Search

CN-122022956-A - Commodity arrival person recommending method based on multi-feature fusion and storage medium

CN122022956ACN 122022956 ACN122022956 ACN 122022956ACN-122022956-A

Abstract

The application provides a commodity speaker recommendation method based on multi-feature fusion and a storage medium. The recommendation method comprises the following steps of S1, extracting the style characteristics of the darers based on historical carried data of the darers, constructing a darer style vector set, S2, generating commodity demand vectors based on commodity information and recommendation demands, S3, calculating matching degree scores of the commodity demand vectors and the darer style vector set, S4, carrying out self-adaptive fusion on the commodity demand vectors and the style characteristics of the darers according to the matching degree scores, and generating a darer matching vector, and S5, generating a darer recommendation list and recommendation reasons based on the darer matching vector.

Inventors

  • GE ZHIHENG
  • XING DONGJIN
  • YANG HONGJIN

Assignees

  • 厦门立马耀网络科技有限公司

Dates

Publication Date
20260512
Application Date
20260209

Claims (10)

  1. 1. The commodity arrival person recommending method based on multi-feature fusion is characterized by comprising the following steps of: s1, extracting style characteristics of a man based on historical cargo data of the man, and constructing a man style vector set; s2, generating commodity demand vectors based on commodity information and recommendation demands; S3, calculating the matching degree score of the commodity demand vector and the darn style vector set; S4, carrying out self-adaptive fusion on the commodity demand vector and the style characteristics of the man according to the matching degree score to generate a man matching vector; And S5, generating a speaker recommendation list and recommendation reasons based on the speaker matching vector.
  2. 2. The commodity pedestrian recommending method based on multi-feature fusion according to claim 1, wherein in step S1, the step of extracting pedestrian style features and constructing a pedestrian style vector set based on historical cargo data of the pedestrian specifically comprises: S11, extracting five types of original features from historical cargo data of a man to form an original feature set F= { F 1 , F 2 , F 3 , F 4 , F 5 }, wherein the first type of features F 1 are cargo capacity feature vectors, the second type of features F 2 are class good feature vectors, the third type of features F 3 are content expression feature vectors, the fourth type of features F 4 are vermicelli feature vectors, and the fifth type of features F 5 are growth potential feature vectors; s12, carrying out feature coding on the five types of original features; S13, fusing and clustering the coded various features to generate a dactylogyrus style vector set.
  3. 3. The method for recommending commodity according to the present invention based on multi-feature fusion of claim 2, wherein said first class feature F 1 comprises at least one vector of a live tape-out force score, a tape-out level, a field-average sales amount, a guest unit price interval, and a tape-out conversion rate; The second class feature F 2 comprises at least one vector of a main product label, a sales volume proportion distribution vector of each product and a history cooperation brand distribution vector; The third type of features F 3 comprises at least one vector of number of people watched in a live broadcast field, average number of comments, average forwarding number, number of praise, average duration of opening and content release frequency; The fourth type of characteristics F 4 comprises at least one vector of a vermicelli gender distribution proportion vector, a vermicelli age group distribution vector, a vermicelli regional distribution vector, a vermicelli city grade distribution vector and an active vermicelli duty ratio; The fifth class of features F 5 includes at least one vector of recent fan growth rate, recent GMV ring ratio change rate, and launch ROI.
  4. 4. The commodity speaker recommendation method according to claim 2, wherein in step S12, said step of feature encoding said five types of original features specifically comprises: Mapping the numerical features to the [0,1] interval x norm = (x-x min )/(x max -x min by using a Min-Max normalization method, wherein x min and x max are the minimum and maximum values of the features in all the owners respectively, and/or The class type features are encoded by One-Hot, the vector length is equal to the total number of classes, the positions of the corresponding classes are 1, the rest are 0, and/or The distribution type characteristic is converted into a probability vector dist= [ p 1 , p 2 , ..., p k ] after K sub-boxes, wherein p i represents the duty ratio of the ith sub-box, and Σp i =1 is satisfied.
  5. 5. The commodity speaker recommendation method based on multi-feature fusion according to claim 4, wherein in step S13, the step of fusing and clustering the encoded various features to generate a speaker style vector set specifically comprises: the coded various features are subjected to weighted splicing according to preset weights to generate a dactylogyrus style vector ：V style = concat(w 1 ·F 1 , w 2 ·F 2 , w 3 ·F 3 , w 4 ·F 4 , w 5 ·F 5 ),, wherein w 1 , w 2 , w 3 , w 4 , w 5 is a weight coefficient of each feature, and the conditions that Sigma w i =1 and concat represents vector splicing operation are met; And carrying out cluster analysis on all the style vector sets { V style-1 , V style-2 , ..., V style-m } of the dado by adopting a K-Means clustering algorithm, wherein the objective function is as follows: j= ΣΣv style-i - μ j || 2 , where μ j is the J-th cluster center, minimizing J by iterative optimization; After the clustering is completed, a style vector set S di ={V style-1 , V style-2 ,...,V style-k of the man d i is formed, wherein k is the style number of the man.
  6. 6. The method for recommending commodity speaker based on multi-feature fusion according to claim 5, wherein in step S2, said generating commodity demand vector based on commodity information and recommendation demand specifically comprises: Defining a commodity as g and a demand vector V demand = [cat g , price g , user g , style g thereof, wherein cat g is a commodity class code vector, price g is a price interval code, user g is a target user portrait vector comprising gender, age, region and city grade distribution, and style g is a brand adjustment demand code; The demand vector V demand =[d 1 , d 2 , ..., d n with the same dimension as the Daren style vector is formed after encoding.
  7. 7. The method for recommending commodity speaker based on multi-feature fusion according to claim 6, wherein in step S3, said step of calculating a matching degree score of said commodity demand vector and said speaker style vector set specifically comprises: for the style vector set S di ={V style-1 ,V style-2 ,...,V style-k of darman d i , the cosine similarity of each style vector to the demand vector V demand is calculated: sim j =cos(V demand ,V style-j )=(V demand ·V style-j )/(||V demand ||×||V style-j |), where j e [1, k ], "·" represents the vector dot product, |·|| represents the L2 norm of the vector; The highest similarity is taken as the match Score for the candidate Score match =max{sim 1 , sim 2 ,...,sim k .
  8. 8. The method for recommending commodity according to the present invention as set forth in claim 7, wherein in step S4, said step of adaptively fusing said commodity demand vector with said style characteristics according to said matching degree score to generate a matching vector of a person comprises: Defining a preset threshold value as theta, wherein a fusion formula is satisfied, V match = α·V style-best + (1-α)·V demand , V style-best is a style vector with highest similarity with a demand vector, and alpha is an adaptive weight coefficient; the calculation method of the weight coefficient alpha comprises the following steps: When Score match is more than or equal to theta, alpha=score match , mainly based on the original style characteristics of the dactylotheca; When Score match < θ, α=score match ×β, where β∈ (0, 1) is the decay factor, more consideration is given to the new demand characteristics of the commodity.
  9. 9. The method for recommending commodity speaker based on multi-feature fusion according to claim 8, wherein said step S4 further comprises: When the matching degree score is lower than a preset value, combining with the growth potential characteristic F 5 , recommending the active and growth potential raters in the near term preferentially.
  10. 10. A computer storage medium storing a computer program executable by a processor of a device in which the computer readable storage medium is located to implement the method of any one of claims 1 to 9.

