Payment label intelligent processing method and system based on multi-dimensional user portraits

Abstract

The invention provides a payment label intelligent processing method and system based on multi-dimensional user portraits, and relates to the technical field of big data analysis, a multi-dimensional user portrait matrix is constructed, space-time self-attention combined extraction is performed on user payment behaviors, combined space-time features are obtained, and a candidate payment label set is obtained through screening; performing semantic matching and rearrangement on the candidate payment tag set through semantic relevancy mapping and a comparative learning mechanism to obtain a semantic tag sequence, and determining a feature contribution degree according to the semantic tag sequence and the combined spatial-temporal features; and performing prediction error feedback on the payment tag of the user based on the feature contribution degree and the historical payment record to obtain a semantic verification result, and performing risk identification and behavior recommendation on the payment tag of the user to obtain a payment decision tag for adaptive recommendation. According to the invention, payment label self-learning and semantic dynamic optimization based on multi-source behavior characteristics can be realized, so that the payment label identification accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data analysis, more specifically, the present application relates to a payment label intelligent processing method and system based on multi-dimensional user portrait. BACKGROUND

[0002] With the rapid development of Internet technology and mobile payment, users generate a large amount of payment behavior data in different terminals and different scenarios; big data analysis technology, as an important basis for supporting intelligent decision-making and precise service, can collect, clean, model and extract features from multi-source heterogeneous data, and then mine user behavior patterns and potential intentions.

[0003] In the payment field, big data analysis usually combines machine learning and artificial intelligence algorithms to identify payment risks, predict payment trends and make personalized recommendations, thereby improving the security and user experience of the payment system. The intelligent processing of payment labels based on multi-dimensional user portraits can analyze multi-dimensional features such as user historical payment data, terminal information, geographic environment and behavior preferences, generate corresponding payment labels to support risk control and behavior recommendation; however, existing payment label intelligent processing technology mostly constructs user portraits in a single dimension, fails to fully integrate spatio-temporal features and multi-source behavior data, resulting in incomplete user portrait expression, and relies on static classification algorithms, lacks semantic relevance modeling and context awareness capabilities, cannot accurately depict the semantic features of user payment behavior, and lacks data interaction for risk identification and behavior recommendation, which can easily cause identification delay and low accuracy. Therefore, how to realize payment label self-learning and semantic dynamic optimization based on multi-source behavior features to improve the accuracy of payment label identification is a difficult problem in the industry. SUMMARY

[0004] The present application provides a payment label intelligent processing method and system based on multi-dimensional user portrait, which can realize payment label self-learning and semantic dynamic optimization based on multi-source behavior features to improve the accuracy of payment label identification.

[0005] In a first aspect, the present application provides a payment label intelligent processing method based on multi-dimensional user portrait, which comprises the following steps: constructing a multi-dimensional user portrait matrix based on multi-source behavior features; performing spatio-temporal self-attention joint extraction on user payment behavior according to the multi-dimensional user portrait matrix to obtain joint spatio-temporal features, and then performing double-channel screening on the joint spatio-temporal features to obtain a candidate payment label set; performing semantic matching and rearrangement on the candidate payment label set through a semantic relevance mapping and contrast learning mechanism to obtain a semantic label sequence, and then determining the feature contribution degree according to the semantic label sequence and the joint spatio-temporal features; obtain a historical payment record of the user, perform prediction error feedback on a payment label of the user based on the feature contribution degree and the historical payment record, and obtain a semantic verification result; perform risk identification and behavior recommendation on the payment label of the user according to the semantic verification result, and obtain a payment decision label for adaptive recommendation.

[0006] In this embodiment, constructing a multi-dimensional user portrait matrix based on multi-source behavior features specifically includes: aligning and normalizing timestamps of the multi-source behavior features to obtain a behavior feature set; performing feature hierarchical coding on the behavior feature set to obtain a correlation degree of each feature layer; constructing a multi-dimensional user portrait matrix based on all correlation degrees and the behavior feature set.

[0007] In this embodiment, performing spatio-temporal self-attention joint extraction on user payment behavior according to the multi-dimensional user portrait matrix to obtain joint spatio-temporal features specifically includes: determining a dependency weight of the user payment behavior; performing time attention dependency modeling and space attention dependency modeling on the multi-dimensional user portrait matrix according to the dependency weight to obtain a time attention feature and a space attention feature of the user payment behavior; performing feature fusion on the time attention feature and the space attention feature to obtain joint spatio-temporal features.

[0008] In this embodiment, performing double-channel screening on the joint spatio-temporal features to obtain a candidate payment label set specifically includes: constructing a semantic feature mapping channel based on a semantic embedding network, and constructing a rule constraint channel based on a pre-set rule set; inputting the joint spatio-temporal features into the semantic feature mapping channel, mapping the joint spatio-temporal features to a label semantic space through the semantic embedding network, and then screening to obtain a candidate label set; performing logical constraint screening on the candidate label set through a rule matching matrix in the rule constraint channel to obtain a candidate payment label set.

