A cross-channel e-commerce user unified portrait dynamic construction method

Abstract

The present application relates to the field of electronic commerce data processing, and more particularly to a cross-channel e-commerce user unified portrait dynamic construction method. It includes: collecting heterogeneous behavior data and standardizing it into an initial behavior feature vector, calculating its spatial covariance in a multi-dimensional feature space for cross-end alignment and fusion to construct a joint feature matrix, inputting it into a deep neural network to extract nonlinear structured features, constructing a time decay function based on fractional calculus to dynamically update the historical features to obtain a dynamic feature tensor, using Dirichlet distribution to model multi-dimensional probability and output a parameterized distribution set representing preference and variance as a portrait label. The present application fits the covariance soft association and fractional non-uniform decay, and outputs a probability distribution with a confidence interval, ensuring high-fidelity smooth evolution of the portrait features and enhancing the noise resistance of the downstream recommendation system.

Description

Technical Field

[0001] This invention relates to the field of e-commerce data processing, and in particular to a method for dynamically constructing a unified profile of e-commerce users across channels. Background Technology

[0002] With the rapid evolution of cross-channel e-commerce, user behavior patterns are widely distributed across multiple touchpoints, including web pages, mini-programs, and standalone clients. To achieve accurate personalized recommendations and refined operations, e-commerce platforms need to build unified user profiles across all platforms. This requires the underlying technical architecture to be capable of deep feature extraction and integration of heterogeneous, massive amounts of discrete interaction data, scientifically aligning fragmented behavioral data distributed in a multi-dimensional feature space, and then constructing a high-information-density joint feature matrix to comprehensively depict users' true consumption intentions and preference structures.

[0003] However, existing cross-platform profile building technologies rely heavily on strong identity identifiers such as mobile phone numbers for deterministic hard association. This approach not only struggles to cover massive amounts of anonymous interaction data from unlogged-in users, but also easily generates feature dimension redundancy during data splicing, leading to a decrease in the accuracy and information density of the underlying data fusion. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a method for dynamically constructing a unified profile of e-commerce users across channels, aiming to improve the problems of difficulty in anonymizing data coverage and low accuracy of feature fusion caused by the reliance on strong identity identifiers in existing technologies.

[0005] This invention provides the following technical solution: a method for dynamically constructing a unified user profile across e-commerce channels, comprising: S1. Collect heterogeneous behavioral data from different e-commerce channels, standardize the heterogeneous behavioral data, and generate an initial behavioral feature vector with a unified dimension. S2. Project the initial behavior feature vector onto a multi-dimensional feature space, calculate the spatial covariance of the initial behavior feature vector in the multi-dimensional feature space, and perform cross-channel alignment and fusion of the initial behavior feature vector based on the spatial covariance to construct a joint feature matrix. S3. Input the joint feature matrix into a deep neural network for nonlinear feature extraction and output a structured feature representation; S4. Construct a time decay function based on fractional calculus, use the time decay function to calculate the evolution weight of the structured feature representation on the time series, and dynamically update the structured feature representation of the historical period through the evolution weight to obtain a dynamic feature tensor. S5. Use the Dirichlet distribution to perform multidimensional probability modeling on the dynamic feature tensor, output the Dirichlet parameterized distribution set representing the multidimensional preference probability tendency and corresponding variance, and output the Dirichlet parameterized distribution set as a unified profile label.

[0006] Preferably, in step S1, the step of generating an initial behavioral feature vector of uniform dimension includes: Extract timestamps, event types, and entity attributes from the heterogeneous behavioral data to construct a multidimensional original behavioral data sequence; Using a pre-defined cross-domain mapping dictionary, the multidimensional original behavioral data sequence is transformed into continuous numerical features, and a global feature coordinate system is established. The continuous numerical features are subjected to missing value imputation and normalization, and then concatenated according to the global feature coordinate system to generate the initial behavioral feature vector of the unified dimension.

[0007] Preferably, in step S2, the step of calculating the spatial covariance of the initial behavioral feature vector in the multidimensional feature space includes: The initial behavioral feature vector is mapped to discrete observation points in the multidimensional feature space to establish the topological structure of the cross-channel feature distribution; Calculate the mean vector of the initial behavioral feature vectors within each channel, and extract the centered deviation matrix by subtracting the mean vector from the initial behavioral feature vectors; Calculate the inner product of the centered deviation matrix and its transpose matrix, and perform degree-of-freedom normalization to obtain the spatial covariance.

