Search

CN-122022864-A - Multi-type user characterization generation method, device, equipment and medium

CN122022864ACN 122022864 ACN122022864 ACN 122022864ACN-122022864-A

Abstract

The application belongs to the field of artificial intelligence and relates to a multi-type user characterization generating method, device, equipment and medium, which comprise the steps of acquiring multi-dimensional associated data of a user, obtaining a multi-dimensional fusion vector sequence through embedding processing and feature fusion, obtaining a behavior time and semantic fusion feature sequence through space-time fusion position coding layer processing, calculating a dimension gating fusion feature of multiple interest dimensions, obtaining a user basic characterization sequence through dimension binding attention fusion and feature enhancement processing, determining a product to be recommended and attribute features of the product, determining a target comprehensive user characterization sequence based on sequence length and the two-way enhancement user interest characterization sequence after the two-way feature fusion processing, and obtaining a multi-type user characterization set through multi-dimensional feature extraction and scene enhancement processing. The application can be applied to the business fields of finance, science and technology and the like, and can obtain accurate multi-type user characterization.

Inventors

  • WANG JIANZONG
  • DENG YUWEI

Assignees

  • 平安科技（深圳）有限公司

Dates

Publication Date
20260512
Application Date
20260109

Claims (10)

  1. 1. A method for generating a multi-type user representation, comprising the steps of: acquiring multi-dimensional associated data of a target user in a target service scene, wherein the multi-dimensional associated data comprises a user behavior sequence, product attribute characteristics and user static characteristics; Embedding processing and feature fusion are carried out on the user behavior sequence, the product attribute features and the user static features to obtain a multi-dimensional fusion vector sequence of the target user; based on the user behavior sequence and the multidimensional fusion vector sequence, performing space-time coding fusion and feature enhancement through a space-time fusion position coding layer to generate a behavior time semantic fusion feature sequence of the target user; Based on the target service scene and the multidimensional fusion vector sequence, calculating dimension gating fusion characteristics of multiple interest dimensions to obtain a user characterization vector sequence fused by the multiple interest dimensions; Performing dimension binding attention fusion and feature enhancement processing on the user characterization vector sequence to obtain a user basic characterization sequence of the target user, and determining a product to be recommended and attribute features of the product to be recommended based on the user basic characterization sequence; performing bidirectional feature fusion processing on the user basic characterization sequence and the attribute features of the product to be recommended to obtain a bidirectional enhanced user interest characterization sequence; and determining a target comprehensive user characterization sequence based on the sequence length of the user behavior sequence and the bidirectional enhanced user interest characterization sequence, and performing multidimensional feature extraction and scene enhancement processing on the target comprehensive user characterization sequence to obtain a multi-type user characterization set.
  2. 2. The method according to claim 1, wherein the step of generating the behavior time semantic fusion feature sequence of the target user by performing space-time coding fusion and feature enhancement through a space-time fusion position coding layer based on the user behavior sequence and the multidimensional fusion vector sequence specifically comprises: Performing day and night period coding on the behavior time stamp in the user behavior sequence to obtain a day and night code; performing periodic coding on the behavior time stamp to obtain Zhou Du codes, and splicing the day and night codes and the Zhou Du codes to obtain absolute time code vectors; calculating a time difference based on the behavior time stamp and the current time stamp, calculating a weight based on the time difference and a preset piecewise attenuation function, and mapping the weight into a relative time attenuation coding vector through a linear layer; carrying out coding fusion on the absolute time coding vector and the relative time attenuation coding vector to obtain a fused vector; And adding the fused vector and the multidimensional fusion vector sequence element by element, and normalizing an addition result to obtain the behavior time semantic fusion feature sequence of the target user.
  3. 3. The method according to claim 1, wherein the step of calculating a dimension-gated fusion feature of multiple dimensions of interest based on the target service scene and the multidimensional fusion vector sequence to obtain a user-characterized vector sequence of multiple dimensions of interest fusion specifically comprises: determining a plurality of interest dimensions of the multi-dimensional fusion vector sequence based on the target service scene and the user behavior sequence, wherein each interest dimension corresponds to a group of exclusive matrixes; Acquiring a target mask of each interest dimension; calculating the interest feature representation of each interest dimension based on the exclusive matrix, the target mask and a multi-dimensional fusion vector sequence; Splicing the interest feature representations of each interest dimension to obtain a multi-interest dimension fusion feature vector; acquiring a gating weight matrix and a biasing item preset in each interest dimension; obtaining the opening degree parameter of each interest dimension based on the multi-interest dimension fusion feature vector, the gating weight matrix and the bias item; Based on the interest feature representation of each interest dimension, a cross-dimension average feature and the opening degree parameter, obtaining a dimension gating fusion feature of each interest dimension, wherein the cross-dimension average feature is obtained by carrying out average pooling on the multi-interest dimension fusion feature vector; and fusing the dimension gating fusion characteristics of each interest dimension to obtain the multidimensional fusion user characterization vector of the target user.
  4. 4. The method according to claim 1, wherein the step of performing dimension binding attention fusion and feature enhancement processing on the user characterization vector sequence to obtain a user base characterization sequence of the target user specifically includes: performing dimension binding attention processing on the multi-dimension fusion user characterization vector to obtain a dimension association attention feature vector; carrying out residual connection and normalization processing on the multi-dimensional fusion user characterization vector and the dimension associated attention feature vector to obtain a residual normalized feature vector; Performing feature transformation and nonlinear enhancement on the residual normalized feature vector to obtain a feedforward enhanced feature vector; And carrying out residual connection and normalization processing on the feedforward enhanced feature vector and the residual normalized feature vector to obtain a user basic characterization sequence of the target user.
