Feature extraction method and related device

Abstract

Embodiments of the present application disclose a feature extraction method and related devices, which can include a feature extraction apparatus, an electronic device, a computer program product, and a computer-readable storage medium. After obtaining a target interaction feature sequence of an interaction behavior, a feature sub-sequence of a preset length is divided in the target interaction feature sequence to obtain at least one interaction feature sub-sequence. Then, based on at least one preset sequence interaction range, features in the time dimension are extracted from the interaction feature sub-sequence to obtain a current feature sequence corresponding to the interaction behavior. Then, the current feature sequence is converted into a candidate feature sequence in the attribute dimension, and features in the attribute dimension are extracted from the candidate feature sequence to obtain a target feature sequence. Then, the target feature sequence is subjected to time-series dimension reduction to obtain an interaction behavior feature of the interaction behavior under a target task. This scheme can improve the accuracy of feature extraction.

Description

Technical Field

[0001] This invention relates to the field of feature processing, and more specifically to a feature extraction method and related equipment, which may include feature extraction devices, electronic devices, computer program products, and computer-readable storage media. Background Technology

[0002] In recent years, with the rapid development of internet technology, various interaction methods have become increasingly convenient. From massive amounts of interactive behavior, feature sequences of these behaviors can be extracted, enabling analysis of these interactions. Current feature extraction methods often employ convolution to perform convolution operations on each attribute in the time dimension to capture the correlation between adjacent behaviors and extract feature sequences. Alternatively, fully connected methods can be used to extract feature sequences by realizing the interaction between sequence attributes.

[0003] In the process of researching and practicing current technologies, the inventors of this application have found that traditional convolutional methods mainly focus on capturing local correlations in the time dimension, and often lack sufficient accuracy in extracting features in the attribute dimension. On the other hand, fully connected methods mainly work on feature extraction in the attribute dimension, and cannot fully capture features in time series data in the time dimension, thus resulting in insufficient accuracy in feature extraction. Summary of the Invention

[0004] This application provides a feature extraction method and related equipment. The related equipment may include a feature extraction device, an electronic device, a computer program product, and a computer-readable storage medium, which can improve the accuracy of feature extraction.

[0005] A feature extraction method, comprising:

[0006] Obtain the target interaction feature sequence of the interaction behavior, wherein the target interaction feature sequence includes the feature sequence of the interaction object corresponding to the interaction behavior during the interaction process;

[0007] Divide the target interactive feature sequence into feature subsequences of a preset length to obtain at least one interactive feature subsequence;

[0008] Based on at least one preset sequence interaction range, features in the time dimension are extracted from the interaction feature subsequence to obtain the current feature sequence corresponding to the interaction behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension.

[0009] The current feature sequence is converted into a candidate feature sequence in the attribute dimension, and the features in the attribute dimension are extracted from the candidate feature sequence to obtain the target feature sequence;

[0010] Temporal dimensionality reduction is performed on the target feature sequence to obtain the interaction behavior features of the interaction behavior under the target task.

[0011] Accordingly, embodiments of this application provide a feature extraction apparatus, including:

[0012] An acquisition unit is used to acquire a target interaction feature sequence of an interaction behavior, wherein the target interaction feature sequence includes the feature sequence of the interaction object corresponding to the interaction behavior during the interaction process.

[0013] A partitioning unit is used to divide the target interactive feature sequence into feature subsequences of a preset length to obtain at least one interactive feature subsequence.

[0014] An extraction unit is configured to extract features in the time dimension from the interaction feature subsequence based on at least one preset sequence interaction range, so as to obtain the current feature sequence corresponding to the interaction behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension.

[0015] A transformation unit is used to convert the current feature sequence into a candidate feature sequence in the attribute dimension, and extract the features in the attribute dimension from the candidate feature sequence to obtain the target feature sequence;

[0016] The dimensionality reduction unit is used to perform temporal dimensionality reduction on the target feature sequence to obtain the interaction behavior features of the interaction behavior under the target task.

[0017] In some embodiments, the extraction unit may be specifically used to extract features in the time dimension from the interaction feature subsequence to obtain an initial time feature sequence; to filter out a target sequence interaction range corresponding to the initial time feature sequence from the preset sequence interaction range; to concatenate the initial time feature sequence based on the target sequence interaction range, and to extract features from the concatenated time feature sequence to obtain the current feature sequence corresponding to the interaction behavior.

[0018] In some embodiments, the extraction unit may be specifically used to extract at least one feature in the time dimension from the interactive feature subsequence to obtain an original time feature sequence; to perform linear activation on the original time feature sequence to obtain an activated time feature sequence; and to scale the activated time feature sequence based on a preset scaling parameter to obtain an initial time feature sequence, wherein the preset scaling parameter indicates the probability that a feature in the activated time feature sequence is set to zero.

[0019] In some embodiments, the extraction unit may be specifically used to obtain the position information of the interaction feature subsequence in the target interaction feature sequence, and sort the initial time feature sequence based on the position information; construct at least one feature sequence group based on the sequence length of the initial time feature sequence and the sorting result, wherein the sum of the sequence lengths of the initial time feature sequences in the feature sequence group is the sequence length corresponding to the target sequence interaction range; and concatenate the initial time feature sequences in the feature sequence group to obtain at least one concatenated time feature sequence corresponding to the target interaction range.

[0020] In some embodiments, the extraction unit may be specifically used to extract features in the time dimension from the spliced ​​time feature sequence to obtain the current time feature sequence, and use the current time feature sequence as the initial time feature sequence; return to the step of filtering the target sequence interaction range corresponding to the initial time feature sequence in the preset sequence interaction range, until the sequence length of the current time feature sequence is the same as the sequence length of the target interaction feature sequence, to obtain the current feature sequence corresponding to the interaction behavior.

[0021] In some embodiments, the transformation unit may be specifically used to obtain the dimensional order in the current feature sequence, the dimensional order including the order between the time dimension and the attribute dimension; based on the dimensional order, the current feature sequence is transposed to obtain a candidate feature sequence on the attribute dimension.

[0022] In some embodiments, the conversion unit may be specifically used to fuse the candidate feature sequence with the target interaction feature sequence to obtain a fused feature sequence; normalize the fused feature sequence to obtain a normalized feature sequence, and extract the features in the attribute dimension from the normalized feature sequence to obtain an attribute feature sequence; and fuse the attribute feature sequence with the fused feature sequence to obtain a target feature sequence.

[0023] In some embodiments, the dimensionality reduction unit may be used to transpose the target feature sequence to obtain a transposed feature sequence; extract at least one dimension of features corresponding to the target task from the transposed feature sequence to obtain at least one task feature sequence; and perform a pooling operation on the task feature sequence to obtain an interaction behavior feature sequence of the interaction behavior under the target task.

[0024] In some embodiments, the acquisition unit may be specifically used to acquire a feature sequence set of at least one feature attribute, the feature sequence set including at least one feature sequence corresponding to the interactive behavior; perform feature interaction on the feature sequences in the feature sequence set to obtain at least one interactive feature sequence; and fuse the interactive feature sequences to obtain the target interactive feature sequence of the interactive behavior.

[0025] In some embodiments, the acquisition unit may be specifically used to acquire at least one interaction data sequence of the interaction behavior, classify the interaction data sequence according to the feature attribute to obtain at least one target interaction data sequence corresponding to the feature attribute; perform feature extraction on the target interaction data sequence to obtain at least one initial feature sequence corresponding to the feature attribute; acquire the interaction location information of the interaction behavior, and perform feature extraction on the interaction location information to obtain the location feature of the interaction behavior; and fuse the initial feature sequence with the location feature to obtain a feature sequence set corresponding to the feature attribute.

[0026] In some embodiments, the acquisition unit may be specifically used to filter out the current feature sequence pair in the feature sequence set, the current feature sequence pair including two feature sequences that need to be interacted with; perform feature interaction on the feature sequences in the current feature sequence pair to obtain the initial interaction feature sequence corresponding to the current feature sequence pair; return to the step of filtering out the current feature sequence pair in the feature sequence set until all feature sequences in the feature sequence set are feature sequences in the current feature sequence pair, thereby obtaining at least one interaction feature sequence.

[0027] In some embodiments, the acquisition unit may be specifically used to extract features from the second feature sequence based on the first feature sequence to obtain a first interactive feature sequence; extract features from the first feature sequence based on the second feature sequence to obtain a second interactive feature sequence; and fuse the first interactive feature sequence and the second interactive feature sequence to obtain an initial interactive feature sequence corresponding to the current feature sequence pair.

[0028] In some embodiments, the acquisition unit may be specifically used to extract feature information in the time dimension from the first feature sequence using a convolutional kernel of at least one scale to obtain a first candidate interaction feature sequence; to extract feature information in the time dimension from the second feature sequence using the convolutional kernel of the at least one scale to obtain a second candidate interaction feature sequence; to perform activation processing on the first candidate interaction feature sequence, and to fuse the activated first candidate interaction feature sequence with the second candidate interaction feature sequence to obtain a first interaction feature sequence.

[0029] In some embodiments, the acquisition unit may be specifically used to activate the second candidate interaction feature sequence to obtain an activated second candidate interaction feature sequence; and to fuse the activated second candidate interaction feature sequence with the first candidate interaction feature to obtain a second interaction feature sequence.

[0030] In some embodiments, the acquisition unit may be specifically used to concatenate the first interaction feature sequence and the second interaction feature sequence to obtain a concatenated interaction feature sequence; perform feature extraction on the concatenated interaction feature sequence to obtain a target concatenated interaction feature sequence; and perform activation processing on the target concatenated interaction feature sequence to obtain the initial interaction feature sequence corresponding to the current feature sequence pair.

[0031] In some embodiments, the acquisition unit may be specifically used to normalize the interaction feature sequence to obtain a normalized interaction feature sequence; fuse the normalized interaction feature sequence to obtain a fused interaction feature sequence; and transpose the fused interaction feature sequence to obtain the target interaction feature sequence of the interaction behavior.

[0032] Furthermore, embodiments of this application also provide an electronic device, including a processor and a memory, wherein the memory stores an application program, and the processor is used to run the application program in the memory to execute the feature extraction method provided in embodiments of this application.

[0033] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute steps in any of the feature extraction methods provided in embodiments of this application.

[0034] Furthermore, embodiments of this application also provide a computer program product, including a computer program or instructions, which, when executed by a processor, implement the steps in the feature extraction method provided in embodiments of this application.

