sequence are obtained by averaging the results after calculating multiple different window lengths. The above acquisition method helps to eliminate the difference in the absolute level of the first and second objects and enhances the robustness of the fluctuation statistics, thereby improving the stability of the matching degree assessment.

[0105] Optionally, the local volatility statistics in the fund volatility sequence obtained by the data processing system are obtained by calculating a sliding window on the fund return sequence. The local volatility within each window is represented as the difference between the maximum and minimum return values ​​within that window. The extreme value decline statistics are obtained by peak detection on the fund return sequence, and are represented as the maximum value of the decline from each peak to the most recent trough. This method is beneficial for capturing the volatility characteristics of the fund return sequence at different time scales, thereby enhancing the ability of the volatility feature vector to represent the fund's risk level.

[0106] In a numerical example, suppose the first time-series data of the first object is [0.001, -0.003, 0.002, ..., 0.012] (a total of 360 values), and the frequency of the behavior in the behavioral statistics sequence is 25 times, corresponding to a high-frequency behavior pattern. The data processing system calculates the dynamic time-warped distance between the first time-series data and the second time-series data of a certain second object, obtaining a distance value of 0.47, and then calculates the matching degree feature data as 0.68. When the preset threshold is 0.7, since the matching degree feature data 0.68 is lower than the preset threshold of 0.7, the data processing system can generate a prompt message to indicate that the matching degree between the first object and the second object is low. After obtaining the fused feature tensor, the data processing system can use cluster analysis to classify the first object into the fourth category, corresponding to the high-frequency behavior pattern category, and based on the matching score, filter out the second objects with higher matching degrees from multiple second objects, such as the second object with a matching score of 0.85, and generate a sorted list. In addition, the data processing system can generate risk warning information based on the difference between the current matching status and the historical matching status of the first object, which can be used to indicate the adjustment of the association between the first object and the second object.

[0107] In some optional embodiments, the first time series data is the transaction amount sequence of the target user within a preset time window, and the behavioral statistics sequence is the frequency of the target user's transaction type and the duration of the transaction interval; the second time series data is the business volume sequence of the branch, and the fluctuation statistics sequence is the fluctuation amplitude statistics and extreme value decline amplitude statistics of the business volume sequence.

[0108] For example, the transaction amount sequence can represent a sequence of transaction amounts for each transaction by a target user within a preset time window, arranged chronologically according to the transaction occurrence time. The transaction type frequency can represent the number of times each different transaction type performed by the target user within the preset time window occurs; different transaction types include deposits, withdrawals, transfers, or wealth management purchases. The transaction interval duration can represent the time interval between two consecutive transactions by the target user. The business volume sequence can represent a sequence of business processing quantities recorded by a branch within a preset time window, arranged chronologically according to the number of days or time periods; the business processing quantity includes the number of customers received, the number of business transactions processed, or the number of transaction requests. The volatility statistic can represent the value obtained after performing volatility analysis on the business volume sequence; the volatility statistic includes the range or standard deviation of the business volume sequence at consecutive time points. The extreme value decline statistic can represent the maximum decline of the business volume sequence from its peak to its subsequent trough. The data processing system obtains the above transaction amount sequence, transaction type frequency, and transaction interval duration as the first time-series data and behavioral statistics sequence, and obtains the above business volume sequence, volatility statistic, and extreme value decline statistic as the second time-series data and volatility statistics sequence. By extracting the time sequence of transaction amounts, the frequency of different transaction types, and the density of transaction intervals from the transaction behavior of target users, the transaction activity and capital flow patterns of target users can be characterized. Similarly, by extracting the fluctuation range and extreme drop range from the business volume sequence of branches, the volatility of branch business load and sudden drop events can be characterized. The data processing system uses these two types of time-series data as input for subsequent feature extraction and matching calculations. This facilitates correlation analysis between the transaction behavior patterns of target users and the fluctuation patterns of branch business volume, providing data support for adjusting branch resource allocation or customer service strategies.