[0008] Preferably, in step S2, the step of aligning and fusing the initial behavioral feature vector across channels based on the spatial covariance includes: The off-diagonal elements of the spatial covariance are numerically compared with a preset association threshold, and cross-channel feature pairs corresponding to the off-diagonal elements that are greater than the preset association threshold are selected. The soft allocation algorithm is used to assign dynamic alignment weights to the cross-channel feature pairs based on the value of the spatial covariance. Based on the dynamic alignment weights, linear superposition and dimensionality reduction are performed on the associated initial behavioral feature vectors to construct the joint feature matrix.

[0009] Preferably, in step S3, the step of outputting the structured feature representation includes: The joint feature matrix is ​​received by the input layer of a deep neural network, and a forward propagation calculation of high-dimensional linear combination is performed through a multilayer perceptron architecture; A standard nonlinear activation function is introduced into the hidden layer to impose transformation constraints and extract nonlinear behavior features. The nonlinear behavioral features are compressed in dimension by the output layer and standardized by combining the layer normalization operator to output the structured feature representation.

[0010] Preferably, in step S4, the step of constructing the time decay function based on fractional calculus includes: The time evolution interval is divided into equally spaced discrete time steps, and the feature is extracted to generate the relative time difference between the current time point and the time point. A fractional integral operator is introduced to establish a memory evolution differential equation with the relative time difference as the independent variable, which characterizes the non-uniform decay characteristics. The memory evolution differential equation is solved using a numerical approximation algorithm to obtain the solution of the continuous decay function and discretize it to construct the time decay function.

[0011] Preferably, in step S4, the step of dynamically updating the structured feature representation of the historical period using the evolution weights includes: Read the structured feature representation of the historical period and parse its associated timeliness attributes; The evolution weights are used as scalar multipliers, and the corresponding elements are multiplied by the structured feature representation of the historical period to obtain the weighted historical matrix. The currently generated structured feature representation is fused with the weighted history matrix by matrix addition to obtain the dynamic feature tensor.

[0012] Preferably, in step S5, the step of performing multidimensional probabilistic modeling of the dynamic feature tensor using the Dirichlet distribution includes: The dynamic feature tensor is transformed into a positive real number vector using a smoothing mapping function, which serves as a prior pseudo-count representing user preferences. The positive real number vector is input as a concentration parameter vector into the standard Dirichlet probability density model to construct a conjugate prior probability model of the multinomial distribution. The latent variables of the conjugate prior probability model are solved using the Bayesian variational inference algorithm to establish the set of distribution parameters that fit the current data.

[0013] Preferably, in step S5, the step of outputting the Dirichlet parameterized distribution set representing the multidimensional preference probability tendency and corresponding variance includes: The mathematical expectation vector is calculated based on the set of distribution parameters obtained from the solution, forming the multidimensional preference probability tendency that describes the intensity of category interest; The autovariance and statistical covariance of each dimension feature are calculated using the set of distribution parameters to generate the corresponding variance of the quantized confidence interval; The multidimensional preference probability tendency and the corresponding variance are numerically packaged and encapsulated to generate a standardized feature package, which is then output as the Dirichlet parameterized distribution set.

[0014] The present invention has the following beneficial effects: 1. In this invention, cross-channel alignment and fusion are performed by calculating the spatial covariance in the multi-dimensional feature space, which improves the problem that traditional cross-end profiling is highly dependent on strong identity identifiers. It uses rigorous probability statistics to realize the underlying soft association of multi-domain heterogeneous data, effectively removes redundant dimensions without the need for hard matching, and improves the information density and fusion accuracy of the joint feature matrix.

[0015] 2. In this invention, fractional calculus operators are introduced to establish memory evolution differential equations to construct time decay functions. This avoids the bias distortion caused by the constant forgetting rate of traditional exponential models, fits the long-tail memory pattern of the non-uniform distribution of consumers' real shopping intentions, and ensures the high-fidelity smooth evolution of user profiles over time.

[0016] 3. In this invention, the Dirichlet distribution is used to probabilistically model dynamic features and output a parameterized distribution set, which improves the problem of overfitting that traditional absolute hard labels are prone to when facing low-frequency cold start users. The final output is upgraded to a joint statistical distribution that includes uncertainty measure and confidence interval, which enhances the noise resistance stability of downstream recommendation system calls. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for dynamically constructing a unified user profile across e-commerce channels, as proposed in this invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In embodiments of the present invention, the present invention provides a method for dynamically constructing a unified user profile across e-commerce channels, such as... Figure 1 As shown, it includes: S1. Collect heterogeneous behavioral data from different e-commerce channels, standardize the heterogeneous behavioral data, and generate an initial behavioral feature vector with a unified dimension. Further, step S1, the step of generating an initial behavioral feature vector of uniform dimension, includes: Extract timestamps, event types, and entity attributes from heterogeneous behavioral data to construct a multidimensional original behavioral data sequence; Using a pre-defined cross-domain mapping dictionary, multidimensional raw behavioral data sequences are transformed into continuous numerical features, and a global feature coordinate system is established. For continuous numerical features, perform missing value imputation and normalization, and then concatenate them according to the global feature coordinate system to generate an initial behavioral feature vector with a unified dimension.