[0009] In this embodiment, performing semantic matching and rearrangement on the candidate payment label set through a semantic correlation degree mapping and contrast learning mechanism to obtain a semantic label sequence specifically includes: performing semantic embedding on the candidate payment label set to obtain a semantic embedding matrix; determining a semantic correlation degree between the semantic embedding matrix and a label semantic space, and then determining a correlation degree ranking result through the semantic correlation degree; The relevance ranking result is compared and learned to obtain a semantic label sequence.

[0010] In this embodiment, the feature contribution degree is determined according to the semantic label sequence and the joint space-time feature, specifically comprising: Based on the correlation, the feature mapping weight of the semantic label sequence and the joint space-time feature is calculated, and then a contribution degree matrix is obtained. The weight distribution vector is determined through the contribution degree matrix, and then the weighted response coefficient of each semantic label in the semantic label sequence and the joint space-time feature is determined according to the weight distribution vector. All weighted response coefficients are aggregated according to time and space dimensions to obtain a feature contribution degree.

[0011] In this embodiment, the historical payment records of the user are obtained through the payment data management database.

[0012] In this embodiment, the payment label of the user is predicted based on the feature contribution degree and the historical payment record to obtain a semantic verification result, specifically comprising: A label entity set is constructed through the payment label of the user. The label entity set is predicted and paired based on the feature contribution degree and the historical payment record to obtain a prediction deviation value. The prediction pairing process is updated through the prediction deviation value to obtain a semantic verification vector. The semantic verification result is determined according to the semantic verification vector and a preset confidence threshold.

[0013] In this embodiment, the multi-source behavior feature refers to a multi-dimensional data set composed of payment behavior attributes of the user under different time, scene and device conditions, and the multi-source behavior feature includes payment features, terminal features and environmental features.

[0014] In a second aspect, the application provides a payment label intelligent processing system based on multi-dimensional user portrait, which is used to execute a payment label intelligent processing method based on multi-dimensional user portrait, and the intelligent processing system comprises: A user portrait construction module is used to construct a multi-dimensional user portrait matrix based on multi-source behavior features. A space-time feature extraction module is used to extract user payment behavior through space-time self-attention joint according to the multi-dimensional user portrait matrix to obtain joint space-time features, and then the joint space-time features are filtered through a double channel to obtain a candidate payment label set. a semantic rearrangement module configured to perform semantic matching rearrangement on the candidate payment label set through a semantic correlation mapping and a contrast learning mechanism to obtain a semantic label sequence, and further determine a feature contribution degree according to the semantic label sequence and the joint spatiotemporal feature; a feedback verification module configured to obtain historical payment records of a user, perform prediction error feedback on a payment label of the user based on the feature contribution degree and the historical payment records, and further obtain a semantic verification result; a risk identification and recommendation module configured to perform risk identification and behavior recommendation on the payment label of the user according to the semantic verification result, and obtain a payment decision label for adaptive recommendation.

[0015] The technical scheme provided by the embodiments disclosed in the present application has the following beneficial effects: A multi-dimensional user portrait matrix is constructed based on multi-source behavior characteristics, spatiotemporal self-attention joint extraction is performed on a user payment behavior according to the multi-dimensional user portrait matrix to obtain a joint spatiotemporal feature, the joint spatiotemporal feature is double-channel filtered to obtain a candidate payment label set, semantic matching rearrangement is performed on the candidate payment label set through a semantic correlation mapping and a contrast learning mechanism to obtain a semantic label sequence, a feature contribution degree is determined according to the semantic label sequence and the joint spatiotemporal feature, historical payment records of a user are obtained, prediction error feedback is performed on a payment label of the user based on the feature contribution degree and the historical payment records, and a semantic verification result is further obtained, risk identification and behavior recommendation are performed on the payment label of the user according to the semantic verification result, and a payment decision label for adaptive recommendation is obtained.

[0016] It can be seen that in the present application, payment label self-learning and semantic dynamic optimization based on multi-source behavior characteristics can be realized. First, by constructing a multi-dimensional user portrait matrix, the payment behavior characteristics of the user under different time, scene and device conditions can be uniformly modeled, so that the user portrait has a high-dimensional and computable behavior semantic expression, which is beneficial to form a complete spatio-temporal description of individual behavior characteristics and provide a unified data input basis for label recognition. Second, according to the multi-dimensional user portrait matrix, the spatio-temporal self-attention of the user payment behavior is extracted, which can capture the dynamic evolution law of the user payment behavior from two dimensions of time sequence and spatial distribution, and through double-channel screening of the joint spatio-temporal characteristics, the semantic feature mapping and rule constraint screening are combined, so as to realize the accurate compression of the label candidate space, reduce invalid label interference, and then provide a high confidence input set for the semantic matching stage. Then, through the semantic correlation mapping and contrast learning mechanism, the candidate payment label set is matched and rearranged, which can realize the dynamic optimization of the label sequence based on the correlation distribution in the semantic vector space. The contrast learning mechanism strengthens the semantic recognition ability through the feature alignment of positive and negative samples, and then calculates the feature contribution degree according to the semantic label sequence and the joint spatio-temporal characteristics, which is beneficial to improve the interpretability and controllability of semantic matching. Then, the self-supervised closed loop of semantic verification is established through the historical payment records and the feature contribution degree, which can correct the prediction results, and is beneficial to the continuous optimization and stable convergence of the actual payment scene, so as to reduce the recognition bias. Finally, through the generation of adaptive payment decision labels, the recommendation strategy can be dynamically adjusted according to the user behavior trend and risk mode, and then the closed-loop control mechanism from semantic recognition to intelligent recommendation is realized, so as to improve the accuracy of payment label recognition.