[0035] In this embodiment, after obtaining the target interaction feature sequence of the interaction behavior, a feature subsequence of preset length is divided from the target interaction feature sequence to obtain at least one interaction feature subsequence. Then, based on at least one preset sequence interaction range, features in the time dimension are extracted from the interaction feature subsequence to obtain the current feature sequence corresponding to the interaction behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension. Then, the current feature sequence is converted into a candidate feature sequence in the attribute dimension, and features in the attribute dimension are extracted from the candidate feature sequence to obtain the target feature sequence. Finally, temporal dimensionality reduction is performed on the target feature sequence to obtain the interaction behavior. This approach identifies interactive behavior features within a target task. Because it divides the target interactive feature sequence into sub-feature sequences of a given length and extracts temporal features from at least one fixed-length sub-sequence by pre-setting the sequence interaction range, it focuses more on regional relational dependencies during feature extraction. This avoids the loss of effective information caused by padding operations at both ends of the sequence, improving the accuracy of temporal feature extraction. Furthermore, by converting temporal features into attribute-dimensional features, it simultaneously mines both attribute-dimensional and temporal features of the behavioral sequence, thus further enhancing the accuracy of feature extraction. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of a scenario for the feature extraction method provided in the embodiments of this application;

[0038] Figure 2 This is a schematic flowchart of the feature extraction method provided in the embodiments of this application;

[0039] Figure 3 This is a schematic diagram of the network structure of the interactive convolutional network provided in the embodiments of this application;

[0040] Figure 4 This is a schematic diagram of the network structure of a fully connected network in the time dimension provided in the embodiments of this application;

[0041] Figure 5 This is a schematic diagram of the network structure of a fully connected network based on the feature (attribute) dimension provided in the embodiments of this application;

[0042] Figure 6This is a schematic diagram of the network structure of the incremental fully connected network provided in the embodiments of this application;

[0043] Figure 7 This is a schematic diagram of the model structure of the feature extraction model provided in the embodiments of this application;

[0044] Figure 8 This is another schematic diagram of the feature extraction method provided in the embodiments of this application;

[0045] Figure 9 This is a schematic diagram of the feature extraction device provided in the embodiments of this application;

[0046] Figure 10 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0047] The technical solutions of the embodiments of this 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 invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] This application provides a feature extraction method and related equipment. The related equipment may include a feature extraction device, an electronic device, a computer program product, and a computer-readable storage medium. The feature extraction device may be integrated into an electronic device, which may be a server or a terminal, etc.

[0049] The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery network (CDN), and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0050] For example, see Figure 1Taking the integration of a feature extraction device into an electronic device as an example, after acquiring the target interaction feature sequence of the interactive behavior, the electronic device can divide the target interaction feature sequence into feature subsequences of a preset length to obtain at least one interaction feature subsequence. Then, based on at least one preset sequence interaction range, features in the time dimension are extracted from the interaction feature subsequence to obtain the current feature sequence corresponding to the interactive behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension. Then, the current feature sequence is converted into a candidate feature sequence in the attribute dimension, and features in the attribute dimension are extracted from the candidate feature sequence to obtain the target feature sequence. Then, temporal dimensionality reduction is performed on the target feature sequence to obtain the interaction behavior features of the interactive behavior under the target task, thereby improving the accuracy of feature extraction.

[0051] It is understood that, in the specific embodiments of this application, the sequence data of the interaction behavior of the object and the target interaction feature sequence and other related data are involved. When the following embodiments of this application are applied to specific products or technologies, permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0052] The following sections provide detailed descriptions of each example. It should be noted that the order in which the embodiments are described is not intended to limit the preferred order of the embodiments.

[0053] This embodiment will be described from the perspective of a feature extraction device, which can be integrated into an electronic device, such as a server or a terminal. The terminal can include tablet computers, laptops, personal computers (PCs), wearable devices, virtual reality devices, or other smart devices capable of feature extraction.

[0054] A feature extraction method, comprising:

[0055] A target interaction feature sequence of interactive behavior is obtained, which includes the feature sequence of the interactive object corresponding to the interactive behavior during the interaction process. A feature subsequence of a preset length is divided in the target interaction feature sequence to obtain at least one interaction feature subsequence. Based on at least one preset sequence interaction range, features in the time dimension are extracted from the interaction feature subsequence to obtain the current feature sequence corresponding to the interactive behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension. The current feature sequence is converted into a candidate feature sequence in the attribute dimension, and features in the attribute dimension are extracted from the candidate feature sequence to obtain the target feature sequence. Temporal dimensionality reduction is performed on the target feature sequence to obtain the interaction behavior features of the interactive behavior under the target task.

[0056] like Figure 2 As shown, Figure 2 This is a schematic diagram of an optional flowchart of the feature extraction method provided in this embodiment. The feature extraction method can be executed by the server or the terminal alone, or it can be executed by the terminal and the server in cooperation. The specific flow of the content search method can be as follows:

[0057] 101. Obtain the target interaction feature sequence of the interaction behavior.

[0058] Interaction behavior can be understood as an interaction between at least two objects. There are many types of interaction behavior, such as payment or transaction behavior for virtual or physical resources / services; browsing, liking, collecting, commenting, forwarding, tipping, or other behaviors related to content; adding / deleting / blocking friends, sending or receiving messages, or other related behaviors, etc.

[0059] The target interaction feature sequence can include the feature sequence of the interaction object corresponding to the interaction behavior during the interaction process. An interaction object can be understood as an object participating in the interaction behavior. For example, taking payment as an example, the interaction objects can include the payer and the receiver; or, taking sending a message as an example, the interaction objects can include the sender and the receiver; or, taking commenting on content as an example, the interaction objects can include the content creator / publisher and the commenter, and so on. The target interaction feature sequence of an interaction behavior can be obtained as follows:

[0060] For example, a feature sequence set of at least one feature attribute can be obtained, and the feature sequences in the feature sequence set can be interacted to obtain at least one interactive feature sequence. The interactive feature sequences can then be fused to obtain the target interactive feature sequence of the interactive behavior. Specifically, this can be done as follows:

[0061] S1. Obtain a set of feature sequences with at least one feature attribute.

[0062] The feature sequence set includes at least one feature sequence corresponding to the interaction behavior. The process of obtaining a feature sequence set for at least one feature attribute can be as follows:

[0063] For example, at least one interactive data sequence of interactive behavior can be obtained, and the interactive data sequence can be classified according to feature attributes to obtain at least one target interactive data sequence corresponding to the feature attributes. Feature extraction can be performed on the target interactive data sequence to obtain at least one initial feature sequence corresponding to the feature attributes. The location information of the interactive behavior can be obtained, and feature extraction can be performed on the location information to obtain the location features of the interactive behavior. The initial feature sequence and the location features can be fused to obtain a set of feature sequences corresponding to the feature attributes.

[0064] Interaction data sequences can be understood as a series of sequential data closely related to each interaction stage, arranged chronologically during the interaction process. Taking commercial transactions as an example, this interaction data sequence can be understood as time-series data of commercial transactions. This time-series data can include basic information comprehensively describing the transaction behavior, as well as information reflecting the dynamic characteristics of merchants in the transaction. Basic information can include transaction amount, transaction channel, business scenario of the transaction, transaction time stamp, merchant category, or other transaction behavior information, etc. Dynamic characteristic information can include the merchant's average transaction amount, maximum transaction amount, or other dynamic transaction information over a past time range, etc. The time range can include five days, one week, ten days, half a month, twenty days, one month, one quarter, half a year, or one year, etc. Taking interactive behavior as an example of transaction behavior, the interactive data sequence contains basic information about a business transaction and the dynamic characteristics of the merchant within the transaction. All possible combinations of attributes for each transaction constitute the value space of the transaction behavior. Each possible value is defined as a token, analogous to a word in a natural language processing task. The sequence length depends on the sequence length of the actual transaction behaviors in the training data. In this scheme, the sequence length can be fixed to a fixed value, such as 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 150, 200, or any other positive integer value, etc.

[0065] There are several ways to obtain at least one interactive data sequence of interactive behavior. For example, at least one interactive data sequence of interactive behavior can be obtained from the content interaction platform. Alternatively, the interactive data of at least one interactive object corresponding to the interactive behavior can be obtained during the interaction process, and time information can be extracted from the interactive data. Based on the time information, the interactive data can be sorted to obtain at least one interactive data sequence. Or, at least one interactive data sequence sent by the interactive terminal of the interactive object corresponding to the interactive behavior can be obtained, and so on.

[0066] After acquiring at least one interactive data sequence of interactive behavior, the interactive data sequence can be classified according to feature attributes to obtain at least one target interactive data sequence corresponding to the feature attributes. The process of classifying interactive data sequences according to feature attribute information may include: acquiring attribute information of the interactive data sequence, and determining the feature attributes corresponding to the interactive data sequence based on the attribute information, thereby obtaining at least one target interactive data sequence corresponding to the feature attributes.

[0067] After classifying the interactive data sequences according to their feature attributes, feature extraction can be performed on at least one target interactive data sequence corresponding to the feature attributes, thereby obtaining at least one initial feature sequence corresponding to the feature attributes. The process of feature extraction on the target interactive data sequence may include: using an embedding network (a feature extraction network) to extract feature embeddings from the target interactive data sequence, thereby obtaining the initial feature sequence of the target interactive feature sequence.

[0068] The location information of an interactive behavior can be understood as location-related information within the interactive behavior. For example, taking a payment behavior as an example, this location information could include the location of the payment device or terminal, or the location of the payment behavior, and so on.

[0069] After extracting features from the target interaction data sequence and from the interaction location information of the acquired interaction behavior, the initial feature sequence and location features can be fused to obtain a set of feature sequences corresponding to the feature attributes. There are several ways to fuse the initial feature sequence and location features. For example, the location features can be accumulated with each feature in the initial feature sequence to obtain a set of feature sequences corresponding to the feature attributes, which can include the feature sequences corresponding to each initial feature sequence. Alternatively, location sub-features corresponding to each feature in the initial feature sequence can be selected from the location features, and these sub-features can be concatenated with the corresponding features in the initial feature sequence to obtain a set of feature sequences corresponding to the feature attributes, which can also include the feature sequences corresponding to each initial feature sequence, and so on.

[0070] S2. Perform feature interaction on the feature sequences in the feature sequence set to obtain at least one interactive feature sequence.

[0071] For example, the current feature sequence pair can be selected from the feature sequence set. The current feature sequence pair includes two feature sequences that need to be interacted. The feature sequences in the current feature sequence pair are interacted to obtain the initial interactive feature sequence corresponding to the current feature sequence pair. Then, the step of selecting the current feature sequence pair from the feature sequence set is returned to be executed until all feature sequences in the feature sequence set are feature sequences in the current feature sequence pair, thus obtaining at least one interactive feature sequence.

[0072] There are several ways to select the current feature sequence pair from the feature sequence set. For example, two feature sequences can be randomly extracted from the feature sequence set without replacement to obtain the current feature sequence pair. Alternatively, the feature sequences in the feature sequence set can be sorted, and based on the sorting result, the first two feature sequences can be extracted from the feature sequence set without replacement to obtain the current feature sequence pair, and so on.