[0020] Specifically, the system uses a distributed log collection engine to access data streams from web pages, third-party mini-programs, and independent mobile clients in real time, collecting heterogeneous behavioral data from different e-commerce channels. The system extracts timestamps, event types, and entity attributes from the underlying business logs to construct a multi-dimensional raw behavioral data sequence. In the e-commerce system, timestamps are recorded precisely to the millisecond level; event types are clearly defined as browsing product details pages, adding items to the cart, submitting orders, and payment settlement; and entity attributes cover the product's general transaction code, its product category level, and unit price. The system serializes all historical interaction actions of a single physical device within a preset time window in chronological order, forming a discrete raw behavioral sequence.

[0021] Subsequently, the system utilizes a pre-defined cross-domain mapping dictionary to transform the multidimensional raw behavioral data sequence into continuous numerical features, thereby establishing a global feature coordinate system. The cross-domain mapping dictionary stores the static correspondence between discrete interaction event labels and basic interest weights across heterogeneous channels at its underlying level. The system extracts the event occurrence frequency and corresponding basic interest weights, quantifying the originally discrete character-based business events into continuous interaction intensity feature values, thus constructing a global feature coordinate system covering all e-commerce product categories.

[0022] Next, the system performs missing value imputation and normalization on this continuous numerical feature. For blank dimensions resulting from no interaction with users in certain non-preferred product categories, the system uses feature zero-value or global mean imputation to fill in the missing values. To eliminate the absolute dimensional differences in unit price or browsing frequency between different product categories, the system uses a linear extremum normalization algorithm to process the features. Let the original continuous feature value of a user's cumulative interaction intensity in a specific e-commerce category be denoted as _____. The system uses the following extreme value normalization formula to map it to normalized eigenvalues. : ; In the above feature normalization formula, This represents the minimum historical statistical value of the interaction intensity of a specific e-commerce category across all channels, within the global feature coordinate system. This represents the maximum historical statistical value of interaction intensity across all channels for a specific e-commerce category within the global feature coordinate system. The system uses this operator to rigorously compress the interaction intensity feature values ​​across all cross-domain dimensions to a continuous real number range of zero to one. In this solution, for a specific digital category, the cumulative original continuous feature values ​​of interaction intensity are... The implementation can be up to 150 times, the minimum historical statistical value across all channels. The highest historical statistical value across all channels is 0 times. The optimal value is 1000 times, from which the normalized eigenvalue is calculated. That is, 0.15.

[0023] Finally, the system arranges the normalized feature values ​​of each dimension in a one-dimensional manner according to the predefined physical dimension order of the global feature coordinate system, and finally generates an initial behavioral feature vector with a fixed length and uniform dimension.

[0024] Through this data preprocessing step, the system completely eliminates the format barriers and absolute dimension interference of heterogeneous data at the underlying level of multiple terminal channels, transforming the loose and disordered original business logs into a data base with consistent structure and smooth values.

[0025] S2. Project the initial behavior feature vector onto the multi-dimensional feature space, calculate the spatial covariance of the initial behavior feature vector in the multi-dimensional feature space, and perform cross-channel alignment and fusion of the initial behavior feature vector based on the spatial covariance to construct a joint feature matrix. Further, step S2, the step of calculating the spatial covariance of the initial behavioral feature vectors in the multidimensional feature space, includes: The initial behavioral feature vector is mapped to discrete observation points in a multi-dimensional feature space to establish the topological structure of the cross-channel feature distribution; Calculate the mean vector of the initial behavioral feature vectors within each channel, and extract the centered deviation matrix by subtracting the mean vector from the initial behavioral feature vectors. Calculate the inner product of the centered deviation matrix and its transpose, and perform degree-of-freedom normalization to obtain the spatial covariance.

[0026] Furthermore, in step S2, the step of aligning and fusing the initial behavioral feature vector across channels based on spatial covariance includes: The off-diagonal elements of the spatial covariance are numerically compared with a preset association threshold, and cross-channel feature pairs corresponding to off-diagonal elements that are greater than the preset association threshold are selected. The soft allocation algorithm is used to assign dynamic alignment weights to cross-channel features based on the spatial covariance value; Based on dynamic alignment weights, linear superposition and dimensionality reduction are performed on the associated initial behavioral feature vectors to construct a joint feature matrix.