[0017] In summary, the technical scheme adopted by the present application can realize payment label self-learning and semantic dynamic optimization based on multi-source behavior characteristics to improve the accuracy of payment label recognition. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only the embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 is a flowchart of a payment label intelligent processing method based on multi-dimensional user portrait according to the present application; Figure 2 is an exemplary flowchart for determining joint spatio-temporal characteristics according to the present application; Figure 3is an example flowchart for determining a semantic label sequence according to the present application; Figure 4 is a module structure diagram of a payment label intelligent processing system based on multi-dimensional user portraits according to the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0021] The embodiments of the present application provide a payment label intelligent processing method and system based on multi-dimensional user portraits, the core of which is to construct a multi-dimensional user portrait matrix based on multi-source behavior characteristics; to perform spatio-temporal self-attention joint extraction on user payment behavior according to the multi-dimensional user portrait matrix to obtain joint spatio-temporal characteristics, and then to perform double-channel screening on the joint spatio-temporal characteristics to obtain a candidate payment label set; to perform semantic matching and rearrangement on the candidate payment label set through a semantic correlation mapping and a contrast learning mechanism to obtain a semantic label sequence, and then to determine a feature contribution degree according to the semantic label sequence and the joint spatio-temporal characteristics; to obtain a user's historical payment record, to perform prediction error feedback on the user's payment label based on the feature contribution degree and the historical payment record, and then to obtain a semantic verification result; to perform risk identification and behavior recommendation on the user's payment label according to the semantic verification result to obtain a payment decision label for adaptive recommendation.

[0022] Embodiment one, in order to better understand the above technical solutions, the above technical solutions will be described in detail below in combination with the drawings in the specification and specific implementation manners, referring to Figure 1 As shown in the figure, the figure is a flowchart of a payment label intelligent processing method based on multi-dimensional user portraits according to the present embodiment of the present application, the intelligent processing method comprising the following steps: In step S1, a multi-dimensional user portrait matrix is constructed based on multi-source behavior characteristics.

[0023] It should be noted that in the embodiment, the multi-source behavior feature refers to a multi-dimensional data set composed of payment behavior attributes of a user under different time, scene and device conditions, and the multi-source behavior feature includes payment features, terminal features and environment features; the payment features are used to describe the amount change, payment frequency and merchant type distribution of the user in the payment process; the terminal features are used to describe the payment terminal type, operating system version, network connection mode and device unique identification information used by the user; and the environment features are used to describe the geographical position information and time period features corresponding to the user payment behavior. In specific implementation, the multi-source behavior feature of the user can be obtained through a payment log database of a payment platform, wherein the payment log database includes payment amount, payment time, merchant number and payment category and the like, and preferably, the payment log database can adopt a distributed data storage structure to facilitate high-concurrency reading and horizontal expansion of behavior data from different sources.

[0024] In the embodiment, the multi-dimensional user portrait matrix based on the multi-source behavior feature can be constructed in the following manner, that is: timestamp alignment and normalization processing are performed on the multi-source behavior feature to obtain a behavior feature set; feature layer coding is performed on the behavior feature set to obtain the relevance of each feature layer; a multi-dimensional user portrait matrix is constructed based on all the relevance and the behavior feature set.

[0025] In specific implementation, the missing time slices of the multi-source behavior feature can be filled by forward padding, and the abnormal values can be replaced by median values, and then the numerical features can be normalized by Z-Score standardization to obtain the behavior feature set; then, the time-continuous multi-source behavior feature is layer-coded according to the attribute type, that is, for the numerical features used to describe the payment quantity and amount change, statistical aggregation is used to generate time window statistics; for the categorical features used to represent the merchant category and payment method, one-hot encoding is used to convert them into vector form; for the text features representing device information and environmental semantics, they are converted into structured data based on a dictionary mapping table, and then the relevance of each feature layer is calculated through the Pearson correlation coefficient; finally, the relevance is used as the feature layer weight, and each feature layer is arranged and spliced into a matrix through the feature layer weight, wherein the rows of the matrix represent user samples, and the columns represent different dimensions of the multi-source behavior feature (such as payment features, terminal features and environment features), and the matrix is used as a multi-dimensional user portrait matrix.

[0026] It should be noted that the multi-dimensional user portrait matrix in the present application refers to a matrix structure with user identification as index and multi-layer encoded features as columns. The multi-dimensional user portrait matrix can provide input data for empty feature extraction and semantic optimization processing. In the present embodiment, feature hierarchical coding refers to dividing behavior features into payment layer, terminal layer and environment layer according to feature semantics, and using common technologies such as numerical aggregation, one-hot coding or embedding mapping to code each layer of data to eliminate feature heterogeneity.

