Search

CN-122022900-A - Method and system for predicting and accurately touching electric business re-purchase behavior by fusing time sequence attention mechanism

CN122022900ACN 122022900 ACN122022900 ACN 122022900ACN-122022900-A

Abstract

The invention belongs to the technical field of behavior analysis of a user of a mobile phone, and particularly relates to a method and a system for predicting and accurately touching the behavior of a user of the mobile phone by fusing a time sequence attention mechanism. The method comprises the steps of obtaining a historical behavior sequence of a target user in a preset statistical period, extracting behavior type, behavior occurrence time and commodity information, constructing a time sequence enhanced behavior representation sequence comprising a relative time difference of adjacent behaviors and a time interval from the last purchasing behavior, carrying out multi-head weighting processing based on a behavior time adjacent relation, a behavior type conversion relation and a commodity semantic similarity relation to obtain a user purchasing intention vector, further carrying out cross-period offset analysis by combining the historical purchasing intention vector of the previous statistical period, generating a repurchase score and an offset risk mark according to an analysis result, and outputting a repurchase prediction result. The invention can simultaneously represent the current behavior characteristics and the cross-period interest migration state of the user, and improves the dynamic property and the distinguishing property of the repurchase identification.

Inventors

  • LI JIAWEI
  • ZHOU XIAYI

Assignees

  • 杭州多熠网络科技有限公司

Dates

Publication Date
20260512
Application Date
20260416

Claims (10)

