Feature derivation method and device, equipment, storage medium and program product
By feature encoding and fusion of transaction basic features, transaction sequence basic features and account basic features, the problem of difficult correlation in multimodal scenarios is solved, more effective derivative features are generated, and the effectiveness of features is improved.
Patent Information
- Application Number
- CN202411092218.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-12-03
AI Technical Summary
Existing feature derivation methods are difficult to capture the correlation between different modes in multimodal scenarios, resulting in the inability to generate effective derivative features.
By obtaining the user's basic trading characteristics, basic trading sequence characteristics and basic account characteristics, tabular feature encoding and transaction sequence time feature encoding, and using low-rank fusion method, the tabular feature representation and transaction sequence feature representation are fused. Generate derivative features.
Effectively capture the correlation between user attributes and behavioral characteristics, generate more effective derivative features, and improve the effectiveness of features.
Smart Images

Figure CN119067776A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of artificial intelligence technology, and in particular to a feature derivation method, device, equipment, storage medium and program product. Background Art
[0002] In the process of data analysis and modeling, the accuracy and robustness of the model can be greatly improved by creating new features or variables from the original data through the reasonable application of feature transformation, feature combination, etc., that is, deriving new feature information from the existing basic features can significantly improve the performance of the model. This model is also often used in the trading process to work efficiently.
[0003] At present, the commonly used feature derivation method is to perform feature modeling on single modal data through neural networks such as multi-layer perceptrons and deep convolutional networks. If this feature derivation method is applied to multimodal scenarios, it is only to encode and splice different modal data based on neural networks. However, the encoding and splicing method makes it difficult to capture the correlation between different modalities, resulting in the inability to generate effective derived features. Summary of the invention
[0004] The main purpose of this application is to provide an invention name, which aims to solve the technical problem that it is difficult to capture the correlation between different modes by using coding splicing, resulting in the inability to generate effective derivative features.
[0005] To achieve the above purpose, the present application proposes a feature derivation method, which comprises:
[0006] Obtaining the basic transaction characteristics and transaction sequence basic characteristics of the user within a set time range, as well as the basic account characteristics of the user;
[0007] Performing table-type feature encoding on the account basic feature and the transaction basic feature to obtain a table-type feature representation, and performing transaction sequence time feature encoding on the transaction sequence basic feature to obtain a transaction sequence type feature representation;
[0008] The table-type feature representation and the transaction sequence-type feature representation are low-rank fused to obtain derived features.
[0009] In one embodiment, the step of performing low-rank fusion on the table-type feature representation and the transaction sequence-type feature representation to obtain derived features includes:
[0010] Based on a preset low-rank weight tensor, feature fusion is performed on the table-type feature representation and the transaction sequence-type feature representation to obtain derived features.
[0011] In one embodiment, before the step of fusing the table-type feature representation and the transaction sequence-type feature representation based on the preset low-rank weight tensor to obtain the derived features, the step includes:
[0012] Perform low-rank decomposition on the preset high-dimensional weight tensor to obtain the preset low-rank weight tensor.
[0013] In one embodiment, the step of encoding the transaction sequence basic features by transaction sequence time features to obtain a transaction sequence type feature representation comprises:
[0014] Performing dimensionless processing on the characteristic values in the basic characteristics of the transaction sequence, and performing multi-channel transaction time characteristic encoding on the timestamps corresponding to each node in the basic characteristics of the transaction sequence to obtain a time characteristic sequence;
[0015] Based on a preset self-attention mechanism, the transaction sequence time feature is encoded on the time feature sequence to obtain a transaction sequence feature representation.
[0016] In one embodiment, the step of performing dimensionless processing on the feature values in the basic features of the transaction sequence, and performing multi-channel transaction time feature encoding on the timestamps corresponding to each node in the basic features of the transaction sequence, and obtaining the time feature sequence includes:
[0017] Performing dimensionless processing on the characteristic values in the basic characteristics of the transaction sequence to obtain dimensionless characteristics;
[0018] Identify the timestamp corresponding to each node from the basic features of the transaction sequence;
[0019] Performing time feature encoding on the timestamp to obtain time features of the time dimension;
[0020] The time feature is concatenated with the dimensionless feature to obtain a time feature sequence.