[0073] After selecting the current feature sequence pair from the feature sequence set, the feature sequences in the current feature sequence pair can be interacted to obtain the initial interactive feature sequence of the current feature sequence pair. The feature sequence pair can include a first feature sequence and a second feature sequence. The process of interacting the feature sequences in the current feature sequence pair can include: extracting features from the second feature sequence based on the first feature sequence to obtain a first interactive feature sequence; extracting features from the first feature sequence based on the second feature sequence to obtain a second interactive feature sequence; and fusing the first and second interactive feature sequences to obtain the initial interactive feature sequence corresponding to the current feature sequence pair. Specifically, it can be as follows:

[0074] (1) Based on the first feature sequence, the second feature sequence is extracted to obtain the first interactive feature sequence.

[0075] For example, a convolutional kernel of at least one scale can be used to extract the feature information in the time dimension of the first feature sequence to obtain the first candidate interaction feature sequence. A convolutional kernel of at least one scale can be used to extract the feature information in the time dimension of the second feature sequence to obtain the second candidate interaction feature sequence. The first candidate interaction feature sequence is activated, and the activated first candidate interaction feature sequence is fused with the second candidate interaction feature sequence to obtain the first interaction feature sequence.

[0076] The process of extracting temporal feature information from the first feature sequence using a convolutional kernel of at least one scale may include: using a one-dimensional convolutional neural network with a convolutional kernel of at least one scale to extract temporal feature information from the first feature sequence, thereby capturing local features and longer-distance dependencies, and thus obtaining the first candidate interactive feature sequence.

[0077] The method of extracting time-dimensional feature information from the second feature sequence using a convolution kernel of at least one scale is similar to the method of extracting time-dimensional feature information from the first feature sequence using a convolution kernel of at least one scale, as described above, and will not be repeated here.

[0078] After extracting the temporal feature information from the first feature sequence, the extracted first candidate interaction feature sequence can be activated to obtain the activated first candidate interaction feature sequence. There are several ways to activate the extracted first candidate interaction feature sequence. For example, GELU (an activation function) can be used, or ReLU (an activation function) can be used, or other activation functions can be used, etc.

[0079] It should be noted that the activation process on the first candidate interactive feature sequence is mainly aimed at removing noise or redundant information at each convolutional scale, thereby obtaining the key feature information in the first candidate interactive feature sequence. This key information can then be used as a mask to extract feature information from the second candidate interactive feature sequence.

[0080] After activating the first candidate interaction feature sequence, the activated first candidate interaction feature sequence can be fused with the second candidate interaction feature sequence to obtain the first interaction feature sequence. The process of fusing the activated first candidate interaction feature sequence with the second candidate interaction feature sequence can include: performing a multiplication operation between the activated first candidate interaction feature sequence and the second candidate interaction feature sequence to obtain the first interaction feature sequence, as shown in formula (1), specifically as follows:

[0081]

[0082] Wherein, C1 is the first interaction feature sequence, S1 is the first candidate interaction feature sequence, and S2 is the second candidate interaction feature sequence. is the activation function, and ⊙ is the multiplication operation.

[0083] (2) Based on the second feature sequence, feature extraction is performed on the first feature sequence to obtain the second interactive feature sequence.

[0084] For example, the second candidate interaction feature sequence can be activated to obtain the activated second candidate interaction feature sequence, and the activated second candidate interaction feature sequence can be fused with the first candidate interaction feature to obtain the second interaction feature sequence.

[0085] The activation process for the second candidate interactive feature sequence can be similar to that for the first candidate interactive feature sequence, as detailed above, and will not be repeated here.

[0086] After activating the second candidate interaction feature sequence, the activated second candidate interaction feature sequence can be fused with the first candidate interaction feature sequence to obtain the second interaction feature sequence. The method of fusing the activated second candidate interaction feature sequence with the first candidate interaction feature sequence can include multiplying the activated second candidate interaction feature sequence with the first candidate interaction feature sequence to obtain the second interaction feature sequence, as shown in formula (2), specifically as follows:

[0087]

[0088] Where C2 is the second interaction feature sequence, S1 is the first candidate interaction feature sequence, and S2 is the second candidate interaction feature sequence. is the activation function, and ⊙ is the multiplication operation.

[0089] (3) The first interactive feature sequence and the second interactive feature sequence are fused together to obtain the initial interactive feature sequence corresponding to the current feature sequence pair.

[0090] For example, the first interaction feature sequence and the second interaction feature sequence can be concatenated to obtain the concatenated interaction feature sequence. Feature extraction can be performed on the concatenated interaction feature sequence to obtain the target concatenated interaction feature sequence. Activation processing can be performed on the target concatenated interaction feature to obtain the initial interaction feature sequence of the current feature sequence pair.

[0091] The process of concatenating the first interactive feature sequence with the second interactive feature sequence may include: adding the features of the first interactive feature sequence and the second interactive feature sequence to obtain the concatenated interactive feature sequence.

[0092] After concatenating the first and second interaction feature sequences, feature extraction can be performed on the concatenated interaction feature sequence to obtain the target concatenated interaction feature sequence. The feature extraction process for the concatenated interaction feature sequence may include using a one-dimensional convolutional network to extract features from the concatenated interaction feature sequence, thereby obtaining the target concatenated interaction feature sequence.

[0093] After feature extraction from the concatenated interactive feature sequence, the extracted target concatenated interactive feature sequence can be activated to obtain the initial interactive feature sequence corresponding to the current feature sequence pair. The activation process of the target concatenated interactive feature sequence can include: using ReLU (an activation function) to activate the target concatenated interactive feature sequence to obtain the initial interactive feature sequence corresponding to the current feature sequence pair, as shown in formula (3), specifically as follows:

[0094] O ICN =ReLU(Conv3(C1+C2)) (3)

[0095] Among them, O ICN C1 is the initial interaction feature sequence, C2 is the first interaction feature sequence, and C3 is the second interaction feature sequence.

[0096] In this scheme, an interactive convolutional network can be used for feature interaction. The network structure of the interactive convolutional network can be as follows: Figure 3 As shown. The interactive convolutional network adopts a two-layer convolutional structure. The first layer uses a one-dimensional convolutional neural network with convolutional kernels of different sizes to capture local features and longer-distance dependencies. The input sequence data can be passed through this series of convolutional networks in parallel, aiming to enhance the model's ability to mine and interpret complex time series patterns with smaller kernels. In contrast, the second layer applies a one-dimensional convolutional network, which is designed to interact with the feature outputs of the previous layer to identify broader and longer-distance dependencies. Between the two layers, different multiplication and addition operations are applied in this scheme to enable the convolutional networks with different kernel sizes to control each other, extract interactive information between features at different scales, and balance the feature extraction of the time dimension of local and global sequence data to produce better complex relationship modeling. In addition, in order to remove noise or redundant information at each convolutional scale, activation functions such as GELU are applied before the multiplication operation. The entire process of feature interaction of the feature sequences in the current feature sequence pair can be shown in formulas (1)-(3).

[0097] After performing feature interaction on the feature sequences in the current feature sequence pair, the process can return to the step of filtering out the current feature sequence pair from the feature sequence set, until all feature sequences in the feature sequence set are feature sequences in the current feature sequence pair, thus obtaining at least one interactive feature sequence.

[0098] S3. Fuse the interaction feature sequences to obtain the target interaction feature sequence of the interaction behavior.

[0099] For example, the interaction feature sequence can be normalized to obtain a normalized interaction feature sequence, the normalized interaction feature sequence can be fused to obtain a fused interaction feature sequence, and the fused interaction feature sequence can be transposed to obtain the target interaction feature sequence of the interaction behavior.

[0100] There are several ways to fuse the normalized interaction feature sequences. For example, the normalized interaction feature sequences can be concatenated to obtain the fused interaction feature sequence, or the normalized interaction feature sequences can be accumulated to obtain the fused interaction feature sequence, and so on.

[0101] After fusing the normalized interaction feature sequences, the fused interaction feature sequences can be transposed to obtain the target interaction feature sequence of the interaction behavior. There are several ways to transpose the fused interaction feature sequences. For example, one can obtain the dimensional order of the fused interaction feature sequences, which can include the order between the time dimension and the attribute dimension. Based on this dimensional order, the fused interaction feature sequences are transposed to obtain the target interaction feature sequence, where the last dimension of the target interaction feature sequence is the time dimension.

[0102] 102. Divide the target interactive feature sequence into feature subsequences of a preset length to obtain at least one interactive feature subsequence.

[0103] For example, at least one segmentation position corresponding to a preset length can be identified in the target interactive feature sequence, and at least one interactive feature subsequence can be divided in the target interactive feature sequence based on the segmentation position.

[0104] The preset length can be a pre-defined length used to divide the subsequences. The sequence length of each interactive feature subsequence is the preset length. For example, if the preset length is n, then the sequence length of the divided interactive feature subsequences can be n.

[0105] 103. Based on at least one preset sequence interaction range, extract the features in the time dimension from the interaction feature subsequence to obtain the current feature sequence corresponding to the interaction behavior.

[0106] The preset sequence interaction range is used to extract the sequence length of the feature sequence that contains features in the time dimension.

[0107] Specifically, the process of extracting features in the time dimension from the interaction feature subsequence based on at least one preset sequence interaction range can be as follows:

[0108] For example, features in the time dimension can be extracted from the interaction feature subsequence to obtain an initial time feature sequence. Then, a target sequence interaction range corresponding to the initial time feature sequence can be selected from a preset sequence interaction range. Based on the target sequence interaction range, the initial time feature sequences are concatenated, and features are extracted from the concatenated time feature sequence to obtain the current feature sequence corresponding to the interaction behavior. Specifically, this can be done as follows:

[0109] (1) Extract the features in the time dimension from the interactive feature subsequence to obtain the initial time feature sequence.

[0110] For example, at least one feature in the time dimension can be extracted from the interactive feature subsequence to obtain the original time feature sequence. The original time feature sequence can be linearly activated to obtain the activated time feature sequence. Based on a preset scaling parameter, the activated time feature sequence can be scaled to obtain the initial time feature sequence.

[0111] There are several ways to extract at least one feature in the time dimension from the interactive feature subsequence. For example, a fully connected network (FCN) can be used to extract at least one feature in the time dimension from the interactive feature subsequence to obtain the initial time feature sequence. Alternatively, other feature extraction networks can be used to extract at least one feature in the time dimension from the interactive feature subsequence to obtain the initial time feature sequence, and so on.

