Dynamic feature interaction recognition method based on third-order tensor
Through the dynamic feature interaction recognition method based on third-order tensors, the dynamic interaction between time-varying data features is captured, and the problem of difficulty in capturing nonlinear dynamic changes is solved in the existing technology, which improves the robustness and generalization ability of the model.
Patent Information
- Application Number
- CN202411714527.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-01-10
- Filing Date
- 2024-11-27
- Publication Date
- 2025-04-25
AI Technical Summary
The prior art is difficult to effectively capture the nonlinear dynamic changes between different individuals and different features of the same unit, and it is impossible to fully capture the dynamic characteristics of complex high-dimensional time sequence data.
A dynamic feature interaction recognition method based on third-order tensors is adopted, and feature interleaving processing is performed by obtaining the eigenvectors of time-varying data, third-order parameter tensors are constructed, dynamic feature interaction weight matrix is calculated, and it is input into the artificial intelligence model as a new feature for training.
Effectively capturing the dynamic interactions between time-varying data features improves the robustness and generalization capabilities of the model, and solves the shortcomings of existing methods in capturing nonlinear dynamic changes.
Smart Images
Figure CN119884701A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a dynamic feature interactive recognition method based on a third-order tensor. Background Art
[0002] The information interaction between different granularities is complex. For example, the interaction relationship between the same features of different individuals is different, and the interaction between different features of the same individual will also change dynamically. For example, for power time series data, it contains a large number of instances with different hierarchical structures. The urban power time series data can be decomposed into regional power time series data, and the regional power time series data can be decomposed into the user power time series data of the area. The power data between different regions will change with time (weather, temperature, season), and the power usage between different users in the same region will also change dynamically with time (weather, temperature, season).
[0003] There are two common methods for capturing the interaction of time series data. One is to construct a custom interaction term (e.g., multiply, differ or otherwise combine the values of two time series to form an interaction term) and add it to the model. Once the model training is completed, the coefficient of the custom interaction term is determined. The interaction either does not change over time or does not change with individuals. It is essentially a static analysis, so this method cannot effectively capture the nonlinear dynamic changes after the interaction and fusion of many factors of different individuals. The second method is dynamic interaction analysis based on sliding windows, that is, calculating the interaction relationship (such as correlation coefficient or covariance) within the sliding window and observing its pattern of change over time. This method is suitable for scenarios where the interaction relationship of time series changes over time, especially sudden changes or periodic changes are obvious. However, this method relies on the design of the sliding window size, is sensitive to parameters and is complex to calculate. Both of the above methods cannot introduce global information well, and it is difficult to capture the dynamic interaction of the characteristics of complex high-dimensional time series data. Summary of the invention
[0004] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a dynamic feature interaction recognition method based on a third-order tensor. The present invention solves the deficiency that the existing method cannot capture the nonlinear dynamic interaction characteristics of a large number of time-varying features.
[0005] To achieve the above purpose, the technical solution adopted by the present invention is: a dynamic feature interaction recognition method based on a third-order tensor, which is used in scenarios where it is necessary to capture the nonlinear dynamic interaction of time-varying data features, and the method comprises the following steps:
[0006] Step 1: Obtain relevant data of various features in the high-dimensional time-varying scene of time-varying data, perform feature processing on the high-dimensional time-varying data, and express it in the form of feature vectors. A certain time-varying data has a feature vector in a certain period, and the dimension of the feature vector is equal to the number of features;
[0007] Step 2: According to the constructed feature vector, all elements of the feature vector of each period of each time-varying data are subjected to pairwise cross-multiplication processing to obtain the feature cross-multiplication matrix of each period of each time-varying data;
[0008] Step 3: Construct a third-order parameter tensor and calculate the dynamic feature interaction weight matrix of each period of each time-varying data;
[0009] Step 4: multiply the feature cross-product matrix by the dynamic feature interaction weight matrix to obtain the strength of the interaction relationship between each feature and other features in each period of each time-varying data;
[0010] Step 5: Use the interaction relationship as a new feature and as an input feature of the artificial intelligence-based time-varying data model, and train them together;
[0011] Step 6: Use the trained artificial intelligence-based time-varying data model to interactively identify dynamic features.
[0012] As a further improvement of the present invention, the step 1 is specifically as follows:
[0013] Obtain relevant data for various features of time-varying data, perform feature processing according to the time-varying data processing method, and obtain the feature vector of a certain time-varying data i at a certain period t. express, Where N is the number of features that affect the data fluctuation of time-varying data, and i∈I means that there are I time-varying data in total.
[0014] As a further improvement of the present invention, the step 2 is specifically as follows:
[0015] According to the constructed feature vector, all elements of the feature vector of each period of each time-varying data are cross-multiplied to obtain the feature cross-multiplication matrix of each period of each time-varying data. Represents, where T is the transpose operation; is the characteristic cross-product matrix The element in the 1st row and the kth column in represents the product of the 1st feature and the kth feature value of the time-varying data i at period t.