[0021] In one embodiment, the step of performing tabular feature encoding on the account basic feature and the transaction basic feature to obtain a tabular feature representation includes:
[0022] The account basic features are combined with the transaction basic features to obtain table basic features;
[0023] The table basic features are encoded into a table-type feature based on a preset multi-layer perceptron to obtain a table-type feature representation.
[0024] In addition, to achieve the above purpose, the present application also proposes a feature derivation device, the feature derivation device comprising:
[0025] The acquisition module is used to obtain the basic characteristics of the user's account, basic characteristics of transactions, and basic characteristics of transaction sequences within a set time range;
[0026] An encoding module, used for performing table-type feature encoding on the account basic feature and the transaction basic feature to obtain a table-type feature representation, and performing transaction sequence time feature encoding on the transaction sequence basic feature to obtain a transaction sequence type feature representation;
[0027] A fusion module is used to perform low-rank fusion on the table-type feature representation and the transaction sequence-type feature representation to obtain derived features.
[0028] In addition, to achieve the above objectives, the present application also proposes a feature derivation device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the feature derivation method described above.
[0029] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the feature derivation method described above are implemented.
[0030] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps of the feature derivation method described above are implemented.
[0031] One or more technical solutions proposed in this application have at least the following technical effects:
[0032] The present application obtains the basic transaction features of the user within the set time range to obtain the information of each transaction of the user within the set time range, and then obtains the basic transaction sequence features of the user within the set time range to obtain the information features of continuous transactions within the set time range, and obtains the account basic features that represent the user information; by performing tabular feature encoding on the account basic features and the transaction basic features, the account basic features and the transaction basic features can be converted into a form acceptable to the neural network layer, which is conducive to the convergence of the neural network; by performing transaction sequence time feature encoding on the transaction sequence basic features, the behavioral characteristics of the transaction sequence basic features in the time dimension can be determined; it can be understood that the tabular features can reflect the attributes of different users corresponding to a certain time, and the transaction sequence basic features can reflect the user behavior characteristics; therefore, by performing low-rank fusion on the tabular feature representation and the transaction sequence feature representation, the correlation between the user's attributes and the user's behavior characteristics can be learned, thereby generating effective derived features; compared with the simple splicing of different modal data in the related art, the present application adopts the above technical solution to improve the effectiveness of the derived features. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.
[0034] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.
[0035] Figure 1 A schematic diagram of a process flow provided for Example 1 of the feature derivation method of the present application;
[0036] Figure 2 A schematic diagram of a flow chart provided for Embodiment 2 of the feature derivation method of the present application;
[0037] Figure 3 A schematic diagram of the first scenario provided for Embodiment 2 of the feature derivation method of the present application;
[0038] Figure 4 A second scenario schematic diagram provided for Embodiment 2 of the feature derivation method of the present application;
[0039] Figure 5 Schematic diagram of the process provided for the third embodiment of the feature derivation method of this application
[0040] Figure 6 A brief flowchart of the feature derivation method provided in Example 2 of the present application;
[0041] Figure 7 This is a schematic diagram of the module structure of the characteristic derivative device of the embodiment of the present application;
[0042] Figure 8 Schematic diagram of the device structure of the hardware operating environment involved in the feature derivation method in the embodiment of the present application.
[0043] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0044] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.
[0045] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.
[0046] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, a feature derivative device, etc. The following takes the feature derivative device as an example to illustrate this embodiment and the following embodiments.
[0047] Based on this, the present application embodiment provides a feature derivation method, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the feature derivation method of the present application.
[0048] In this embodiment, the feature derivation method includes steps S10 to S30:
[0049] Step S10, obtaining the basic transaction characteristics and transaction sequence basic characteristics of the user within a set time range, as well as the basic account characteristics of the user;
[0050] It should be noted that, since the analysis of the user's account transactions requires determining the time period that the user wants to analyze so as to perform the analysis according to the user's requirements, the time range can be the time period of the transactions that the user wants to analyze, wherein the time range can be the most recent time period or a certain time period in history. For example, the time range can be the most recent month, or it can be from May to July last year, etc., which is specifically set by the user according to needs.
