Merchant marketing accounting processing method and system based on AI

Abstract

The invention relates to the field of intelligent marketing accounting, and provides an AI-based merchant marketing accounting processing method and system, and the method comprises the steps: carrying out the embedded coding and time sequence aggregation of transaction data of a first commodity through employing an AI-based data analysis and processing technology, carrying out the semantic coding of the information description of the first commodity, and carrying out the semantic coding of the information description of the first commodity; therefore, whether the first commodity is a high-profit commodity or not is intelligently assessed according to a principal component multi-modal fusion representation feature between a feature after transaction data time sequence aggregation and a first commodity description semantic feature, and whether the account division proportion is increased or not is judged based on an assessment result. Therefore, market changes can be quickly responded, manual operation can be reduced while commodity characteristics and market performance are deeply understood, and the intelligent degree of commercial tenant marketing account division processing can be improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent marketing account splitting, and more specifically, to an AI-based merchant marketing account splitting processing method and system. Background Art

[0002] In modern e-commerce and retail industries, merchant marketing account splitting is an important link, which involves how to reasonably allocate sales revenue to different channels, partners or marketing activities. This process is not only related to the merchant's profit model, but also directly affects the effect of marketing promotion and the enthusiasm of partners.

[0003] However, traditional merchant marketing account splitting processing methods are usually based on manual operations and simple data analysis methods. Specifically, in traditional merchant marketing account splitting processing, since it mainly relies on manual operations, the entire process is not only cumbersome and error-prone, but also difficult to achieve comprehensive and in-depth analysis due to the limitations of human eyes and brains in processing information. In addition, traditional manual account splitting relies on historical data and preset rules, which limits the ability to respond immediately to market dynamics. For example, when demand for goods surges due to holiday promotions or emergencies, manual account splitting cannot be adjusted quickly to reflect these changes, which may miss opportunities to optimize profit distribution. At the same time, this method usually only focuses on intuitive indicators such as sales and profits, ignoring the semantic associations in transaction data and product descriptions. This one-sided data analysis cannot fully tap the deep-level characteristics and potential value behind the data, thus affecting the accuracy and fairness of account splitting.

[0004] Therefore, an optimized merchant marketing account splitting processing solution is desired. Summary of the invention

[0005] In response to the shortcomings of the prior art, the present application provides an AI-based merchant marketing account splitting processing method and system.

[0006] According to one aspect of the present application, a merchant marketing account splitting processing method based on AI is provided, which includes:

[0007] Extracting the time queue of the transaction data of the first commodity;

[0008] Extracting information description of the first product;

[0009] Performing semantic coding on the information description of the first product to obtain a semantic coding feature vector of the information description of the first product;

[0010] Using the transaction embedding coding matrix, each transaction data in the time queue of the transaction data of the first commodity is respectively embedded and coded to obtain a time queue of transaction data embedding coding vectors;

[0011] The time queue of the transaction data embedded in the coding vector is input into a feature aggregation analysis network guided by node feature significance decay to obtain a first commodity transaction time series aggregation representation vector;

[0012] Inputting the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector into a saliency-driven principal component feature fusion network to obtain a first commodity transaction-attribute principal component multimodal fusion representation vector;

[0013] Based on the first commodity transaction-attribute principal component multimodal fusion representation vector, an evaluation result is obtained, and based on the evaluation result, an account splitting recommendation suggestion is generated.

[0014] In the above-mentioned AI-based merchant marketing account splitting processing method, the time queue of the transaction data embedded coding vector is input into a feature aggregation analysis network guided by node feature significance attenuation to obtain a first commodity transaction time series aggregation representation vector, including: calculating the feature significance description factor of each transaction data embedded coding vector in the time queue of the transaction data embedded coding vector, wherein the feature significance description factor is related to the mean and variance of each transaction data embedded coding vector; constructing the feature significance attenuation factor of each transaction data embedded coding vector based on the distance span between each transaction data embedded coding vector in the time queue of the transaction data embedded coding vector and the current transaction data embedded coding vector; modulating the feature significance description factor of each transaction data embedded coding vector based on the feature significance attenuation factor of each transaction data embedded coding vector to obtain a sequence of transaction data feature significance attenuation description factors; and The sequence of the feature attenuation description factors of the transaction data is input into the gated mask module to obtain the sequence of the feature significance attenuation weight factors of the transaction data; the sequence of the feature significance attenuation weight factors of the transaction data is used as the sequence of weights, and the weighted sum of the time queue of the transaction data embedded coding vector is calculated to obtain the first commodity transaction time series aggregation representation vector; wherein, the sequence of the feature significance attenuation description factors of the transaction data is input into the gated mask module to obtain the sequence of the feature significance attenuation weight factors of the transaction data, including: performing normalization processing based on the Sigmoid function on each feature significance attenuation description factor in the sequence of the feature significance attenuation description factors of the transaction data to obtain a set of normalized feature significance attenuation description factors of the transaction data; each normalized feature significance attenuation description factor in the set of normalized feature significance attenuation description factors of the transaction data is input into the gated function for masking processing to obtain the sequence of the feature significance attenuation weight factors of the transaction data.

[0015] In the above-mentioned AI-based merchant marketing account splitting processing method, the characteristic significance description factor of each transaction data embedded coding vector in the time queue of the transaction data embedded coding vector is calculated, including: respectively calculating the mean and variance of the transaction data embedded coding vector to obtain the transaction data characteristic mean and the transaction data characteristic variance; calculating the positional difference between the transaction data embedded coding vector and the transaction data characteristic mean to obtain a transaction data differential vector; calculating the fourth power of each eigenvalue in the transaction data differential vector to obtain a transaction data modulation difference vector; calculating the expected value of the transaction data modulation difference vector to obtain a transaction data expected value; dividing the transaction data expected value by the square of the transaction data characteristic variance to obtain the characteristic significance description factor corresponding to the transaction data embedded coding vector.

[0016] In the above-mentioned AI-based merchant marketing account splitting processing method, based on the distance span between each transaction data embedded coding vector in the time queue of the transaction data embedded coding vector and the current transaction data embedded coding vector, the feature significance attenuation factor of each transaction data embedded coding vector is constructed, including: extracting the maximum value in each transaction data embedded coding vector to obtain a sequence of transaction data feature maximum values; extracting the maximum value of the current transaction data embedded coding vector to obtain the current transaction data feature maximum value; calculating the positional subtraction between the current transaction data feature maximum value and the sequence of transaction data feature maximum values ​​to obtain a sequence of transaction data feature offset values; counting the distance span values ​​between each transaction data embedded coding vector and the current transaction data embedded coding vector to obtain a sequence of transaction data distance span values, and calculating the squares of each value in the sequence of transaction data distance span values ​​to obtain a sequence of transaction data distance span modulation values; dividing the sequence of transaction data feature offset values ​​by the sequence of transaction data distance span modulation values ​​by position to obtain the feature significance attenuation factor of each transaction data embedded coding vector.