[0027] In step S2, the user payment behavior is extracted by time-space self-attention joint extraction according to the multi-dimensional user portrait matrix to obtain joint time-space features, and then the joint time-space features are screened by double channels to obtain a candidate payment label set.

[0028] Preferably, in the present embodiment, reference is made to FIG. 2, which is an exemplary flowchart for determining joint time-space features according to the present application. In the present embodiment, the joint time-space features can be obtained by the following steps: Figure 2 First, in step S21, the dependence weight of the user payment behavior is determined. Then, in step S22, the multi-dimensional user portrait matrix is respectively modeled by time attention dependence and space attention dependence according to the dependence weight to obtain time attention features and space attention features of the user payment behavior. Finally, in step S23, the time attention features and the space attention features are fused to obtain joint time-space features.

[0029] It should be noted that the user payment behavior refers to behavior features describing payment records, including payment amount, payment frequency and average payment interval.

[0030] ​In a specific implementation, first, a sliding time window is divided according to a time dimension (for example, the past 7 days), a multi-dimensional user portrait matrix is linearly transformed through the sliding time window to obtain a time query vector, a time key vector and a time value vector, then the time query vector, the time key vector and the time value vector are normalized by using Softmax to obtain a dependency weight; then, the dependency weight is used as a convolution kernel of a spatio-temporal convolution network, and then the multi-dimensional user portrait matrix is spatio-temporally convolved through the spatio-temporal convolution network with the convolution kernel set, and a time attention feature is output through a time channel of the spatio-temporal convolution network, and a space attention feature is output through a space channel of the spatio-temporal convolution network, preferably, in the two types of attention calculation in the embodiment, the dimensions of the time query vector, the time key vector and the time value vector can be adjusted through a learnable linear transformation, and a residual connection can be added after the attention calculation, which is conducive to stable training; finally, the time attention feature and the space attention feature can be fused through a fully connected layer, and a layer normalization is performed for regularization processing to prevent overfitting, preferably, a multi-head cross attention mechanism can be used to calculate an interaction attention weight between the time attention feature and the space attention feature to enhance the coupled expression of the spatio-temporal features, that is, a high-dimensional vector subjected to fusion and normalization processing can be obtained, and the high-dimensional vector is used as a joint spatio-temporal feature, wherein the multi-head cross attention mechanism can be realized by constructing query, key and value vectors, and then performing steps such as scaled dot-product attention and linear transformation.

[0031] It should be noted that the joint spatio-temporal feature in the present application is a high-dimensional vector for simultaneously representing the coupling relationship between time and space, and the high-dimensional vector can be directly used for semantic screening and label prediction; in the embodiment, the dependency weight refers to a numerical measure of quantifying the importance of user historical behavior in the current label identification task, and the most influential spatio-temporal segment of the current payment semantics can be extracted in the spatio-temporal modeling process; the time attention feature refers to a vector representation describing the time evolution law of user behavior obtained after time attention weighting and aggregation, which can depict the dynamic mode of payment behavior over time; the space attention feature refers to a vector representation obtained based on the weighting of geographic location, terminal type and neighborhood similarity, which can depict the aggregation characteristics of payment behavior in spatial distribution and terminal difference.

[0032] In the embodiment, the joint spatio-temporal feature is double-channel screened to obtain a candidate payment label set, which can be realized in the following manner, that is: a semantic feature mapping channel is constructed based on a semantic embedding network, and a rule constraint channel is constructed based on a pre-set rule set; the joint spatio-temporal feature is input into the semantic feature mapping channel, the joint spatio-temporal feature is mapped to a label semantic space through the semantic embedding network, and then a candidate label set is screened; The candidate payment label set is obtained by logically constraining and screening the candidate label set through a rule matching matrix in a rule constraint channel.

[0033] In a specific implementation, first, a semantic embedding network based on a Transformer structure is used to forward encode the joint space-time feature, extract a joint space-time semantic vector, and map the semantic vector to a label semantic space through a semantic embedding layer to form a semantic feature mapping channel; payment rules obtained according to experience are used to obtain restriction conditions including a payment amount interval, a payment frequency threshold, a device type restriction, and a region whitelist, and the restriction conditions are loaded into a rule matching matrix using a Transformer structure, and the rule matching matrix is constructed into a rule constraint channel; then, the semantic feature mapping channel can measure and screen label semantics through a cosine similarity, thereby obtaining a candidate label set; finally, the rule matching matrix in the rule constraint channel is used to perform rule matching operations on the candidate label set, that is, if a rule corresponding to a label is not satisfied, the label is removed from the set; if a label satisfies multiple rules, the label is added with a weight score, and a candidate payment set is obtained through screening, where the rule matching operation process can be implemented through a Pandas library.

[0034] It should be noted that the candidate payment label set in the present application is a payment label obtained through semantic and rule screening, and is used to describe key payment information of a user; the semantic embedding network in the present embodiment is a deep encoding network used to uniformly map structured joint space-time features and textual label descriptions to the same vector space, and is used to measure semantic correlation between payment behavior features and label descriptions; the label semantic space is a multi-dimensional vector space generated by the semantic embedding network, and is used to represent semantic differences between labels, and a smaller semantic distance indicates that a behavior and a label are more consistent; the rule matching matrix is a structured constraint table in which business rules are expressed and stored in the form of a matrix, and business logic constraints can be applied to a semantic candidate set; the candidate label set refers to a label set obtained through preliminary screening based on semantic correlation.