[0112] After extracting at least one feature in the time dimension from the interactive feature subsequence, the extracted original time feature sequence can be linearly activated to obtain the activated time feature sequence. There are several ways to linearly activate the original time feature sequence. For example, the ReLU activation function can be used, or other activation functions can be used, and so on.

[0113] After linear activation of the original time feature sequence, the activated time feature sequence can be scaled based on a preset scaling parameter to obtain the initial time feature sequence. The preset scaling parameter indicates the probability that a feature in the activated time feature sequence will be set to zero. The process of scaling the activated time feature sequence based on the preset scaling parameter can include multiplying the preset scaling parameter by the feature value of the activated time feature sequence to obtain the initial time feature sequence.

[0114] (2) Filter out the target sequence interaction range corresponding to the initial time feature sequence from the preset sequence interaction range.

[0115] For example, the preset sequence interaction ranges can be sorted, and based on the sorting results, the smallest preset sequence interaction range can be selected from the preset sequence interaction sorted ranges to obtain the target sequence interaction range corresponding to the initial time feature sequence. Alternatively, the sequence length of the initial time feature sequence can be obtained, and the preset sequence interaction ranges with sequence lengths adjacent to that sequence length can be selected from the preset sequence interaction ranges to obtain the target sequence interaction range corresponding to the initial time feature sequence. Or, the target sequence length corresponding to each preset sequence interaction range can be obtained, and based on the target sequence length, the preset sequence interaction ranges with the sequence length of the preset initial time feature sequence can be selected from the preset sequence interaction ranges to obtain the target sequence interaction range corresponding to the initial time feature sequence, and so on.

[0116] (3) Based on the target sequence interaction range, the initial time feature sequence is spliced ​​together, and the spliced ​​time feature sequence is used to extract features to obtain the current feature sequence corresponding to the interaction behavior.

[0117] The process of concatenating the initial time feature sequence features based on the target sequence interaction range can be as follows:

[0118] For example, the position information of the interactive feature subsequence in the target interactive feature sequence can be obtained, and the initial time feature sequence can be sorted based on the position information. Based on the sequence length of the initial time feature sequence and the sorting result, at least one feature sequence group can be constructed. The sum of the sequence lengths of the initial time feature sequences in the feature sequence group is the sequence length corresponding to the target sequence interaction range. The initial time feature sequences in the feature sequence group can be concatenated to obtain at least one concatenated time feature sequence corresponding to the target interaction range.

[0119] There are several ways to construct at least one feature sequence group based on the sequence length and sorting result of the initial time feature sequence. For example, if the initial time feature sequence includes feature sequence 1, feature sequence 2, feature sequence 3, and feature sequence 4, and the sorting of each feature sequence is as described above, with each feature sequence having a length of 4 and the target interaction range having a sequence length of 8, then feature sequence 1 and feature sequence 2 can be considered as one feature sequence group, and feature sequence 3 and feature sequence 4 can be considered as another feature sequence group. Another example is if the initial time feature sequence includes feature sequence 1, feature sequence 2, feature sequence 3, feature sequence 4, feature sequence 5, and feature sequence 6, and the sorting of each feature sequence is as described above, with each feature sequence having a length of 4 and the target interaction range having a sequence length of 12, then feature sequence 1, feature sequence 2, and feature sequence 3 can be considered as one feature sequence group, and feature sequence 4, feature sequence 5, and feature sequence 6 can be considered as another feature sequence group, and so on.

[0120] It should be noted that each feature sequence group can contain at least two initial time feature sequences, and the sum of the sequence lengths of each feature sequence group is the same.

[0121] After constructing at least one feature sequence group based on the sequence length and sorting result of the initial time feature sequence, the initial time feature sequences in the feature sequence group can be spliced ​​together to obtain at least one spliced ​​time feature sequence corresponding to the target interaction range.

[0122] After concatenating the initial time feature sequences based on the target sequence interaction range, feature extraction can be performed on the concatenated time features to obtain the current feature sequence corresponding to the interaction behavior. The feature extraction process for the concatenated feature sequence can include: extracting features in the time dimension from the concatenated feature sequence to obtain the current time feature sequence; using the current time feature sequence as the initial time feature sequence; and returning to execute the step of filtering the target sequence interaction range corresponding to the initial time feature sequence within the preset sequence interaction range until the sequence length of the current time feature sequence is the same as the sequence length of the target interaction feature sequence, thus obtaining the current feature sequence corresponding to the interaction behavior.

[0123] The method of extracting features in the time dimension from the spliced ​​time feature sequence is similar to the method of extracting features in the time dimension from the interactive feature subsequence, as described above, and will not be repeated here.

[0124] After extracting the time dimension features from the concatenated time feature sequences, the extracted current time feature sequence can be used as the initial time feature sequence. Then, the process of filtering the target sequence interaction range corresponding to the initial time feature sequence within the preset sequence interaction range is repeated until the sequence length of the current time feature sequence is the same as the sequence length of the target interaction feature sequence. This yields the current feature sequence corresponding to the interaction behavior.

[0125] In this scheme, a time-dimensional fully connected network can be used to extract time-dimensional features from the target interaction sub-features. The time-dimensional fully connected network aims to achieve the functionality of a convolutional network using a linear fully connected network. However, unlike convolutional kernels which capture adjacency dependencies, the time-dimensional fully connected network focuses more on region-based dependencies and does not require padding operations to maintain data dimensional consistency. The network structure of the time-dimensional fully connected network can be as follows: Figure 4 As shown, the time-dimensional fully connected network consists of n fully connected layers, and the number of preset sequence interaction ranges can also be n. Each fully connected layer sequentially contains a feature subsequence of fixed length. Figure 4In this process, a first-layer fully connected network extracts initial temporal feature sequences one by one from multiple interaction feature sub-sequences of given length. Based on a preset sequence interaction range, the output initial temporal feature sequences are concatenated and used as input to the second-layer fully connected network. This is analogous to increasing the receptive field in convolutional networks. The second-layer fully connected network increases the sequence interaction range based on the sequence length of the initial temporal feature sequence output by the first layer, and so on until the entire feature sequence passes through the final fully connected network layer, thus extracting the current feature sequence of the interaction behavior. The length of the extracted current feature sequence can then be the same as the length of the target interaction feature sequence.

[0126] 104. Convert the current feature sequence into a candidate feature sequence in the attribute dimension, and extract the features in the attribute dimension from the candidate feature sequence to obtain the target feature sequence.

[0127] Here, the attribute dimension can be understood as the dimension of the feature attributes, or it can be seen as the feature dimension of the feature sequence corresponding to the feature attributes. The process of transforming the current feature sequence into a candidate feature sequence in terms of the attribute dimension can be as follows:

[0128] For example, the dimensional order in the current feature sequence can be obtained, which may include the order between the time dimension and the attribute dimension. Based on the dimensional order, the current feature sequence can be transposed to obtain the candidate feature sequence in the attribute dimension.

[0129] After converting the current feature sequence into a candidate feature sequence along the attribute dimension, features along the attribute dimension can be extracted from the candidate feature sequence to obtain the target feature sequence. The process of extracting features along the attribute dimension from the candidate feature sequence can include: fusing the candidate sequence features with the target interaction feature sequence to obtain a fused feature sequence; normalizing the fused feature sequence to obtain a normalized feature sequence; extracting features along the attribute dimension from the normalized feature sequence to obtain an attribute feature sequence; and fusing the attribute feature sequence with the fused feature sequence to obtain the target feature sequence.

[0130] There are several ways to fuse candidate sequence features with target interaction feature sequences. For example, a residual network can be used to fuse candidate sequence features with target interaction feature sequences to obtain a fused feature sequence. Alternatively, candidate sequence features can be directly concatenated or added to target interaction feature sequences to obtain a fused feature sequence, and so on.

[0131] After fusing the candidate feature sequence and the target interaction feature sequence, the fused feature sequence can be normalized to obtain a normalized feature sequence. Then, features along the attribute dimension are extracted from the normalized feature sequence to obtain the attribute feature sequence. The process of extracting features along the attribute dimension from the normalized feature sequence can include: extracting at least one feature along the attribute dimension from the normalized feature sequence to obtain the original attribute feature sequence; performing linear activation on the original attribute feature sequence to obtain the activated attribute feature sequence; and scaling the activated attribute feature sequence based on a preset scaling parameter to obtain the target feature sequence.

[0132] The method of extracting the attribute feature sequence in the attribute dimension from the normalized feature sequence is similar to the method of extracting the initial time feature in the time dimension from the interaction feature subsequence, as described above, and will not be repeated here.

[0133] This scheme employs a fully connected network along the feature (attribute) dimension to extract attribute feature sequences from a normalized feature sequence. The network structure of this fully connected network along the feature (attribute) dimension can be as follows: Figure 5 As shown, this fully connected network can include an FC layer, a ReLU activation layer, and a Dropout layer. The FC layer extracts at least one feature in the attribute dimension from the normalized feature sequence to obtain the original attribute feature sequence. Then, the ReLU activation layer linearly activates the original attribute feature sequence to obtain the activated attribute feature sequence. Based on a preset scaling parameter, the Dropout layer scales the activated attribute feature sequence to obtain the target feature sequence.

[0134] 105. Perform temporal dimensionality reduction on the target feature sequence to obtain the interactive behavior features under the target task.

[0135] Temporal dimensionality reduction can be understood as mapping high-dimensional sequence data to a low-dimensional space, thereby obtaining the interactive behavior characteristics under the target task. The target task can be understood as the downstream task that requires the interactive behavior feature sequence. The target task can be of various types, such as classifying or identifying risks in interactive behaviors, or classifying or identifying risks in the interactive objects corresponding to the interactive behaviors, and so on. Taking classification as the type of target task and transaction behavior as an example, the target task could include tasks in the Few-Shot classification scenario, as well as classification tasks in other transaction fraud risk control scenarios, and so on. Taking transaction behavior as an example, the target task could include at least one task in the transaction risk control scenario. Transaction risk control is a crucial link in the commercial transaction process. Its main task is to accurately identify and effectively prevent fraudulent behavior by constructing a feature model of user and merchant transaction behavior. This process involves in-depth analysis of the purchasing behavior and transaction characteristics of consumers and merchants, thereby inferring the behavioral motivations and habit patterns of both parties in the transaction. By leveraging advanced algorithms and technologies, and continuously optimizing and improving feature models, transaction risk control can reveal abnormal trading patterns, thereby promptly identifying and intercepting problematic accounts. This not only helps prevent various fraudulent activities such as fake investments but also significantly enhances the security and stability of the entire trading ecosystem.

[0136] The process of performing temporal dimensionality reduction on the target feature sequence can be summarized as follows:

[0137] For example, the target feature sequence can be transposed to obtain a transposed feature sequence. At least one dimension of features corresponding to the target task can be extracted from the transposed feature sequence to obtain at least one task feature sequence. The task feature sequence can be pooled to obtain the interaction behavior feature sequence under the target task.