[0017] In the above-mentioned AI-based merchant marketing account splitting processing method, the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector are input into the significance-driven principal component feature fusion network to obtain the first commodity transaction-attribute principal component multimodal fusion representation vector, including: standardizing the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector to obtain a standardized first commodity transaction time series aggregation representation vector and a standardized first commodity information description semantic encoding feature vector; calculating the standardized first commodity transaction time series aggregation representation vector and the standardized first commodity information description semantic encoding feature vector. The sample covariance matrix of the encoded feature vector is used to obtain the sample covariance matrix of the first commodity transaction time series aggregation and the semantic sample covariance matrix of the first commodity information description; the feature vector extraction based on matrix decomposition is performed on the sample covariance matrix of the first commodity transaction time series aggregation and the semantic sample covariance matrix of the first commodity information description to obtain the set of principal component feature vectors of the first commodity transaction time series aggregation and the set of principal component feature vectors of the first commodity information description; the set of principal component feature vectors of the first commodity transaction time series aggregation and the set of principal component feature vectors of the first commodity information description are input into the maximum approximate query matching network to obtain the principal component feature vectors of the first commodity transaction time series aggregation. The best matching pairs of the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector are obtained by combining the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector; the best matching pairs of each first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector in the set of the best matching pairs of the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector are input into the semantic fine-grained gated joint module to obtain the set of first commodity transaction-information description principal component fusion feature vectors; the set of the first commodity transaction-information description principal component fusion feature vectors is cascaded to obtain the first commodity transaction-attribute principal component multimodal fusion representation vector; wherein, the calculation The sample covariance matrix of the standardized first commodity transaction time series aggregation representation vector and the standardized first commodity information description semantic encoding feature vector is obtained to obtain the first commodity transaction time series aggregation sample covariance matrix and the first commodity information description semantic sample covariance matrix, including: multiplying the transposed vector of the standardized first commodity transaction time series aggregation representation vector with the standardized first commodity transaction time series aggregation representation vector, and then dividing the obtained standardized first commodity transaction time series aggregation association matrix by the value obtained by subtracting one from the length of the standardized first commodity transaction time series aggregation representation vector by position to obtain the first commodity transaction time series aggregation sample covariance matrix;After multiplying the transposed vector of the standardized first product information description semantic coding feature vector by the standardized first product information description semantic coding feature vector, the obtained standardized first product information description semantic association matrix is ​​divided by the value obtained by subtracting one from the length of the standardized first product information description semantic coding feature vector to obtain the first product information description semantic sample covariance matrix. ;

[0018] In the above-mentioned AI-based merchant marketing account splitting processing method, the first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector are standardized to obtain a standardized first commodity transaction time series aggregation representation vector and a standardized first commodity information description semantic coding feature vector, including: respectively calculating the mean and standard deviation of the first commodity transaction time series aggregation representation vector to obtain the mean of the first commodity transaction time series aggregation feature and the standard deviation of the first commodity transaction time series aggregation feature; subtracting the first commodity transaction time series aggregation representation vector from the mean of the first commodity transaction time series aggregation feature by position, and calculating the first commodity transaction time series deviation. The shift vector is divided by the standard deviation of the first commodity transaction time series aggregation feature by position to obtain the standardized first commodity transaction time series aggregation representation vector; the mean and standard deviation of the first commodity information description semantic coding feature vector are calculated respectively to obtain the first commodity information description semantic feature mean and the first commodity information description semantic feature standard deviation; after subtracting the first commodity information description semantic coding feature vector from the first commodity information description semantic feature mean by position, the calculated first commodity information description semantic offset vector is divided by the standard deviation of the first commodity information description semantic feature to obtain the standardized first commodity information description semantic coding feature vector.

[0019] In the above-mentioned AI-based merchant marketing account splitting processing method, the set of the first commodity transaction time series aggregated principal component feature vectors and the set of the first commodity information description principal component feature vectors are input into the maximum approximate query matching network to obtain a set of best matching pairs of the first commodity transaction time series aggregated principal component feature vectors and the first commodity information description principal component feature vectors, including: extracting the predetermined first commodity transaction time series aggregated principal component feature vectors from the set of the first commodity transaction time series aggregated principal component feature vectors; calculating the cosine similarity between the predetermined first commodity transaction time series aggregated principal component feature vectors and each first commodity information description principal component feature vector in the set of the first commodity information description principal component feature vectors to obtain a set of matching query similarities; and taking the first commodity information description principal component feature vector corresponding to the largest matching query similarity in the set of matching query similarities and the predetermined first commodity transaction time series aggregated principal component feature vector as the best matching pair of the predetermined first commodity transaction time series aggregated principal component feature vector and the first commodity information description principal component feature vector.

[0020] In the above-mentioned AI-based merchant marketing account splitting processing method, each best matching pair of the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector in the set of the best matching pairs of the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector is input into the semantic fine-grained gated joint module to obtain a set of first commodity transaction-information description principal component fusion feature vectors, including: respectively calculating the position difference, position dot multiplication and position addition between the best matching pairs of the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector to obtain the first Commodity transaction-information description principal component difference vector, the first commodity transaction-information description principal component dot product vector and the first commodity transaction-information description principal component sum vector; cascade the first commodity transaction-information description principal component difference vector, the first commodity transaction-information description principal component dot product vector and the first commodity transaction-information description principal component sum vector, and then perform one-dimensional convolution encoding to obtain the first commodity transaction-information description principal component multi-dimensional fusion vector; perform local window-based maximum pooling processing on the first commodity transaction-information description principal component multi-dimensional fusion vector to obtain the first commodity transaction-information description principal component fusion feature vector.

[0021] In the above-mentioned AI-based merchant marketing account splitting processing method, an evaluation result is obtained based on the first commodity transaction-attribute principal component multimodal fusion representation vector, and an account splitting recommendation suggestion is generated based on the evaluation result, including: inputting the first commodity transaction-attribute principal component multimodal fusion representation vector into a classifier-based commodity type evaluator to obtain the evaluation result, and the evaluation result is used to indicate whether the first commodity is a high-profit commodity; in response to the evaluation result that the first commodity is a high-profit commodity, the account splitting recommendation suggestion is generated, and the account splitting recommendation suggestion is used to indicate an increase in the account splitting ratio.

[0022] According to another aspect of the present application, there is provided an AI-based merchant marketing account splitting processing system, which includes:

[0023] This application has significant technical effects due to the adoption of the above technical solutions:

[0024] A commodity transaction time queue data extraction module, used to extract the time queue of the transaction data of the first commodity;

[0025] A commodity information description extraction module, used to extract the information description of the first commodity;

[0026] A commodity information description semantic coding module, used for semantically coding the information description of the first commodity to obtain a semantic coding feature vector of the first commodity information description;

[0027] A commodity transaction time queue data embedding coding module, used to use a transaction embedding coding matrix to embed code each transaction data in the time queue of the transaction data of the first commodity to obtain a time queue of transaction data embedding coding vectors;

[0028] A commodity transaction time series feature set analysis module, used to embed the transaction data into the time queue of the coding vector and input it into a feature aggregation analysis network guided by node feature significance decay to obtain a first commodity transaction time series aggregation representation vector;

[0029] A commodity multimodal feature fusion module is used to input the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector into a saliency-driven principal component feature fusion network to obtain a first commodity transaction-attribute principal component multimodal fusion representation vector; an account splitting recommendation suggestion generation module is used to obtain an evaluation result based on the first commodity transaction-attribute principal component multimodal fusion representation vector, and generate an account splitting recommendation suggestion based on the evaluation result.

[0030] The AI-based merchant marketing split-account processing method and system provided in this application adopts AI-based data analysis and processing technology to embed and time-series aggregate the transaction data of the first commodity, and semantically encode the information description of the first commodity, so as to intelligently evaluate whether the first commodity is a high-profit commodity based on the characteristics of the transaction data after time-series aggregation and the principal component multimodal fusion representation characteristics between the semantic characteristics of the first commodity description, and judge whether to increase the split-account ratio based on the evaluation results. In this way, it is possible to quickly respond to market changes, and reduce manual operations while deeply understanding the characteristics and market performance of commodities, which is conducive to improving the intelligence of merchant marketing split-account processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0032] Figure 1 It is a flowchart of an AI-based merchant marketing account splitting processing method according to an embodiment of the present application.