[0109] In some embodiments, the transaction amount sequence obtained by the data processing system is a daily cumulative value sequence within a preset time window. The frequency of transaction types includes the frequency of a first behavior type, the frequency of a second behavior type, and the frequency of a third behavior type. The transaction interval is represented as the average of multiple transaction intervals. Based on the aforementioned transaction amount sequence, transaction type frequency, and average transaction interval, the data processing system determines the behavioral feature vector of the first object. This behavioral feature vector reflects whether the first object belongs to different categories such as high-frequency small-amount behavior, low-frequency large-amount behavior, or balanced behavior. This acquisition method helps to refine the behavioral profile of the first object and improve the accuracy of subsequent matching scores in characterizing the relationship between the first object type and the service node's business volume. For example, the transaction amount sequence obtained by the data processing system is a daily cumulative transaction amount sequence within a preset time window. The frequency of transaction types includes the frequency of deposit transactions, withdrawal transactions, and transfer transactions. The transaction interval is represented as the average of multiple transaction intervals. The data processing system determines the behavioral feature vector of the target user based on the above transaction amount sequence, transaction type frequency and average transaction interval. This behavioral feature vector can reflect whether the target user belongs to different categories such as high-frequency small deposit type, low-frequency large transfer type or balanced transaction type.

[0110] In other embodiments, the data processing system acquires a business volume sequence as the daily task processing quantity sequence of a service node within a preset time window. The fluctuation amplitude statistics are calculated using a sliding window, where the fluctuation amplitude within each window is represented by the difference between the maximum and minimum task processing quantities within that window, divided by the average task processing quantity within that window. The extreme value decline amplitude statistics are obtained by detecting local peaks in the business volume sequence, and are represented by the maximum value among the ratios of the decline amplitude from each peak to the most recent trough to the peak value. Based on the aforementioned fluctuation amplitude statistics and extreme value decline amplitude statistics, the data processing system determines the fluctuation feature vector of the service node. This fluctuation feature vector reflects the relative fluctuation intensity and the severity of sudden drops in the service node's business load. This acquisition method helps reduce the impact of differences in the absolute value of business volume among different service nodes on the fluctuation measurement results, enabling the fluctuation feature vector to more accurately reflect the relative fluctuation intensity of the business volume sequence, thereby improving the comparability between service nodes of different sizes.

[0111] For example, the business volume sequence acquired by the data processing system is the daily customer reception sequence of a branch within a preset time window. The fluctuation amplitude statistic is calculated using a sliding window, where the fluctuation amplitude within each window is represented by the difference between the maximum and minimum customer reception numbers within that window, divided by the average customer reception number within that window. The extreme value decline amplitude statistic is obtained by detecting local peaks in the business volume sequence, and is represented by the maximum value among the ratios of the decline amplitude from each peak to the most recent trough to the peak value. Based on the above fluctuation amplitude and extreme value decline amplitude statistics, the data processing system determines the branch's fluctuation characteristic vector, which reflects the relative fluctuation intensity and the severity of sudden declines in the branch's business load.

[0112] See Figure 2 According to another aspect of the embodiments of this application, a time-series data processing apparatus is also provided, comprising: an acquisition unit, configured to acquire first time-series data, a behavioral statistics sequence, second time-series data, and a fluctuation statistics sequence, wherein the first time-series data is a time-series data sequence concerning a first object, and the behavioral statistics sequence includes at least the occurrence frequency of various business behaviors of the first object; the second time-series data is a time-series data sequence concerning a second object, and the fluctuation statistics sequence includes at least the fluctuation amplitude and extreme value decrease amplitude of the second time-series data, and the first object and the second object are different objects with business dealings; a first feature vector determination unit, configured to determine a behavioral feature vector of the first object based on the first time-series data and the behavioral statistics sequence, wherein the behavioral feature vector is used to characterize the time-series behavioral pattern and behavioral statistics characteristics of the first object; a second feature vector determination unit, configured to determine a fluctuation feature vector of the second object based on the second time-series data and the fluctuation statistics sequence, wherein the fluctuation feature vector is used to characterize the frequency domain characteristics and data fluctuation patterns of the second object over time; and a feature vector processing unit, configured to determine a matching score between the first object and the second object based on the behavioral feature vector and the fluctuation feature vector, and output the matching score to a display device or generate a control signal based on the matching score.

[0113] Optionally, the first feature vector determination unit includes: a feature extraction subunit, used to input the first temporal data and behavioral statistics sequence into a bidirectional long short-term memory network, extract the temporal dependency features between the first temporal data and the behavioral statistics sequence through the bidirectional long short-term memory network, and determine the hidden state vector of the temporal dependency features at each time step, wherein the hidden state vector is used to represent the cumulative temporal features from the starting time step to the current time step; a weight acquisition subunit, used to acquire the attention weight at each time step, wherein the attention weight is used to represent the degree of contribution of the hidden state vector at the corresponding time step to the behavioral feature vector; and a weighted summation subunit, used to perform weighted summation of the hidden state vector at each time step according to the attention weight at each time step to obtain the behavioral feature vector.

