Search

CN-121998715-A - Advertisement pushing method based on E-commerce big data feedback

CN121998715ACN 121998715 ACN121998715 ACN 121998715ACN-121998715-A

Abstract

The invention belongs to the technical field of advertisement pushing, in particular to an advertisement pushing method based on big data feedback of electronic commerce, which can extract user intention snapshot in real time by constructing a multidimensional space-time behavior matrix and combining a sliding time window dynamic slicing technology, the intention transition detection model which is completed by training is used for accurately identifying intention transition events such as interest point transition, decision strength change and the like, so that the user demand fluctuation can be followed in real time, the advertisement pushing and the user real-time intention are highly matched, the problem of dislocation between pushing and real-time demand is effectively solved, and the accuracy of advertisement pushing is remarkably improved.

Inventors

  • CAO HANMEI

Assignees

  • 重庆叶凡科技有限公司

Dates

Publication Date
20260508
Application Date
20260128

Claims (10)

  1. 1. The advertisement pushing method based on the big data feedback of the E-commerce is characterized by comprising the following steps of: s1, acquiring a real-time behavior data stream of a user on an e-commerce platform, and constructing a multidimensional space-time behavior matrix based on the real-time behavior data stream of the user on the e-commerce platform; s2, dynamically slicing the multidimensional space-time behavior matrix by adopting a sliding time window, wherein each slice corresponds to one user intention snapshot, and identifying intention transition events between adjacent slices through an intention transition detection model, wherein the intention transition events comprise interest point transfer, decision strength change and purchase intention generation or regression; S3, generating a group of candidate sets comprising real push advertisements and virtual advertisements aiming at the user intention snapshot; S4, pushing the candidate set to the user, capturing interactive feedback of the candidate set, and quantifying certainty and openness of user intention by comparing feedback differences of the user on the real advertisement and the virtual advertisement; and S5, adjusting a pushing strategy based on the nature of the intention of the user and the intention transition event.
  2. 2. The advertisement pushing method based on e-commerce big data feedback of claim 1, wherein in step S1, the method for constructing the multidimensional space-time behavior matrix is as follows: Receiving a real-time behavior data stream from an e-commerce platform, analyzing each behavior event to generate a standard event tuple, wherein the standard event tuple comprises a user identifier, a behavior type, an absolute timestamp of the occurrence of the behavior, a page uniform resource identifier, commodity category codes and context information; Mapping the absolute time stamp into a serial number of a continuous second-level time slice from a reference time by taking seconds as a basic unit, and classifying a plurality of behavior events occurring in the same second into the same second-level time slice; Constructing a unified virtual shopping space coordinate system, mapping different page uniform resource identifiers into first dimension coordinate values, mapping different commodity category codes into second dimension coordinate values, and expressing a physical position of a user in one-time browsing action as one two-dimension coordinate point in the virtual shopping space; and converting each standard event tuple into a multi-dimensional feature vector, taking the user identifier as a first dimension index, taking the second-level time slice sequence number as a second dimension index, and taking the multi-dimensional feature vector as a third dimension element to generate a three-dimensional space-time behavior matrix.
  3. 3. The advertisement pushing method based on E-commerce big data feedback of claim 2, wherein the feature vector comprises a space coordinate vector formed by combining page coordinates and category coordinates, a behavior type vector for performing single-hot coding on a behavior type, and a context feature vector for performing numerical coding on context information.
  4. 4. The advertisement pushing method based on big data feedback of electronic commerce according to claim 3, wherein in step S2, the generating a user intention snapshot specifically comprises: Presetting a window size with a fixed length and a sliding step length in a sliding window, starting from a starting time point of the multidimensional space-time behavior matrix, orderly intercepting behavior data with continuous time length at intervals of the sliding step length to form a slice, forming a submatrix by behavior event data in each slice, taking the submatrix as input of an intention transition detection model at a corresponding time point, and generating a user intention snapshot of the time point through the intention transition detection model.
  5. 5. The advertisement pushing method based on electronic commerce big data feedback of claim 4, wherein the training method of the intention transition detection model is as follows: Slicing historical user behaviors according to a sliding window from a historical user behavior log to obtain time sequence behavior slice pairs, and dividing the slice pairs into a training set and a verification set; Determining the main intention type of each slice based on the finally generated behaviors in each time window slice, and generating a main intention label, carrying out quantitative assignment according to the density, duration and sequence consistency of behaviors in the slices to obtain an intention strength label; initializing parameters of an intention transition detection model, selecting a slice and corresponding main intention labels, intention intensity labels and intention transition labels from a training set, calculating main intention classification loss, intention intensity regression loss and intention transition detection loss for the main intention labels, the intention intensity labels and the intention transition labels respectively, linearly combining the main intention classification loss, the intention intensity regression loss and the intention transition detection loss to obtain a joint loss value, sequentially calculating partial derivatives of the joint loss value to the trainable parameters in the intention transition detection model through gradient back propagation to form gradient vectors, and updating the intention transition detection model parameters by adopting an estimation algorithm; and judging that the model completes convergence when the improvement of the comprehensive performance index of the intent transition detection model on the verification set in 10 continuous periods does not exceed a preset threshold value.