[0033] Figure 2 It is a system block diagram of an AI-based merchant marketing account splitting processing system according to an embodiment of the present application. DETAILED DESCRIPTION

[0034] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0035] In modern e-commerce and retail, merchant marketing account splitting is a key link, which determines how to reasonably distribute sales revenue to various channels and partners. This process is crucial to the merchant's profit model and marketing effectiveness, and also affects the enthusiasm of partners. However, traditional account splitting methods mainly rely on manual operations and basic data analysis, which is not only inefficient and error-prone, but also difficult to conduct comprehensive and in-depth analysis. In addition, traditional methods are limited by historical data and fixed rules, lack the ability to respond immediately to market dynamics, especially in special circumstances such as surges in demand, it is difficult to adjust the account splitting strategy in time, and may miss the opportunity to optimize profit distribution. At the same time, these methods often only focus on superficial indicators such as sales and profits, ignoring the deep semantic associations in transaction data and product descriptions, resulting in the accuracy and fairness of account splitting being affected.

[0036] Based on this, the present application proposes an AI-based merchant marketing account splitting processing solution, which extracts the time queue of the transaction data of the first commodity, extracts the information description of the first commodity, and uses AI-based data analysis and processing technology to embed code and time-series aggregation of each transaction data, and semantically encodes the information description of the first commodity, so as to intelligently evaluate whether the first commodity is a high-profit commodity based on the principal component multimodal fusion representation features between the time-series aggregation features of the transaction data and the semantic features of the first commodity description, and judge whether to increase the account splitting ratio based on the evaluation results. In this way, it is possible to respond quickly to market changes, so that account splitting decisions can keep up with market dynamics and improve the timeliness of account splitting. At the same time, it is possible to understand the characteristics and market performance of commodities from multiple dimensions, provide a more comprehensive basis for account splitting, thereby reducing the need for manual operation and improving the intelligence of merchant marketing account splitting processing.

[0037] Figure 1 Flow chart of the merchant marketing account splitting processing method based on AI according to an embodiment of the present application. Figure 1 As shown, according to the embodiment of the present application, the merchant marketing account splitting processing method based on AI includes: S110, extracting the time queue of the transaction data of the first commodity; S120, extracting the information description of the first commodity; S130, semantically encoding the information description of the first commodity to obtain the semantic encoding feature vector of the first commodity information description; S140, using the transaction embedding coding matrix to embed the respective transaction data in the time queue of the transaction data of the first commodity to obtain the time queue of the transaction data embedding coding vector; S150, inputting the time queue of the transaction data embedding coding vector into the feature aggregation analysis network guided by the significance attenuation of the node features to obtain the first commodity transaction time series aggregation representation vector; S160, inputting the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector into the significance-driven principal component feature fusion network to obtain the first commodity transaction-attribute principal component multimodal fusion representation vector; S170, obtaining an evaluation result based on the first commodity transaction-attribute principal component multimodal fusion representation vector, and generating an account splitting recommendation suggestion based on the evaluation result.

[0038] In step S110 and step S120, the time queue of the transaction data of the first commodity and the information description of the first commodity are extracted respectively. It should be understood that the time queue of the transaction data of the first commodity contains the transaction records of the commodity within a certain time range. These records describe the specific circumstances of each transaction, including information such as transaction time, transaction amount, and transaction quantity; and the information description of the first commodity contains information such as the material, purpose, type, and price of the commodity. In general, the time queue of the first commodity transaction data provides real-time information about the transaction dynamics of the commodity, which can help analyze the current market demand and sales trends. The information description of the first commodity provides basic information and market positioning of the commodity, which helps to understand the performance of different commodities under different market conditions. By performing deep feature extraction and fusion on these two types of data, a scientific basis can be provided for the final account splitting decision, thereby optimizing the merchant's income distribution strategy.

[0039] In step S130, the information description of the first commodity is semantically encoded to obtain a semantically encoded feature vector of the first commodity information description. Accordingly, considering that the transaction data of the first commodity contains semantic information about the commodity, such as the type, purpose, material, etc. of the commodity. In order to better extract the key semantic information in the commodity description, in the technical solution of the present application, the information description of the first commodity is semantically encoded to obtain a semantically encoded feature vector of the first commodity information description, so that the keywords and concepts in the commodity description can be understood more accurately, thereby better evaluating the potential value of the commodity.

[0040] In step S140, each transaction data in the time queue of the transaction data of the first commodity is embedded and encoded using a transaction embedding coding matrix to obtain a time queue of transaction data embedded coding vectors. It should be understood that each transaction data in the time queue of the transaction data of the first commodity contains key transaction information, such as transaction volume, transaction frequency, transaction time, etc., and the transaction data usually contains a large number of dimensions. Therefore, in order to reduce the computational complexity of the data and to be able to mine and extract key transaction data information, in the technical solution of the present application, each transaction data in the time queue of the transaction data of the first commodity is embedded and encoded using a transaction embedding coding matrix to obtain a time queue of transaction data embedded coding vectors. The transaction embedding coding matrix is ​​a dynamically learned matrix that can map discrete commodity transaction data to a continuous vector space to obtain a low-dimensional embedded coding vector representation. By processing the transaction embedding coding matrix, a dense representation of the original commodity transaction data can be obtained, which makes subsequent calculations and analysis more efficient.

[0041] In step S150, the time queue of the transaction data embedded in the coding vector is input into a feature aggregation analysis network guided by the node feature significance decay to obtain the first commodity transaction time series aggregation representation vector. Accordingly, considering that the degree of correlation and influence between the transaction data embedded in the coding vector at different time points in the time queue of the transaction data embedded in the coding vector are different, and in order to more accurately and meticulously characterize the significant features of the transaction data between different time spans, so as to capture the overall trend and pattern of the entire time series, in the technical solution of the present application, the time queue of the transaction data embedded in the coding vector is input into a feature aggregation analysis network guided by the node feature significance decay to obtain the first commodity transaction time series aggregation representation vector. In particular, the feature aggregation analysis network guided by the node feature significance decay is a technology that extracts and integrates features using a method based on feature significance analysis. In detail, first, the feature significance attenuation factor of each transaction data embedding coding vector is calculated. In particular, the feature significance attenuation factor is calculated based on the mean, expected value and variance of each transaction data embedding coding vector. In this way, the contribution and importance of each transaction data feature to the overall data can be quantified. The mean, variance and expected value are key statistics for describing data distribution. By calculating the attenuation factor through these statistics, representative transaction data features can be identified, thereby providing a basis for subsequent feature aggregation. Then, based on the distance span between each transaction data embedding coding vector and the current transaction data embedding coding vector, a feature significance attenuation factor is constructed to indicate that with the continuous changes over time, recent transaction data can better reflect the current market conditions and consumer preferences, and give recent transactions a higher weight. Then, the feature significance attenuation factor of each feature vector is applied to its feature significance description factor to modulate it to reflect the importance of the feature at the current moment, thereby dynamically adjusting the contribution of each feature. Finally, each modulated factor is input into the gated mask module to convert it into a weight factor to perform weighted summation on the time queue of the transaction data embedded in the encoding vector to obtain the first commodity transaction time series aggregation representation vector that finally integrates the entire time domain.