[0035] In step S3, the candidate payment label set is semantically matched and rearranged through a semantic correlation mapping and a contrastive learning mechanism, to obtain a semantic label sequence, and then a feature contribution degree is determined according to the semantic label sequence and the joint space-time feature.

[0036] Preferably, in the present embodiment, as shown in Figure 3 The figure is an example flowchart for determining a semantic label sequence according to the present application, and in the present embodiment, the candidate payment label set is semantically matched and rearranged through a semantic correlation mapping and a contrastive learning mechanism to obtain a semantic label sequence, which can be implemented through the following steps: Firstly, in step S31, the candidate payment label set is subjected to semantic embedding to obtain a semantic embedding matrix; Then, in step S32, the semantic correlation degree of the semantic embedding matrix and the label semantic space is determined, and then the correlation degree ranking result is determined through the semantic correlation degree; Finally, in step S33, the correlation degree ranking result is subjected to contrast learning optimization to obtain a semantic label sequence.

[0037] In a specific implementation, firstly, the text description of the candidate label can be subjected to word segmentation, stop word removal and lemmatization processing, and then the processed text is input into a text encoder based on the Transformer structure for encoding, so as to obtain a text semantic vector corresponding to each label, and all label semantic vectors are combined into a matrix structure to obtain a semantic embedding matrix; then, the semantic similarity between the text semantic vector and the semantic vector in the label semantic space can be calculated by using the cosine similarity, the average of all semantic similarities is taken as the semantic correlation degree, and the semantic vector with a higher semantic similarity than the semantic correlation degree is replaced with the semantic vector in the label semantic space to obtain a correlation degree ranking result; finally, the loss value is calculated by using a contrast loss function (i.e., a cross-entropy contrast loss), the parameters of the semantic embedding network are updated through back propagation, the similarity score between the semantic vector and the text semantic vector is calculated, and the ranking is updated until the loss value reaches Top-1 stability, and then the label sequence arranged from high to low in terms of semantic matching strength after optimization is output as a semantic label sequence, which can be directly input for feature contribution degree calculation.

[0038] It should be noted that the semantic label sequence in the present application refers to an ordered label list obtained by arranging the labels from high to low in terms of semantic correlation after contrast learning optimization, which is used to provide a high-confidence label input for feature contribution degree calculation and semantic verification; the semantic embedding matrix refers to a matrix structure formed after the candidate label text description is mapped into a continuous vector set by a language encoding model, which is used to convert discrete label semantics into a high-dimensional vector representation that can be calculated; the semantic correlation degree refers to a numerical measure of the closeness of the behavior semantic vector and the label semantic vector in the semantic space; the contrast learning optimization in the present embodiment refers to an optimization unit with the core of training positive and negative samples and minimizing the contrast loss, which is beneficial to improving the discriminability of semantic matching by shortening the distance of positive samples and distancing the distance of negative samples.

[0039] In the present embodiment, the feature contribution degree can be determined according to the semantic label sequence and the joint spatio-temporal feature in the following manner, i.e.: The feature mapping weight of the semantic label sequence and the joint spatio-temporal feature is calculated based on the correlation, and then a contribution degree matrix is obtained; determine a weight distribution vector according to the weight distribution vector, and determine a weighted response coefficient of each semantic label in the semantic label sequence and the joint space-time feature; aggregate all the weighted response coefficients according to the time and space dimensions to obtain feature contribution degrees.

[0040] In a specific implementation, first, similarity measurement is performed on each label semantic vector in the semantic label sequence and each space-time feature vector in the joint space-time feature, preferably, vector dot product is used as the measurement manner, and the obtained similarity value is used as a feature mapping weight, which is filled as a matrix element according to the [semantic label, joint space-time feature] index position to obtain a contribution degree matrix; then, each row and each column of the contribution degree matrix is normalized by applying, for example, a Softmax function normalization function, to obtain a weight distribution vector, the weight distribution vector and the corresponding joint space-time feature vector are multiplied element by element, and summation is performed in the dimension to obtain a weighted response coefficient of each semantic label relative to the space-time feature; finally, the weighted response coefficient can be averaged according to a set time window (for example, the past 7 days) to obtain a time aggregation result, and the weighted response coefficient is spatially aggregated according to spatial grouping (for example, divided according to terminal type), and the feature contribution degree is generated.

[0041] It should be noted that the feature contribution degree in the present application refers to the weighted response intensity of the semantic label in the joint space-time feature space, which is used to comprehensively describe an index of the overall influence of the semantic label in the space-time layer, and can provide a quantitative reference for prediction error correction; the contribution degree matrix refers to a two-dimensional matrix structure with semantic labels as rows and joint space-time feature dimensions or time windows as columns, and the matrix elements are the values obtained by similarity measurement between the labels and the features, which are used to reflect the response intensity of the semantic label to each space-time feature; the weight distribution vector represents the relative contribution proportion of each space-time feature under a specific semantic label, which is used to map the similarity to an interpretable weight distribution; the weighted response coefficient represents the response intensity of a specific semantic label in the current space-time context, which is used to convert the statistical weight into a semantic response index.

