Method and device for recording historical data

Abstract

The embodiments of this specification describe a method and device for recording historical data. According to the method of the embodiment, the historical data within the current data recording period is first obtained, and then the historical data is subjected to feature extraction to obtain a historical feature vector corresponding to the current data recording period. The historical feature vector of the current data recording period is further used to update the historical feature vector obtained before the data recording period, and then the updated historical feature vector can be used to characterize all historical data of the previous data recording periods. In this way, by recording the feature vector that characterizes the characteristic performance of the event instead of recording the original data of the characteristic performance of each event, the storage resources of the device can be saved. Furthermore, in each data recording period, the historical feature vector generated by the subsequent data is used to update the previous historical feature vector, instead of storing the historical feature vectors obtained in all periods, which can further save the storage resources of the device.

Description

Technical Field

[0001] One or more embodiments of the present specification relate to the field of computer technology, and in particular, to a method and device for recording historical data. Background Art

[0002] In the field of financial risk control, it is often necessary to store a large amount of historical data to focus on certain behavioral characteristics, such as violations that have occurred in historical financial transactions.

[0003] However, when paying attention to certain behavior characteristics, the generated historical data needs to be continuously stored, and the stored historical data will inevitably include a large amount of data on behavior characteristics that do not need to be paid attention to, which causes a large amount of device storage resources to be occupied. Summary of the invention

[0004] One or more embodiments of this specification describe a method and apparatus for recording historical data, which can save storage resources of a device.

[0005] According to a first aspect, a method for recording historical data is provided, comprising:

[0006] For each data logging cycle:

[0007] Get historical data within the current data recording period;

[0008] Extracting features from the historical data to obtain at least one historical feature vector corresponding to a current data recording period;

[0009] Using the at least one historical feature vector to update the historical feature vector obtained in the first data recording period before the current data recording period; wherein each value in the historical feature vector represents a feature value of an event feature performance;

[0010] The updated historical feature vector is used as a record of all historical data in each data recording period.

[0011] In a possible implementation, obtaining at least one historical feature vector corresponding to the current data recording period includes:

[0012] Determine at least one convolution kernel according to the event feature performance to be recorded; wherein each event feature performance corresponds to a convolution kernel;

[0013] For each event characteristic, execute:

[0014] Using the convolution kernel corresponding to the event feature expression to extract features from the sequence of historical data, a feature vector corresponding to the event feature expression is obtained;

[0015] Determine the maximum value in the feature vector as the feature value corresponding to the feature performance of the current event in the current recording period;

[0016] A vector composed of feature values ​​corresponding to the feature performance of each event is determined as a historical feature vector corresponding to the current data recording period.

[0017] In a possible implementation, the convolution kernel is obtained by learning the historical data represented by the corresponding event features through an end-to-end neural network.

[0018] In a possible implementation, determining the maximum value in the feature vector as the feature value corresponding to the feature representation of the current event in the current recording period includes:

[0019] The Max-Pooling mechanism of the convolutional neural network is used to determine the maximum value in the feature vector, and the maximum value is determined as the feature value corresponding to the feature performance of the current event in the current recording period.

[0020] In a possible implementation, the updating of the historical feature vector obtained in the first data recording period before the current data recording period by using the at least one historical feature vector includes:

[0021] For each event characteristic performance, the larger value of the historical characteristic vector of the current data recording period and the historical characteristic vector of the first data recording period corresponding to the event characteristic performance is used as the characteristic value representing the event characteristic performance;

[0022] The vector formed by the obtained feature values ​​corresponding to the feature performance of each event is determined as the historical feature vector of the current data recording period.

[0023] In a possible implementation manner, the first data recording period includes a data recording period before a current data recording period.

[0024] According to a second aspect, there is provided a device for recording historical data, comprising:

[0025] A historical data acquisition module, configured to acquire historical data within a current data recording period for each data recording period;

[0026] a feature extraction module configured to perform feature extraction on the historical data acquired by the historical data acquisition module to obtain at least one historical feature vector corresponding to a current data recording period;

[0027] A feature vector updating module, configured to update a historical feature vector obtained in a first data recording period before a current data recording period using the at least one historical feature vector obtained by the feature extraction module; wherein each value in the historical feature vector represents a feature value of a feature performance of an event;

[0028] The historical data recording module is configured to use the historical feature vectors obtained after being updated by the feature vector updating module as records of all historical data in each data recording period.