[0138] The transpose of the target feature sequence can be similar to that of the current feature sequence, as described above, and will not be repeated here.

[0139] After transposing the target feature sequence, at least one dimension of features corresponding to the target task can be extracted from the transposed feature sequence. The process of extracting at least one dimension of features corresponding to the target task from the transposed feature sequence may include: obtaining task feature dimension information of the target task, which indicates the feature dimensions under the target task; and using a fully connected network to extract at least one dimension of features corresponding to the target task from the transposed feature sequence based on the task feature dimension information, thereby obtaining at least one task feature sequence.

[0140] After extracting at least one dimension of features corresponding to the target task from the transposed feature sequence, pooling operations can be performed on the extracted at least one task feature sequence to obtain the interaction behavior feature sequence under the target task. There are several ways to perform pooling operations on the at least one task feature sequence. For example, max pooling can be performed on the at least one task feature sequence to obtain the interaction behavior feature sequence under the target task; average pooling can be performed on the at least one task feature sequence to obtain the interaction behavior feature sequence under the target task; min pooling can be performed on the at least one task feature sequence to obtain the interaction behavior feature sequence under the target task; or other pooling operations can be used to pool the at least one task feature sequence to obtain the interaction behavior feature sequence under the target task, and so on.

[0141] In the process of temporal dimensionality reduction of the target feature sequence, this scheme first uses a fully connected network to extract at least one task feature sequence under the target task, and then uses pooling operation to fuse the extracted task feature sequences. This not only learns the sequence pattern, but also maps the sequence from the original input length to the embedding layer dimension required by the target task, thereby improving the accuracy of feature extraction.

[0142] Optionally, in some embodiments, after performing temporal dimensionality reduction on the target feature sequence to obtain the interaction behavior features under the target task, the target task can also be completed using the interaction behavior features. For example, taking the target task as risk identification of the interaction behavior, the risk category of the interaction behavior can be determined based on the interaction behavior features. When the risk category indicates that the interaction behavior is risky, the interaction behavior can be intercepted or frozen. As another example, taking the target task as risk identification of the interaction object corresponding to the interaction behavior, the risk category of the interaction object corresponding to the interaction behavior can be determined based on the interaction behavior features. When the risk category indicates that the interaction object is risky, the interaction object can be processed.

[0143] There are various ways to process interactive objects. For example, you can freeze or ban the interactive object's account on the interactive platform, or restrict other interactive behaviors of the interactive object on the interactive platform, or send risk warning information to other objects during the interaction of the interactive object, or freeze the interactive object's resources (virtual or physical resources) on the interactive platform, and so on.

[0144] In this scheme, after performing feature interaction on the feature sequences in the feature sequence set, an incremental fully connected network can be used to fuse the interaction feature sequences, and then extract the interaction behavior feature sequences of the target task from the fused target interaction feature sequences. The network structure of the incremental fully connected network can be as follows: Figure 6 As shown, the interaction feature sequence is normalized and then transposed, with the dimension to be processed placed at the end of the data block dimension. Then, the transposed target interaction feature sequence is input into a time-dimensional fully connected network (FC_time). The n-layer fully connected network in (FC_time) processes the fixed-length word sequences sequentially to extract time-dimensional features, thus obtaining the current feature sequence corresponding to the interaction behavior. The current feature sequence is then transposed to obtain candidate feature sequences in the attribute dimension. These candidate feature sequences are fused with the target interaction feature sequence to obtain a fused feature sequence. This fused feature sequence is then normalized to obtain a normalized feature sequence. A feature (attribute)-dimensional fully connected network (FC_feature) is used to extract attribute-dimensional features from the normalized feature sequence, thus obtaining an attribute feature sequence. Finally, the attribute feature sequence is fused with the fused feature sequence to obtain the target feature sequence. The target feature sequence is transposed to obtain the transposed feature sequence. A fully connected layer in a temporal dimensionality reduction network is used to extract features of at least one dimension corresponding to the target task from the transposed feature sequence, thereby obtaining at least one task feature sequence. A pooling layer in a temporal dimensionality reduction network is used to perform pooling operations on the task feature sequence, thereby obtaining the interaction behavior feature sequence under the target task.

[0145] Optionally, in some embodiments, during the process of extracting the interaction behavior feature sequence of the target task using an incremental fully connected network, a feature (attribute) dimension fully connected network can first be used to extract the attribute dimension features from the current interaction features after feature interaction, thereby obtaining an attribute feature sequence. The attribute feature sequence is then transposed to obtain a transposed attribute feature sequence. A time dimension fully connected network is used to extract the time dimension features from the transposed feature sequence to obtain a time feature sequence. Finally, a time-series dimensionality reduction network is used to perform time-series dimensionality reduction on the time feature sequence to obtain the interaction behavior feature sequence under the target task.

[0146] In this scheme, a feature extraction model can be used to extract interactive behavior features under the target task from at least one interactive data sequence of interactive behaviors. The feature extraction model can include a feature interaction network and an incremental fully connected network. Taking three feature attributes as an example, the feature extraction model can be specifically as follows: Figure 7As shown. The input to the feature extraction model can be a four-dimensional data set (batch_size, sequence_length, dimensionof attributes, dimension of features). First, it can be divided into multiple data blocks (interactive data sequences) of size (batch_size, sequence_length, dimension of features) based on the feature attributes (attr). The initial feature sequences and positional features corresponding to these blocks are accumulated and then input into an interactive convolutional network (Interactive Convolutional Network) to capture local correlations in the temporal dimension (sequence_length), thus obtaining the interactive convolution result, i.e., at least one interactive feature sequence (O). ICN The results of the interactive convolutions are first concatenated into a data format of (batch_size, sequence_length, dimension of concat_features) (i.e., the target interactive feature sequence). It's important to note that due to variations in the number of convolutional kernels and attributes in the convolutional network, the dimension of the concatenated features will differ from the input data. Finally, the concatenated result (the target interactive feature sequence) is input into an incremental fully connected network to obtain the interactive behavior feature sequence required for the target task.

[0147] This solution employs two main approaches. First, it designs an incremental fully connected network, implementing the functionality of a convolutional network while preserving the original data size and avoiding the loss of valuable information. It cleverly utilizes the transpose of sequence data to simultaneously extract attribute and temporal features from behavioral sequences on top of a traditional fully connected network. Second, by introducing an interactive convolutional network, it effectively removes noisy behaviors and redundant information from the data, while accurately preserving the effective interaction information between behaviors. This improvement significantly enhances the feature extraction model's ability to understand and predict complex transaction behaviors, thereby improving the accuracy of feature extraction.

[0148] As can be seen from the above, in this embodiment, after obtaining the target interaction feature sequence of the interaction behavior, a feature subsequence of a preset length is divided from the target interaction feature sequence to obtain at least one interaction feature subsequence. Then, based on at least one preset sequence interaction range, features in the time dimension are extracted from the interaction feature subsequence to obtain the current feature sequence corresponding to the interaction behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension. Then, the current feature sequence is converted into a candidate feature sequence in the attribute dimension, and features in the attribute dimension are extracted from the candidate feature sequence to obtain the target feature sequence. Finally, temporal dimensionality reduction is performed on the target feature sequence to obtain... Interactive behavior characteristics under the target task; because this scheme can divide the target interactive feature sequence into sub-feature sequences of a given length, and extract the temporal dimension features from at least one fixed-length sub-sequence by pre-setting the sequence interaction range, the feature extraction process focuses more on regional relational dependencies, thereby avoiding the problem of effective information loss caused by padding operations at both ends of the sequence, and improving the accuracy of feature extraction in the temporal dimension. In addition, by converting the features in the temporal dimension into features in the attribute dimension, it is possible to simultaneously mine the attribute dimension features and the temporal dimension features of the behavioral sequence, thus improving the accuracy of feature extraction.

[0149] Based on the method described in the above embodiments, the following examples will provide further detailed explanations.

[0150] In this embodiment, the feature extraction device will be specifically integrated into an electronic device, with the electronic device serving as a server and the interaction behavior being a transaction behavior, as an example for explanation.

[0151] like Figure 8 As shown, a feature extraction method is described below, with the specific process as follows:

[0152] 201. The server obtains a set of feature sequences for at least one feature attribute.

[0153] For example, the server can obtain at least one transaction data sequence of a transaction behavior in the content interaction platform, or it can obtain transaction data of at least one transaction object corresponding to the transaction behavior during the interaction process, extract time information from the transaction data, and sort the transaction data based on the time information to obtain at least one transaction data sequence, or it can also obtain at least one transaction data sequence sent by the interaction terminal of the transaction object corresponding to the transaction behavior, and so on.

[0154] The server acquires the attribute information of the transaction data sequence and, based on this information, determines the corresponding feature attributes of the transaction data sequence, thereby obtaining at least one target transaction data sequence corresponding to the feature attributes. The server can use an embedding network to extract feature embeddings from the target transaction data sequence, thus obtaining the initial feature sequence of the target transaction feature sequence. The server also acquires the location information of the transaction behavior and extracts features from this location information to obtain the location features of the transaction behavior.

[0155] The server can accumulate the location features with the features in each initial feature sequence to obtain a set of feature sequences corresponding to the feature attributes. This set of feature sequences can include the feature sequences corresponding to each initial feature sequence. Alternatively, the server can filter out the location sub-features corresponding to each feature in the initial feature sequence from the location features, and concatenate the location sub-features with the corresponding features in the initial feature sequence to obtain a set of feature sequences corresponding to the feature attributes. This set of feature sequences can include the feature sequences corresponding to each initial feature sequence, and so on.

[0156] 202. The server performs feature interaction on the feature sequences in the feature sequence set to obtain at least one transaction feature sequence.

[0157] For example, the server can randomly extract two feature sequences without replacement from the feature sequence set to obtain the current feature sequence pair. Alternatively, it can sort the feature sequences in the feature sequence set and extract the first two feature sequences without replacement based on the sorting result to obtain the current feature sequence pair, and so on.

[0158] The server can use a one-dimensional convolutional neural network with at least one scale of convolutional kernels to extract temporal feature information from the first feature sequence, thereby capturing local features and longer-distance dependencies, and thus obtaining the first candidate transaction feature sequence. Then, it can use convolutional kernels with at least one scale to extract temporal feature information from the second feature sequence, obtaining the second candidate transaction feature sequence.

[0159] The server can use GELU to activate the first candidate transaction feature sequence to obtain the activated first candidate transaction feature sequence, or it can use RELU to activate the first candidate transaction feature sequence to obtain the activated first candidate transaction feature sequence, or it can use other activation functions to activate the first candidate transaction feature sequence to obtain the activated first candidate transaction feature sequence, and so on.