  1. 1. The method for predicting and accurately touching the repurchase behavior of the electric user by fusing the time sequence attention mechanism is characterized by comprising the following steps of: S1, acquiring a historical behavior sequence of a target user in a preset statistical period, and extracting behavior types, behavior occurrence time and commodity information corresponding to each behavior; S2, converting each behavior in the historical behavior sequence into a basic behavior representation, and performing time enhancement processing on the basic behavior representation based on the relative time difference between adjacent behaviors and the time interval between each behavior and the latest purchasing behavior to construct a time sequence enhancement behavior representation sequence; S3, inputting the time sequence enhancement behavior representation sequence into a multi-head attention network, respectively carrying out independent weighted calculation by different attention heads based on a behavior time adjacent relation, a behavior type conversion relation and a commodity semantic similarity relation, and fusing the output results of the attention heads to obtain a user purchase intention vector in the current statistical period; S4, extracting a historical purchase intention vector of a previous statistical period of the target user, and carrying out cross-period offset analysis on the user purchase intention vector of the current statistical period and the historical purchase intention vector of the previous statistical period, wherein the cross-period offset analysis at least comprises vector similarity calculation, vector offset calculation and vector offset direction judgment; s5, inputting the cross-period offset analysis result and the user purchase intention vector of the current statistical period into a repeat purchase score generation step together to generate a corresponding repeat purchase score and an offset risk mark; s6, outputting a repurchase prediction result of the target user according to the comparison result of the repurchase score and a preset threshold value and combining the offset risk mark.
  2. 2. The method for predicting and accurately touching the electric user's repurchase behavior by fusing time sequence attention mechanisms according to claim 1, wherein the method comprises the following steps: in step S2, the method for constructing the time sequence enhancement behavior representation sequence includes: mapping the behavior type into a behavior type vector; Mapping commodity information into commodity semantic vectors; The relative time difference between adjacent behaviors and the time interval from the last purchase behavior are respectively mapped into time coding vectors; and fusing the behavior type vector, the commodity semantic vector and the two time coding vectors to obtain the time sequence enhanced behavior representation of the corresponding behavior.
  3. 3. The method for predicting and accurately touching the electric user's repurchase behavior by fusing time sequence attention mechanisms according to claim 1, wherein the method comprises the following steps: In step S2, the relative time difference between the adjacent behaviors is encoded in segments according to a plurality of time intervals, the time interval from the last purchase behavior is encoded in layers according to a preset post-purchase return visit stage, and the relative time difference and the time interval from the last purchase behavior are used as two independent time references to participate in the construction of the time sequence enhancement behavior representation.
  4. 4. The method for predicting and accurately touching the electric user's repurchase behavior by fusing time sequence attention mechanisms according to claim 1, wherein the method comprises the following steps: In step S3, the multi-head attention network includes a time relationship attention header, a behavior transformation attention header, and a commodity semantic attention header; The time relation attention head calculates time relation weight according to the time interval between the behaviors, and performs weighted summation on time sequence enhancement behavior representations of the behaviors according to the time relation weight to obtain a time relation feature vector; The behavior transformation attention head calculates behavior transformation weights according to transformation sequences among behavior types, and performs weighted summation on time sequence enhancement behavior representations of the behaviors according to the behavior transformation weights to obtain behavior transformation feature vectors; The commodity semantic attention head calculates commodity semantic weights according to semantic similarity among commodity information, and performs weighted summation on time sequence enhancement behavior representations of all behaviors according to the commodity semantic weights to obtain commodity semantic feature vectors; and vector splicing is carried out on the time relation feature vector, the behavior transformation feature vector and the commodity semantic feature vector according to a preset sequence, and linear transformation is carried out on the spliced result to obtain the user purchase intention vector in the current statistical period.
  5. 5. The method for predicting and accurately touching the electric user's repurchase behavior by fusing time sequence attention mechanisms according to claim 1, wherein the method comprises the following steps: In step S4, the cross-period offset analysis includes the steps of: S41, calculating cosine similarity between a user purchase intention vector in a current statistical period and a historical purchase intention vector in a previous statistical period; S42, calculating the vector offset of the user purchase intention vector in the current statistical period relative to the historical purchase intention vector in the previous statistical period; s43, determining a vector offset direction according to the coordinate difference value of the user purchase intention vector in the current statistical period and the historical purchase intention vector in the previous statistical period; And S44, outputting the cosine similarity, the vector offset and the vector offset direction as cross-period offset analysis results.
  6. 6. The method for predicting and accurately touching a user' S repurchase behavior of an electric motor with a fusion time sequence attention mechanism according to claim 5, wherein in step S5, the generation method of the repurchase score and the offset risk mark comprises the following steps: Forming a scoring input sequence by each dimension value of the user purchase intention vector in the current statistical period, the cosine similarity and the vector offset according to a preset sequence, and carrying out weighted summation on the scoring input sequence to obtain an initial repeat purchase score; performing direction correction on the initial repurchase score according to the vector offset direction to obtain a corrected repurchase score; generating an offset risk flag when the vector offset is above a first threshold and the cosine similarity is below a second threshold; And when the vector offset is lower than or equal to the first threshold value or the cosine similarity is higher than or equal to the second threshold value, canceling an offset risk mark, and outputting the corrected buyback score as the buyback score.
  7. 7. The method for predicting and accurately touching the user's repurchase behavior of the electric motor fused with the time sequence attention mechanism of claim 6, wherein the method comprises the following steps: The cross-period offset analysis further includes a continuous period stability analysis including the steps of: obtaining vector offset directions of two continuous statistical periods; When the vector offset directions of two continuous statistical periods are reversed, the risk level corresponding to the offset risk mark is improved; and when the cosine similarity of two continuous statistical periods is higher than the second threshold value and the vector offset is lower than the first threshold value, reducing the risk level corresponding to the offset risk mark.
  8. 8. The method for predicting and accurately touching the electric user's repurchase behavior by fusing time sequence attention mechanisms according to claim 1, wherein the method comprises the following steps: In step S6, the multiple purchase prediction result at least includes a multiple purchase determination result, multiple purchase scores and dominant commodity categories, where the dominant commodity categories are determined according to commodity semantic weights corresponding to commodity categories in the current statistical period.
  9. 9. An electric user repurchase behavior prediction and accurate touch system incorporating a time-series attention mechanism for implementing the method of any of claims 1-8, comprising: the behavior acquisition unit is used for acquiring a historical behavior sequence of a target user in a preset statistical period and extracting behavior types, behavior occurrence time and commodity information corresponding to each behavior; The time sequence enhancement representation construction unit is used for carrying out time enhancement processing on the basic behavior representation based on the relative time difference between adjacent behaviors and the time interval between each behavior and the latest purchasing behavior, and generating a time sequence enhancement behavior representation sequence; The multi-head relation modeling unit is used for carrying out independent weighted calculation on the time sequence enhanced behavior expression sequence based on a behavior time adjacent relation, a behavior type conversion relation and a commodity semantic similarity relation respectively, and generating a user purchase intention vector in the current statistical period; The cross-period offset analysis unit is used for carrying out vector similarity calculation, vector offset calculation and vector offset direction judgment on the user purchase intention vector in the current statistical period and the historical purchase intention vector in the previous statistical period, and outputting a cross-period offset analysis result; The score generating unit is used for generating a repurchase score and an offset risk mark according to the cross-period offset analysis result and the user purchase intention vector of the current statistical period; And the result output unit is used for outputting the repurchase prediction result of the target user according to the comparison result of the repurchase score and the preset threshold value and the combination of the offset risk mark.
  10. 10. The e-commerce user repurchase behavior prediction system incorporating a time-sequential attention mechanism of claim 9, wherein: The multi-head relation modeling unit comprises a time relation weighting subunit, a behavior transformation weighting subunit and a commodity semantic weighting subunit; The cross-period offset analysis unit comprises a similarity calculation subunit, an offset direction judgment subunit and a continuous period stability analysis subunit; The score generation unit includes an initial score generation subunit and a risk adjustment subunit.