[0029] In a possible implementation, the method further includes: a feature vector determination module, and the feature vector determination module is configured to perform the following operations when obtaining at least one historical feature vector corresponding to the current data recording period:

[0030] Determine at least one convolution kernel according to the event feature performance to be recorded; wherein each event feature performance corresponds to a convolution kernel;

[0031] For each event characteristic, execute:

[0032] Using the convolution kernel corresponding to the event feature expression to extract features from the sequence of historical data, a feature vector corresponding to the event feature expression is obtained;

[0033] Determine the maximum value in the feature vector as the feature value corresponding to the feature performance of the current event in the current recording period;

[0034] A vector composed of feature values ​​corresponding to the feature performance of each event is determined as a historical feature vector corresponding to the current data recording period.

[0035] In a possible implementation, when the feature vector determination module determines the maximum value in the feature vector as the feature value corresponding to the feature representation of the current event in the current recording period, it is configured to perform the following operations:

[0036] The Max-Pooling mechanism of the convolutional neural network is used to determine the maximum value in the feature vector, and the maximum value is determined as the feature value corresponding to the feature performance of the current event in the current recording period.

[0037] According to a third aspect, a computing device is provided, comprising: a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method described in any one of the first aspects is implemented.

[0038] According to the method and device provided in the embodiments of this specification, when recording historical data, for each data recording period, the historical data within the data recording period can be first obtained, and then the historical data can be feature extracted to obtain the historical feature vector corresponding to the data recording period. The historical feature vector is further used to update the historical feature vector obtained in the previous recording period, and the historical feature vector obtained after the update is used as a record of all historical data in each data recording period. It can be seen that first, this scheme considers extracting features from the historical data of each data recording period, rather than recording the original data of the historical data, which can greatly save the resources of the storage device. Secondly, when the historical feature vector of the current data recording period is obtained, the historical feature vector is further used to update the historical feature vector obtained previously. In this way, there is no need to store the historical feature vectors obtained in each data recording period, but the features that can better characterize the event feature performance are recorded and stored through updating, which further saves the resources of the storage device. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 is a flow chart of a method for recording historical data provided by one embodiment of this specification;

[0041] Figure 2 is a flow chart of a method for determining a historical feature vector provided by one embodiment of this specification;

[0042] Figure 3 is a schematic diagram of feature extraction provided by an embodiment of this specification;

[0043] Figure 4 is a flow chart of a method for updating a historical feature vector provided by an embodiment of this specification;

[0044] Figure 5 is a schematic diagram of updating a feature vector provided by an embodiment of this specification;

[0045] Figure 6 is a flow chart of a method for recording historical data provided by another embodiment of the present specification;

[0046] Figure 7 It is a schematic diagram of a device for recording historical data provided by yet another embodiment of the present specification. DETAILED DESCRIPTION

[0047] As mentioned above, in the field of financial risk control, it is often necessary to store a large amount of historical data to focus on certain behavioral characteristics. For example, it is necessary to focus on accounts that have changed or reissued cards to determine whether users have used new or old accounts to conduct some abnormal transactions; for example, it is necessary to focus on accounts that have violated regulations in the past to determine whether the account will violate regulations in the future.

[0048] However, when paying attention to some behavioral characteristics, a large amount of historical data needs to be recorded. For example, when an account has a violation, all the historical data of the account must be recorded to achieve constant attention to the account. Obviously, in practice, there will be a large number of characteristics and data that do not need to be paid attention to. If all the historical data corresponding to each behavioral characteristic are stored, it will inevitably occupy a large amount of device resources and bring a great burden to the device. At present, when solving this problem, the usual way is to expand device resources to store more data. However, when storing all historical data, it is bound to include a large amount of data that does not need to be paid attention to. For example, after an account has violated the rules once, it has not violated the rules again for a long time. If all the historical data are recorded, it is obvious that there is a large amount of normal behavior data, which does not need to be paid attention to, which will cause a large amount of device resources to be occupied, resulting in a waste of storage resources.

[0049] Based on this, this solution considers first extracting features from historical data, saving equipment resources by storing features that need to be paid attention to instead of raw data. Secondly, by using the features obtained from subsequent data to update the previously obtained features, the purpose of further saving equipment resources is achieved.