[0042] Specifically, in an embodiment of the present application, the time queue of the transaction data embedded coding vector is input into a feature aggregation analysis network guided by node feature significance decay to obtain a first commodity transaction time series aggregation representation vector, including: calculating the feature significance description factor of each transaction data embedded coding vector in the time queue of the transaction data embedded coding vector, wherein the feature significance description factor is related to the mean and variance of each transaction data embedded coding vector; constructing the feature significance decay factor of each transaction data embedded coding vector based on the distance span between each transaction data embedded coding vector in the time queue of the transaction data embedded coding vector and the current transaction data embedded coding vector; modulating the feature significance description factor of each transaction data embedded coding vector based on the feature significance decay factor of each transaction data embedded coding vector to obtain a sequence of transaction data feature significance decay description factors; and modulating the transaction data feature significance decay description factor. The sequence of the factors is input into the gated mask module to obtain a sequence of transaction data feature significance attenuation weight factors; the sequence of transaction data feature significance attenuation weight factors is used as the sequence of weights, and the weighted sum of the time queue of the transaction data embedded coding vector is calculated to obtain the first commodity transaction time series aggregation representation vector; wherein, the sequence of transaction data feature significance attenuation description factors is input into the gated mask module to obtain a sequence of transaction data feature significance attenuation weight factors, including: performing normalization processing based on the Sigmoid function on each transaction data feature significance attenuation description factor in the sequence of transaction data feature significance attenuation description factors to obtain a set of normalized transaction data feature significance attenuation description factors; each normalized transaction data feature significance attenuation description factor in the set of normalized transaction data feature significance attenuation description factors is input into the gated function for masking processing to obtain the sequence of transaction data feature significance attenuation weight factors.

[0043] More specifically, in an embodiment of the present application, the characteristic significance description factor of each transaction data embedded coding vector in the time queue of the transaction data embedded coding vector is calculated, including: respectively calculating the mean and variance of the transaction data embedded coding vector to obtain the transaction data characteristic mean and the transaction data characteristic variance; calculating the positional difference between the transaction data embedded coding vector and the transaction data characteristic mean to obtain a transaction data differential vector; calculating the fourth power of each eigenvalue in the transaction data differential vector to obtain a transaction data modulation difference vector; calculating the expected value of the transaction data modulation difference vector to obtain a transaction data expected value; dividing the transaction data expected value by the square of the transaction data characteristic variance to obtain the characteristic significance description factor corresponding to the transaction data embedded coding vector.

[0044] More specifically, in an embodiment of the present application, based on the distance span between each transaction data embedded coding vector in the time queue of the transaction data embedded coding vector and the current transaction data embedded coding vector, a feature significance attenuation factor of each transaction data embedded coding vector is constructed, including: extracting the maximum value in each transaction data embedded coding vector to obtain a sequence of transaction data feature maximum values; extracting the maximum value of the current transaction data embedded coding vector to obtain the current transaction data feature maximum value; calculating the positional subtraction between the current transaction data feature maximum value and the sequence of transaction data feature maximum values ​​to obtain a sequence of transaction data feature offset values; counting the distance span values ​​between each transaction data embedded coding vector and the current transaction data embedded coding vector to obtain a sequence of transaction data distance span values, and calculating the squares of each value in the sequence of transaction data distance span values ​​to obtain a sequence of transaction data distance span modulation values; dividing the sequence of transaction data feature offset values ​​by the sequence of transaction data distance span modulation values ​​by position to obtain the feature significance attenuation factor of each transaction data embedded coding vector.

[0045] In the embodiment of the present application, specifically, the time queue of the transaction data embedded in the coding vector is input into the feature aggregation analysis network guided by the node feature significance attenuation to obtain the first commodity transaction time series aggregation representation vector, which is used to: input the time queue of the transaction data embedded in the coding vector into the feature aggregation analysis network guided by the node feature significance attenuation, and process it with the following feature aggregation formula to obtain the first commodity transaction time series aggregation representation vector; wherein the feature aggregation formula is:

[0046] X={x1,x2,...,x k ,...,x n}

[0047]

[0048] w i =sigmod(k i )

[0049] ws i =mask(w i )

[0050]

[0051] Where X represents the time queue of the transaction data embedded in the encoding vector, x k 、x n and x iare the kth, nth and ith transaction data embedding encoding vectors in the time queue of the transaction data embedding encoding vector respectively, max(·) is the maximum value of the extracted vector, Count(x n →x i ) represents x n With x i The distance span between i is the feature significance attenuation factor of the i-th transaction data embedding encoding vector, x i (j) is the eigenvalue of the jth position of the i-th transaction data embedding encoding vector, μ i and σ i 4 are the mean and square of the variance of the i-th transaction data embedding encoding vector, k i is the transaction data feature significance attenuation description factor corresponding to the i-th transaction data embedding encoding vector, sigmoid(·) is the sigmoid function, and w i is the ith normalized transaction data feature significance attenuation description factor in the set of normalized transaction data feature significance attenuation description factors, mask(·) is masking processing, ws i is the i-th transaction data feature significance attenuation weight factor in the sequence of transaction data feature significance attenuation weight factors, θ is a predetermined threshold, and V1 is the first commodity transaction time series aggregation representation vector.

[0052] In step S160, the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector are input into the saliency-driven principal component feature fusion network to obtain the first commodity transaction-attribute principal component multimodal fusion representation vector. It should be understood that the first commodity transaction time series aggregation representation vector expresses the salient transaction feature information of the transaction data in the entire time domain. The first commodity information description semantic encoding feature vector expresses the semantic feature information related to the commodity description. In order to fuse the features from different data sources (transaction data and commodity descriptions) so as to be able to more comprehensively understand the attributes and market performance of the commodity, to capture the complex characteristics of the commodity, and to improve the depth and breadth of the analysis. In the technical solution of the present application, the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector are input into the saliency-driven principal component feature fusion network to obtain the first commodity transaction-attribute principal component multimodal fusion representation vector. It is worth mentioning that the saliency-driven principal component feature fusion network aims to create a multimodal feature relationship representation of the key feature vector by using feature sparse processing, principal component analysis and saliency feature fusion technology. Specifically, the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector are first standardized to eliminate the influence of different dimensions and magnitudes. Then, the standardized sample covariance matrix is ​​calculated to capture the autocorrelation feature information of each, and the first commodity transaction time series aggregation sample covariance matrix and the first commodity information description semantic sample covariance matrix are obtained, and the sample covariance matrix is ​​subjected to matrix decomposition-based feature vector extraction to obtain a set of first commodity transaction time series aggregation principal component feature vectors and a set of first commodity information description principal component feature vectors. Subsequently, the cosine similarity between each first commodity transaction time series aggregation principal component feature vector and each first commodity information description principal component feature vector is calculated to take the maximum cosine similarity as the best matching pair of the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector. Furthermore, the semantic association relationship between each best matching pair is calculated to capture and extract the semantic interaction relationship between the transaction data and the commodity description data to obtain a set of first commodity transaction-information description principal component fusion feature vectors. Finally, the set of principal component feature vectors is cascaded to generate a first commodity transaction-attribute principal component multimodal fusion representation vector containing commodity transaction semantic information and description semantic information.