[0042] In step S4, the historical payment records of the user are obtained, prediction error feedback is performed on the payment label of the user based on the feature contribution degree and the historical payment records, and then a semantic verification result is obtained.

[0043] It should be noted that in the present embodiment, the historical payment record of the user is obtained through the payment data management database, and the payment data management database refers to a structured data management system for centrally storing and managing user payment data, account information and payment behavior logs. The payment data management database can be a distributed database, and the historical payment record recorded in the payment data management database includes user identification, payment time, payment amount, payment type, merchant number, payment method and payment terminal, etc. Field information, which can completely reflect the payment behavior track of the user; in actual implementation, the payment data management database can obtain the historical payment record by calling the log interface of the payment system, and in the present embodiment, the historical payment record can select the payment record from the current payment label acquisition time to the past 7 days, which is beneficial to provide reliable data support for prediction error feedback and semantic verification.

[0044] In the present embodiment, the prediction error feedback of the payment label of the user is carried out based on the feature contribution degree and the historical payment record, and then the semantic verification result is obtained. The following methods can be used, that is: A label entity set is constructed through the payment label of the user; The label entity set is paired for prediction based on the feature contribution degree and the historical payment record, and a prediction deviation value is obtained; The process of prediction pairing is updated through the prediction deviation value, and a semantic verification vector is obtained; According to the semantic verification vector and the preset confidence threshold, a semantic verification result is determined.

[0045] It should be noted that the payment label of the user includes payment amount, payment times and average payment interval, etc. In specific implementation, firstly, the payment label of the user can be preprocessed through text cleaning (such as removing spaces, unifying cases, synonym mapping), and a label entity set for comparison is constructed; secondly, the label entity set, the semantic label sequence and the feature contribution degree are mapped and associated, that is, the historical appearance frequency and the weighted score of the feature contribution degree of each label entity are compared, and the weighted score is compared with the historical true statistical value to obtain a prediction deviation value. Preferably, the mean square error can be used as the difference comparison method, which is beneficial to synchronous statistics and difference comparison.

[0046] In a specific implementation, the feedback update of the prediction pairing process can adopt an error back propagation algorithm based on gradient descent. In an actual implementation, the prediction deviation value can be input into a feedback update model of the error back propagation algorithm based on gradient descent, a mean square error can be used as a loss function, a gradient direction between a model prediction score and a historical true statistical value can be calculated, the semantic mapping parameters and the label weight can be adjusted according to the gradient direction, so that the output distribution of the model gradually approximates to the statistical distribution of the historical true behavior. Preferably, the Adam optimization algorithm can be used for adaptive adjustment of the parameters to avoid oscillation caused by an excessively large learning rate. A response score of each label is obtained, all response scores are converted into a score vector between 0 and 1, and the confidence score vector is used as a semantic verification vector. In the embodiment, the prediction pairing process refers to a process of matching the label prediction result output by the model with the historical true label one by one and calculating the difference between them. Finally, the semantic verification vector is determined according to a preset confidence threshold. The value of the confidence threshold can be determined based on the statistical distribution of the historical samples. For example, the mean value of all label historical confidence is calculated, and the mean value is used as the confidence threshold to balance the recall rate and the accuracy. In an actual implementation, the semantic confidence of the semantic verification vector output by the feedback update model is compared with the confidence threshold. That is, when the semantic confidence is greater than or equal to the confidence threshold, it is determined that the semantic verification of the semantic verification vector is verified. Otherwise, it is recorded as a low-confidence semantic verification vector. A set composed of all the semantic verification vectors that pass the verification is used as a semantic verification result set.

[0047] It should be noted that the semantic verification vector in the present application refers to a label confidence score set after feedback update, which has the same dimension as the payment label and can be used to convert the semantic matching result into a quantifiable confidence indicator. The confidence threshold is a boundary value for distinguishing between high-confidence labels and low-confidence labels. The semantic verification result refers to a label set obtained by comparing the semantic verification vector with the confidence threshold, which includes labels that the model considers to be reliable in semantic matching and consistent with the user's true payment behavior, and can be used for risk identification and adaptive recommendation process. The label entity set refers to a label set formed after standardization and deduplication processing, which can be used to establish a unified label benchmark for prediction comparison, and facilitate the calculation of the deviation between the historical statistical characteristics and the model prediction result.

[0048] In step S5, the payment label of the user is identified and recommended according to the semantic verification result, and a payment decision label for adaptive recommendation is obtained.

[0049] In the embodiment, the payment label of the user is identified and recommended according to the semantic verification result, and a payment decision label for adaptive recommendation is obtained. Specifically, the following methods can be used: The semantic verification result is analyzed to obtain a risk feature matrix. filter the payment label of the user through the risk feature matrix to obtain a filtered payment label; use the adaptive optimization strategy network to perform recommendation sorting on the filtered payment label to obtain a payment decision label for adaptive recommendation.