[0050] like Figure 1 As shown, the embodiment of this specification provides a method for recording historical data. The method may include the following steps for each data recording cycle:

[0051] Step 101: Obtain historical data within the current data recording period;

[0052] Step 103: extracting features from the historical data to obtain at least one historical feature vector corresponding to the current data recording period;

[0053] Step 105: using at least one historical feature vector to update the historical feature vector obtained in the first data recording period before the current data recording period; wherein each value in the historical feature vector represents a feature value of an event feature performance;

[0054] Step 107: The updated historical feature vector is used as a record of all historical data in each data recording period.

[0055] In this embodiment, when recording historical data, it can be considered to first extract features from the historical data of the current data recording period in each set data recording period, and then use the obtained historical feature vector to update the historical feature vector obtained in the previous data recording period, and then use the updated historical feature vector as a record of all previous historical data. It can be seen that this solution is not a complete record of historical data, but stores the features of event feature performance that need to be paid attention to, thereby saving a large amount of device storage resources. In addition, this solution further uses the historical feature vector obtained subsequently to update the historical feature vector obtained previously, that is, it is not necessary to store the historical feature vector obtained in each data recording period, but to continuously update the previously obtained features according to the subsequent features, thereby further saving the storage resources of the device.

[0056] Usually, when recording data, the original data will be recorded. For example, for a financial account, the account log will record all its transactions, login time, logout time, login device, etc., so that when analyzing some behaviors of the account later, the original account log can be searched and analyzed. However, the event features that need to be paid attention to are different for different scenarios. For example, when you want to pay attention to whether the account is logged in on different devices at the same time, then the data related to the event feature performance is the data of the login device, and other transaction data does not need to be recorded; for another example, when you want to pay attention to whether a certain account conducts large transactions, you need to pay attention to whether large transactions appear in the transaction data, and the login device data does not need to be paid attention to. Therefore, different event feature performances can be included in different historical data, or in the same set of historical data.

[0057] The following is a detailed description of the embodiments of the present invention. Figure 1 The steps in are described separately.

[0058] First, in step 101, historical data in the current data recording period is obtained.

[0059] In this step, the data recording cycle can be set according to specific needs or the frequency of historical data generation. For example, when the frequency of historical data generation is low, a week or even a month can be used as a data recording cycle; for another example, when the frequency of historical data generation is high, a day or even an hour can be used as a data recording cycle. In addition, the first data recording cycle can be the previous data recording cycle of the current data recording cycle, or any one or more data recording cycles before the current data recording cycle.

[0060] Next, in step 103, feature extraction is performed on the historical data to obtain at least one historical feature vector corresponding to the current data recording period.

[0061] When paying attention to some events, the points of concern will have certain characteristic manifestations. For example, when you need to pay attention to a certain account, the common characteristic manifestation of illegal transactions may be abnormal transaction amounts, etc. Based on this, consider using the feature extraction method in step 103 to extract the features you want to extract and record them. In other words, each behavioral feature that you want to pay attention to will correspond to an event characteristic manifestation. When extracting features from historical data, you can extract features for each event characteristic manifestation separately. To extract features for different event characteristic manifestations, you need to use different convolution kernels to achieve this. Therefore, if Figure 2 As shown, this step 103 can be implemented by the following steps when extracting features from historical data to obtain at least one historical feature vector corresponding to the current data recording period:

[0062] Step 201: determining at least one convolution kernel according to the event feature performance to be recorded; wherein each event feature performance corresponds to a convolution kernel;

[0063] For each event characteristic, execute:

[0064] Step 203: extracting features from the sequence of historical data using the convolution kernel corresponding to the event feature expression, and obtaining a feature vector corresponding to the event feature expression;

[0065] Step 205: determining the maximum value in the feature vector as the feature value corresponding to the feature performance of the current event in the current recording period;

[0066] Step 207: Determine a vector consisting of feature values ​​corresponding to each event feature as a historical feature vector corresponding to the current data recording period.

[0067] In this embodiment, when determining the historical feature vector, the convolution kernel can be first determined according to the event features to be recorded, and different event feature expressions usually correspond to different convolution kernels. In this way, by using different convolution kernels to extract features from the sequence of historical data, the feature vectors corresponding to each convolution kernel can be obtained respectively, that is, the feature vectors corresponding to various event feature expressions can be obtained.