[0016] As a further improvement of the present invention, the step 3 is specifically as follows:
[0017] Construct the third-order parameter tensor Γ, Calculate the dynamic feature interaction weight matrix of each time-varying data in each period, and use express, At period t, the dynamic interaction weight vector of the kth feature of time-varying data i and all features including itself is:
[0018]
[0019] In order to capture the dynamic interaction, at time t, the observable eigenvector of the time-varying data i is calculated using the kth slice Γ[k,1:N,1:N] of the third-order parameter tensor Γ Perform linear transformation operation and use an activation function tanh() to perform nonlinear transformation on the feature vector after linear transformation of the third-order parameter tensor, and obtain where · represents the dot product.
[0020] As a further improvement of the present invention, the step 4 is specifically as follows:
[0021] Multiply the feature matrix Interacting weight vectors with dynamic features Multiply them together to get the strength of the interaction between the kth feature of the time-varying data i in period t and other features including itself in,
[0022]
[0023] Thus, the strength of the interaction between all features of time-varying data i and other features including itself in period t is expressed by a vector Indicates that At this point, the dynamic change in the strength of the dynamic interaction between each feature and other features in each period of each time-varying data can be captured.
[0024] As a further improvement of the present invention, the time-varying data is financial data, earthquake data, power data, weather forecast data and traffic flow data.
[0025] The beneficial effects of the present invention are:
[0026] The present invention is applicable to all kinds of scenarios that need to capture the interaction of time-varying data features, and can effectively capture the dynamic interaction that affects the ever-changing time-varying data features, and solve the problem that the existing methods actively ignore or cannot capture the nonlinear dynamic interaction characteristics of a large number of time-varying features, thereby making up for the fact that the existing low signal-to-noise ratio and data characteristics that are difficult to meet stability are significantly different from other fields (such as natural language processing, speech recognition, image recognition, autonomous driving, etc.). Directly calling machine learning or deep learning models developed from other fields cannot effectively capture the laws of time-varying data fluctuations, resulting in the current situation where the model robustness and generalization ability are poor when using existing artificial intelligence technology to solve time-varying data. The present invention captures the new characteristics of the dynamic interaction that changes constantly due to operation between time-varying data features, uses it as the input feature of the existing artificial intelligence-based model and trains it together with the existing model, which can prompt the model to better learn the inherent laws of time-varying data fluctuations, thereby making the trained artificial intelligence-based time-varying data model more robust and generalizable. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 is a flow chart of an embodiment of the present invention;
[0028] Figure 2 This is an algorithm framework diagram of a dynamic feature interactive recognition method based on a third-order tensor of the present invention. DETAILED DESCRIPTION
[0029] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0030] Example
[0031] like Figure 1 and Figure 2 As shown, a dynamic feature interactive recognition method based on a third-order tensor is described. This embodiment takes financial data as an example. The method includes five steps:
[0032] Step 1: Obtain relevant data on various features in high-dimensional time-varying scenarios such as financial asset pricing or financial asset risk assessment, perform feature processing on the high-dimensional time-varying data, and express it in the form of a feature vector. A certain asset has a feature vector in a certain period, and the dimension of the feature vector is equal to the number of features.
[0033] Step 2: Based on the constructed feature vector, perform pairwise cross-multiplication on all elements of the feature vector of each asset in each period to obtain the feature cross-multiplication matrix of each asset in each period.
[0034] Step 3: Construct a third-order parameter tensor and calculate the dynamic feature interaction weight matrix for each asset in each period
[0035] Step 4: Multiply the feature cross-product matrix obtained in step 2 by the dynamic feature interaction weight matrix obtained in step 3 to obtain the strength of the interaction between each feature and other features of each asset in each period.
[0036] Step 5: Use the interaction relationship as a new feature and use it as the input feature of the subsequent model.
[0037] Step 6: Steps 1-5 should be trained as a module along with the entire model framework to solve the problem.
[0038] The following is a detailed description of each step:
[0039] Step 1 is as follows: obtain relevant data of various features used for asset pricing or asset risk assessment, perform feature processing according to the general financial data processing method, obtain the feature vector of a certain asset i in a certain period t, and use express, Where N is the number of features that affect asset price fluctuations, and i∈I means there are I assets in total.
[0040] Step 2 is as follows: Based on the constructed feature vector, all elements of the feature vector of each asset in each period are cross-multiplied to obtain the feature cross-multiplication matrix of each asset in each period. Represents, where T is the transpose operation. is the characteristic cross-product matrix The element in the 1st row and the kth column represents the product of the 1st characteristic and the kth characteristic value of asset i in period t.