[0051] It should be noted that the basic transaction feature may be a feature corresponding to an account, and the basic transaction feature may be obtained by manually calculating the transaction information of the account within a period of time based on prior knowledge, that is, the basic transaction feature may be obtained by preprocessing the transaction information of the account within a period of time, wherein the transaction information may include account transaction data, account balance and transaction time. The basic transaction sequence feature may be a feature obtained based on a series of continuous transaction information, and the feature may include not only the feature of each transaction information, but also the feature of the arrangement order of each feature. The basic account feature corresponds to the feature of the account, and includes basic information such as the user's gender, age, account level, etc.
[0052] In addition, it should be noted that the account level allows the bank to divide the customer's account into different levels according to factors such as the customer's financial assets, transaction frequency, consumption behavior, etc. in the bank, so as to provide differentiated services and rights.
[0053] It is understandable that in order to help banks better serve customers and promote refined management of banking business, banks generally classify user accounts into account levels. Therefore, in order to obtain the characteristics of account transactions more accurately, the account level needs to be included in the basic account characteristics.
[0054] It can be understood that, since the account's transaction basic characteristics and transaction sequence basic characteristics contain the user's consumption information within a set time range, and the account basic characteristics contain the user's basic identity information and the basic information recorded in the bank for the account, by obtaining the account's transaction basic characteristics and transaction sequence basic characteristics, as well as the account basic characteristics, more abundant features related to the account transactions can be obtained, thereby increasing the amount of information of the original features used to derive features, allowing feature derivation to have more derivative schemes, and improving the correlation between the final derived features and the account transactions.
[0055] Step S20, performing table-type feature encoding on the account basic feature and the transaction basic feature to obtain a table-type feature representation, and performing transaction sequence time feature encoding on the transaction sequence basic feature to obtain a transaction sequence feature representation;
[0056] It should be noted that tabular feature encoding is a technique that converts input features into a form acceptable to the neural network layer through operations such as normalization (for example, through normalization operations, numerical data is converted into a value between 0 and 1, or into data that conforms to a normal distribution, etc.), which is conducive to the convergence of the neural network; transaction sequence time feature encoding is a technique that uses multiple kernel functions to perform time feature mapping function encoding on the time in the basic features of the transaction sequence, and splices it with other features in the basic features of the transaction sequence.
[0057] It can be understood that since the account basic features and transaction basic features contain numerical features of multiple different units, such as transaction amount and transaction time, the account basic features and transaction basic features can be converted into a form acceptable to the neural network layer by performing tabular feature encoding on the account basic features and the transaction basic features, that is, a tabular feature representation is obtained, which is conducive to the convergence of the neural network, and can increase the richness of the original features required for the derived features, and increase the stability of the correlation between the features, so as to broaden the feature derivation scheme.
[0058] In a specific implementation, the tabular feature encoding of the account basic features and the transaction basic features can be performed in parallel with the transaction sequence time feature encoding of the transaction sequence basic features. That is, when performing the tabular feature encoding, the transaction sequence time feature encoding can be performed as required. The two encodings can be synchronously encoded as required, or can be parallel asynchronously encoded to save the time for obtaining the tabular feature representation and the transaction sequence feature representation required for the low-rank fusion feature.
[0059] Step S30, performing low-rank fusion on the table-type feature representation and the transaction sequence-type feature representation to obtain derived features.
[0060] It should be noted that low-rank fusion is to decompose high-dimensional tensors into low-rank factors to avoid the problem of high computational complexity caused by "dimensionality explosion" in other tensor methods. Specifically, it can use the parallel decomposition of modality-specific low-rank weight tensors and input tensors to effectively calculate tensor-based fusion, thereby reducing the time complexity from exponential to linear, reducing the number of model parameters, and reducing the model computational complexity.
[0061] It can be understood that since the tabular features can reflect the attributes of different users corresponding to a certain time, the basic features of the transaction sequence can reflect the user behavior characteristics. By performing low-rank fusion of the tabular feature representation and the transaction sequence feature representation, the correlation between the user's attributes and the user's behavior characteristics can be learned, that is, the tabular feature representation and the transaction sequence feature representation can interact deeply, thereby deriving new features.