[0053] Specifically, in an embodiment of the present application, the first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector are input into a saliency-driven principal component feature fusion network to obtain a first commodity transaction-attribute principal component multimodal fusion representation vector, including: standardizing the first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector to obtain a standardized first commodity transaction time series aggregation representation vector and a standardized first commodity information description semantic coding feature vector; calculating the sample covariance matrix of the standardized first commodity transaction time series aggregation representation vector and the standardized first commodity information description semantic coding feature vector to obtain a first commodity transaction time series aggregation sample covariance matrix and a first commodity information description semantic sample covariance matrix; performing matrix decomposition-based feature vector extraction on the first commodity transaction time series aggregation sample covariance matrix and the first commodity information description semantic sample covariance matrix to obtain a set of first commodity transaction time series aggregation principal component feature vectors and a set of first commodity information description principal component feature vectors; inputting the set of the first commodity transaction time series aggregation principal component feature vectors and the set of the first commodity information description principal component feature vectors into a maximum approximate query matching network to obtain the first commodity transaction time series aggregation principal component feature vectors. The best matching pairs of the vector and the first commodity information description principal component feature vector are obtained; each best matching pair of the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector in the set of the best matching pairs of the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector is input into the semantic fine-grained gated joint module to obtain a set of first commodity transaction-information description principal component fusion feature vectors; the set of the first commodity transaction-information description principal component fusion feature vectors is cascaded to obtain the first commodity transaction-attribute principal component multimodal fusion representation vector; wherein, the calculation of the The sample covariance matrix of the standardized first commodity transaction time series aggregation representation vector and the standardized first commodity information description semantic encoding feature vector is obtained to obtain the first commodity transaction time series aggregation sample covariance matrix and the first commodity information description semantic sample covariance matrix, including: multiplying the transposed vector of the standardized first commodity transaction time series aggregation representation vector with the standardized first commodity transaction time series aggregation representation vector, and then dividing the obtained standardized first commodity transaction time series aggregation association matrix by the value obtained by subtracting one from the length of the standardized first commodity transaction time series aggregation representation vector by position to obtain the first commodity transaction time series aggregation sample covariance matrix;After multiplying the transposed vector of the standardized first product information description semantic coding feature vector by the standardized first product information description semantic coding feature vector, the obtained standardized first product information description semantic association matrix is ​​divided by the value obtained by subtracting one from the length of the standardized first product information description semantic coding feature vector to obtain the first product information description semantic sample covariance matrix. ;

[0054] More specifically, in an embodiment of the present application, the first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector are standardized to obtain a standardized first commodity transaction time series aggregation representation vector and a standardized first commodity information description semantic coding feature vector, including: respectively calculating the mean and standard deviation of the first commodity transaction time series aggregation representation vector to obtain the mean of the first commodity transaction time series aggregation feature and the standard deviation of the first commodity transaction time series aggregation feature; subtracting the first commodity transaction time series aggregation representation vector from the first commodity transaction time series aggregation feature mean by position, and calculating the first commodity transaction time series offset vector The standardized first commodity transaction time series aggregation representation vector is obtained by positionally dividing the standard deviation of the first commodity transaction time series aggregation feature; the mean and standard deviation of the first commodity information description semantic coding feature vector are calculated respectively to obtain the first commodity information description semantic feature mean and the first commodity information description semantic feature standard deviation; after subtracting the first commodity information description semantic coding feature vector from the first commodity information description semantic feature mean by position, the calculated first commodity information description semantic offset vector is divided by the standard deviation of the first commodity information description semantic feature to obtain the standardized first commodity information description semantic coding feature vector.

[0055] More specifically, in an embodiment of the present application, a set of the first commodity transaction time series aggregated principal component feature vectors and a set of the first commodity information description principal component feature vectors are input into a maximum approximate query matching network to obtain a set of best matching pairs of the first commodity transaction time series aggregated principal component feature vectors and the first commodity information description principal component feature vectors, including: extracting a predetermined first commodity transaction time series aggregated principal component feature vector from the set of the first commodity transaction time series aggregated principal component feature vectors; calculating the cosine similarity between the predetermined first commodity transaction time series aggregated principal component feature vector and each first commodity information description principal component feature vector in the set of the first commodity information description principal component feature vectors to obtain a set of matching query similarities; and taking the first commodity information description principal component feature vector corresponding to the largest matching query similarity in the set of matching query similarities and the predetermined first commodity transaction time series aggregated principal component feature vector as the best matching pair of the predetermined first commodity transaction time series aggregated principal component feature vector and the first commodity information description principal component feature vector.

[0056] More specifically, in an embodiment of the present application, each best matching pair of the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector in the set of best matching pairs of the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector is input into the semantic fine-grained gated joint module to obtain a set of first commodity transaction-information description principal component fusion feature vectors, including: respectively calculating the position difference, position dot multiplication and position addition between the best matching pairs of the first commodity transaction time series aggregation principal component feature vector and the first commodity information description principal component feature vector to obtain the first commodity transaction. Yi - information description principal component difference vector, the first commodity transaction - information description principal component dot product vector and the first commodity transaction - information description principal component sum vector; the first commodity transaction - information description principal component difference vector, the first commodity transaction - information description principal component dot product vector and the first commodity transaction - information description principal component sum vector are cascaded and then one-dimensional convolutional coding is performed to obtain the first commodity transaction - information description principal component multi-dimensional fusion vector; the first commodity transaction - information description principal component multi-dimensional fusion vector is subjected to local window-based maximum pooling processing to obtain the first commodity transaction - information description principal component fusion feature vector.

[0057] In the embodiment of the present application, specifically, the first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector are input into the saliency-driven principal component feature fusion network to obtain the first commodity transaction-attribute principal component multimodal fusion representation vector, which is used to: input the first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector into the saliency-driven principal component feature fusion network, and process them with the following principal component feature fusion formula to obtain the first commodity transaction-attribute principal component multimodal fusion representation vector; wherein the principal component feature fusion formula is:

[0058]

[0059] C1=U1Λ1U1 T

[0060] C2=U2Λ2U2 T

[0061] U1=[v 11 ,v 12 ,…,v 1m ]

[0062] Λ1=diag(λ 11 ,λ 12 ,…,λ 1m )

[0063] U2=[v 21 ,v 22 ,…,v 2m ]

[0064] Λ2=diag(λ 21 ,λ 22 ,…,λ 2m )

[0065]

[0066] v f =[v p1 ;v p2 ;...v pm ]

[0067] Wherein, V1 and V2 represent the first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector, respectively; μ1 and μ2 represent the means of the first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector, respectively; σ1 and σ2 represent the standard deviations of the first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector, respectively; V 1norm and V 2normare the standardized first commodity transaction time series aggregation representation vector and the standardized first commodity information description semantic encoding feature vector, respectively, 1norm T and V 2norm T V 1norm and V 2norm The transposed vector of the first commodity transaction time series aggregation representation vector and the length of the first commodity information description semantic encoding feature vector are n1 and n2 respectively. C1 and C2 are the first commodity transaction time series aggregation sample covariance matrix and the first commodity information description semantic sample covariance matrix respectively. U1 and U2 are the first commodity transaction time series aggregation principal component orthogonal matrix and the first commodity information description principal component orthogonal matrix respectively. Λ1 and Λ2 are the first commodity transaction time series aggregation diagonal matrix and the first commodity information description diagonal matrix respectively. diag(λ 11 ,λ 12 ,…,λ 1m ) is the matrix whose diagonal elements are λ 11 ,λ 12 ,…,λ 1m The first commodity transaction time series aggregation diagonal matrix, λ 11 ,λ 12 ,…,λ 1m are the weight values ​​of the principal component eigenvectors of the first commodity transaction time series aggregation, diag(λ 21 ,λ 22 ,…,λ 2m ) is the matrix whose diagonal elements are λ 21 ,λ 22 ,…,λ 2m The first product information description diagonal matrix, λ 21 ,λ 22 ,…,λ 2m are the weight values ​​of the principal component feature vectors describing the first product information, U1 T and U2 T are the transposed matrices of U1 and U2, respectively, 11 ,v 12 ,…,v 1m is the principal component feature vector of each first commodity transaction time series aggregation in the set of principal component feature vectors of the first commodity transaction time series aggregation, v 21 ,v 22 ,…,v 2m is each first product information description principal component feature vector in the set of the first product information description principal component feature vectors, <v 1i ·v 2j >To calculate the v 1i and v 2jThe inner product of the vectors between ||·|| is the norm of the calculated vector, To return the j value corresponding to the maximum value, k is the maximum approximate matching value. ⊙ and are position difference, position point multiplication and position addition respectively, [·,·,·] is cascade processing, conv1D(·) is a one-dimensional convolutional encoding operation, MaxPool(·) is a maximum pooling operation, v pi is the i-th first commodity transaction-information description principal component fusion feature vector in the set of the first commodity transaction-information description principal component fusion feature vectors, m is the number of feature vectors in the set of the first commodity transaction-information description principal component fusion feature vectors, v f It is the first commodity transaction-attribute principal component multimodal fusion representation vector.