[0068] In order to further save the storage resources of the device, it is considered to use the most obvious value in the feature vector of each event feature to replace the event feature, that is, take the maximum value in the feature vector of each event feature to represent the feature value of the event feature. In this way, the vector composed of each obtained feature value can be used to represent the historical feature vector of the data recording period. It can be seen that when determining the historical feature vector of each data recording period, the maximum value in the feature vector corresponding to each event feature is used to represent each event feature to be concerned, and the vector composed of each maximum value is used as the historical feature vector of the data recording period. In this way, instead of recording the feature vector of each event feature, only the maximum value is recorded, which ensures that the event feature can be recorded with the most prominent performance, and at the same time can further reduce the equipment resources occupied during record storage.

[0069] like Figure 3 The schematic diagram of feature extraction shown in the figure shows that, for the three groups of historical data A, B, and C obtained by three operations in a certain data recording cycle, there are four event feature expressions that need to be paid attention to, that is, there are four corresponding convolution kernels used to extract features from these three groups of historical data (convolution kernel 1 and convolution kernel 4 are used to illustrate in the figure). For each convolution kernel, it slides through the three groups of historical data in turn, and each convolution kernel obtains a corresponding feature vector, that is, each feature vector corresponds to an event feature expression, which are feature vector 1, feature vector 2, feature vector 3, and feature vector 4. Further, for each feature vector obtained by the convolution kernel, the maximum value of the eigenvalue in the feature vector is determined, and the maximum value is used to characterize the event feature expression corresponding to the convolution kernel. That is, the historical feature vector of the data recording cycle finally obtained contains 4 eigenvalues.

[0070] like Figure 3 In the feature extraction shown in the figure, feature extraction can be achieved through convolutional neural networks, and the weights of the convolution kernels can be learned by end-to-end neural network using historical data of event feature representation. The convolution kernel corresponds to the event feature representation to be extracted, such as "a small transaction is made quickly after account registration", "the device used for the two logins before and after is inconsistent", etc.

[0071] In step 205, when determining the maximum value in the feature vector as the feature value corresponding to the feature representation of the current event in the current recording period, the Max-Pooling mechanism of the neural network can be considered to determine the maximum value in the feature vector, and determine the maximum value as the feature value corresponding to the feature representation of the current event in the current recording period.

[0072] Considering that some users may want to use some good operations to cover up some illegal behaviors after performing some illegal behaviors. Based on this, this embodiment considers using the Max-Pooling feature aggregation method to take the maximum value of each feature vector obtained, so that by obtaining the maximum performance of the event feature in history, even if a user subsequently uses some operations that comply with regulations to cover up the illegal behavior, it can be recorded. In addition, the method of only recording the maximum value in each feature vector can also greatly save the storage resources of the device.

[0073] Of course, in other possible implementations, when determining the feature values ​​to be recorded that can characterize the feature performance of each event, step 205 can also take the average value, median, etc. of each feature value in the feature vector of each event feature performance. For example, if you want to focus on the overall performance of a certain feature performance in a certain period, you can take the average value of each feature value in the feature vector of the feature performance as the feature value that characterizes the feature performance of the event. There is no special limitation on the method of obtaining the feature value that can characterize the feature performance of each event.

[0074] Next, in step 105, at least one historical feature vector is used to update the historical feature vector obtained in the first data recording period before the current data recording period.

[0075] After extracting features from historical data, you can consider storing the historical feature vectors obtained in each data recording period. Of course, in order to further reduce the pressure on the storage device caused by data storage, you can also use the historical feature vectors obtained in the current data recording period to update the historical feature vectors obtained before the data recording period. In this way, there is no need to store the historical feature vectors of each data recording period, thereby achieving the purpose of further saving device storage resources. For example, Figure 4 As shown, in a possible implementation, step 105 can be implemented by the following steps:

[0076] Step 401: for each event characteristic performance, the larger value of the historical characteristic vector of the current data recording period and the historical characteristic vector of the first data recording period corresponding to the event characteristic performance is used as a characteristic value representing the event characteristic performance;

[0077] Step 403: Determine the vector formed by the obtained feature values ​​corresponding to the feature performance of each event as the historical feature vector of the current data recording period.