  5. 5. The method of claim 1, wherein the step of performing a bi-directional feature fusion process on the user base characterization sequence and the product attribute feature to be recommended to obtain a bi-directional enhanced user interest characterization sequence specifically comprises: Splicing the attribute characteristics of the product to be recommended to obtain multi-attribute splicing characteristics of the product; Performing feature mapping and enhancement on the multi-attribute splicing features of the product to obtain a product induction vector; Calculating the association degree of each user behavior characterization in the user basic characterization sequence and the product induction vector to obtain a product association degree weight; Carrying out weighted summation on each user behavior representation and the product association degree weight to obtain a new representation of the user fused with the product induction information; and fusing the user basic characterization sequence with the new characterization of the user to obtain a bidirectional enhanced user interest characterization sequence.
  6. 6. The method according to claim 1, wherein the step of determining a target integrated user characterization sequence based on the sequence length of the user behavior sequence and the bi-directional enhanced user interest characterization sequence, in particular comprises: If the sequence length of the user behavior sequence is smaller than or equal to a length threshold value, determining the bidirectional enhanced user interest characterization sequence as a target comprehensive user characterization sequence; If the sequence length of the user behavior sequence is greater than the length threshold, extracting a behavior feature vector of each behavior corresponding to each interest dimension from a bidirectional enhanced user interest characterization sequence, and acquiring a feature mean value of each behavior feature vector, wherein each interest dimension is determined based on the target service scene and the user behavior sequence; based on the characteristic mean value, calculating a relative importance value of each behavior in the corresponding interest dimension; calculating the dimension contribution degree of each behavior in the corresponding interest dimension based on the characteristic mean value and the relative importance value to obtain a dimension behavior contribution degree set of each behavior; Adding each dimension contribution degree in the dimension behavior contribution degree set, and calculating the comprehensive score of each behavior to obtain a comprehensive score set; calculating the mean value and standard deviation of the comprehensive score set, and acquiring adjustment parameters which are determined based on the target service scene; Calculating a behavior screening threshold based on the mean value, the standard deviation and the adjustment parameter, and screening behaviors corresponding to the comprehensive scoring set based on the behavior screening threshold and the comprehensive scoring set to obtain an index set of high-importance behaviors; extracting behavior characterization of behaviors corresponding to the index set from the bidirectional enhanced user interest characterization sequence, and calculating the attention weight between every two high-importance behaviors based on the behavior characterization; Based on the attention weight, carrying out weighted summation on the bidirectional enhanced user interest characterization sequence to obtain high-importance behavior attention fusion characterization; And performing dimension mapping and restoration on the high-importance behavior attention fusion characterization through a preset linear layer to obtain a sparse user characterization sequence, and determining the sparse user characterization sequence as a target comprehensive user characterization sequence.
  7. 7. The method according to claim 1, wherein the step of performing multidimensional feature extraction and scenerising enhancement processing on the target comprehensive user characterization sequence to obtain a multi-type user characterization set specifically comprises: carrying out maximum pooling and average pooling treatment on the target comprehensive user characterization sequence to obtain a first interest feature vector and a second interest feature vector; Splicing the first interest feature vector and the second interest feature vector to obtain an initial global feature vector, and carrying out feature enhancement on the initial global feature vector to obtain global long-term characterization; Extracting features of corresponding interest dimensions from the target comprehensive user characterization sequence aiming at each interest dimension, wherein the features are feature intervals exclusive to each dimension after being divided according to the interest dimensions from the target comprehensive user characterization sequence, and each interest dimension is determined based on the target service scene and the user behavior sequence; Performing dimension reduction on the features to obtain dimension preference characterization corresponding to each interest dimension, and determining the dimension preference characterization as a dimension exclusive short-term characterization; Dividing the target comprehensive user characterization sequence according to time windows, and calculating the average characteristic of each time window to obtain a time sequence primary characterization; performing time sequence enhancement on the time sequence preliminary characterization to obtain a time sequence dynamic characterization; And constructing a multi-type user characterization set based on the global long-term characterization, the dimension exclusive short-term characterization and the time sequence dynamic characterization.
  8. 8. A multi-type user token generation apparatus, comprising: The system comprises an acquisition module, a processing module and a processing module, wherein the acquisition module is used for acquiring multi-dimensional associated data of a target user in a target service scene, wherein the multi-dimensional associated data comprises a user behavior sequence, product attribute characteristics and user static characteristics; The first fusion module is used for carrying out embedding processing and feature fusion on the user behavior sequence, the product attribute features and the user static features to obtain a multi-dimensional fusion vector sequence of the target user; the coding module is used for carrying out space-time coding fusion and feature enhancement through a space-time fusion position coding layer based on the user behavior sequence and the multidimensional fusion vector sequence to generate a behavior time semantic fusion feature sequence of the target user; The computing module is used for computing the dimension gating fusion characteristics of a plurality of interest dimensions based on the target service scene and the multidimensional fusion vector sequence to obtain a user characterization vector sequence fused by the multiple interest dimensions; The dimension binding module is used for carrying out dimension binding attention fusion and feature enhancement processing on the user characterization vector sequence to obtain a user basic characterization sequence of the target user, and determining a product to be recommended and attribute features of the product to be recommended based on the user basic characterization sequence; The second fusion module is used for carrying out bidirectional feature fusion processing on the user basic characterization sequence and the attribute features of the product to be recommended to obtain a bidirectional enhanced user interest characterization sequence; The extraction module is used for determining a target comprehensive user characterization sequence based on the sequence length of the user behavior sequence and the bidirectional enhanced user interest characterization sequence, and carrying out multidimensional feature extraction and scenerification enhancement processing on the target comprehensive user characterization sequence to obtain a multi-type user characterization set.
  9. 9. A computer device comprising a memory and a processor, the memory having stored therein computer readable instructions which when executed by the processor implement the steps of the method of generating a user profile of the type of any one of claims 1 to 7.
  10. 10. A computer readable storage medium having stored thereon computer readable instructions which when executed by a processor implement the steps of the multi-type user token generation method of any one of claims 1 to 7.