[0062] The present embodiment provides a feature derivation method, which obtains the basic transaction features of the user within the set time range to obtain the information of each transaction of the user within the set time range, and then obtains the basic transaction sequence features of the user within the set time range to obtain the information features of continuous transactions within the set time range, and obtains the account basic features that characterize the user information; by performing tabular feature encoding on the account basic features and the transaction basic features, the account basic features and the transaction basic features can be converted into a form acceptable to the neural network layer, which is conducive to the convergence of the neural network; by performing transaction sequence time feature encoding on the transaction sequence basic features, the behavioral characteristics of the transaction sequence basic features in the time dimension can be determined; it can be understood that the tabular feature representation can reflect the attributes of different users corresponding to a certain time, and the transaction sequence basic features can reflect the user behavior characteristics; therefore, by performing low-rank fusion on the tabular feature representation and the transaction sequence feature representation, the correlation between the user's attributes and the user's behavior characteristics can be learned, thereby generating effective derived features; compared with the simple splicing of different modal data in the related art, the above method adopted in the present embodiment can improve the effectiveness of the derived features.
[0063] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can be referred to the above introduction, and will not be repeated in the following. Figure 2 , step S30 further includes step S31:
[0064] Step S31, based on a preset low-rank weight tensor, feature fusion is performed on the table-type feature representation and the transaction sequence-type feature representation to obtain derived features.
[0065] It should be noted that the preset low-rank weight tensor is a weight tensor with lower dimension than the high-dimensional weight tensor.
[0066] Reference Figure 3 In order to fuse the tabular feature representation with the transaction sequence feature representation, it is necessary to perform outer products on each dimension of the tabular feature representation and each dimension of the transaction sequence feature representation to obtain a matrix Z. When it is necessary to further perform deep fusion on the matrix Z, it is necessary to multiply the matrix Z with a preset high-dimensional weight tensor to obtain a derived feature. It can be understood that the computational complexity of this process is very large.
[0067] In order to solve the problem of large amount of calculation mentioned above, the tabular feature representation and the transaction sequence feature representation are subjected to feature fusion based on the preset low-rank weight tensor. Before the step of obtaining the derived features, the preset high-dimensional weight tensor can be subjected to low-rank decomposition to obtain the preset low-rank weight tensor.
[0068] Specifically, Figure 4 As shown, the preset high-dimensional weight tensor can be decomposed into multiple low-dimensional weight tensors, that is, decomposed into multiple preset low-rank weight tensors (W v *W l ); wherein the number of the preset low-rank weight tensors can be determined based on the hyperparameter r set when decomposing the preset high-dimensional weight tensor.
[0069] After obtaining the preset low-rank weight tensor, each dimension of the tabular feature representation is multiplied by the preset low-rank weight tensor to obtain a first calculation result, and then each dimension of the transaction sequence feature representation is multiplied by the preset low-rank weight tensor to obtain a second calculation result. Finally, the first calculation result and the second calculation result are outer-producted to complete the deep fusion of the tabular feature representation and the transaction sequence feature representation, thereby obtaining derived features. Compared with feature fusion based on high-dimensional weight tensors, this embodiment uses the preset low-rank weight tensor to fuse the tabular feature representation and the transaction sequence feature representation, which has a smaller amount of calculation and improves calculation efficiency.
[0070] In this embodiment, a preset high-dimensional weight tensor is subjected to low-rank decomposition to obtain a preset low-rank weight tensor. Based on the preset low-rank weight tensor, feature fusion is performed on the table-type feature representation and the transaction sequence-type feature representation to obtain derived features, thereby improving computational efficiency.
[0071] Based on the first and second embodiments of the present application, in the third embodiment of the present application, the same or similar contents as those in the above embodiments can be referred to the above introduction, and will not be described in detail later. Figure 5 , step S20 also includes steps S01-S02:
[0072] Step S01, performing dimensionless processing on the feature values in the basic features of the transaction sequence, and performing multi-channel transaction time feature encoding on the timestamps corresponding to each node in the basic features of the transaction sequence to obtain a time feature sequence;
[0073] Step S02: encoding the transaction sequence time features of the time feature sequence based on a preset self-attention mechanism to obtain a transaction sequence feature representation.