[0068] In step S170, based on the first commodity transaction-attribute principal component multimodal fusion representation vector, an evaluation result is obtained, and based on the evaluation result, a split-account recommendation suggestion is generated. Specifically, in an embodiment of the present application, based on the first commodity transaction-attribute principal component multimodal fusion representation vector, an evaluation result is obtained, and based on the evaluation result, a split-account recommendation suggestion is generated, including: inputting the first commodity transaction-attribute principal component multimodal fusion representation vector into a classifier-based commodity type evaluator to obtain the evaluation result, and the evaluation result is used to indicate whether the first commodity is a high-profit commodity; in response to the evaluation result that the first commodity is a high-profit commodity, the split-account recommendation suggestion is generated, and the split-account recommendation suggestion is used to indicate an increase in the split-account ratio. That is, the first commodity transaction-attribute principal component multimodal fusion representation vector obtained by performing principal component multimodal significant fusion using the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector is classified, so as to intelligently evaluate whether the first commodity is a high-profit commodity, and based on the evaluation result, determine whether to increase the split-account ratio. In this way, it is possible to respond quickly to market changes, so that account splitting decisions can keep up with market dynamics and improve the timeliness of account splitting. At the same time, it is possible to understand product characteristics and market performance from multiple dimensions, provide a more comprehensive basis for account splitting, thereby reducing the need for manual operations and improving the intelligence of merchant marketing account splitting processing.

[0069] Preferably, considering that the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector respectively represent the global time domain guided aggregation features of the transaction data of the first commodity based on the semantic embedding encoding feature time series node significance attenuation and the text semantic context encoding features of the information description of the first commodity, therefore, when performing the principal component feature fusion based on significance-driven, the feature principal component query matching significance difference caused by the feature modality and feature dimension differences will lead to the semantic aggregation fine-grained alignment conflict, thereby causing the loss of aggregation key information and affecting the expression effect of the first commodity transaction-attribute principal component multimodal fusion representation vector.

[0070] Based on this, in the preferred example, when the first commodity transaction-attribute principal component multimodal fusion representation vector is input into the classifier-based commodity type evaluator to obtain an evaluation result, the first commodity transaction-attribute principal component multimodal fusion representation vector is optimized. Specifically, the optimization process includes the following steps: clustering all feature values ​​of the first commodity transaction-attribute principal component multimodal fusion representation vector based on the distance between feature values, such as the absolute value distance of the difference, and determining the ratio of the number of features of the clustering features to the number of features of the first commodity transaction-attribute principal component multimodal fusion representation vector; arranging the clustering features into a first commodity transaction-attribute principal component multimodal fusion clustering vector, and dividing the binary norm of the first commodity transaction-attribute principal component multimodal fusion clustering vector by the binary norm of the first commodity transaction-attribute principal component multimodal fusion representation vector to obtain the first commodity transaction-attribute principal component multimodal fusion clustering related weight value; dividing the first power value of the first commodity transaction-attribute principal component multimodal fusion clustering vector with the first norm as the exponent of the ratio value by the A norm of the first commodity transaction-attribute principal component multimodal fusion representation vector is raised to a second power value with the ratio value as an exponent to obtain a second commodity transaction-attribute principal component multimodal fusion clustering-related weight value; for each eigenvalue of the first commodity transaction-attribute principal component multimodal fusion clustering vector, multiply it by the inverse of the difference between the first and second commodity transaction-attribute principal component multimodal fusion clustering-related weight values ​​to obtain the optimized eigenvalue of the first commodity transaction-attribute principal component multimodal fusion clustering vector; for each eigenvalue outside the cluster in the first commodity transaction-attribute principal component multimodal fusion representation vector, multiply it by the inverse of the sum of the first and second commodity transaction-attribute principal component multimodal fusion clustering-related weight values ​​to obtain the optimized out-of-class eigenvalue of the first commodity transaction-attribute principal component multimodal fusion representation vector; the optimized eigenvalue of the first commodity transaction-attribute principal component multimodal fusion clustering vector and the optimized out-of-class eigenvalue of the first commodity transaction-attribute principal component multimodal fusion representation vector are combined to form an optimized first commodity transaction-attribute principal component multimodal fusion representation vector.

[0071] That is, the optimization process is expressed as:

[0072]

[0073] Wherein, f is each eigenvalue of the first commodity transaction-attribute principal component multimodal fusion clustering vector, V n is the first commodity transaction-attribute principal component multimodal fusion representation vector, is the first commodity transaction-attribute principal component multimodal fusion clustering vector, n is the number of features of the first commodity transaction-attribute principal component multimodal fusion representation vector, k is the number of features of the first commodity transaction-attribute principal component multimodal fusion clustering vector, λ is the ratio of the number of features of the first commodity transaction-attribute principal component multimodal fusion clustering vector to the number of features of the first commodity transaction-attribute principal component multimodal fusion representation vector, represents the clustering feature set corresponding to the first commodity transaction-attribute principal component multimodal fusion clustering vector, ||·||2 and They respectively represent the λth power of the second norm and the first norm of the vector, and f' is each eigenvalue of the optimized first commodity transaction-attribute principal component multimodal fusion representation vector.

[0074] Here, in order to avoid the loss of key suffix semantic information relative to the original feature set as a whole due to aggregation conflict because the first commodity transaction-attribute principal component multimodal fusion feature vector is based on aggregated features, the clustering ratio of the number of features of the first commodity transaction-attribute principal component multimodal fusion clustering vector relative to the number of features of the first commodity transaction-attribute principal component multimodal fusion representation vector is used as a decision function to perform an adversarial decision on the set absolute representation of the one-norm of the first commodity transaction-attribute principal component multimodal fusion clustering vector and the first commodity transaction-attribute principal component multimodal fusion representation vector, and respectively compare the first commodity transaction-attribute principal component multimodal fusion clustering vector and the first commodity transaction-attribute principal component multimodal fusion representation vector with the clustering ratio of the number of features of the first commodity transaction-attribute principal component multimodal fusion clustering vector and the first commodity transaction-attribute principal component multimodal fusion representation vector. The clustering intrinsic conflict representation of the second norm of the first commodity transaction-attribute principal component multimodal fusion representation vector is positively and negatively interacted to construct the optimized first commodity transaction-attribute principal component multimodal fusion representation vector based on the aggregated features and the solid alignment guardrail of the original feature set as a whole, so as to achieve the mitigation of the intention of harmful information loss based on the aggregated risk transferability of the optimized first commodity transaction-attribute principal component multimodal fusion representation vector, and improve the expression effect of the optimized first commodity transaction-attribute principal component multimodal fusion representation vector, so as to improve the accuracy of the evaluation result obtained by inputting the first commodity transaction-attribute principal component multimodal fusion feature vector into the classifier-based commodity type evaluator. In this way, it is possible to respond quickly to market changes, so that the account splitting decision can keep up with market dynamics and improve the timeliness of account splitting. At the same time, it is possible to understand the characteristics of commodities and market performance from multiple dimensions, provide a more comprehensive basis for account splitting, thereby reducing the need for manual operation and improving the intelligence of merchant marketing account splitting processing.