[0078] In this embodiment, when the historical feature vector of the previous data recording period is updated by using the historical feature vector of the current data recording period, it is possible to consider comparing the feature values ​​representing the same event feature performance in the feature vectors of the current data recording period and the previous first data recording period, and then determine the feature value of the same event feature performance in the comparison result as the feature value corresponding to the event feature performance in the updated historical feature vector. It can be seen that this scheme not only considers using the historical feature vector obtained later to update the historical feature vector obtained previously to save the storage resources of the storage device. Moreover, when updating the historical feature vector, it is considered to update and replace the feature value with the maximum feature performance for each event feature performance, so that the event feature performance can be maximized to avoid being covered by normal historical data. For example, there are several transactions with violations in a certain account, and the degree of these violations is different. After the characteristic behavior is recorded, the account subsequently performs a large number of normal operations in order to cover up the violation. Through this embodiment, when recording historical data, it will record the behavior with the most prominent violation performance in the account, that is, the maximum value of the feature value in the historical feature vector obtained in each recording period. Therefore, its illegal behavior is difficult to cover up.

[0079] In such Figure 5 The updating schematic diagram of the feature vector shown in the figure, the historical feature vector obtained in the first data recording period before the current data recording period is A, and the feature values ​​corresponding to its various event feature expressions are a1, a2, a3, a4, a5, a6 respectively; the historical feature vector obtained in the current data recording period is B, and the feature values ​​of its various event feature expressions are b1, b2, b3, b4, b5, b6 respectively. When updating, by comparing two values ​​representing the feature expression of the same event, the larger value of the two values ​​is determined as the feature value corresponding to the feature expression of the event, that is, the feature vector C[c1, c2, c3, c4, c5, c6] is obtained, wherein c1 is the larger value of a1 and b1, c2 is the larger value of a2 and b2, c3 is the larger value of a3 and b3, c4 is the larger value of a4 and b4, c5 is the larger value of a5 and b5, and c6 is the larger value of a6 and b6. For example, [a1, a2, a3, a4, a5, a6] is [0.5, 6, 4, 9, 20, 1], and [b1, b2, b3, b4, b5, b6] is [6, 0.5, 8, 3, 18, 5], then by comparison, the final historical feature vector c is [6, 6, 8, 9, 20, 5].

[0080] In some possible embodiments, if the feature values ​​of each event feature in the historical feature vector obtained in the first data recording period and the current data recording period do not correspond one to one, then Figure 6As shown in the schematic diagram, the historical feature vector obtained in the current data recording period and the historical feature vector obtained in the first data recording period can be aligned before comparison. That is, the position of the feature value in the feature vector can be adjusted according to the event feature performance, so that the feature vector A[a1, a2, a3, a4, a5, a6] of the first data recording period corresponds to the feature vector B[b1, b2, b3, b4, b5, b6] of the current data recording period, so that the maximum value of the corresponding two can be directly determined to obtain the final historical feature vector.

[0081] Of course, it is worth pointing out that when step 401 compares the feature values ​​representing the same event feature performance in the feature vectors of the current data recording period and the first data recording period, the maximum value of the feature values ​​representing the same event feature performance can be determined as the feature value representing the event feature performance in the historical feature vector, and the average value of the feature values ​​representing the same event feature performance in the two historical feature vectors can be calculated, and the average value can be used as the feature value representing the event feature performance in the historical feature vector. For example, [a1, a2, a3, a4, a5, a6] is [1, 6, 4, 9, 20, 1], [b1, b2, b3, b4, b5, b6] is [6, 1, 8, 3, 18, 5], and the final historical feature vector c is [3.5, 3.5, 6, 6, 19, 3]. Of course, it can also be determined by assigning different weights to the historical feature vectors of the current data recording period and the historical feature vectors of the first data recording period according to actual applications and using a certain calculation method. No specific limitation is made here on how to determine the feature value corresponding to the feature performance of the event in the finally generated historical feature vector based on the two feature values ​​corresponding to the feature performance of the same event.

[0082] Furthermore, if the historical feature vector obtained in the current data recording period includes the first event feature expression; wherein the historical feature vector obtained in the first data recording period does not include the first event feature expression;

[0083] After determining the feature vector composed of the larger value of the feature values ​​corresponding to the feature expressions of the same event as the historical feature vector of the current recording period in step 403, the method further includes:

[0084] The feature value corresponding to the first event feature expression is added to the historical feature vector of the current recording period.