[0027] Specifically, the system first receives an initial behavioral feature vector of uniform dimension and maps it to discrete observation points in a multi-dimensional feature space, thereby establishing the topological structure of the cross-channel feature distribution. In a specific e-commerce operating environment, suppose the system extracts a total of... User behavior samples, each sample contains These are continuous features, encompassing specific business metrics such as web-based digital product browsing frequency and mobile-based beauty product add-to-cart intensity, thus forming a dimension of [missing information]. Initial behavioral feature matrix In this scheme, the total number of user behavior samples... The preferred value is 10000, a continuous feature dimension. The preferred dimension is 128. The system targets each dimension within the feature space. Calculate the feature mean of all observed samples in this dimension. : ; In the formula Indicates the first The user observation sample at the th The exact value of each e-commerce interaction feature dimension. In this solution, if the first... The user in the first Interaction values ​​in each dimension The feature mean was calculated from all samples 50 times. It can be 20 times. The system utilizes the initial behavioral feature matrix. Subtract the corresponding feature mean from each element in the matrix. Extract the centered deviation matrix to characterize the degree of data deviation. .

[0028] Subsequently, the system calculates the spatial covariance of the initial behavioral feature vectors within the multidimensional feature space. The system then extracts the aforementioned centered bias matrix. Its transpose Perform matrix inner product operations and divide by statistical degrees of freedom for normalization to obtain the spatial covariance matrix of the global features. : ; In the formula For dimension size equal to A symmetric square matrix, These represent the degrees of freedom in the unbiased estimation calculation. The off-diagonal elements of this spatial covariance matrix... The first was rigorously quantified The characteristics of the first e-commerce channel and the second The spatial distribution of the characteristics of individual e-commerce channels shows a fluctuating correlation.

[0029] After obtaining the spatial covariance matrix, the system begins cross-channel alignment and fusion. The system pre-sets a positive real-valued correlation threshold. The spatial covariance matrix All off-diagonal elements are compared to the threshold in absolute value. The system filters out noisy data with weak correlation and specifically selects those that meet the threshold. The off-diagonal elements are used to combine the two channel features corresponding to these elements into cross-channel feature pairs. In this scheme, the positive real number association threshold is... The optimal value is 0.65. For each selected cross-channel feature pair, the system introduces a soft allocation algorithm to assign dynamic alignment weights. The system uses an exponential normalization function with a smoothing coefficient to calculate the feature mapping weights. : ; In the formula Indicates the first Information from the first feature dimension is mapped to the first... The dynamic alignment weights used when considering each feature dimension. A scaling factor set for the business side to control the smoothness of weight allocation. This indicates that after pre-threshold filtering, it is compared with the first... A set of indices for all target feature dimensions where each feature exhibits a significant statistical correlation. In this scheme, a scaling factor is used to control the smoothness of weight allocation. The preferred value is 0.8.

[0030] The system performs linear superposition and dimensionality reduction on correlated initial behavioral feature vectors based on calculated dynamic alignment weights. The system then centers the deviation matrix. With all dynamically aligned weights Weight mapping matrix formed by aggregation Perform matrix multiplication: ; In the formula This is the final joint feature matrix. The calculation process uses linear projection mapping in a high-dimensional space to condense the originally sparse and scattered multi-channel discrete behavioral features into a low-dimensional dense representation space.

[0031] By performing the above-mentioned data alignment and fusion steps based on spatial covariance, the system completes the underlying soft connection of heterogeneous channel features using pure data-driven probabilistic association without relying on strong identity identifiers such as mobile phone numbers. This effectively removes redundant dimensions of cross-terminal behavioral data and improves the information density of the feature matrix.

[0032] S3. Input the joint feature matrix into a deep neural network for nonlinear feature extraction and output a structured feature representation; Further, step S3, the step of outputting the structured feature representation, includes: The joint feature matrix is ​​received by the input layer of a deep neural network, and a forward propagation calculation of high-dimensional linear combination is performed through a multilayer perceptron architecture. A standard nonlinear activation function is introduced into the hidden layer to impose transformation constraints and extract nonlinear behavior features. The nonlinear behavioral features are compressed in dimension by the output layer and standardized by the layer normalization operator, resulting in a structured feature representation.