[0075] In summary, the AI-based merchant marketing split-account processing method based on the embodiment of the present application is explained, which uses AI-based data analysis and processing technology to embed the first commodity transaction data and time-series aggregation, and semantically encode the information description of the first commodity, so as to intelligently evaluate whether the first commodity is a high-profit commodity based on the principal component multimodal fusion representation features between the time-series aggregation features of the transaction data and the semantic features of the first commodity description, and judge whether to increase the split-account ratio based on the evaluation results. In this way, it is possible to quickly respond to market changes, and reduce manual operations while deeply understanding the characteristics and market performance of the commodities, which is conducive to improving the intelligence of merchant marketing split-account processing.

[0076] Figure 2 : is a system block diagram of an AI-based merchant marketing account splitting processing system according to an embodiment of the present application. Figure 2As shown, according to the embodiment of the present application, the merchant marketing account splitting processing system 100 based on AI includes: a commodity transaction time queue data extraction module 110, which is used to extract the time queue of the transaction data of the first commodity; a commodity information description extraction module 120, which is used to extract the information description of the first commodity; a commodity information description semantic coding module 130, which is used to semantically encode the information description of the first commodity to obtain the semantic coding feature vector of the first commodity information description; a commodity transaction time queue data embedding coding module 140, which is used to use the transaction embedding coding matrix to embed the respective transaction data in the time queue of the transaction data of the first commodity to obtain the time queue of the transaction data embedding coding vector; a merchant A commodity transaction time series feature set analysis module 150 is used to input the time queue of the transaction data embedded in the coding vector into a feature aggregation analysis network guided by node feature significance attenuation to obtain a first commodity transaction time series aggregation representation vector; a commodity multimodal feature fusion module 160 is used to input the first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector into a significance-driven principal component feature fusion network to obtain a first commodity transaction-attribute principal component multimodal fusion representation vector; and a split-account recommendation suggestion generation module 170 is used to obtain an evaluation result based on the first commodity transaction-attribute principal component multimodal fusion representation vector, and generate a split-account recommendation suggestion based on the evaluation result.

[0077] Here, those skilled in the art can understand that the specific functions and operations of each unit and module in the above-mentioned AI-based merchant marketing account splitting processing system 100 have been referred to above. Figure 1 The AI-based merchant marketing account splitting processing method has been introduced in detail, and therefore, its repeated description will be omitted.

[0078] In summary, the AI-based merchant marketing split-account processing system 100 based on the embodiment of the present application is explained, which uses AI-based data analysis and processing technology to embed the first commodity transaction data and time series aggregation, and semantically encode the information description of the first commodity, so as to intelligently evaluate whether the first commodity is a high-profit commodity based on the characteristics of the transaction data after time series aggregation and the principal component multimodal fusion representation characteristics between the semantic features of the first commodity description, and judge whether to increase the split-account ratio based on the evaluation results. In this way, it is possible to respond quickly to market changes, and reduce manual operations while deeply understanding the characteristics and market performance of the commodities, which is conducive to improving the intelligence of merchant marketing split-account processing.

Claims

1. A merchant marketing account splitting processing method based on AI, characterized in that: include: Extracting the time queue of the transaction data of the first commodity; Extracting information description of the first product; Performing semantic coding on the information description of the first product to obtain a semantic coding feature vector of the information description of the first product; Using the transaction embedding coding matrix, each transaction data in the time queue of the transaction data of the first commodity is respectively embedded and coded to obtain a time queue of transaction data embedding coding vectors; The time queue of the transaction data embedded in the coding vector is input into a feature aggregation analysis network guided by node feature significance decay to obtain a first commodity transaction time series aggregation representation vector; Inputting the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector into a saliency-driven principal component feature fusion network to obtain a first commodity transaction-attribute principal component multimodal fusion representation vector; Based on the first commodity transaction-attribute principal component multimodal fusion representation vector, an evaluation result is obtained, and based on the evaluation result, an account splitting recommendation suggestion is generated.

2. The AI-based merchant marketing account splitting processing method according to claim 1, characterized in that: The time queue of embedding the transaction data into the encoding vector is input into a feature aggregation analysis network guided by node feature significance decay to obtain a first commodity transaction time series aggregation representation vector, including: Calculating a feature significance description factor of each transaction data embedding code vector in the time queue of the transaction data embedding code vector, wherein the feature significance description factor is related to the mean and variance of each transaction data embedding code vector; Constructing a feature significance attenuation factor of each transaction data embedding coding vector based on a distance span between each transaction data embedding coding vector in the time queue of the transaction data embedding coding vector and the current transaction data embedding coding vector; Modulating the feature significance description factor of each transaction data embedded coding vector based on the feature significance attenuation factor of each transaction data embedded coding vector to obtain a sequence of transaction data feature significance attenuation description factors; Inputting the sequence of transaction data feature significance attenuation description factors into a gated mask module to obtain a sequence of transaction data feature significance attenuation weight factors; Taking the sequence of the transaction data feature significance attenuation weight factors as the sequence of weights, calculating the weighted sum of the time queues of the transaction data embedding encoding vectors to obtain the first commodity transaction time series aggregation representation vector; The sequence of transaction data feature significance attenuation description factors is input into a gated mask module to obtain a sequence of transaction data feature significance attenuation weight factors, including: Performing normalization processing based on a Sigmoid function on each transaction data feature significance attenuation description factor in the sequence of transaction data feature significance attenuation description factors to obtain a set of normalized transaction data feature significance attenuation description factors; Each normalized transaction data feature significance attenuation description factor in the set of normalized transaction data feature significance attenuation description factors is input into a gating function for masking processing to obtain a sequence of transaction data feature significance attenuation weight factors.

3. The AI-based merchant marketing account splitting processing method according to claim 2, characterized in that: Calculating the feature significance description factor of each transaction data embedding code vector in the time queue of the transaction data embedding code vector comprises: Respectively calculating the mean and variance of the transaction data embedding coding vector to obtain the transaction data feature mean and the transaction data feature variance; Calculating the positional difference between the transaction data embedding encoding vector and the transaction data feature mean to obtain a transaction data differential vector; Calculating the fourth power of each eigenvalue in the transaction data differential vector to obtain a transaction data modulation difference vector; Calculating the expected value of the transaction data modulation difference vector to obtain the transaction data expected value; The transaction data expected value is divided by the square of the transaction data feature variance to obtain a feature significance description factor corresponding to the transaction data embedding coding vector.

4. The AI-based merchant marketing account splitting processing method according to claim 3 is characterized in that: Based on the distance span between each transaction data embedding coding vector in the time queue of the transaction data embedding coding vector and the current transaction data embedding coding vector, constructing a feature significance attenuation factor of each transaction data embedding coding vector, including: Extracting the maximum value of each transaction data embedded in the encoding vector to obtain a sequence of maximum values ​​of transaction data features; Extracting the maximum value of the current transaction data embedding coding vector to obtain the maximum value of the current transaction data feature; Calculating the positional subtraction between the current transaction data feature maximum value and the sequence of transaction data feature maximum values ​​to obtain a sequence of transaction data feature offset values; Counting the distance span values ​​between each transaction data embedding coding vector and the current transaction data embedding coding vector to obtain a sequence of transaction data distance span values, and calculating the square of each value in the sequence of transaction data distance span values ​​to obtain a sequence of transaction data distance span modulation values; The sequence of the transaction data feature offset values ​​is divided by the sequence of the transaction data distance span modulation values ​​by position to obtain the feature significance attenuation factor of each transaction data embedded coding vector.