[0114] Optionally, the second feature vector determination unit includes: a sequence transformation subunit for converting the second time-series data and fluctuation statistics sequence into a time-series feature sequence; a convolutional feature extraction subunit for inputting the time-series feature sequence into a convolutional neural network, wherein the convolutional neural network extracts local features of the time-series feature sequence under different window lengths through multiple convolutional kernels, wherein the local features include the upward trend, downward trend, or fluctuation amplitude of the time-series feature sequence at consecutive time points; a feature combination subunit for combining all local features into a convolutional feature sequence; and a time-series encoding subunit for inputting the convolutional feature sequence into a gated recurrent unit in time step order, and outputting a fluctuation feature vector after processing the convolutional feature sequence by the gated recurrent unit.

[0115] Optionally, the temporal coding subunit includes: an input acquisition module, used to acquire the feature value corresponding to the (i+1)th time step in the convolutional feature sequence as the input information of the (i+1)th time step, where i is an integer greater than 1; a history state acquisition module, used to acquire the hidden state vector of the (i)th time step output by the gated recurrent unit, where the hidden state vector of the (i)th time step is used to characterize the feature representation obtained by the gated recurrent unit after encoding the input information from the first time step to the (i)th time step; and a gating calculation module, used to determine the output value of the reset gate and the output value of the update gate according to the input information of the (i+1)th time step and the hidden state vector of the (i)th time step, where the output value of the reset gate is used to control the... The following modules are defined: a hidden state vector at time step i; an update gate; a candidate state determination module; a current state determination module; and an output module. The update gate's output value controls the proportion of information ignored in the hidden state vector at time step i. The output of the update gate controls the data fusion ratio between the input information at time step i+1 and the hidden state vector at time step i. The candidate hidden state vector is used to characterize the current time step feature representation of the candidate. The current state determination module is used to determine the hidden state vector at time step i+1 based on the update gate's output value, the hidden state vector at time step i, and the candidate hidden state vector. Finally, the output module uses the hidden state vector at the last time step as the fluctuation feature vector.

[0116] Optionally, the feature vector processing unit includes: a derivative calculation subunit, used to calculate the first derivative value of the first time series data at each time point to obtain a first derivative sequence; and to calculate the first derivative value of the second time series data at each time point to obtain a second derivative sequence; a distance acquisition subunit, used to acquire the dynamic time warping distance between the first derivative sequence and the second derivative sequence, wherein the dynamic time warping distance is the minimum cumulative sum of the distances between corresponding points of the first derivative sequence and the second derivative sequence under the alignment path that satisfies the boundary conditions, continuity, and monotonicity constraints; a matching degree determination subunit, used to determine the sequence matching degree based on the dynamic time warping distance, wherein the sequence matching degree is inversely proportional to the dynamic time warping distance; a matching degree feature generation subunit, used to generate matching degree feature data based on the sequence matching degree, wherein the matching degree feature data is used to characterize the degree of matching between the first object and the second object; and a score determination subunit, used to determine the matching score based on the behavioral feature vector, the fluctuation feature vector, and the matching degree feature data.

[0117] Optionally, the score determination subunit includes: a feature fusion module, used to fuse behavioral feature vectors, fluctuation feature vectors, and matching degree feature data to obtain a fused feature tensor; a gated cross-network module, used to input the fused feature tensor into a gated cross-network, wherein the gated cross-network includes a first branch and a second branch, the first branch generates a gate signal through a gated activation function, and the second branch generates residual features through a linear transformation, wherein the gate signal is used to control the retention ratio of different feature dimensions in the fused feature tensor, and the residual features are feature vectors obtained after linear mapping of the fused feature tensor, used to retain the original information of the fused feature tensor; and a matching score calculation module, used to determine the matching score based on the gate signal and the residual features.

[0118] Optionally, the time-series data processing device further includes: a multi-object matching unit, used to obtain the matching score between the first object and each second object for multiple second objects after the feature vector processing unit determines the matching score between the first object and the second object based on the behavioral feature vector and the fluctuation feature vector; and a sorting list generation unit, used to sort the second objects from high to low according to the matching score, and select the top K second objects to generate a sorting list, where K is a preset positive integer.

[0119] Optionally, the first time series data is the total user return sequence of the target user within a preset time window, and the behavioral statistics sequence is the transaction behavior statistics sequence of the target user, which includes the frequency of transactions and the duration of holding positions of the target user; the second time series data is the fund return sequence of the fund to be recommended, and the volatility statistics sequence is the fund volatility statistics sequence of the fund to be recommended, which includes the local volatility amplitude statistics and extreme value decline amplitude statistics of the fund return sequence.