[0085] Next, in step 107, the updated historical feature vector may be used as a record of all historical data in each data recording period.

[0086] In this embodiment, in addition to considering specific event characteristic manifestations that have been paid attention to in advance, when it is necessary to pay attention to other event characteristic manifestations at a certain moment, or an event characteristic manifestation that needs to be paid attention to appears in a certain data recording cycle, the characteristic value obtained for the first time for the characteristic manifestation can be directly added to the obtained historical characteristic vector, and the data containing the event characteristic manifestation in subsequent cycles can be recorded according to the various embodiments of this specification.

[0087] Obviously, when the historical feature vector obtained in the current data recording period is used to update the historical feature vector obtained in the first data recording period, the feature vector obtained in the first data recording period may be the data recording period before the current data recording period, or may be any one or more data recording periods before the current data recording period. The following takes the first data recording period as the data recording period before the current data recording period as an example to illustrate the embodiments of this specification. Figure 6 As shown, the method for recording historical data provided in the embodiment of this specification may include the following steps:

[0088] Step 601: Obtain historical data within the current data recording period;

[0089] The historical data is data related to the behavioral characteristics of certain events that need attention. For example, the historical data may be the transaction records of a certain financial account, the login records of a certain account, or the usage records of a certain system.

[0090] Step 603: Determine the convolution kernel according to the event feature performance to be recorded;

[0091] The convolution kernel is used to extract the corresponding features of historical data, that is, each convolution kernel corresponds to a characteristic performance of an event to be concerned. For example, convolution kernel A corresponds to the characteristic performance of the user's bad transaction behavior, convolution kernel B corresponds to the characteristic performance of the user's card replacement or re-issuance, and convolution kernel C corresponds to the characteristic performance of the user's login level of a certain account. The convolution kernel can be learned from the historical data of the corresponding event characteristic performance. For example, convolution kernel A can be learned through an end-to-end neural network using the historical data of the user's bad transaction behavior; convolution kernel B can be learned through an end-to-end neural network using the historical data of the user's card replacement or re-issuance.

[0092] Step 605: Use the convolution kernels corresponding to the various event feature expressions to perform feature extraction on the sequence of historical data to obtain feature vectors corresponding to the feature expressions of each event.

[0093] In this embodiment, when the convolution kernel corresponding to each event feature is performing feature extraction on the historical data, the convolution kernel slides through the sequence of historical data in turn, and extracts the features of the event feature corresponding to the convolution kernel by performing convolution operation with the historical data in each sliding frame. In this way, by performing feature extraction on all historical data, a feature vector representing the feature of the event can be obtained.

[0094] Step 607: For each feature vector of the event feature representation, determine the maximum value in the feature vector as the feature value representing the event feature representation.

[0095] In this embodiment, in order to further save the storage resources of the device and maximize the characteristic performance of the characteristic behavior to be concerned, the maximum value in the characteristic vector of each event characteristic performance can be used as the characteristic value representing the event characteristic performance. In this way, even if the user can use some compliant operations to cover up previous violations, this method can also record the violations. For example, the Max-Pooling aggregation mechanism of the convolutional neural network can be used to determine the maximum value in the characteristic vector, and the maximum value of each event characteristic performance is determined as the characteristic value of the event characteristic performance in the current recording cycle.

[0096] Of course, here we can also calculate the average value of each eigenvalue in the eigenvector, or assign a certain weight to each eigenvalue, so as to obtain the eigenvalue representing the characteristic performance of each event in this way.

[0097] Step 609: determining the vector formed by the characteristic values ​​of each event characteristic as the historical characteristic vector of the current data recording period;

[0098] Step 611: comparing the feature values ​​representing the same event feature performance in the feature vectors of the current data recording period and the previous data recording period;

[0099] In this embodiment, when comparing feature values, two feature values ​​representing the same event feature performance in two sets of feature vectors are compared. For example, 1 to 6 correspond to 6 different event feature performances, and when the feature vector A[a1, a2, a3, a4, a5, a6] of the current data recording period is compared with the feature vector B[b1, b2, b3, b4, b5, b6] of the previous data recording period, a1 is compared with b1, a2 is compared with b2, a3 is compared with b3, a4 is compared with b4, a5 is compared with b5, and a6 is compared with b6.