[0033] Specifically, the input layer of the deep neural network receives the joint feature matrix generated in the previous processing stage. In this scheme, the specific architecture of the deep neural network is configured to include... A fully connected hidden layer, where the number of hidden layers is [number]. The system preferably uses a 4-layer architecture, with each hidden layer ideally containing 64 neurons. In specific cross-channel e-commerce scenarios, this joint feature matrix serves as an input parameter, containing aligned multi-terminal interaction data, such as a high-dimensional fusion value of the user's browsing time on a product details page on a webpage and the frequency of adding items to the cart on a mobile device. The system performs forward propagation computation of high-dimensional linear combinations using a multilayer perceptron architecture, denoted as... Through the first hidden layer weight matrix of the input layer With bias vector Calculate the output of the linear mapping The corresponding linear forward propagation formula is: ; In the above formula, The feature mapping weights, pre-trained from massive historical e-commerce transaction records, are used to quantify the deep linear cross-importance between behaviors across different channels. This represents the network's fundamental bias constant vector. In this scheme, the first hidden layer weight matrix... The optimal dimension is 128×64, and the network's basic bias constant vector is... The preferred length is 64 dimensions. Subsequently, the system introduces a standard nonlinear activation function into the computational flow of the hidden layer to transform and constrain the above linear combination results. To filter out low-level data noise generated in cross-domain mapping and highlight the user's genuine positive purchase intent, the system uses a linear rectified function to extract nonlinear behavioral features. Its activation transformation formula is: ; In this formula, mathematical operators This is used to compare the zero value with the output value of a linear mapping and output the larger one. The system utilizes this non-linear truncation mechanism to extract deep, abstract, and non-linear e-commerce interaction behavior features contained in multi-channel data. The system then processes these non-linear behavior features... As input to the next layer, according to the aforementioned forward propagation and activation formula, in subsequent... The system iteratively calculates within each hidden layer, extracting deeper-dimensional behavioral association features layer by layer. Finally, the system compresses these nonlinear behavioral features through the network output layer, linearly projecting them into a low-dimensional, compact feature space vector. In this scheme, the compact feature space vector The optimal dimension is 32. The system performs standardization processing using a layer normalization operator, outputting the final structured feature representation. Its standardized calculation formula is: ; In this formula, Represents a low-dimensional feature space vector The arithmetic mean of all internal feature dimensions. This represents the corresponding characteristic variance. It is a very small positive real number set to prevent the denominator from being zero. and These represent the global scaling and translation parameters learned by the deep neural network during backpropagation, respectively. This represents the element-wise multiplication operation of a matrix. In this scheme, it prevents the division of extremely small positive real numbers with a denominator of zero. Preferably 1e-5, global scaling parameter The initial value is preferably 1, and the translation parameter is... The initial value is preferably 0. In e-commerce operations, the system uses this normalization mechanism to force the feature representations of high-frequency active users and low-frequency long-tail users to be flattened into the same statistical distribution space, thus completely eliminating the covariate bias problem caused by differences in user activity.

[0034] Through this step of computation, the system transforms the loosely structured cross-channel behavior space matrix into dense and stably distributed low-dimensional user representation features, significantly improving the feature representation capability and noise robustness of the profile data.

[0035] S4. Construct a time decay function based on fractional calculus, use the time decay function to calculate the evolution weight of the structured feature representation on the time series, and dynamically update the structured feature representation of the historical period through the evolution weight to obtain the dynamic feature tensor. Furthermore, step S4, the step of constructing the time decay function based on fractional calculus, includes: The time evolution interval is divided into equally spaced discrete time steps, and the feature is extracted to generate the relative time difference between the current time point and the time point. By introducing a fractional integral operator, a memory evolution differential equation with relative time difference as the independent variable is established to characterize the non-uniform decay characteristics. The memory evolution differential equation is solved using a numerical approximation algorithm to obtain the solution of the continuous decay function and discretize it to construct a time decay function.

[0036] Furthermore, step S4, which involves dynamically updating the structured feature representation of historical cycles using evolutionary weights, includes: Read the structured feature representation of historical cycles and parse its associated timeliness attributes; The evolution weights are used as scalar multipliers, and the corresponding elements are multiplied with the structured feature representation of the historical period to obtain the weighted historical matrix. The currently generated structured feature representation is fused with the weighted historical matrix by matrix addition to obtain the dynamic feature tensor.

[0037] Specifically, the preset user profile time evolution interval is divided into equally spaced discrete time steps. In actual e-commerce platform operation, this time step is usually set to one hour or one calendar day. The system extracts the historical time point when the user generates a specific e-commerce interaction and the current system operation time point, and calculates the relative time difference between the two. .