[0041] Step 3 is as follows: construct the third-order parameter tensor Γ, Calculate the dynamic feature interaction weight matrix for each asset in each period, using express, At period t, the dynamic interaction weight vector between the kth feature of asset i and all features (including itself) is:
[0042]
[0043] Here, in order to capture the dynamic interaction, at period t, the kth slice Γ[k,1:N,1:N] of the third-order parameter tensor Γ is used to map the observable feature vector of asset i to Perform a linear transformation operation, and further use an activation function tanh() to perform a nonlinear transformation on the feature vector that has undergone a linear transformation of the third-order parameter tensor, and obtain where · represents the dot product.
[0044] Step 4 is as follows: multiply the feature matrix obtained in step 2 Interaction weight vector with the dynamic feature obtained in step 3 Multiply them together to get the strength of the interaction between the kth feature of asset i and other features (including itself) in period t in,
[0045]
[0046] Therefore, the strength of the interaction between all the characteristics of asset i and other characteristics (including itself) in period t can be expressed by a vector Indicates that At this point, the dynamically changing strength of the dynamic interaction between each feature of each asset and other features in each period can be captured.
[0047] Step 5: Get the interaction relationship between all features of asset i and other features (including itself) in period t After that, As new features, they are used as input features for subsequent models.
[0048] Step 6: Steps 1-5 should be trained as a module along with the entire model framework to solve the problem.
[0049] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.
Claims
1. A dynamic feature interactive recognition method based on third-order tensor, characterized in that: For use in scenarios where it is necessary to capture nonlinear dynamic interactions of time-varying data features, the method comprises the following steps: Step 1: Obtain relevant data of various features in the high-dimensional time-varying scene of time-varying data, perform feature processing on the high-dimensional time-varying data, and express it in the form of feature vectors. A certain time-varying data has a feature vector in a certain period, and the dimension of the feature vector is equal to the number of features; Step 2: According to the constructed feature vector, all elements of the feature vector of each period of each time-varying data are subjected to pairwise cross-multiplication processing to obtain the feature cross-multiplication matrix of each period of each time-varying data; Step 3: Construct a third-order parameter tensor and calculate the dynamic feature interaction weight matrix of each period of each time-varying data; Step 4: multiply the feature cross-product matrix by the dynamic feature interaction weight matrix to obtain the strength of the interaction relationship between each feature and other features in each period of each time-varying data; Step 5: Use the interaction relationship as a new feature and as an input feature of the artificial intelligence-based time-varying data model, and train them together; Step 6: Use the trained artificial intelligence-based time-varying data model to interactively identify dynamic features.
2. The dynamic feature interactive recognition method based on third-order tensor according to claim 1 is characterized in that: The step 1 is specifically as follows: Obtain relevant data for various features of time-varying data, perform feature processing according to the time-varying data processing method, and obtain the feature vector of a certain time-varying data i at a certain period t. express, Where N is the number of features that affect the data fluctuation of time-varying data, and i∈I means that there are I time-varying data in total.
3. The dynamic feature interactive recognition method based on third-order tensor according to claim 2 is characterized in that: The step 2 is specifically as follows: According to the constructed feature vector, all elements of the feature vector of each period of each time-varying data are cross-multiplied to obtain the feature cross-multiplication matrix of each period of each time-varying data. Represents, where T is the transpose operation; is the characteristic cross-product matrix The element in the 1st row and the kth column in represents the product of the 1st feature and the kth feature value of the time-varying data i at period t.
4. The dynamic feature interactive recognition method based on third-order tensor according to claim 3 is characterized in that: The step 3 is as follows: Construct the third-order parameter tensor Γ, Calculate the dynamic feature interaction weight matrix of each time-varying data in each period, and use express, At period t, the dynamic interaction weight vector of the kth feature of time-varying data i and all features including itself is: In order to capture the dynamic interaction, at time t, the observable eigenvector of the time-varying data i is calculated using the kth slice Γ[k,1:N,1:N] of the third-order parameter tensor Γ Perform linear transformation operation and use an activation function tanh () to perform nonlinear transformation on the feature vector after linear transformation of the third-order parameter tensor, and obtain where · represents the dot product.
5. The dynamic feature interactive recognition method based on third-order tensor according to claim 4 is characterized in that: The step 4 is specifically as follows: Multiply the feature matrix Interacting weight vectors with dynamic features Multiply them together to get the strength of the interaction between the kth feature of the time-varying data i in period t and other features including itself in, Thus, the strength of the interaction between all features of time-varying data i and other features including itself in period t is expressed by a vector Indicates that At this point, the dynamic change in the strength of the dynamic interaction between each feature and other features in each period of each time-varying data can be captured.
6. The dynamic feature interactive recognition method based on third-order tensor according to any one of claims 1 to 5, characterized in that: The time-varying data include financial data, earthquake data, power data, weather forecast data and traffic flow data.