Description

Method and system for predicting and accurately touching electric business re-purchase behavior by fusing time sequence attention mechanism Technical Field The invention belongs to the technical field of behavior analysis of a user of a mobile phone, and particularly relates to a method and a system for predicting and accurately touching the behavior of a user of the mobile phone by fusing a time sequence attention mechanism. Background Along with the continuous development of an electronic commerce platform, a great amount of behavior data is continuously precipitated in links of browsing, clicking, collecting, purchasing, ordering, evaluating, after-sales and the like by a user, so that the development of user consumption trend identification and marketing decision support based on data analysis becomes an important research direction in the field of electronic commerce. The user repurchase behavior prediction is one of key contents in the user fine operation, and the result can directly influence the judgment of the service life cycle value of the user, the distribution efficiency of marketing resources and the conversion performance of the whole platform. Modeling analysis is performed around the possibility of user repurchase, and differentiated user management is performed accordingly, so that the method becomes an important component in an intelligent electronic commerce operation system gradually. In the prior art, analysis for the re-purchase behavior of the user of the electric user is generally based on the historical order record of the user, the commodity browsing record, the access frequency, the stay time, the collection and purchase condition, the promotion response condition and the basic user portrait information, and a mode of rule screening, statistical analysis or machine learning modeling is adopted to judge whether the user has a re-purchase tendency. The partial scheme is used for carrying out classified management on high-activity users, high-value users or loss risk users by constructing a user tag system, and introducing a regression model, a tree model or a neural network model to calculate the re-purchase probability and using a prediction result in marketing touch scenes such as coupon throwing, short message pushing, in-station message reminding and the like. The technical route still mainly has the following technical problems that firstly, time sequence relation, stage change characteristics and inter-behavior association in a user behavior sequence are insufficiently mined, discrete behaviors are easy to conduct static processing, so that the representation capability of real consumption intention is insufficient, secondly, the importance degree of behavior information under different time windows is lack of effective distinction, key behavior signals which have stronger influence on a repurchase result are difficult to accurately identify, thirdly, the prediction result is often not tightly connected with a follow-up touching strategy, and the situation that mutual rupture between prediction and marketing execution exists, so that the application effect and conversion value of the repurchase prediction result in an actual operation scene are influenced. Therefore, it is necessary to provide a scheme for predicting and accurately touching the user's repurchase behavior facing the e-commerce scene, so as to improve the identification capability of the user's dynamic behavior characteristics, enhance the synergy between the repurchase judgment result and the operation touch, and thereby better satisfy the development requirement of the refined operation of the e-commerce platform. Disclosure of Invention Aiming at the problems, the invention aims to provide a method for predicting and accurately touching the repurchase behavior of an electric user by fusing a time sequence attention mechanism, which comprises the following steps: S1, acquiring a historical behavior sequence of a target user in a preset statistical period, and extracting behavior types, behavior occurrence time and commodity information corresponding to each behavior; S2, converting each behavior in the historical behavior sequence into a basic behavior representation, and performing time enhancement processing on the basic behavior representation based on the relative time difference between adjacent behaviors and the time interval between each behavior and the latest purchasing behavior to construct a time sequence enhancement behavior representation sequence; S3, inputting the time sequence enhancement behavior representation sequence into a multi-head attention network, respectively carrying out independent weighted calculation by different attention heads based on a behavior time adjacent relation, a behavior type conversion relation and a commodity semantic similarity relation, and fusing the output results of the attention heads to obtain a user purchase intention vector in the current statistical perio