[0050] In a specific implementation, first, the historical payment records corresponding to the semantic verification result are called from the payment data management database, structured query statements are used to extract the data to the Pandas analysis environment, the extracted data is calculated according to the label dimension to obtain risk factors (the risk factors include: semantic confidence mean value, label historical occurrence frequency, payment amount mean value and variance, expected deviation rate, abnormal payment rate, etc.), and the multiple risk factors of each label are organized as matrix rows according to a fixed column order, so as to obtain the risk feature matrix; then, the risk filtering is performed on the payment label of the user based on the risk feature matrix, that is, the risk feature matrix is input into the Gaussian mixture model, and the filtered payment label can be obtained through the filtering of the Gaussian mixture model, that is, the risk confidence score of each payment label is calculated through the Gaussian mixture model and compared with a preset threshold, the labels with a risk confidence score higher than the preset threshold are marked and removed, the labels with a risk confidence score lower than the preset threshold and meeting the business whitelist are retained, and for the labels in the intermediate risk interval, a secondary judgment can be made according to rules (for example, the one with a small amount fluctuation and a high historical frequency is preferred), wherein the preset threshold can be set according to the actual quantity demand and amount demand of the payment label; finally, the adaptive optimization strategy network is a multi-layer perceptron network, wherein the input of the multi-layer perceptron network includes a user portrait vector (read from a multi-dimensional user portrait matrix), a semantic embedding vector of each candidate label, a behavior matching score and a corresponding risk confidence score, and a cross-entropy and Adam optimizer are used for batch training, and the filtered payment label is input into the adaptive optimization strategy network, so that the payment decision label for adaptive recommendation can be output through the adaptive optimization strategy network.

[0051] It should be noted that in the present application, the payment decision label refers to a label set with a confidence score and a risk identifier output after risk filtering and adaptive sorting, which can be used as a decision basis for the front end of the system for personalized display or real-time push; the risk feature matrix refers to a two-dimensional numerical matrix organized by taking the label as the row and various risk factors as the column, which is used to express the risk performance of each candidate label in the dimensions of semantic confidence, frequency, amount stability and abnormality in a structured form.

[0052] In summary, the technical scheme adopted by the present application can realize payment label self-learning and semantic dynamic optimization based on multi-source behavior characteristics, so as to improve the payment label recognition accuracy.

[0053] Embodiment two, the application provides a payment label intelligent processing system based on multi-dimensional user portrait, referring to Figure 4 As shown in the figure, it is a module structure diagram of a payment label intelligent processing system based on multi-dimensional user portrait provided by the application, and the intelligent processing system comprises: A user portrait construction module 100 is configured to construct a multi-dimensional user portrait matrix based on multi-source behavior characteristics. A space-time feature extraction module 200 is configured to perform space-time self-attention joint extraction on user payment behavior according to the multi-dimensional user portrait matrix to obtain joint space-time features, and then perform double-channel screening on the joint space-time features to obtain a candidate payment label set. A semantic rearrangement module 300 is configured to perform semantic matching rearrangement on the candidate payment label set through a semantic correlation mapping and a contrast learning mechanism to obtain a semantic label sequence, and then determine a feature contribution degree according to the semantic label sequence and the joint space-time features. A feedback verification module 400 is configured to obtain historical payment records of a user, perform prediction error feedback on the payment label of the user based on the feature contribution degree and the historical payment records, and then obtain a semantic verification result. A risk identification and recommendation module 500 is configured to perform risk identification and behavior recommendation on the payment label of the user according to the semantic verification result to obtain a payment decision label for adaptive recommendation.

[0054] The application is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of the flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one flow or multiple flows and / or blocks Figure 1 The functions specified in one flow or multiple flows and / or blocks

[0055] Those skilled in the art can understand that all or part of the steps of various methods in the above embodiments can be completed by instructing the relevant hardware by means of a program, and the program can be stored in a computer readable storage medium, including Read-Only Memory (ROM), Random Access Memory (RAM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), One-time Programmable Read-Only Memory (OTPROM), Electrically-Erasable Programmable Read-Only Memory (EEPROM), Compact Disc Read-Only Memory (CD-ROM) or other optical disk storage, magnetic disk storage, magnetic tape storage, or any other medium that can be used to carry or store data in a computer readable manner.

[0056] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements, but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the statement "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the element.

Claims

1. A payment tag intelligent processing method based on multi-dimensional user profiles, characterized in that, The intelligent processing method includes the following steps: Construct a multi-dimensional user profile matrix based on multi-source behavioral features; Based on the multi-dimensional user profile matrix, spatiotemporal self-attention joint extraction of user payment behavior is performed to obtain joint spatiotemporal features. Then, the joint spatiotemporal features are filtered through dual channels to obtain a candidate payment tag set. The candidate payment tag set is semantically matched and rearranged through semantic relevance mapping and contrastive learning mechanisms to obtain a semantic tag sequence, and then the feature contribution is determined based on the semantic tag sequence and the joint spatiotemporal features. Obtain the user's historical payment records, and based on the feature contribution and the historical payment records, perform prediction error feedback on the user's payment tags to obtain semantic verification results; Based on the semantic verification results, risk identification and behavior recommendation are performed on the user's payment tags to obtain payment decision tags for adaptive recommendation.