[0100] Of course, you can also consider aligning the two vectors to make a more direct comparison. If the eigenvalue ratios representing the characteristic performance of each event are not one-to-one corresponding, you can also consider adjusting the positions of the eigenvalues ​​in the eigenvectors to make them one-to-one corresponding according to the event characteristic performance.

[0101] Step 613: Determine a vector consisting of the maximum values ​​of the feature vectors corresponding to each event in the feature vectors of the current data recording period and the previous data recording period as a historical feature vector;

[0102] In this embodiment, the feature vector obtained in the previous data recording period is updated by continuously using the feature vector obtained in the current data recording period, and the maximum value of the feature vector in the current data recording period and the previous data recording period is used for updating. In this way, the characteristic behavior to be concerned can be maximized, preventing users from covering up previous violations through compliant operations.

[0103] Step 615: The obtained historical feature vector is used as a record of all data within each data recording period ending at the current time.

[0104] like Figure 7 As shown, an embodiment of the present specification further provides a device for recording historical data, which may include:

[0105] The historical data acquisition module 701 is configured to acquire historical data in the current data recording period for each data recording period;

[0106] The feature extraction module 702 is configured to perform feature extraction on the historical data acquired by the historical data acquisition module 701 to obtain at least one historical feature vector corresponding to the current data recording period;

[0107] The feature vector updating module 703 is configured to update the historical feature vector obtained in the first data recording period before the current data recording period using at least one historical feature vector obtained by the feature extraction module 702; wherein each value in the historical feature vector represents a feature value of a feature performance of an event;

[0108] The historical data recording module 704 is configured to use the historical feature vectors obtained after being updated by the feature vector updating module 703 as records of all historical data in each data recording period.

[0109] In a possible implementation, the historical data recording device provided in the embodiments of this specification may further include: a feature vector determination module, and the feature vector determination module is configured to perform the following operations when obtaining at least one historical feature vector corresponding to the current data recording period:

[0110] Determine at least one convolution kernel according to the event feature performance to be recorded; wherein each event feature performance corresponds to a convolution kernel;

[0111] For each event characteristic, execute:

[0112] The convolution kernel corresponding to the event feature expression is used to extract features from the sequence of historical data to obtain a feature vector corresponding to the event feature expression;

[0113] The maximum value in the feature vector is determined as the feature value corresponding to the feature performance of the current event in the current recording period;

[0114] A vector composed of feature values ​​corresponding to the feature performance of each event is determined as a historical feature vector corresponding to the current data recording period.

[0115] In one possible implementation, the convolution kernel can be learned by an end-to-end neural network using historical data represented by corresponding event features.

[0116] In a possible implementation, when the feature vector determination module determines the maximum value in the feature vector as the feature value corresponding to the feature representation of the current event in the current recording period, it is configured to perform the following operations:

[0117] The Max-Pooling mechanism of the convolutional neural network is used to determine the maximum value in the feature vector, and the maximum value is determined as the feature value corresponding to the feature performance of the current event in the current recording period.

[0118] In a possible implementation, when the feature vector updating module 703 uses at least one historical feature vector to update the historical feature vector obtained in the first data recording period before the current data recording period, it is configured to perform the following operations:

[0119] For each event characteristic performance, the larger value of the historical characteristic vector of the current data recording period and the historical characteristic vector of the first data recording period corresponding to the event characteristic performance is used as the characteristic value representing the event characteristic performance;

[0120] The vector formed by the obtained feature values ​​corresponding to the feature performance of each event is determined as the historical feature vector of the current data recording period.

[0121] In a possible implementation manner, the first data recording period includes a data recording period before the current data recording period.

[0122] The present specification also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed in a computer, the computer is caused to execute a method in any one of the embodiments of the present specification.

[0123] The present specification also provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method in any one of the embodiments in the specification is implemented.

[0124] It is to be understood that the structures illustrated in the embodiments of this specification do not constitute a specific limitation on the recording device of historical data. In other embodiments of the specification, the recording device of historical data may include more or fewer components than those illustrated, or combine certain components, or separate certain components, or arrange the components differently. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0125] The information interaction, execution process, etc. between the units in the above-mentioned device are based on the same concept as the method embodiments of this specification. For specific contents, please refer to the description in the method embodiments of this specification and will not be repeated here.