[0038] The system introduces fractional-order calculus operators to establish a memory evolution differential equation with relative time difference as the independent variable. This differential equation is mainly used to characterize the non-uniform decay of consumer shopping interest over time. Its business logic lies in reflecting that the rate of decay of long-term attention preference for high-value digital products is much lower than the rate of decay of short-term impulse purchases of fast-moving consumer goods. The evolution equation constructed by the system is as follows: ; In the formula, Represents a continuously decaying function. That is, the relative time difference extracted above. , The underlying data table is pre-set with a basic decay rate constant for a specific e-commerce product category. The parameter is a fractional-order parameter with a value between zero and one, used to characterize the long-tail retention strength of e-commerce users' historical behavior memories. This represents the fractional derivative operator. In this scheme, the relative time difference... The preferred implementation unit is 24 hours, specifically the basic attenuation rate constant for a particular type of e-commerce. The preferred value is 0.05, the fractional order parameter. The preferred value is 0.85.

[0039] The system utilizes a numerical approximation algorithm to solve the memory evolution differential equation, obtaining the exact solution for the continuously decaying function. The system also uses the Mita Greifler function to obtain the analytical form of the equation and represents the relative time difference. Substituting these values ​​into the discretization sample, we can ultimately construct the time decay function for subsequent calculations: ; In the formula, The evolution weights for a specific relative time difference are the final output of the system. This represents the single-parameter Mitagrefleur function. Represents the gamma function. is the non-negative integer index label in the formula for the expansion of an infinite series.

[0040] After calculating the evolution weights of each historical feature, the system reads the structured feature representation of the previous historical period from the graph storage cluster and extracts its associated generation timeliness attributes. The system then assigns the evolution weights... As a scalar multiplier, it is used in conjunction with the extracted structured feature representation matrix of historical cycles. Perform scalar multiplication on the corresponding elements to calculate and obtain the weighted history matrix. : ; Finally, the system generates a structured feature representation matrix based on new user clicks and add-to-cart actions within the latest time period. , and the calculated weighted history matrix Perform matrix addition and fusion in the corresponding dimension: ; In the formula, This is the dynamic feature tensor output by the system in its final iteration.

[0041] Through this step, the system overcomes the bottleneck of the traditional exponential decay model's rapid forgetting, and rigorously fits the long-tail memory pattern of e-commerce users' real shopping intentions, ensuring the high-fidelity dynamic and smooth evolution of profile features over time.

[0042] S5. Use the Dirichlet distribution to perform multidimensional probability modeling on the dynamic feature tensor, output the Dirichlet parameterized distribution set representing the multidimensional preference probability tendency and corresponding variance, and output the Dirichlet parameterized distribution set as a unified profile label.

[0043] Furthermore, step S5, which involves using the Dirichlet distribution to perform multidimensional probabilistic modeling of the dynamic feature tensor, includes: The dynamic feature tensor is transformed into a positive real number vector using a smooth mapping function, which serves as a prior pseudo-count representing user preferences. By inputting positive real number vectors as concentration parameter vectors into the standard Dirichlet probability density model, a conjugate prior probability model of multinomial distribution is constructed. The latent variables of the conjugate prior probability model are solved using the Bayesian variational inference algorithm to establish the set of distribution parameters that fit the current data.

[0044] Further, step S5, which outputs the Dirichlet parameterized distribution set representing the multidimensional preference probability tendency and corresponding variance, includes: The mathematical expectation vector is calculated based on the set of distribution parameters obtained from the solution, forming a multidimensional preference probability tendency that describes the intensity of interest in product categories; The autovariance and statistical covariance of each dimension feature are calculated using the set of distribution parameters, and the corresponding variance of the quantized confidence interval is generated. The multidimensional preference probability tendency and its corresponding variance are numerically packaged and encapsulated to generate a standardized feature package, which is then output as a Dirichlet parameterized distribution set.

[0045] Specifically, the system receives a dynamic feature tensor obtained through time-series updates. To meet the strict mathematical requirement of parameter nonnegativity in the probability distribution model, the system uses a smoothing mapping function to transform this dynamic feature tensor into a vector of positive real numbers. For dynamic feature tensors composed of features from different categories in e-commerce scenarios, the system employs a soft positive PlaS activation function for transformation. Let the total number of e-commerce categories in the system be... The dynamic feature tensor in the first... The values ​​for each product category are The system uses the following formula to calculate the transformed positive real number prior pseudo-count. : ; In this formula, A small positive real constant is preset to prevent the occurrence of extreme values ​​of zero. This prior pseudo-count represents the initial expected base of the model's prediction of user preferences in various e-commerce categories such as mobile phones and digital products, and clothing and bags, before observing the latest round of explicit behavior. In this scheme, the dynamic feature tensor value for a certain category... The embodiment can be 2.5, a small positive real constant. The preferred value is 1e-6.