[0074] It should be noted that the basic features of the transaction sequence are composed of multiple tuples, each of which corresponds to a timestamp t i For example, the basic features of a transaction sequence can be expressed as S = {s1, s2, ..., s L}, the basic feature of the transaction sequence is a sequence of length L, and the tuple can be represented as s i (t i , a i , bi ), each tuple is a node in the sequence S; where a i , b i are all feature values in the basic features of the transaction sequence, for example, a i , b i are the transaction amount and the account balance after the transaction, etc. i The timestamp corresponding to each node.
[0075] Specifically, dimensionless processing can be achieved by eliminating the dimensional effects between different features or indicators through data transformation to make the data comparable. Since the dimensions of different feature values may be different, the feature values in the basic features of the transaction sequence can be dimensionlessly processed to unify the dimensions of the feature values, so that the neural network can converge better.
[0076] Multi-channel transaction time feature encoding can be achieved by mapping the timestamp corresponding to the dimensionless feature into a vector based on multiple channels when processing transaction data, so that the mapped vector can express the information represented by the transaction sequence feature in the time dimension; it can be understood that different channels correspond to different kernels, and different kernels have different kernel functions. By mapping the timestamps corresponding to the dimensionless features through the different kernel functions corresponding to each channel, the distribution of the basic features of the transaction sequence in the time dimension can be better learned.
[0077] Based on this, the feature values in the basic features of the transaction sequence are dimensionally processed, and the timestamps corresponding to each node in the basic features of the transaction sequence are encoded with multi-channel transaction time features. The specific implementation method for obtaining the time feature sequence may be: dimensionally processing the feature values in the basic features of the transaction sequence to obtain dimensionless features; identifying the timestamps corresponding to each node from the basic features of the transaction sequence; encoding the timestamps with time features to obtain time features of the time dimension; and concatenating the time features with the dimensionless features to obtain a time feature sequence.
[0078] It should be noted that the timestamp can be the time node of each transaction.
[0079] It is understandable that since transaction time is an important feature of transaction information, in order to add behavioral features of time dimensions such as transaction intervals and frequencies in the feature derivation process to solve the problem that it is difficult to capture behavioral features such as periodicity and time-varying in transaction sequences in common technologies, it is possible to perform multi-channel transaction time feature encoding on the timestamps corresponding to each node in the basic features of the transaction sequence, and concatenate the time features with the dimensionless features to obtain a time feature sequence.
[0080] Specifically, the timestamp corresponding to each node in the basic features of the transaction sequence can be calculated based on the kernel function corresponding to each channel to obtain the weight coefficient of each node, and the basic features of the transaction sequence can be aggregated based on the weight coefficient of each node to obtain the time series features.
[0081] Among them, when calculating the timestamp corresponding to each node in the basic features of the transaction sequence and obtaining the weight coefficient of each node, it is necessary to assume that the change of each weight coefficient in the time dimension obeys a certain prior distribution (such as normal distribution, Poisson distribution, etc.) or the change in the time dimension conforms to a certain change law (for example, linear change, exponential change, etc.); since the distribution of the transaction sequence feature representation in the time dimension is uncertain, if the mapped vector is to be able to express the information of the transaction sequence feature representation in the time dimension, multiple kernel functions can be used to combine multiple change methods (for example, changes that conform to the normal distribution or Poisson distribution, linear changes, exponential changes, etc.), and the change law of the transaction sequence feature representation in the time dimension can be learned through a neural network, so that the obtained time series features can more comprehensively describe the distribution of the transaction sequence feature representation in the time dimension.
[0082] Furthermore, in order to help the model better understand and process the time feature sequence, the time feature sequence can be encoded into transaction sequence time features based on a preset self-attention mechanism to capture the dependencies between tuples in the time feature sequence and obtain a transaction sequence feature representation; wherein the preset self-attention mechanism can be a sequence model architecture that can calculate the attention score of each value in the sequence by specifying a query and a key and sum the sequence values with the score as the weight, thereby modeling the sequence.