Description

Multi-type user characterization generation method, device, equipment and medium Technical Field The application relates to the technical field of artificial intelligence, is applied to online processing of business scenes in financial science and technology, and particularly relates to a method, a device, equipment and a medium for generating multi-type user characterization. Background In the application fields of mainstream recommendation systems such as e-commerce, short video content recommendation and local life service, a user characterization modeling method based on a transform architecture becomes a mainstream technical direction, and a core of the method encodes a user behavior sequence by means of a self-attention mechanism so as to mine potential interest preference of a user. In the prior art, a standard self-attention mechanism is adopted by a conventional transducer recommendation model, time sequence information is modeled only through a position index corresponding to a behavior sequence, and the characteristic of the dimension of an actual time stamp is not introduced, so that the generated user characterization is difficult to capture the actual relevance of the behavior time sequence. The BERT derived recommendation model directly multiplexes the text coding structure, and the user behavior sequence is equivalent to natural language sentences for processing, and attention calculating logic is not customized for the core characteristics of user and product interaction, so that the user characterization can not accurately match the interaction requirement of the recommendation scene. The space-time fusion analogies model tries to integrate time/space features, but only completes simple splicing at a feature layer, does not embed space-time information into a matrix calculation process of a self-attention mechanism, belongs to shallow layer sub-feature fusion, and causes insufficient space-time dimension fusion effect of user characterization. Overall, the lack of a custom Transformer structural design for user characterization learning in the prior art ultimately results in significant drawbacks in the accuracy and suitability of user interest characterization. In summary, in the user characterization modeling process based on the transducer, the user characterization generated in the prior art has the problems of time sequence relevance deficiency, insufficient adaptation of interaction requirements, shallow space-time fusion and the like, so that the real interest preference of the user cannot be accurately and dynamically reflected, and the real-time high-precision user requirement matching and content pushing of the recommendation system are difficult to support. Disclosure of Invention The embodiment of the application aims to provide a multi-type user characterization generation method, a device, computer equipment and a storage medium, so as to solve the problem that the prior art cannot obtain accurate multi-type user characterization and is difficult to dynamically reflect the real interest preference of a user. In a first aspect, a method for generating multiple types of user characterizations is provided, which adopts the following technical scheme: The method comprises the steps of obtaining multidimensional associated data of a target user in a target service scene, wherein the multidimensional associated data comprise a user behavior sequence, product attribute characteristics and user static characteristics, carrying out embedding processing and characteristic fusion on the user behavior sequence, the product attribute characteristics and the user static characteristics to obtain a multidimensional fusion vector sequence of the target user, carrying out space-time coding fusion and characteristic enhancement through a space-time fusion position coding layer based on the user behavior sequence and the multidimensional fusion vector sequence to generate a behavior time semantic fusion characteristic sequence of the target user, calculating dimension gating fusion characteristics of a plurality of interest dimensions based on the target service scene and the multidimensional fusion vector sequence to obtain a user characterization vector sequence fused with the multiple interest dimensions, carrying out dimension binding attention fusion and characteristic enhancement processing on the user characterization vector sequence to obtain a user basic characterization sequence of the target user, determining attribute characteristics of a product to be recommended and a product to be recommended based on the user basic characterization sequence, carrying out bidirectional characteristic fusion processing on the user basic characterization sequence and the attribute characteristics of the product to be recommended to obtain a bidirectional enhancement user interest characterization sequence, determining a target comprehensive user characterization sequence based on the sequence lengt