5. The AI-based merchant marketing account splitting processing method according to claim 4, characterized in that: Inputting the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector into a saliency-driven principal component feature fusion network to obtain a first commodity transaction-attribute principal component multimodal fusion representation vector, including: Standardizing the first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector to obtain a standardized first commodity transaction time series aggregation representation vector and a standardized first commodity information description semantic coding feature vector; Calculate the sample covariance matrix of the standardized first commodity transaction time series aggregate representation vector and the standardized first commodity information description semantic encoding feature vector to obtain the first commodity transaction time series aggregate sample covariance matrix and the first commodity information description semantic sample covariance matrix; Performing matrix decomposition-based feature vector extraction on the first commodity transaction time series aggregated sample covariance matrix and the first commodity information description semantic sample covariance matrix to obtain a set of first commodity transaction time series aggregated principal component feature vectors and a set of first commodity information description principal component feature vectors; Inputting the set of the first commodity transaction time series aggregated principal component feature vectors and the set of the first commodity information description principal component feature vectors into a maximum approximate query matching network to obtain a set of best matching pairs of the first commodity transaction time series aggregated principal component feature vectors and the first commodity information description principal component feature vectors; Input each best matching pair of the first commodity transaction time series aggregated principal component feature vector and the first commodity information description principal component feature vector in the set of best matching pairs of the first commodity transaction time series aggregated principal component feature vector and the first commodity information description principal component feature vector into the semantic fine-grained gated joint module to obtain a set of first commodity transaction-information description principal component fusion feature vectors; Cascading the set of the first commodity transaction-information description principal component fusion feature vectors to obtain the first commodity transaction-attribute principal component multimodal fusion representation vector; The step of calculating the sample covariance matrix of the standardized first commodity transaction time series aggregate representation vector and the standardized first commodity information description semantic encoding feature vector to obtain the first commodity transaction time series aggregate sample covariance matrix and the first commodity information description semantic sample covariance matrix includes: After multiplying the transposed vector of the standardized first commodity transaction time series aggregation representation vector by the standardized first commodity transaction time series aggregation representation vector, the obtained standardized first commodity transaction time series aggregation association matrix is ​​divided by the value obtained by subtracting one from the length of the standardized first commodity transaction time series aggregation representation vector to obtain the first commodity transaction time series aggregation sample covariance matrix; After multiplying the transposed vector of the standardized first product information description semantic encoding feature vector by the standardized first product information description semantic encoding feature vector, the obtained standardized first product information description semantic association matrix is ​​divided by the value obtained by subtracting one from the length of the standardized first product information description semantic encoding feature vector to obtain the first product information description semantic sample covariance matrix.

6. The AI-based merchant marketing account splitting processing method according to claim 5, characterized in that: The first commodity transaction time series aggregation representation vector and the first commodity information description semantic coding feature vector are standardized to obtain a standardized first commodity transaction time series aggregation representation vector and a standardized first commodity information description semantic coding feature vector, including: Respectively calculating the mean and standard deviation of the first commodity transaction time series aggregation representation vector to obtain the first commodity transaction time series aggregation feature mean and the first commodity transaction time series aggregation feature standard deviation; After subtracting the first commodity transaction time series aggregation representation vector from the first commodity transaction time series aggregation feature mean by position, the first commodity transaction time series offset vector calculated and the first commodity transaction time series aggregation feature standard deviation are divided by position to obtain the standardized first commodity transaction time series aggregation representation vector; Respectively calculating the mean and standard deviation of the semantic encoding feature vector of the first product information description to obtain the mean of the first product information description semantic feature and the standard deviation of the first product information description semantic feature; After subtracting the first product information description semantic coding feature vector from the first product information description semantic feature mean by position, the calculated first product information description semantic offset vector and the first product information description semantic feature standard deviation are divided by position to obtain the standardized first product information description semantic coding feature vector.

7. The AI-based merchant marketing account splitting processing method according to claim 6, characterized in that: Inputting the set of the first commodity transaction time series aggregated principal component feature vectors and the set of the first commodity information description principal component feature vectors into a maximum approximate query matching network to obtain a set of best matching pairs of the first commodity transaction time series aggregated principal component feature vectors and the first commodity information description principal component feature vectors, including: Extracting a predetermined first commodity transaction time series aggregated principal component feature vector from a set of the first commodity transaction time series aggregated principal component feature vectors; Calculating the cosine similarity between the predetermined first commodity transaction time series aggregated principal component feature vector and each first commodity information description principal component feature vector in the set of the first commodity information description principal component feature vectors to obtain a set of matching query similarities; The first product information description principal component feature vector corresponding to the largest matching query similarity in the set of matching query similarities and the predetermined first product transaction time series aggregated principal component feature vector are taken as the best matching pair of the predetermined first product transaction time series aggregated principal component feature vector and the first product information description principal component feature vector.

8. The AI-based merchant marketing account splitting processing method according to claim 7, characterized in that: Inputting each best matching pair of the first commodity transaction time series aggregated principal component feature vector and the first commodity information description principal component feature vector in the set of best matching pairs of the first commodity transaction time series aggregated principal component feature vector and the first commodity information description principal component feature vector into the semantic fine-grained gating joint module to obtain a set of first commodity transaction-information description principal component fusion feature vectors, including: Respectively calculating the position difference, position dot product and position addition between the best matching pair of the first commodity transaction time series aggregated principal component feature vector and the first commodity information description principal component feature vector to obtain the first commodity transaction-information description principal component difference vector, the first commodity transaction-information description principal component dot product vector and the first commodity transaction-information description principal component sum vector; The first commodity transaction-information description principal component difference vector, the first commodity transaction-information description principal component dot product vector and the first commodity transaction-information description principal component sum vector are cascaded and then one-dimensional convolutional encoded to obtain a first commodity transaction-information description principal component multi-dimensional fusion vector; The first commodity transaction-information description principal component multi-dimensional fusion vector is subjected to local window-based maximum pooling processing to obtain the first commodity transaction-information description principal component fusion feature vector.

9. The AI-based merchant marketing account splitting processing method according to claim 8, characterized in that: Based on the first commodity transaction-attribute principal component multimodal fusion representation vector, an evaluation result is obtained, and based on the evaluation result, a split account recommendation is generated, including: Inputting the first commodity transaction-attribute principal component multimodal fusion representation vector into a classifier-based commodity type evaluator to obtain the evaluation result, wherein the evaluation result is used to indicate whether the first commodity is a high-profit commodity; In response to the evaluation result that the first commodity is a high-profit commodity, the account sharing recommendation suggestion is generated, and the account sharing recommendation suggestion is used to indicate an increase in the account sharing ratio.

10. An AI-based merchant marketing account splitting processing system, characterized in that: include: A commodity transaction time queue data extraction module, used to extract the time queue of the transaction data of the first commodity; A commodity information description extraction module, used to extract the information description of the first commodity; A commodity information description semantic coding module, used for semantically coding the information description of the first commodity to obtain a semantic coding feature vector of the first commodity information description; A commodity transaction time queue data embedding coding module, used to use a transaction embedding coding matrix to embed code each transaction data in the time queue of the transaction data of the first commodity to obtain a time queue of transaction data embedding coding vectors; A commodity transaction time series feature set analysis module, used to embed the transaction data into a time queue of a coding vector and input it into a feature aggregation analysis network guided by node feature significance decay to obtain a first commodity transaction time series aggregation representation vector; A commodity multimodal feature fusion module, used for inputting the first commodity transaction time series aggregation representation vector and the first commodity information description semantic encoding feature vector into a saliency-driven principal component feature fusion network to obtain a first commodity transaction-attribute principal component multimodal fusion representation vector; The account splitting recommendation suggestion generation module is used to obtain an evaluation result based on the first commodity transaction-attribute principal component multimodal fusion representation vector, and generate an account splitting recommendation suggestion based on the evaluation result.

Images