[0083] In summary, the specific implementation process of obtaining the transaction sequence feature representation in this embodiment can be: dimensionless processing is performed on the feature values in the basic features of the transaction sequence to obtain dimensionless features, that is, S is converted into a dimensionless feature f S , for each node s in the sequence S i , we can use Layer Norm (Layer Normalization, a technique used to normalize layer inputs during neural network training) to normalize the feature values (a i 、b i ) are dimensionless processed to obtain a i ', b i ', then each node in the sequence can be expressed as (t i , a i ', b i'); and perform multi-channel transaction time feature encoding on the timestamps corresponding to each node in the basic features of the transaction sequence to obtain a time feature sequence, that is, for each node s in the sequence S i , using the time feature mapping function to convert the timestamp t i Mapped into a vector β i ; Further, the time feature is concatenated with the dimensionless feature to obtain a time feature sequence, and each node in the time feature sequence can be expressed as (β i , a i ', b i ').
[0084] Finally, based on the preset self-attention mechanism, the time feature sequence is encoded into the transaction sequence time feature, and the feature fusion is performed through the convolutional neural network to obtain the transaction sequence feature representation. The specific implementation method can be: based on the preset self-attention mechanism, each node s in the time feature sequence is encoded into the transaction sequence time feature. i are all related to other nodes {s1, s2, ..., s i} to perform feature fusion and obtain the vector h corresponding to each node i , and finally the matrix [h1, h2, h3, ..., h L ] to obtain the transaction sequence feature representation f S .
[0085] Optionally, step S20 further includes steps S1 to S2:
[0086] Step S1, concatenating the account basic features with the transaction basic features to obtain table basic features;
[0087] Step S2, performing table-type feature encoding on the table basic features based on a preset multi-layer perceptron to obtain a table-type feature representation.
[0088] It should be noted that concatenating the account basic features with the transaction basic features is to concatenate the x-dimensional account basic features with the y-dimensional transaction basic features to obtain the (x+y)-dimensional table basic features.
[0089] It can be understood that since splicing the account basic features with the transaction basic features can more completely reflect the account's transaction records, the account basic features can be spliced with the transaction basic features. After obtaining the table basic features, in order to normalize the table basic features, the table basic features can be encoded into table-type features to obtain a table-type feature representation.
[0090] For example, to help understand the implementation process of the feature derivation method obtained by combining this embodiment with the above embodiments, please refer to Figure 6 , Figure 6 A brief flowchart of a feature derivation method is provided, specifically:
[0091] 1) Account transaction data preprocessing: Given an account u and a time window T at time t0, obtain the account u in [t 0-T , t0] time range and intercept or fill it to a fixed length L, and obtain the account transaction sequence type basic feature S0 = {s1, s2, ..., s L}, and obtain the basic account characteristics h of account u at the same time u And the trading basis characteristics formulated by experts u , h u With g u The basic features f of the account transaction table are obtained by splicing t ;
[0092] 2) Table feature encoding: For the basic features of the account transaction table type f t First, the input features are normalized through the BatchNorm layer, and then the processed features are input into the multi-layer perceptron for further encoding to obtain the tabular feature representation f t ';
[0093] 3) Trading sequence temporal feature encoding: Setting multiple kernel functions K j , perform Fourier transform on each kernel function and extract the real part to discretize and construct the time feature mapping function Φ j , for s i (t i , a i , b i )∈S0, the timestamp t i By splicing the Φ encoding results, s is obtained i The temporal feature encoding ti;
[0094] 4) Sequence feature encoding based on attention mechanism: First, LayerNorm is used to encode the a in the transaction sequence. i 、b i The processed time feature sequences are then input into a multi-layer masked self-attention layer, and the output of the last layer is encoded through a convolutional neural network to obtain a transaction sequence feature representation f s ;
[0095] 5) Multimodal feature derivation based on low-rank fusion: First, one-dimensional unit features are added to the tabular feature representation and the transaction sequence feature representation as two modal features and input into the low-rank fusion layer for feature fusion. Then, the fused features are input into the fully connected layer of the specified size to obtain the final derived feature f.
[0096] Among them, Batch Norm is an algorithm widely used in deep learning. The Batch Norm layer mainly standardizes the data of each batch to speed up the training speed of the neural network and improve the generalization ability of the model. Therefore, the Batch Norm layer can be used to normalize a certain feature of all samples in a batch. Layer Norm is an algorithm used to standardize the activation of neural network layers in deep learning. It aims to stabilize the gradient flow during training and accelerate model convergence. Therefore, the Batch Norm layer can be used to normalize a part of the features within a sample.