[0046] Subsequently, the system concatenates the prior pseudo-counts corresponding to all categories into a complete positive real number vector, which is then input as the concentration parameter vector into the standard Dirichlet probability density model. The system uses this to construct a conjugate prior probability model of a multinomial distribution. The probability density function of this model is defined as follows: ; In this formula, Indicates that the user is The true potential preference probability distribution vector for each e-commerce product category Purchase or click for users The actual intention probability of a product category Represents the standard gamma function. This is the concentration parameter vector containing all prior pseudo-counts. In this scheme, the total number of e-commerce categories... The preferred value is 50.

[0047] After establishing the aforementioned conjugate prior probability model, the system uses a Bayesian variational inference algorithm to approximate the latent variables to be solved. The system constructs a variational lower bound by combining newly observed e-commerce interaction frequency data within the current time window. By calculating the relative entropy between the approximate posterior distribution and the true posterior distribution and performing gradient descent iterations, the system solves for the variational parameters that maximize the evidence lower bound. After iterative convergence, the system establishes the final set of distribution parameters that fit the current e-commerce business data. This set contains the final concentration parameters of all categories updated with posterior data. .

[0048] The system calculates the expected value vector based on the obtained distribution parameter set. The system calculates the user's expected value vector at the [number]th [timeframe] using the following expected value formula. Multidimensional preference probability tendencies for specific e-commerce product categories : ; This calculation step transforms the abstract model parameters into business preference percentages that can be directly parsed by the recommendation engine and summed to one. In this scheme, if the final concentration parameter of a specific category after iterative convergence... The sum of concentration parameters for all categories is 5.2. If the value is 100, then the multidimensional preference probability calculated using the above expectation formula is... A value of 0.052 indicates a business preference of 5.2%.

[0049] To quantify the confidence level of the recommendation system when invoking the features of this profile, the system further utilizes the distribution parameter set to calculate the autovariance and statistical covariance of each dimension of the features. Let the sum of the elements in the distribution parameter set be... The system calculates the first number using the following formulas respectively. The independent variance of each feature dimension and the The first feature dimension and the first feature dimension Covariance between each feature dimension : ; ; The calculated autovariance and covariance together form the corresponding variance matrix, which is used to accurately characterize the confidence interval of the system's prediction of user intent and the inverse relationship between different category preferences.

[0050] Finally, the system numerically packages and encapsulates the calculated multidimensional preference probability vectors and their corresponding variance matrices to generate a standardized feature package with a unified structure. This standardized feature package is then directly pushed to the underlying caching system as a Dirichlet parameterized distribution set for use by the product recall algorithm.

[0051] The implementation of this step ensures that the output user profile is no longer a single absolute business label, but a joint statistical distribution that includes uncertainty measures and probability boundaries, effectively enhancing the system's recommendation stability when facing low-frequency e-commerce users and cold start data.

[0052] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamically constructing a unified user profile across e-commerce channels, characterized in that, include: S1. Collect heterogeneous behavioral data from different e-commerce channels, standardize the heterogeneous behavioral data, and generate an initial behavioral feature vector with a unified dimension. S2. Project the initial behavior feature vector onto a multi-dimensional feature space, calculate the spatial covariance of the initial behavior feature vector in the multi-dimensional feature space, and perform cross-channel alignment and fusion of the initial behavior feature vector based on the spatial covariance to construct a joint feature matrix. S3. Input the joint feature matrix into a deep neural network for nonlinear feature extraction and output a structured feature representation; S4. Construct a time decay function based on fractional calculus, use the time decay function to calculate the evolution weight of the structured feature representation on the time series, and dynamically update the structured feature representation of the historical period through the evolution weight to obtain a dynamic feature tensor. S5. Use the Dirichlet distribution to perform multidimensional probability modeling on the dynamic feature tensor, output the Dirichlet parameterized distribution set representing the multidimensional preference probability tendency and corresponding variance, and output the Dirichlet parameterized distribution set as a unified profile label.