[0126] Those skilled in the art should be aware that in one or more of the above examples, the functions described in this specification can be implemented by hardware, software, widgets, or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0127] The specific implementation methods described above further illustrate the purpose, technical solutions and beneficial effects described in this specification. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. Methods for recording historical data, including: For each data logging cycle: Get historical data within the current data recording period; Extracting features from the historical data to obtain at least one historical feature vector corresponding to a current data recording period; Using the at least one historical feature vector to update the historical feature vector obtained in the first data recording period before the current data recording period; wherein each value in the historical feature vector represents a feature value of an event feature performance; The updated historical feature vector is used as a record of all historical data in each data recording period.

2. The method according to claim 1, wherein: The obtaining of at least one historical feature vector corresponding to the current data recording period includes: Determine at least one convolution kernel according to the event feature performance to be recorded; wherein each event feature performance corresponds to a convolution kernel; For each event characteristic, execute: Using the convolution kernel corresponding to the event feature expression to extract features from the sequence of historical data, a feature vector corresponding to the event feature expression is obtained; Determine the maximum value in the feature vector as the feature value corresponding to the feature performance of the current event in the current recording period; A vector composed of feature values ​​corresponding to the feature performance of each event is determined as a historical feature vector corresponding to the current data recording period.

3. The method according to claim 2, wherein: The convolution kernel is obtained by learning the historical data represented by the corresponding event characteristics through an end-to-end neural network.

4. The method according to claim 2, wherein: The step of determining the maximum value in the feature vector as the feature value corresponding to the feature performance of the current event in the current recording period includes: The Max-Pooling mechanism of the convolutional neural network is used to determine the maximum value in the feature vector, and the maximum value is determined as the feature value corresponding to the feature performance of the current event in the current recording period.

5. The method according to claim 1, wherein: The updating of the historical feature vector obtained in the first data recording period before the current data recording period by using the at least one historical feature vector comprises: For each event characteristic performance, the larger value of the historical characteristic vector of the current data recording period and the historical characteristic vector of the first data recording period corresponding to the event characteristic performance is used as the characteristic value representing the event characteristic performance; The vector formed by the obtained feature values ​​corresponding to the feature performance of each event is determined as the historical feature vector of the first data recording period.

6. The method according to any one of claims 1 to 5, wherein: The first data recording period includes a data recording period before the current data recording period.

7. A device for recording historical data, including: A historical data acquisition module, configured to acquire historical data within a current data recording period for each data recording period; a feature extraction module configured to perform feature extraction on the historical data acquired by the historical data acquisition module to obtain at least one historical feature vector corresponding to a current data recording period; A feature vector updating module, configured to update a historical feature vector obtained in a first data recording period before a current data recording period using the at least one historical feature vector obtained by the feature extraction module; wherein each value in the historical feature vector represents a feature value of a feature performance of an event; The historical data recording module is configured to use the historical feature vectors obtained after being updated by the feature vector updating module as records of all historical data in each data recording period.

8. The apparatus according to claim 7, further comprising: The feature vector determination module is configured to perform the following operations when obtaining at least one historical feature vector corresponding to the current data recording period: Determine at least one convolution kernel according to the event feature performance to be recorded; wherein each event feature performance corresponds to a convolution kernel; For each event characteristic, execute: Using the convolution kernel corresponding to the event feature expression to extract features from the sequence of historical data, a feature vector corresponding to the event feature expression is obtained; Determine the maximum value in the feature vector as the feature value corresponding to the feature performance of the current event in the current recording period; A vector composed of feature values ​​corresponding to the feature performance of each event is determined as a historical feature vector corresponding to the current data recording period.

9. The device according to claim 8, wherein: When the feature vector determination module determines the maximum value in the feature vector as the feature value corresponding to the feature representation of the current event in the current recording period, the feature vector determination module is configured to perform the following operations: The Max-Pooling mechanism of the convolutional neural network is used to determine the maximum value in the feature vector, and the maximum value is determined as the feature value corresponding to the feature performance of the current event in the current recording period.

10. A computing device comprising a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method according to any one of claims 1 to 6 is implemented.

Images