[0097] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the method of deriving the features of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.
[0098] This application also provides a feature derivative device, please refer to Figure 7 , the feature derivation device comprises:
[0099] The acquisition module 10 is used to acquire the basic account characteristics, basic transaction characteristics and basic transaction sequence characteristics of the user within a set time range;
[0100] The encoding module 20 is used to perform table-type feature encoding on the account basic feature and the transaction basic feature to obtain a table-type feature representation, and to perform transaction sequence time feature encoding on the transaction sequence basic feature to obtain a transaction sequence feature representation;
[0101] The fusion module 30 is used to perform low-rank fusion on the table-type feature representation and the transaction sequence-type feature representation to obtain derived features.
[0102] Optionally, the fusion module 30 is further used to perform feature fusion on the table-type feature representation and the transaction sequence-type feature representation based on a preset low-rank weight tensor to obtain derived features.
[0103] Optionally, before the step of fusing the table-type feature representation and the transaction sequence-type feature representation based on a preset low-rank weight tensor to obtain derived features, the feature derivation device further includes:
[0104] The decomposition module is used to perform low-rank decomposition on a preset high-dimensional weight tensor to obtain the preset low-rank weight tensor.
[0105] Optionally, the encoding module 20 is used to perform dimensionless processing on the feature values in the basic features of the transaction sequence, and to perform multi-channel transaction time feature encoding on the timestamps corresponding to each node in the basic features of the transaction sequence to obtain a time feature sequence; and to perform transaction sequence time feature encoding on the time feature sequence based on a preset self-attention mechanism to obtain a transaction sequence type feature representation.
[0106] Optionally, the encoding module 20 is also used to perform dimensionless processing on the feature values in the basic features of the transaction sequence to obtain dimensionless features; identify the timestamps corresponding to each node from the basic features of the transaction sequence; perform time feature encoding on the timestamps to obtain time features of the time dimension; and concatenate the time features with the dimensionless features to obtain a time feature sequence.
[0107] Optionally, the encoding module 20 is further used to concatenate the account basic features with the transaction basic features to obtain table basic features; and perform table-type feature encoding on the table basic features based on a preset multi-layer perceptron to obtain a table-type feature representation.
[0108] The feature derivation device provided by the present application adopts the feature derivation method in the above embodiment, which can solve the technical problem that effective derived features cannot be generated. Compared with the prior art, the beneficial effects of the feature derivation device provided by the present application are the same as the beneficial effects of the feature derivation method provided by the above embodiment, and other technical features in the feature derivation device are the same as the features disclosed in the above embodiment method, which will not be described in detail here.
[0109] The present application provides a feature derivation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the feature derivation method in the above-mentioned embodiment one.
[0110] Reference below Figure 8, which shows a schematic diagram of the structure of a feature-derived device suitable for implementing the embodiment of the present application. The feature-derived device in the embodiment of the present application may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 8 The characteristic derivative device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.
[0111] like Figure 8 As shown, the feature derivative device may include a processing device 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the feature derivative device are also stored. The processing device 1001, ROM1002, and RAM1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the feature derivative device to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows a feature derivative device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or have alternatively.
[0112] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.
[0113] The feature derivation device provided by the present application adopts the feature derivation method in the above embodiment, which can solve the technical problem of being unable to generate effective derived features. Compared with the prior art, the beneficial effects of the feature derivation device provided by the present application are the same as the beneficial effects of the feature derivation method provided by the above embodiment, and the other technical features in the feature derivation device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.
[0114] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.
[0115] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
[0116] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer program) stored thereon, wherein the computer-readable program instructions are used to execute the feature derivation method in the above-mentioned embodiment.
[0117] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.
[0118] The computer-readable storage medium may be included in the feature-derivative device; or may exist independently without being assembled into the feature-derivative device.
[0119] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the feature derivation device, the feature derivation device: obtains the user's transaction basic features and transaction sequence basic features within a set time range, as well as the user's account basic features; performs tabular feature encoding on the account basic features and the transaction basic features to obtain a tabular feature representation, and performs transaction sequence time feature encoding on the transaction sequence basic features to obtain a transaction sequence feature representation; performs low-rank fusion on the tabular feature representation and the transaction sequence feature representation to obtain a derived feature.