[0120] Optionally, the first time series data is the transaction amount sequence of the target user within a preset time window, and the behavioral statistics sequence is the frequency of the target user's transaction type and the duration of the transaction interval; the second time series data is the business volume sequence of the branch, and the fluctuation statistics sequence is the fluctuation amplitude statistics and extreme value decline amplitude statistics of the business volume sequence.

[0121] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described timing data processing method.

[0122] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when one or more programs are executed by one or more processors, the one or more processors cause the one or more processors to perform the timing data processing method described above.

[0123] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which implement the above-described timing data processing method when executed by a processor.

[0124] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the above embodiments of this application, the descriptions of each embodiment have their own emphasis, and for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0125] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0126] The units described as separate components may or may not be physically separate. 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 can be selected to achieve the purpose of this embodiment according to actual needs.

[0127] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0128] If the integrated unit is implemented as 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 this application, in essence, 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. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0129] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A time-series data processing method, characterized in that, include: Acquire first time-series data, behavioral statistics sequence, second time-series data, and fluctuation statistics sequence, wherein the first time-series data is a time-series data sequence about a first object, and the behavioral statistics sequence includes at least the occurrence frequency of various business behaviors of the first object; the second time-series data is a time-series data sequence about a second object, and the fluctuation statistics sequence includes at least the fluctuation amplitude and extreme value decrease amplitude of the second time-series data, and the first object and the second object are different objects with business dealings. The behavioral feature vector of the first object is determined based on the first time-series data and the behavioral statistical sequence, wherein the behavioral feature vector is used to characterize the time-series behavioral pattern and behavioral statistical features of the first object; The fluctuation feature vector of the second object is determined based on the second time series data and the fluctuation statistical sequence, wherein the fluctuation feature vector is used to characterize the frequency domain characteristics and data fluctuation patterns of the second object over time; The matching score between the first object and the second object is determined based on the behavioral feature vector and the fluctuation feature vector, and the matching score is output to the display device or a control signal is generated based on the matching score.

2. The method according to claim 1, characterized in that, Determining the behavioral feature vector of the first object based on the first time-series data and the behavioral statistical sequence includes: The first time series data and the behavioral statistics sequence are input into a bidirectional long short-term memory network. The bidirectional long short-term memory network is used to extract the temporal dependency features between the first time series data and the behavioral statistics sequence, and the hidden state vector of the temporal dependency features at each time step is determined. The hidden state vector is used to characterize the cumulative temporal features from the starting time step to the current time step. Obtain the attention weight for each time step, wherein the attention weight is used to characterize the contribution of the hidden state vector at the corresponding time step to the behavior feature vector; The hidden state vector at each time step is weighted and summed according to the attention weight at each time step to obtain the behavior feature vector.

3. The method according to claim 1, characterized in that, Determining the volatility feature vector of the second object based on the second time-series data and the volatility statistical sequence includes: The second time-series data and the fluctuation statistics sequence are converted into a time-series feature sequence; The temporal feature sequence is input into a convolutional neural network, wherein the convolutional neural network extracts local features of the temporal feature sequence under different window lengths through multiple convolutional kernels, wherein the local features include the upward trend, downward trend or fluctuation amplitude of the temporal feature sequence at consecutive time points; Combine all local features into a convolutional feature sequence; The convolutional feature sequence is input into the gated recurrent unit in time step order, and the wave feature vector is output after the convolutional feature sequence is processed by the gated recurrent unit.

4. The method according to claim 3, characterized in that, The wave feature vector is output after the convolutional feature sequence is processed by the gated recurrent unit, including: Obtain the feature value corresponding to the (i+1)th time step in the convolutional feature sequence, and use it as the input information for the (i+1)th time step, where i is an integer greater than 1; Obtain the hidden state vector of the i-th time step output by the gated recurrent unit, wherein the hidden state vector of the i-th time step is used to characterize the feature representation obtained by the gated recurrent unit after encoding the input information from the first time step to the i-th time step; Based on the input information at the (i+1)th time step and the hidden state vector at the ith time step, the output values ​​of the reset gate and the update gate are determined. The output value of the reset gate is used to control the proportion of information ignored in the hidden state vector at the ith time step; the output value of the update gate is used to control the data fusion ratio between the input information at the (i+1)th time step and the hidden state vector at the ith time step. Based on the output value of the reset gate and the input information of the (i+1)th time step, a candidate hidden state vector is determined, wherein the candidate hidden state vector is used to characterize the current time step feature representation of the candidate. The hidden state vector for the (i+1)th time step is determined based on the output value of the update gate, the hidden state vector at the i-th time step, and the candidate hidden state vector. The hidden state vector of the last time step is used as the fluctuation feature vector.