Description

Commodity arrival person recommending method based on multi-feature fusion and storage medium Technical Field The application relates to the technical field of information, in particular to a commodity speaker recommendation method based on multi-feature fusion and a storage medium. Background With the rapid development of live electronic commerce, how to accurately match most suitable with-product owners for products has become a core problem facing brands and merchants. The effective recommended product can not only improve commodity conversion rate, but also enhance brand influence. The existing Daren recommendation method mainly has the following technical defects: 1. For example, the prior art mainly relies on simple statistical indexes such as vermicelli quantity, historical sales quantity and the like to recommend, and the matching degree of the content styles of the users and the commodity characteristics is not considered. For example, if a high-end skin care brand is selected to be fun to the person to carry goods only according to the amount of vermicelli, the conversion rate may be low due to the fact that the styles are not consistent. 2. The man-in-the-field multifarious can not be identified; a typical style of a person is many, such as a person may have both "professional component analysis" and "daily skin care sharing" styles. In the prior art, the man is often regarded as a single feature vector, and the multifaceted nature of the man cannot be identified and utilized, so that the recommendation result is one-sided. 3. The existing recommendation system cannot flexibly adjust the recommendation strategy when facing new commodities or special classes. When the merchandise demand does not exactly match the existing style of the rater, the system is unable to effectively identify raters with learning ability and growth potential. 4. The recommendation result cannot be explained, most of the existing recommendation systems are black box models, recommendation reasons cannot be explained for brands, the brands cannot understand recommendation logic, and the trust degree and the adoption will of the recommendation systems are reduced. Aiming at the problems, the invention provides a commodity pedestrian recommending method based on multi-feature fusion, which comprehensively characterizes pedestrian styles through five-dimensional features, dynamically balances matching degree by adopting a self-adaptive fusion mechanism, generates interpretable recommending reasons and remarkably improves recommending accuracy and practicality. Disclosure of Invention The invention aims to solve the problems and provide a commodity speaker recommendation method based on multi-feature fusion and a storage medium. The technical scheme of the application is realized as follows: the invention provides a commodity speaker recommendation method based on multi-feature fusion, which comprises the following steps: s1, extracting style characteristics of a man based on historical cargo data of the man, and constructing a man style vector set; s2, generating commodity demand vectors based on commodity information and recommendation demands; S3, calculating the matching degree score of the commodity demand vector and the darn style vector set; S4, carrying out self-adaptive fusion on the commodity demand vector and the style characteristics of the man according to the matching degree score to generate a man matching vector; And S5, generating a speaker recommendation list and recommendation reasons based on the speaker matching vector. As a further improvement, in step S1, the step of extracting the style characteristics of the David based on the historical tape-up data of the David, and constructing the style vector set of the David specifically includes: S11, extracting five types of original features from historical cargo data of a man to form an original feature set F= { F 1, F2, F3, F4, F5 }, wherein the first type of features F 1 are cargo capacity feature vectors, the second type of features F 2 are class good feature vectors, the third type of features F 3 are content expression feature vectors, the fourth type of features F 4 are vermicelli feature vectors, and the fifth type of features F 5 are growth potential feature vectors; s12, carrying out feature coding on the five types of original features; S13, fusing and clustering the coded various features to generate a dactylogyrus style vector set. As a further improvement, the first class characteristic F 1 specifically includes at least one vector of a live tape-out force score, a tape-out grade, a field average sales amount, a guest unit price interval, and a tape-out conversion rate; the second type of characteristics F 2 specifically comprises at least one vector of a main product type label, a distribution vector of sales of each product type and a distribution vector of historical cooperation brands; The third type of characteristics F 3 specifically compri