[0120] Computer program code for performing the operations of the present application may be written in one or more programming languages or a combination thereof, including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0121] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.
[0122] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.
[0123] The readable storage medium provided in the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned feature derivation method, and can solve the technical problem that effective derived features cannot be generated. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in the present application are the same as the beneficial effects of the feature derivation method provided in the above-mentioned embodiment, and will not be repeated here.
[0124] The present application also provides a computer program product, including a computer program, which implements the steps of the feature derivation method as described above when executed by a processor.
[0125] The computer program product provided by the present application can solve the technical problem of being unable to generate effective derived features. Compared with the prior art, the beneficial effects of the computer program product provided by the present application are the same as the beneficial effects of the feature derivation method provided by the above embodiment, which will not be repeated here.
[0126] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.
Claims
1. A feature derivation method, characterized in that: The method includes: Obtaining the basic transaction characteristics and transaction sequence basic characteristics of the user within a set time range, as well as the basic account characteristics of the user; Performing table-type feature encoding on the account basic feature and the transaction basic feature to obtain a table-type feature representation, and performing transaction sequence time feature encoding on the transaction sequence basic feature to obtain a transaction sequence type feature representation; The table-type feature representation and the transaction sequence-type feature representation are low-rank fused to obtain derived features.
2. The method according to claim 1, characterized in that The step of performing low-rank fusion on the table-type feature representation and the transaction sequence-type feature representation to obtain derived features comprises: Based on a preset low-rank weight tensor, feature fusion is performed on the table-type feature representation and the transaction sequence-type feature representation to obtain derived features.
3. The method according to claim 2, characterized in that Before the step of fusing the table-type feature representation and the transaction sequence-type feature representation based on the preset low-rank weight tensor to obtain the derived features, the step includes: Perform low-rank decomposition on the preset high-dimensional weight tensor to obtain the preset low-rank weight tensor.
4. The method according to any one of claims 1 to 3, characterized in that The step of encoding the transaction sequence time feature of the transaction sequence basic feature to obtain the transaction sequence type feature representation comprises: Performing dimensionless processing on the characteristic values in the basic characteristics of the transaction sequence, and performing multi-channel transaction time characteristic encoding on the timestamps corresponding to each node in the basic characteristics of the transaction sequence to obtain a time characteristic sequence; Based on a preset self-attention mechanism, the transaction sequence time feature is encoded on the time feature sequence to obtain a transaction sequence feature representation.
5. The method according to claim 4, characterized in that The step of performing dimensionless processing on the characteristic values in the basic characteristics of the transaction sequence, and performing multi-channel transaction time characteristic encoding on the timestamps corresponding to each node in the basic characteristics of the transaction sequence to obtain the time characteristic sequence comprises: Performing dimensionless processing on the characteristic values in the basic characteristics of the transaction sequence to obtain dimensionless characteristics; Identify the timestamp corresponding to each node from the basic features of the transaction sequence; Performing time feature encoding on the timestamp to obtain time features of the time dimension; The time feature is concatenated with the dimensionless feature to obtain a time feature sequence.
6. The method according to claim 1, characterized in that The step of performing tabular feature encoding on the account basic features and the transaction basic features to obtain tabular feature representation comprises: The account basic features are combined with the transaction basic features to obtain table basic features; The table basic features are encoded into a table-type feature based on a preset multi-layer perceptron to obtain a table-type feature representation.
7. A feature derivation device, characterized in that: The device comprises: The acquisition module is used to obtain the basic characteristics of the user's account, basic characteristics of transactions, and basic characteristics of transaction sequences within a set time range; An encoding module, used for performing table-type feature encoding on the account basic feature and the transaction basic feature to obtain a table-type feature representation, and performing transaction sequence time feature encoding on the transaction sequence basic feature to obtain a transaction sequence type feature representation; A fusion module is used to perform low-rank fusion on the table-type feature representation and the transaction sequence-type feature representation to obtain derived features.
8. A feature derivation device, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the feature derivation method according to any one of claims 1 to 6.
9. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the feature derivation method according to any one of claims 1 to 6 are implemented.
10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the feature derivation method according to any one of claims 1 to 6 are implemented.