2. The method for dynamically constructing a unified user profile across e-commerce channels according to claim 1, characterized in that, Step S1, the step of generating an initial behavioral feature vector of uniform dimension, includes: Extract timestamps, event types, and entity attributes from the heterogeneous behavioral data to construct a multidimensional original behavioral data sequence; Using a pre-defined cross-domain mapping dictionary, the multidimensional original behavioral data sequence is transformed into continuous numerical features, and a global feature coordinate system is established. The continuous numerical features are subjected to missing value imputation and normalization, and then concatenated according to the global feature coordinate system to generate the initial behavioral feature vector of the unified dimension.

3. The method for dynamically constructing a unified user profile across e-commerce channels according to claim 1, characterized in that, Step S2, the step of calculating the spatial covariance of the initial behavioral feature vector in the multidimensional feature space, includes: The initial behavioral feature vector is mapped to discrete observation points in the multidimensional feature space to establish the topological structure of the cross-channel feature distribution; Calculate the mean vector of the initial behavioral feature vectors within each channel, and extract the centered deviation matrix by subtracting the mean vector from the initial behavioral feature vectors; Calculate the inner product of the centered deviation matrix and its transpose matrix, and perform degree-of-freedom normalization to obtain the spatial covariance.

4. The method for dynamically constructing a unified user profile across e-commerce channels according to claim 1, characterized in that, In step S2, the step of aligning and fusing the initial behavioral feature vector across channels based on the spatial covariance includes: The off-diagonal elements of the spatial covariance are numerically compared with a preset association threshold, and cross-channel feature pairs corresponding to the off-diagonal elements that are greater than the preset association threshold are selected. The soft allocation algorithm is used to assign dynamic alignment weights to the cross-channel feature pairs based on the value of the spatial covariance. Based on the dynamic alignment weights, linear superposition and dimensionality reduction are performed on the associated initial behavioral feature vectors to construct the joint feature matrix.

5. The method for dynamically constructing a unified user profile across e-commerce channels according to claim 1, characterized in that, Step S3, the step of outputting the structured feature representation, includes: The joint feature matrix is ​​received by the input layer of a deep neural network, and a forward propagation calculation of high-dimensional linear combination is performed through a multilayer perceptron architecture; A standard nonlinear activation function is introduced into the hidden layer to impose transformation constraints and extract nonlinear behavior features. The nonlinear behavioral features are compressed in dimension by the output layer and standardized by combining the layer normalization operator to output the structured feature representation.

6. The method for dynamically constructing a unified user profile across e-commerce channels according to claim 1, characterized in that, Step S4, the step of constructing the time decay function based on fractional calculus, includes: The time evolution interval is divided into equally spaced discrete time steps, and the feature is extracted to generate the relative time difference between the current time point and the time point. A fractional integral operator is introduced to establish a memory evolution differential equation with the relative time difference as the independent variable, which characterizes the non-uniform decay characteristics. The memory evolution differential equation is solved using a numerical approximation algorithm to obtain the solution of the continuous decay function and discretize it to construct the time decay function.

7. The method for dynamically constructing a unified user profile across e-commerce channels according to claim 1, characterized in that, Step S4, the step of dynamically updating the structured feature representation of the historical period using the evolution weights, includes: Read the structured feature representation of the historical period and parse its associated timeliness attributes; The evolution weights are used as scalar multipliers, and the corresponding elements are multiplied by the structured feature representation of the historical period to obtain the weighted historical matrix. The currently generated structured feature representation is fused with the weighted history matrix by matrix addition to obtain the dynamic feature tensor.

8. The method for dynamically constructing a unified user profile across e-commerce channels according to claim 1, characterized in that, Step S5, the step of performing multidimensional probabilistic modeling of the dynamic feature tensor using the Dirichlet distribution, includes: The dynamic feature tensor is transformed into a positive real number vector using a smoothing mapping function, which serves as a prior pseudo-count representing user preferences. The positive real number vector is input as a concentration parameter vector into the standard Dirichlet probability density model to construct a conjugate prior probability model of the multinomial distribution. The latent variables of the conjugate prior probability model are solved using the Bayesian variational inference algorithm to establish the set of distribution parameters that fit the current data.

9. The method for dynamically constructing a unified user profile across e-commerce channels according to claim 1, characterized in that, Step S5, the step of outputting the Dirichlet parameterized distribution set representing the multidimensional preference probability tendency and corresponding variance, includes: The mathematical expectation vector is calculated based on the set of distribution parameters obtained from the solution, forming the multidimensional preference probability tendency that describes the intensity of category interest; The autovariance and statistical covariance of each dimension feature are calculated using the set of distribution parameters to generate the corresponding variance of the quantized confidence interval; The multidimensional preference probability tendency and the corresponding variance are numerically packaged and encapsulated to generate a standardized feature package, which is then output as the Dirichlet parameterized distribution set.

Images