[0160] The server will perform a multiplication operation between the first candidate transaction feature sequence and the second candidate transaction feature after activation to obtain the first transaction feature sequence, as shown in formula (1).

[0161] The server can activate the second candidate transaction feature sequence to obtain the activated second candidate transaction feature sequence. The activated second candidate transaction feature sequence is then multiplied by the first candidate transaction feature sequence to obtain the second transaction feature sequence, as shown in formula (2).

[0162] The server can add the features of the first transaction feature sequence and the second transaction feature sequence to obtain the concatenated transaction feature sequence. A one-dimensional convolutional network is used to extract features from the concatenated transaction feature sequence to obtain the target concatenated transaction feature sequence. ReLU is used to activate the target concatenated transaction feature sequence to obtain the initial transaction feature sequence corresponding to the current feature sequence, as shown in formula (3).

[0163] The server can return to the step of filtering the current feature sequence pair from the feature sequence set until all feature sequences in the feature sequence set are feature sequences in the current feature sequence pair, thus obtaining at least one transaction feature sequence.

[0164] 203. Merge the transaction feature sequences to obtain the target transaction feature sequence of the transaction behavior.

[0165] For example, the server can use a BN layer (normalization layer) to normalize the transaction feature sequence to obtain a normalized transaction feature sequence.

[0166] The server can concatenate the normalized transaction feature sequences to obtain the fused transaction feature sequence, or it can accumulate the normalized transaction feature sequences to obtain the fused transaction feature sequence, and so on.

[0167] The server can obtain the dimensional order of the merged transaction feature sequence, which can include the order between the time dimension and the attribute dimension. Based on this dimensional order, the merged transaction feature sequence is transposed to obtain the target transaction feature sequence, where the last dimension of the target transaction feature sequence is the time dimension.

[0168] 204. The server divides the target transaction feature sequence into feature subsequences of a preset length to obtain at least one transaction feature subsequence.

[0169] For example, the server can identify at least one segmentation position corresponding to a preset length in the target transaction feature sequence, and divide at least one transaction feature subsequence in the target transaction feature sequence based on the segmentation position.

[0170] 205. Based on at least one preset sequence interaction range, the server extracts features in the time dimension from the transaction feature subsequence to obtain the current feature sequence corresponding to the transaction behavior.

[0171] For example, the server can use a fully connected network (FCN) to extract at least one feature in the time dimension from the transaction feature subsequence to obtain an initial time feature sequence. Alternatively, it can use other feature extraction networks to extract at least one feature in the time dimension from the transaction feature subsequence to obtain an initial time feature sequence, and so on.

[0172] The server can use the ReLU activation function to linearly activate the original time feature sequence to obtain the activated time feature sequence. Alternatively, it can use other activation functions to linearly activate the original time feature sequence to obtain the activated time feature sequence, and so on.

[0173] The server multiplies the preset scaling parameters with the feature values ​​of the features in the activated time feature sequence to obtain the initial time feature sequence.

[0174] The server can sort the preset sequence interaction ranges and, based on the sorting results, filter out the smallest preset sequence interaction range from the sorted preset sequence interaction ranges to obtain the target sequence interaction range corresponding to the initial time feature sequence. Alternatively, it can obtain the sequence length of the initial time feature sequence and filter out the preset sequence interaction ranges with sequence lengths adjacent to that sequence length to obtain the target sequence interaction range corresponding to the initial time feature sequence. Or, it can obtain the target sequence length corresponding to each preset sequence interaction range and, based on the target sequence length, filter out the preset sequence interaction ranges with the sequence length of the preset initial time feature sequence to obtain the target sequence interaction range corresponding to the initial time feature sequence, and so on.

[0175] The server can obtain the position information of the transaction feature subsequence in the target transaction feature sequence, and sort the initial time feature sequence based on the position information. Based on the sequence length of the initial time feature sequence and the sorting result, at least one feature sequence group is constructed. The sum of the sequence lengths of the initial time feature sequences in the feature sequence group is the sequence length corresponding to the target sequence interaction range. The initial time feature sequences in the feature sequence group are concatenated to obtain at least one concatenated time feature sequence corresponding to the target interaction range.

[0176] The server extracts the time dimension features from the concatenated feature sequence to obtain the current time feature sequence. It then uses the current time feature sequence as the initial time feature sequence and returns to execute the step of filtering the target sequence interaction range corresponding to the initial time feature sequence within the preset sequence interaction range until the sequence length of the current time feature sequence is the same as the sequence length of the target transaction feature, thus obtaining the current feature sequence corresponding to the transaction behavior.

[0177] 206. The server converts the current feature sequence into a candidate feature sequence in the attribute dimension, and extracts the features in the attribute dimension from the candidate feature sequence to obtain the target feature sequence.

[0178] For example, the server can obtain the dimensional order in the current feature sequence, which may include the order between the time dimension and the attribute dimension. Based on the dimensional order, the current feature sequence can be transposed to obtain the candidate feature sequence in the attribute dimension.

[0179] The server can use a residual network to fuse the candidate sequence features with the target transaction feature sequence to obtain the fused feature sequence. Alternatively, it can directly concatenate or add the candidate sequence features with the target transaction feature sequence to obtain the fused feature sequence, and so on.

[0180] The server can normalize the fused feature sequence to obtain a normalized feature sequence. At least one feature in the attribute dimension can be extracted from the normalized feature sequence to obtain the original attribute feature sequence. The original attribute feature sequence can be linearly activated to obtain the activated attribute feature sequence. Based on a preset scaling parameter, the activated attribute feature sequence can be scaled to obtain the target feature sequence.

[0181] 207. The server performs temporal dimensionality reduction on the target feature sequence to obtain the transaction behavior features under the target task.

[0182] For example, the server can transpose the target feature sequence to obtain a transposed feature sequence. It then obtains the task feature dimension information of the target task, which indicates the feature dimensions under the target task. Based on this task feature dimension information, a fully connected network is used to extract at least one dimension of features corresponding to the target task from the transposed feature sequence, thereby obtaining at least one task feature sequence.

[0183] The server can perform max pooling on at least one task feature sequence to obtain a transaction behavior feature sequence under the target task; or, it can perform average pooling on at least one task feature sequence to obtain a transaction behavior feature sequence under the target task; or, it can perform min pooling on at least one task feature sequence to obtain a transaction behavior feature sequence under the target task; or, it can use other pooling operations to pool at least one task feature sequence to obtain a transaction behavior feature sequence under the target task, and so on.

[0184] Optionally, in some embodiments, taking the target task of identifying risks in transaction behavior as an example, the server can determine the risk category of the transaction behavior based on the characteristics of the transaction behavior, and when the risk category indicates that the transaction behavior is risky, the transaction behavior can be intercepted or frozen; for example, taking the target task of identifying risks in the transaction object corresponding to the transaction behavior as an example, the risk category of the transaction object corresponding to the transaction behavior can be determined based on the characteristics of the transaction behavior, and when the risk category indicates that the transaction object is risky, the transaction object can be processed.

[0185] Optionally, in some embodiments, during the process of extracting the transaction behavior feature sequence of the target task using an incremental fully connected network, the server may first use a feature (attribute) dimension fully connected network to extract features in the attribute dimension from the current transaction features after feature interaction, thereby obtaining an attribute feature sequence. The attribute feature sequence is then transposed to obtain a transposed attribute feature sequence. A time dimension fully connected network is then used to extract features in the time dimension from the transposed feature sequence to obtain a time feature sequence. Finally, a time-series dimensionality reduction network is used to perform time-series dimensionality reduction on the time feature sequence to obtain the transaction behavior feature sequence under the target task.

[0186] As can be seen from the above, in this embodiment of the application, after the server obtains the target transaction feature sequence of the transaction behavior, it divides the target transaction feature sequence into feature subsequences of a preset length to obtain at least one transaction feature subsequence. Then, based on at least one preset sequence interaction range, it extracts features in the time dimension from the transaction feature subsequence to obtain the current feature sequence corresponding to the transaction behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension. Then, it converts the current feature sequence into a candidate feature sequence in the attribute dimension and extracts features in the attribute dimension from the candidate feature sequence to obtain the target feature sequence. Finally, it performs temporal dimensionality reduction on the target feature sequence to obtain... The scheme identifies the characteristics of transaction behavior under the target task. Because it can divide the target transaction feature sequence into sub-feature sequences of a given length and extract time-dimensional features from at least one fixed-length sub-sequence through a preset sequence interaction range, the feature extraction process focuses more on regional relational dependencies, thus avoiding the loss of effective information caused by padding operations at both ends of the sequence. This improves the accuracy of time-dimensional feature extraction. Furthermore, by converting time-dimensional features into attribute-dimensional features, it simultaneously mines both attribute-dimensional and time-dimensional features of the behavioral sequence, thereby enhancing the accuracy of feature extraction.

[0187] To better implement the above methods, this application also provides a feature extraction device that can be integrated into an electronic device, such as a server or terminal. The terminal may include a tablet computer, a laptop computer, and / or a personal computer.

[0188] For example, such as Figure 9 As shown, the feature extraction device may include an acquisition unit 301, a segmentation unit 302, an extraction unit 303, a transformation unit 304, and a dimensionality reduction unit 305, as follows:

[0189] (1) Obtain unit 301;

[0190] The acquisition unit 301 is used to acquire the target interaction feature sequence of the interaction behavior, which includes the feature sequence of the interaction object corresponding to the interaction behavior in the interaction process.

[0191] For example, the acquisition unit 301 can be specifically used to acquire a feature sequence set of at least one feature attribute, filter out the current feature sequence pair in the feature sequence set, the current feature sequence pair includes two feature sequences that need to be interacted with, perform feature interaction on the feature sequences in the current feature sequence pair to obtain the initial interaction feature sequence corresponding to the current feature sequence pair, return to execute the step of filtering out the current feature sequence pair in the feature sequence set until all feature sequences in the feature sequence set are feature sequences in the current feature sequence pair, obtain at least one interaction feature sequence, and fuse the interaction feature sequences to obtain the target interaction feature sequence of the interaction behavior.

[0192] (2) Divide the unit into 302;

[0193] The segmentation unit 302 is used to segment a feature subsequence of a preset length in the target interactive feature sequence to obtain at least one interactive feature subsequence.

[0194] For example, the segmentation unit 302 can be used to identify at least one segmentation position corresponding to a preset length in the target interactive feature sequence, and based on the segmentation position, to divide at least one interactive feature subsequence in the target interactive feature sequence.

[0195] (3) Extraction unit 303;

[0196] Extraction unit 303 is used to extract features in the time dimension from the interaction feature subsequence based on at least one preset sequence interaction range, so as to obtain the current feature sequence corresponding to the interaction behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension.