2. The intelligent processing method for payment tags based on multi-dimensional user profiles as described in claim 1, characterized in that, Constructing a multi-dimensional user profile matrix based on multi-source behavioral features specifically includes: The multi-source behavioral features are time-stamp aligned and normalized to obtain a behavioral feature set; The behavioral feature set is subjected to feature hierarchical encoding to obtain the relevance of each feature layer; A multi-dimensional user profile matrix is ​​constructed based on all the relevances and the behavioral feature set.

3. The intelligent processing method for payment tags based on multi-dimensional user profiles as described in claim 1, characterized in that, Based on the multi-dimensional user profile matrix, spatiotemporal self-attention joint extraction is performed on user payment behavior to obtain joint spatiotemporal features, specifically including: Determine the dependency weights of user payment behavior; Based on the dependency weights, temporal attention dependency modeling and spatial attention dependency modeling are performed on the multi-dimensional user profile matrix to obtain the temporal attention features and spatial attention features of user payment behavior. The temporal attention features and the spatial attention features are fused to obtain joint spatiotemporal features.

4. The intelligent processing method for payment tags based on multi-dimensional user profiles as described in claim 1, characterized in that, The candidate payment tag set obtained by performing dual-channel filtering on the joint spatiotemporal features specifically includes: A semantic feature mapping channel is constructed based on a semantic embedding network, and a rule constraint channel is constructed based on a preset rule set; The joint spatiotemporal features are input into the semantic feature mapping channel, and the joint spatiotemporal features are mapped to the label semantic space through the semantic embedding network, thereby obtaining a set of candidate labels. The candidate tag set is logically constrained and filtered using the rule matching matrix in the rule constraint channel to obtain the candidate payment tag set.

5. The intelligent processing method for payment tags based on multi-dimensional user profiles as described in claim 1, characterized in that, The candidate payment tag set is semantically matched and rearranged using semantic relevance mapping and contrastive learning mechanisms to obtain a semantic tag sequence, specifically including: Semantic embedding is performed on the candidate payment tag set to obtain a semantic embedding matrix; The semantic relevance between the semantic embedding matrix and the label semantic space is determined, and then the relevance ranking result is determined based on the semantic relevance. The relevance ranking results are compared and optimized to obtain a semantic label sequence.

6. The intelligent processing method for payment tags based on multi-dimensional user profiles as described in claim 1, characterized in that, Determining the feature contribution based on the semantic label sequence and the joint spatiotemporal features specifically includes: Based on the correlation, the feature mapping weights of the semantic label sequence and the joint spatiotemporal features are calculated, and then the contribution matrix is ​​obtained; The weight distribution vector is determined by the contribution matrix, and then the weighted response coefficient of each semantic tag and the joint spatiotemporal feature in the semantic tag sequence is determined based on the weight distribution vector. All weighted response coefficients are aggregated according to the time and spatial dimensions to obtain the feature contribution.

7. The intelligent processing method for payment tags based on multi-dimensional user profiles as described in claim 1, characterized in that, The payment data management database retrieves users' historical payment records.

8. The intelligent processing method for payment tags based on multi-dimensional user profiles as described in claim 1, characterized in that, Based on the feature contribution and the historical payment records, the user's payment tag is predicted with error feedback, and the semantic verification result is obtained, specifically including: Construct a tag entity set based on the user's payment tags; Based on the feature contribution and the historical payment records, the tag entity set is predicted and paired to obtain the prediction deviation value; The prediction pairing process is updated by feeding back the prediction deviation value to obtain the semantic verification vector. The semantic verification result is determined based on the semantic verification vector and the preset confidence threshold.

9. The intelligent processing method for payment tags based on multi-dimensional user profiles as described in claim 1, characterized in that, The multi-source behavioral features refer to a multi-dimensional data set composed of user payment behavior attributes under different time, scenario and device conditions. The multi-source behavioral features include: payment features, terminal features and environmental features.

10. A payment tag intelligent processing system based on multi-dimensional user profiles, used to execute the payment tag intelligent processing method based on multi-dimensional user profiles as described in any one of claims 1 to 9, characterized in that, The intelligent processing system includes: The user profile building module is used to construct a multi-dimensional user profile matrix based on multi-source behavioral features. The spatiotemporal feature extraction module is used to perform spatiotemporal self-attention joint extraction of user payment behavior based on the multi-dimensional user profile matrix to obtain joint spatiotemporal features, and then perform dual-channel filtering on the joint spatiotemporal features to obtain a candidate payment tag set; The semantic reordering module is used to perform semantic matching and reordering on the candidate payment tag set through semantic relevance mapping and contrastive learning mechanism to obtain a semantic tag sequence, and then determine the feature contribution based on the semantic tag sequence and the joint spatiotemporal features. The feedback verification module is used to obtain the user's historical payment records, and based on the feature contribution and the historical payment records, to provide prediction error feedback for the user's payment tags, thereby obtaining the semantic verification result; The risk identification and recommendation module is used to identify risks and recommend behaviors for users' payment tags based on the semantic verification results, and obtain payment decision tags for adaptive recommendation.

Images