5. The method according to claim 1, characterized in that, Determining the matching score between the first object and the second object based on the behavioral feature vector and the fluctuation feature vector includes: Calculate the first derivative value of the first time series data at each time point to obtain the first derivative sequence; Calculate the first derivative value of the second time series data at each time point to obtain the second derivative sequence; Obtain the dynamic time warping distance between the first derivative sequence and the second derivative sequence, wherein the dynamic time warping distance is the minimum cumulative sum of the distances between corresponding points between the first derivative sequence and the second derivative sequence under the alignment path that satisfies the boundary conditions, continuity and monotonicity constraints; The sequence matching degree is determined based on the dynamic time warping distance, wherein the sequence matching degree is inversely proportional to the dynamic time warping distance; Based on the sequence matching degree, matching degree feature data is generated, wherein the matching degree feature data is used to characterize the degree of matching between the first object and the second object; The matching score is determined based on the behavioral feature vector, the fluctuation feature vector, and the matching degree feature data.

6. The method according to claim 5, characterized in that, Determining the matching score based on the behavioral feature vector, the fluctuation feature vector, and the matching degree feature data includes: The behavioral feature vector, the fluctuation feature vector, and the matching degree feature data are fused to obtain a fused feature tensor. The fused feature tensor is input into a gated cross-network, wherein the gated cross-network includes a first branch and a second branch. The first branch generates a gate signal through a gated activation function, and the second branch generates residual features through a linear transformation. The gate signal is used to control the retention ratio of different feature dimensions in the fused feature tensor, and the residual features are feature vectors obtained by linearly mapping the fused feature tensor, used to retain the original information of the fused feature tensor. The matching score is determined based on the gating signal and the residual characteristics.

7. The method according to claim 1, characterized in that, After determining the matching score between the first object and the second object based on the behavioral feature vector and the fluctuation feature vector, the method further includes: For multiple second objects, obtain the matching score between the first object and each second object; The second objects are sorted from high to low according to the matching scores, and the top K second objects are selected to generate a sorted list, where K is a preset positive integer.

8. The method according to claim 1, characterized in that, The first time series data is the total user revenue sequence of the target user within a preset time window, and the behavior statistics sequence is the transaction behavior statistics sequence of the target user, which includes the transaction frequency and holding duration of the target user; The second time series data is the fund return series of the fund to be recommended, and the volatility statistics series is the fund volatility statistics series of the fund to be recommended. The fund volatility statistics series includes the local volatility amplitude statistics and the extreme value decline amplitude statistics of the fund return series.

9. The method according to claim 1, characterized in that, The first time series data is the transaction amount sequence of the target user within a preset time window, and the behavior statistics sequence is the frequency of the transaction type and the duration of the transaction interval of the target user; the second time series data is the business volume sequence of the branch, and the fluctuation statistics sequence is the fluctuation amplitude statistics and extreme value decline amplitude statistics of the business volume sequence.

10. A time-series data processing apparatus, characterized in that, include: The acquisition unit is used to acquire first time-series data, behavioral statistics sequence, second time-series data, and fluctuation statistics sequence, wherein the first time-series data is a time-series data sequence about a first object, and the behavioral statistics sequence includes at least the occurrence frequency of various business behaviors of the first object; the second time-series data is a time-series data sequence about a second object, and the fluctuation statistics sequence includes at least the fluctuation amplitude and extreme value decrease amplitude of the second time-series data, and the first object and the second object are different objects with business dealings. The first feature vector determination unit is used to determine the behavioral feature vector of the first object based on the first time series data and the behavioral statistical sequence, wherein the behavioral feature vector is used to characterize the time series behavioral pattern and behavioral statistical features of the first object. The second feature vector determination unit is used to determine the fluctuation feature vector of the second object based on the second time series data and the fluctuation statistical sequence, wherein the fluctuation feature vector is used to characterize the frequency domain characteristics and data fluctuation patterns of the second object over time. The feature vector processing unit is used to determine the matching score between the first object and the second object based on the behavior feature vector and the fluctuation feature vector, and output the matching score to the display device or generate a control signal based on the matching score.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device in which the computer-readable storage medium is located performs the timing data processing method according to any one of claims 1 to 9.

12. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being 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 cause the one or more processors to perform the timing data processing method according to any one of claims 1 to 9.

13. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processor, implement the timing data processing method of any one of claims 1 to 9.

Images