[0197] For example, extraction unit 303 can be used to extract features in the time dimension from the interaction feature subsequence to obtain an initial time feature sequence, filter out the target sequence interaction range corresponding to the initial time feature sequence from the preset sequence interaction range, concatenate the initial time feature sequence based on the target sequence interaction range to obtain at least one concatenated time feature sequence corresponding to the target interaction range, extract features in the time dimension from the concatenated feature sequence to obtain the current time feature sequence, and use the current time feature sequence as the initial time feature sequence, return to execute the step of filtering out the target sequence interaction range corresponding to the initial time feature sequence from the preset sequence interaction range, until the sequence length of the current time feature sequence is the same as the sequence length of the target interaction feature, and obtain the current feature sequence corresponding to the interaction behavior.

[0198] (4) Conversion unit 304;

[0199] The transformation unit 304 is used to transform the current feature sequence into a candidate feature sequence in the attribute dimension, and extract the features in the attribute dimension from the candidate feature sequence to obtain the target feature sequence.

[0200] For example, the transformation unit 304 can be used to convert the current feature sequence into a candidate feature sequence in the attribute dimension, fuse the candidate sequence features with the target interaction feature sequence to obtain a fused feature sequence, normalize the fused feature sequence to obtain a normalized feature sequence, extract the attribute dimension features from the normalized feature sequence to obtain an attribute feature sequence, and fuse the attribute feature sequence with the fused feature sequence to obtain the target feature sequence.

[0201] (5) Dimensionality reduction unit 305;

[0202] Dimensionality reduction unit 305 is used to perform temporal dimensionality reduction on the target feature sequence to obtain the interactive behavior features of the interactive behavior under the target task.

[0203] For example, the dimensionality reduction unit 305 can be used to transpose the target feature sequence to obtain the transposed feature sequence, extract at least one dimension of features corresponding to the target task from the transposed feature sequence to obtain at least one task feature sequence, and perform pooling operation on the task feature sequence to obtain the interaction behavior feature sequence under the target task.

[0204] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0205] As can be seen from the above, in this embodiment of the application, after the acquisition unit 301 acquires the target interaction feature sequence of the interaction behavior, the segmentation unit 302 divides the target interaction feature sequence into feature subsequences of a preset length to obtain at least one interaction feature subsequence. Then, the extraction unit 303 extracts features in the time dimension from the interaction feature subsequence based on at least one preset sequence interaction range to obtain the current feature sequence corresponding to the interaction behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension. Then, the conversion unit 304 converts the current feature sequence into a candidate feature sequence in the attribute dimension and extracts features in the attribute dimension from the candidate feature sequence to obtain the target feature sequence. Finally, the dimensionality reduction unit 305... Temporal dimensionality reduction is performed on the target feature sequence to obtain the interactive behavior features under the target task. Since this scheme can divide the target interactive feature sequence into sub-feature sequences of a given length and extract the temporal dimension features from at least one fixed-length sub-sequence by pre-setting the sequence interaction range, the feature extraction process focuses more on regional relational dependencies, thereby avoiding the problem of effective information loss caused by padding operations at both ends of the sequence and improving the accuracy of feature extraction in the temporal dimension. In addition, by converting the features in the temporal dimension into features in the attribute dimension, the feature extraction of the behavioral sequence in both the attribute dimension and the temporal dimension can be achieved simultaneously, thus improving the accuracy of feature extraction.

[0206] This application also provides an electronic device, such as... Figure 10 As shown, it illustrates a structural schematic diagram of the electronic device involved in the embodiments of this application, specifically:

[0207] The electronic device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will understand that... Figure 10 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0208] The processor 401 is the control center of the electronic device, connecting various parts of the device via various interfaces and lines. It executes software programs and / or modules stored in the memory 402, and calls data stored in the memory 402, to perform various functions and process data. Optionally, the processor 401 may include one or more processing cores; preferably, the processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 401.

[0209] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 402 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

[0210] The electronic device also includes a power supply 403 that supplies power to the various components. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 403 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0211] The electronic device may also include an input unit 404, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0212] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 401 in the electronic device loads the executable files corresponding to the processes of one or more applications into the memory 402 according to the following instructions, and the processor 401 runs the applications stored in the memory 402 to realize various functions, as follows:

[0213] A target interaction feature sequence of interactive behavior is obtained, which includes the feature sequence of the interactive object corresponding to the interactive behavior during the interaction process. A feature subsequence of a preset length is divided in the target interaction feature sequence to obtain at least one interaction feature subsequence. Based on at least one preset sequence interaction range, features in the time dimension are extracted from the interaction feature subsequence to obtain the current feature sequence corresponding to the interactive behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension. The current feature sequence is converted into a candidate feature sequence in the attribute dimension, and features in the attribute dimension are extracted from the candidate feature sequence to obtain the target feature sequence. Temporal dimensionality reduction is performed on the target feature sequence to obtain the interaction behavior features of the interactive behavior under the target task.

[0214] For example, an electronic device can acquire a set of feature sequences of at least one characteristic attribute, filter out a current feature sequence pair from the feature sequence set, the current feature sequence pair includes two feature sequences that need to interact, perform feature interaction on the feature sequences in the current feature sequence pair to obtain an initial interactive feature sequence corresponding to the current feature sequence pair, return to the step of filtering out the current feature sequence pair from the feature sequence set, until all feature sequences in the feature sequence set are feature sequences of the current feature sequence pair, obtaining at least one interactive feature sequence, and fuse the interactive feature sequences to obtain a target interactive feature sequence for the interactive behavior. At least one segmentation position corresponding to a preset length is identified in the target interactive feature sequence, and based on the segmentation position, at least one interactive feature subsequence is divided in the target interactive feature sequence. The process involves extracting time-dimensional features from the interaction feature subsequence to obtain an initial time feature sequence. Then, a target sequence interaction range corresponding to the initial time feature sequence is selected from a preset sequence interaction range. Based on the target sequence interaction range, the initial time feature sequence is concatenated to obtain at least one concatenated time feature sequence corresponding to the target interaction range. Time-dimensional features are extracted from the concatenated feature sequence to obtain the current time feature sequence. This current time feature sequence is then used as the initial time feature sequence. The process of selecting the target sequence interaction range corresponding to the initial time feature sequence from the preset sequence interaction range is repeated until the sequence length of the current time feature sequence is the same as the sequence length of the target interaction feature sequence, thus obtaining the current feature sequence corresponding to the interaction behavior. The current feature sequence is then converted into a candidate feature sequence in the attribute dimension. The candidate sequence features are fused with the target interaction feature sequence to obtain a fused feature sequence. The fused feature sequence is then normalized to obtain a normalized feature sequence. Attribute-dimensional features are extracted from the normalized feature sequence to obtain an attribute feature sequence. The attribute feature sequence is then fused with the fused feature sequence to obtain the target feature sequence. The target feature sequence is transposed to obtain the transposed feature sequence. At least one dimension of features corresponding to the target task is extracted from the transposed feature sequence to obtain at least one task feature sequence. The task feature sequence is pooled to obtain the interaction behavior feature sequence under the target task.

[0215] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0216] As can be seen from the above, in this embodiment, after obtaining the target interaction feature sequence of the interaction behavior, a feature subsequence of a preset length is divided from the target interaction feature sequence to obtain at least one interaction feature subsequence. Then, based on at least one preset sequence interaction range, features in the time dimension are extracted from the interaction feature subsequence to obtain the current feature sequence corresponding to the interaction behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension. Then, the current feature sequence is converted into a candidate feature sequence in the attribute dimension, and features in the attribute dimension are extracted from the candidate feature sequence to obtain the target feature sequence. Finally, temporal dimensionality reduction is performed on the target feature sequence to obtain... Interactive behavior characteristics under the target task; because this scheme can divide the target interactive feature sequence into sub-feature sequences of a given length, and extract the temporal dimension features from at least one fixed-length sub-sequence by pre-setting the sequence interaction range, the feature extraction process focuses more on regional relational dependencies, thereby avoiding the problem of effective information loss caused by padding operations at both ends of the sequence, and improving the accuracy of feature extraction in the temporal dimension. In addition, by converting the features in the temporal dimension into features in the attribute dimension, it is possible to simultaneously mine the attribute dimension features and the temporal dimension features of the behavioral sequence, thus improving the accuracy of feature extraction.

[0217] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0218] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the feature extraction methods provided in embodiments of this application. For example, the instructions can execute the following steps:

[0219] A target interaction feature sequence of interactive behavior is obtained, which includes the feature sequence of the interactive object corresponding to the interactive behavior during the interaction process. A feature subsequence of a preset length is divided in the target interaction feature sequence to obtain at least one interaction feature subsequence. Based on at least one preset sequence interaction range, features in the time dimension are extracted from the interaction feature subsequence to obtain the current feature sequence corresponding to the interactive behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension. The current feature sequence is converted into a candidate feature sequence in the attribute dimension, and features in the attribute dimension are extracted from the candidate feature sequence to obtain the target feature sequence. Temporal dimensionality reduction is performed on the target feature sequence to obtain the interaction behavior features of the interactive behavior under the target task.

[0220] For example, a feature sequence set of at least one feature attribute is obtained. The current feature sequence pair, comprising two feature sequences that need to interact, is selected from this set. The feature sequences in the current feature sequence pair are then interacted to obtain the initial interactive feature sequence corresponding to the current feature sequence pair. The process of selecting the current feature sequence pair from the feature sequence set is repeated until all feature sequences in the set are from the current feature sequence pair, resulting in at least one interactive feature sequence. These interactive feature sequences are then fused to obtain the target interactive feature sequence for the interactive behavior. At least one segmentation position corresponding to a preset length is identified in the target interactive feature sequence. Based on the segmentation position, at least one interactive feature subsequence is divided in the target interactive feature sequence. The process involves extracting time-dimensional features from the interaction feature subsequence to obtain an initial time feature sequence. Then, a target sequence interaction range corresponding to the initial time feature sequence is selected from a preset sequence interaction range. Based on the target sequence interaction range, the initial time feature sequence is concatenated to obtain at least one concatenated time feature sequence corresponding to the target interaction range. Time-dimensional features are extracted from the concatenated feature sequence to obtain the current time feature sequence. This current time feature sequence is then used as the initial time feature sequence. The process of selecting the target sequence interaction range corresponding to the initial time feature sequence from the preset sequence interaction range is repeated until the sequence length of the current time feature sequence is the same as the sequence length of the target interaction feature sequence, thus obtaining the current feature sequence corresponding to the interaction behavior. The current feature sequence is then converted into a candidate feature sequence in the attribute dimension. The candidate sequence features are fused with the target interaction feature sequence to obtain a fused feature sequence. The fused feature sequence is then normalized to obtain a normalized feature sequence. Attribute-dimensional features are extracted from the normalized feature sequence to obtain an attribute feature sequence. The attribute feature sequence is then fused with the fused feature sequence to obtain the target feature sequence. The target feature sequence is transposed to obtain the transposed feature sequence. At least one dimension of features corresponding to the target task is extracted from the transposed feature sequence to obtain at least one task feature sequence. The task feature sequence is pooled to obtain the interaction behavior feature sequence under the target task.