  6. 6. The advertisement pushing method based on electronic commerce big data feedback according to claim 5, wherein in step S3, the real advertisement generating process is as follows: Inputting the user intention snapshot into a global advertisement pool, searching a plurality of candidate advertisements with highest semantic matching degree with the current intention snapshot, calculating matching degree scores of the plurality of candidate advertisements and the user intention snapshot, and selecting the advertisement with the highest score as a real push advertisement.
  7. 7. The advertisement pushing method based on electronic commerce big data feedback of claim 6, wherein the virtual advertisement generating process is: Applying offset to the user intention snapshot, enabling the user intention snapshot to offset towards the adjacent intention direction to obtain the offset intention snapshot, taking the offset intention snapshot as input of a global advertisement pool, retrieving and outputting a preselected virtual advertisement set from the global advertisement pool, selecting advertisements which are different from the real push advertisements in terms of category, price band or brand but have comparability with specific commodities from the preselected virtual advertisement set, taking the advertisements as virtual advertisements, and combining the real push advertisements with the virtual advertisements to form a candidate set which is finally pushed to the user.
  8. 8. The advertisement pushing method based on electronic commerce big data feedback according to claim 7, wherein in step S4, the certainty and openness of the quantified user intention are specifically: Capturing the interaction behavior of a user on each advertisement in a candidate set, accumulating the scores of the interaction behavior of each advertisement according to different scores of each interaction behavior of the interaction behavior to obtain a comprehensive feedback score of the advertisement, calculating the average comprehensive feedback score of the real push advertisement and the average comprehensive feedback score of all virtual advertisements, and calculating a feedback difference metric value according to the average comprehensive feedback score of the real push advertisement and the average comprehensive feedback score of all virtual advertisements The calculation formula is as follows: Wherein the method comprises the steps of Average composite feedback score for true push advertisement, the For an average composite feedback score for all virtual advertisements, For a minimum positive number, the feedback difference measurement value Obtaining the intent deterministic quantization value through Sigmoid function mapping Calculating intent openness quantization value from intent deterministic quantization value The calculation formula is as follows: Deterministic quantization of values according to intent And comparing the current intention of the user with the first threshold value and the second threshold value.
  9. 9. The advertisement pushing method based on e-commerce big data feedback of claim 8, wherein the user current intention comprises: When intention is deterministic quantization value When the current intention of the user is higher than a first preset threshold value, judging that the certainty of the current intention of the user is high, namely, the current intention is better for real advertisements; When intention is deterministic quantization value When the current intention openness of the user is higher than a second preset threshold value, the virtual advertisement acceptance is better; When intention is deterministic quantization value And when the user intention is between the first threshold value and the second threshold value, determining that the user intention is in a fuzzy transition state.
  10. 10. Advertisement push system based on electronic commerce big data feedback, which is characterized by comprising: the intention flow modeling module is used for collecting real-time behavior data flow of the user on the e-commerce platform and constructing a multidimensional space-time behavior matrix based on the real-time behavior data flow of the user on the e-commerce platform; The dynamic slice analysis module is used for dynamically slicing the multidimensional space-time behavior matrix by adopting a sliding time window, each slice corresponds to one user intention snapshot, and an intention transition event between adjacent slices is identified through an intention transition detection model, wherein the intention transition event comprises interest point transfer, decision strength change and purchase intention generation or regression; the advertisement generation module is used for generating a group of candidate sets comprising real push advertisements and virtual advertisements aiming at the user intention snapshot; The interactive pushing and feedback capturing module is used for pushing the candidate set to the user and capturing interactive feedback of the candidate set, and the certainty and openness of the user intention are quantified by comparing feedback differences of the user on the real advertisement and the virtual advertisement; And the strategy module is used for adjusting the pushing strategy based on the nature of the intention of the user and the intention transition event.