[0221] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0222] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0223] Since the instructions stored in the computer-readable storage medium can execute the steps of any feature extraction method provided in the embodiments of this application, the beneficial effects that any feature extraction method provided in the embodiments of this application can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0224] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the methods provided in various alternative implementations of the aforementioned feature extraction aspect or interactive behavior feature extraction (modeling).

[0225] The foregoing has provided a detailed description of a feature extraction method and related equipment provided in the embodiments of this application. The related equipment may include feature extraction devices, electronic devices, computer program products, and computer-readable storage media. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A feature extraction method, characterized in that, include: Obtain the target interaction feature sequence of the interaction behavior, wherein the target interaction feature sequence includes the feature sequence of the interaction object corresponding to the interaction behavior during the interaction process; Divide the target interactive feature sequence into feature subsequences of a preset length to obtain at least one interactive feature subsequence; Based on at least one preset sequence interaction range, features in the time dimension are extracted from the interaction feature subsequence to obtain the current feature sequence corresponding to the interaction behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension. The current feature sequence is converted into a candidate feature sequence in the attribute dimension, and the features in the attribute dimension are extracted from the candidate feature sequence to obtain the target feature sequence; Temporal dimensionality reduction is performed on the target feature sequence to obtain the interaction behavior features of the interaction behavior under the target task.

2. The feature extraction method according to claim 1, characterized in that, The step of extracting time-dimensional features from the interaction feature subsequence based on at least one preset sequence interaction range to obtain the current feature sequence corresponding to the interaction behavior includes: The features in the time dimension are extracted from the interactive feature subsequence to obtain the initial time feature sequence; Filter out the target sequence interaction range corresponding to the initial time feature sequence from the preset sequence interaction range; Based on the target sequence interaction range, the initial time feature sequence is concatenated, and features are extracted from the concatenated time feature sequence to obtain the current feature sequence corresponding to the interaction behavior.

3. The feature extraction method according to claim 2, characterized in that, The step of extracting features along the time dimension from the interactive feature subsequence to obtain an initial time feature sequence includes: At least one feature in the time dimension is extracted from the interactive feature subsequence to obtain the original time feature sequence; The original time feature sequence is linearly activated to obtain the activated time feature sequence; Based on a preset scaling parameter, the activated time feature sequence is scaled to obtain an initial time feature sequence. The preset scaling parameter indicates the probability that a feature in the activated time feature sequence is set to zero.

4. The feature extraction method according to claim 2, characterized in that, The step of concatenating the initial time feature sequence based on the target sequence interaction range includes: Obtain the position information of the interaction feature subsequence in the target interaction feature sequence, and sort the initial time feature sequence based on the position information; Based on the sequence length and sorting result of the initial time feature sequence, at least one feature sequence group is constructed, wherein the sum of the sequence lengths of the initial time feature sequences in the feature sequence group is the sequence length corresponding to the target sequence interaction range; The initial time feature sequences in the feature sequence group are concatenated to obtain at least one concatenated time feature sequence corresponding to the target interaction range.

5. The feature extraction method according to claim 2, characterized in that, The step of extracting features from the concatenated time feature sequence to obtain the current feature sequence corresponding to the interaction behavior includes: The features in the time dimension are extracted from the spliced ​​time feature sequence to obtain the current time feature sequence, and the current time feature sequence is used as the initial time feature sequence. Return to the step of filtering out the target sequence interaction range corresponding to the initial time feature sequence in the preset sequence interaction range, until the sequence length of the current time feature sequence is the same as the sequence length of the target interaction feature sequence, and obtain the current feature sequence corresponding to the interaction behavior.

6. The feature extraction method according to any one of claims 1 to 5, characterized in that, The step of converting the current feature sequence into a candidate feature sequence along the attribute dimension includes: Obtain the dimensional order in the current feature sequence, where the dimensional order includes the order between the time dimension and the attribute dimension; Based on the dimensional order, the current feature sequence is transposed to obtain the candidate feature sequence on the attribute dimension.

7. The feature extraction method according to any one of claims 1 to 5, characterized in that, The step of extracting features along the attribute dimension from the candidate feature sequence to obtain the target feature sequence includes: The candidate feature sequence is fused with the target interaction feature sequence to obtain the fused feature sequence; The fused feature sequence is normalized to obtain a normalized feature sequence, and features on the attribute dimension are extracted from the normalized feature sequence to obtain an attribute feature sequence. The attribute feature sequence is fused with the fused feature sequence to obtain the target feature sequence.

8. The feature extraction method according to any one of claims 1 to 5, characterized in that, The step of performing temporal dimensionality reduction on the target feature sequence to obtain the interaction behavior features of the interaction behavior under the target task includes: The target feature sequence is transposed to obtain the transposed feature sequence; Extract at least one dimension of features corresponding to the target task from the transposed feature sequence to obtain at least one task feature sequence; The task feature sequence is pooled to obtain the interaction behavior feature sequence under the target task.

9. The feature extraction method according to any one of claims 1 to 5, characterized in that, The target interaction feature sequence for obtaining interactive behavior includes: Obtain a feature sequence set of at least one feature attribute, the feature sequence set including at least one feature sequence corresponding to the interaction behavior; The feature sequences in the feature sequence set are subjected to feature interaction to obtain at least one interactive feature sequence; The interaction feature sequences are fused to obtain the target interaction feature sequence of the interaction behavior.

10. The feature extraction method according to claim 9, characterized in that, The step of obtaining a feature sequence set with at least one feature attribute includes: At least one interaction data sequence of the interaction behavior is obtained, and the interaction data sequence is classified according to the feature attribute to obtain at least one target interaction data sequence corresponding to the feature attribute. Feature extraction is performed on the target interactive data sequence to obtain at least one initial feature sequence corresponding to the feature attribute; The interaction location information of the interaction behavior is obtained, and the interaction location information is used to extract features to obtain the location features of the interaction behavior. The initial feature sequence is fused with the location feature to obtain the feature sequence set corresponding to the feature attribute.

11. The feature extraction method according to claim 9, characterized in that, The step of performing feature interaction on the feature sequences in the feature sequence set to obtain at least one interactive feature sequence includes: The current feature sequence pair is selected from the feature sequence set, and the current feature sequence pair includes the two feature sequences that need to interact with each other. The feature sequences in the current feature sequence pair are subjected to feature interaction to obtain the initial interaction feature sequence corresponding to the current feature sequence pair; Return to the step of filtering the current feature sequence pair in the feature sequence set until all feature sequences in the feature sequence set are feature sequences in the current feature sequence pair, and obtain at least one interactive feature sequence.

12. The feature extraction method according to claim 11, characterized in that, The feature sequence pair includes a first feature sequence and a second feature sequence. The step of performing feature interaction on the feature sequences in the current feature sequence pair to obtain the initial interaction feature sequence corresponding to the current feature sequence pair includes: Based on the first feature sequence, feature extraction is performed on the second feature sequence to obtain the first interactive feature sequence; Based on the second feature sequence, feature extraction is performed on the first feature sequence to obtain the second interactive feature sequence; The first interaction feature sequence and the second interaction feature sequence are fused to obtain the initial interaction feature sequence corresponding to the current feature sequence pair.

13. The feature extraction method according to claim 12, characterized in that, The step of extracting features from the second feature sequence based on the first feature sequence to obtain the first interactive feature sequence includes: The first candidate interaction feature sequence is obtained by extracting the temporal dimension feature information from the first feature sequence using a convolutional kernel of at least one scale. The feature information in the time dimension is extracted from the second feature sequence using a convolutional kernel of at least one scale to obtain a second candidate interaction feature sequence; The first candidate interaction feature sequence is activated, and the activated first candidate interaction feature sequence is fused with the second candidate interaction feature sequence to obtain the first interaction feature sequence.

14. The feature extraction method according to claim 13, characterized in that, The step of extracting features from the first feature sequence based on the second feature sequence to obtain the second interactive feature sequence includes: The second candidate interaction feature sequence is activated to obtain the activated second candidate interaction feature sequence. The activated second candidate interaction feature sequence is fused with the first candidate interaction feature to obtain the second interaction feature sequence.

15. The feature extraction method according to claim 12, characterized in that, The step of fusing the first interaction feature sequence and the second interaction feature sequence to obtain the initial interaction feature sequence corresponding to the current feature sequence pair includes: The first interaction feature sequence and the second interaction feature sequence are concatenated to obtain the concatenated interaction feature sequence. Feature extraction is performed on the spliced ​​interaction feature sequence to obtain the target spliced ​​interaction feature sequence; The target concatenated interactive feature sequence is activated to obtain the initial interactive feature sequence corresponding to the current feature sequence pair.

16. The feature extraction method according to claim 9, characterized in that, The step of fusing the interaction feature sequences to obtain the target interaction feature sequence of the interaction behavior includes: The interaction feature sequence is normalized to obtain a normalized interaction feature sequence. The normalized interaction feature sequences are fused to obtain the fused interaction feature sequence. The fused interaction feature sequence is transposed to obtain the target interaction feature sequence of the interaction behavior.

17. A feature extraction device, characterized in that, include: An acquisition unit is used to acquire a target interaction feature sequence of an interaction behavior, wherein the target interaction feature sequence includes the feature sequence of the interaction object corresponding to the interaction behavior during the interaction process. A partitioning unit is used to divide the target interactive feature sequence into feature subsequences of a preset length to obtain at least one interactive feature subsequence. An extraction unit is configured to extract features in the time dimension from the interaction feature subsequence based on at least one preset sequence interaction range, so as to obtain the current feature sequence corresponding to the interaction behavior. The preset sequence interaction range indicates the sequence length of the feature sequence used to extract features in the time dimension. A transformation unit is used to convert the current feature sequence into a candidate feature sequence in the attribute dimension, and extract the features in the attribute dimension from the candidate feature sequence to obtain the target feature sequence; The dimensionality reduction unit is used to perform temporal dimensionality reduction on the target feature sequence to obtain the interaction behavior features of the interaction behavior under the target task.

18. An electronic device, characterized in that, It includes a processor and a memory, the memory storing an application program, and the processor running the application program within the memory to perform the steps of the feature extraction method according to any one of claims 1 to 16.

19. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the feature extraction method according to any one of claims 1 to 16.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the steps of the feature extraction method according to any one of claims 1 to 16.

Images