Description

Advertisement pushing method based on E-commerce big data feedback Technical Field The invention belongs to the technical field of advertisement pushing, and particularly relates to an advertisement pushing method based on electronic commerce big data feedback. Background In the prior art, most of advertisement pushing is based on user history behavior labels to carry out immobilized advertisement matching, dynamic changes of user intention cannot be captured in real time, so that an advertisement pushing system can misuse that the intention of a user is unchanged to continue pushing the original product, and the pushing is disjointed with real-time requirements of the user, and therefore an advertisement pushing method based on electronic commerce big data feedback is improved and designed. Disclosure of Invention Aiming at the defects in the prior art, the invention provides an advertisement pushing method based on big data feedback of electronic commerce, which is used for solving the problems in the background art. In order to solve the technical problems, the invention adopts the following technical scheme: An advertisement pushing method based on big data feedback of an electronic commerce comprises the following steps: s1, acquiring a real-time behavior data stream of a user on an e-commerce platform, and constructing a multidimensional space-time behavior matrix based on the real-time behavior data stream of the user on the e-commerce platform; s2, dynamically slicing the multidimensional space-time behavior matrix by adopting a sliding time window, wherein each slice corresponds to one user intention snapshot, and identifying intention transition events between adjacent slices through an intention transition detection model, wherein the intention transition events comprise interest point transfer, decision strength change and purchase intention generation or regression; S3, generating a group of candidate sets comprising real push advertisements and virtual advertisements aiming at the user intention snapshot; S4, pushing the candidate set to the user, capturing interactive feedback of the candidate set, and quantifying certainty and openness of user intention by comparing feedback differences of the user on the real advertisement and the virtual advertisement; and S5, adjusting a pushing strategy based on the nature of the intention of the user and the intention transition event. Preferably, in step S1, the method for constructing the multidimensional space-time behavior matrix includes: Receiving a real-time behavior data stream from an e-commerce platform, analyzing each behavior event to generate a standard event tuple, wherein the standard event tuple comprises a user identifier, a behavior type, an absolute timestamp of the occurrence of the behavior, a page uniform resource identifier, commodity category codes and context information; Mapping the absolute time stamp into a serial number of a continuous second-level time slice from a reference time by taking seconds as a basic unit, and classifying a plurality of behavior events occurring in the same second into the same second-level time slice; Constructing a unified virtual shopping space coordinate system, mapping different page uniform resource identifiers into first dimension coordinate values, mapping different commodity category codes into second dimension coordinate values, and expressing a physical position of a user in one-time browsing action as one two-dimension coordinate point in the virtual shopping space; and converting each standard event tuple into a multi-dimensional feature vector, taking the user identifier as a first dimension index, taking the second-level time slice sequence number as a second dimension index, and taking the multi-dimensional feature vector as a third dimension element to generate a three-dimensional space-time behavior matrix. Preferably, the feature vector comprises a space coordinate vector formed by combining page coordinates and category coordinates, a behavior type vector for performing single-heat coding on the behavior type, and a context feature vector for performing numerical coding on the context information. Preferably, in step S2, the generating a user intention snapshot is specifically: Presetting a window size with a fixed length and a sliding step length in a sliding window, starting from a starting time point of the multidimensional space-time behavior matrix, orderly intercepting behavior data with continuous time length at intervals of the sliding step length to form a slice, forming a submatrix by behavior event data in each slice, taking the submatrix as input of an intention transition detection model at a corresponding time point, and generating a user intention snapshot of the time point through the intention transition detection model. Preferably, the training method of the intent transition detection model is as follows: Slicing historical user behaviors according to a sliding