Method and system for precise matching of experts and brands based on machine learning

Abstract

The present application relates to brand matching technical field, specifically, it relates to a method and system for precise matching of experts and brands based on machine learning. Including the following steps: S1, collecting audience data of the expert live room, and screening the audience data with the viewing time; S2, collecting the fan data of the expert, collecting the fan list according to the fan data, and obtaining the like data and comment data of the expert's works, collecting the potential user list according to the like data and comment data; by dividing the data into continuous periods according to the collection time to generate user sub-portrait of each period, capturing the evolution of user behavior over time, summarizing the user portrait of each period to form the total user portrait, and quantifying the dynamic trend of user groups through difference analysis to provide dynamic map observation for brand matching, and then introducing network hotspot data to quantify the influence factor of hotspot on user portrait change and identify the emerging trend.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brand matching, in particular to a method and system for precise matching of KOLs and brands based on machine learning. BACKGROUND

[0002] In the field of the integration of social media and e-commerce, precise matching of KOLs and brands is a core link to improve marketing efficiency, mainly by filtering potential partners through the overlap of user portraits and the basic characteristics of brand target audiences, to reduce marketing costs and improve conversion efficiency.

[0003] However, most of the existing technologies are only applicable to scenarios where user groups are stable and market trends change slowly, such as static user portrait analysis of traditional e-commerce platforms. In the current short video live streaming scenario with strong real-time performance and fragmented user behavior, relying solely on static fan portraits or single interaction data cannot capture dynamic preference changes of users due to network hotspots and time differences, resulting in a lag in matching results relative to market trends, which causes a lack of timeliness in KOL and brand matching marketing. Therefore, a method and system for precise matching of KOLs and brands based on machine learning are proposed. SUMMARY

[0004] The present application aims to provide a method and system for precise matching of KOLs and brands based on machine learning to solve the problems raised in the background.

[0005] To achieve the above technical problems, one of the purposes of the present application is to provide a method for precise matching of KOLs and brands based on machine learning, comprising the following steps:

[0006] S1, collecting audience data of a KOL live streaming room, and performing effective screening of the audience list based on the viewing time;

[0007] S2, collecting fan data of the KOL, collecting fan lists based on the fan data, and collecting like data and comment data of the KOL's works, and collecting potential user lists based on the like data and comment data;

[0008] S3, generating network portraits for each user based on the audience list, fan list and potential user list of different time periods, and summarizing the network portraits of users corresponding to the KOL in different time periods, and analyzing user difference data in different time periods;

[0009] S4, obtaining network hotspot data, recording the change data of each user's network portrait, and then analyzing the influence of hotspots on each user based on the network hotspot data, and then establishing a machine learning model based on the network hotspot data, user difference data and hotspot influence data, and outputting the real-time user total portrait and predicted user total portrait of the KOL by the machine learning model.

[0010] S5, acquire the order history data, and match the historical purchase records of the corresponding users according to the order history data, and then combine the machine learning model with the historical purchase records of each user to analyze the predicted purchase preferences;

[0011] S6, according to the real-time user total portrait of the influencer, the predicted user total portrait, and the predicted purchase preference, perform correlation analysis on the to-be-matched brand, and perform brand matching recommendation for the influencer according to the correlation.

[0012] As a further improvement of the technical solution, the steps of S1 are as follows:

[0013] S1.1, when the influencer enters the live mode, the audience data of the live room is collected in real time, and the real-time viewing time of each audience is calculated according to the audience data;

[0014] S1.2, according to the set effective viewing time, then compare the real-time viewing time of each audience with the effective viewing time, if the real-time viewing time exceeds the effective viewing time, the audience is determined to be effective, otherwise, if the real-time viewing time does not exceed the effective viewing time, the monitoring is kept;

[0015] S1.3, collect the audiences determined to be effective in S1.2, and obtain the audience list according to the collection result.

[0016] As a further improvement of the technical solution, the steps of S2 are as follows:

[0017] S2.1, log in to the network account of the influencer, and then collect the fan data in the network account to obtain the fan data of the influencer, and collect the fan list according to the fan data to obtain the fan list corresponding to the influencer;

[0018] S2.2, use the network account of the influencer to collect the like data and comment data of the network works of the influencer, and then set the like frequency threshold and comment frequency threshold according to the number of network works of the influencer;

[0019] S2.3, according to the like data and comment data, perform user attribution analysis to obtain the like frequency and comment frequency of each user, and then compare the like frequency and comment frequency of the user with the like frequency threshold and comment frequency threshold;

[0020] When one of the like frequency and comment frequency of the user exceeds the like frequency threshold and comment frequency threshold, the user is determined to be a potential user, and the potential user list is recorded;

[0021] When the like frequency and comment frequency of the user do not exceed the like frequency threshold and comment frequency threshold, the user is kept under monitoring.

[0022] As a further improvement of the technical solution, the S3 is as follows:

[0023] Fan list > audience list;

[0024] Audience list > potential users.

[0025] As a further improvement of the technical solution, the S3 is as follows:

[0026] S3.1, according to the collection time, the audience list, the fan list and the potential user list are divided into time periods, the network portrait of each user in the audience list, the fan list and the potential user list in different time periods is generated, and the network portrait corresponding to each user is obtained according to the generation result;

[0027] S3.2, the network portrait of the user is summarized and divided into time periods, so as to obtain the portrait corresponding to each user in different time periods and the total portrait of the user corresponding to the influencer in different time periods;

[0028] S3.3, the total network portrait of all users corresponding to the influencer in different time periods is analyzed, and the user difference data corresponding to the influencer is obtained.

[0029] As a further improvement of the technical solution, the S4 is as follows:

[0030] S4.1, the network hotspot data in different time periods is obtained, the network portrait of each user in different time periods is analyzed, the change data corresponding to the network portrait of the user in different time periods is obtained, and the change data in different time periods is combined with the network hotspot data to analyze the hotspot influence data, and the hotspot influence data corresponding to the user is obtained;

[0031] S4.2, a machine learning model is established according to the real-time network hotspot data, the change data, the user difference data and the hotspot influence data, and then the real-time total user portrait and the predicted total user portrait corresponding to the influencer are output after the machine learning model learns the data.

[0032] As a further improvement of the technical solution, the user total portrait of S3.2 and the real-time user total portrait of S4.2 are different, the user total portrait of S3.2 is defined as the basic user total portrait, and the real-time user total portrait of S4.2 is defined after the basic user total portrait is corrected.

[0033] As a further improvement of the technical solution, the S5 is as follows:

[0034] S5.1, obtain order history data in the network account of the celebrity and the brand sales end, match the historical purchase record of the corresponding user according to the order history data, and obtain the historical purchase record corresponding to each user;

[0035] S5.2, calculate the chain value between each product purchase, then input the historical purchase record of each user and the chain data into the machine learning model, and output the predicted purchase preference of each user by the machine learning model.

[0036] The second purpose of the application is to provide a celebrity and brand precise matching system based on machine learning, which comprises the machine learning based celebrity and brand precise matching method described in any one of the above embodiments, and comprises a list collection module, a list analysis module and a brand matching module.

[0037] The list collection module is used for collecting audience data for effective screening of the audience list, collecting fan data for collection of the fan list, collecting like data and comment data for collection of the potential user list, generating a network portrait for each user according to the audience list, the fan list and the potential user list in different time periods, and analyzing the user difference data in different time periods.

[0038] The list analysis module is used for obtaining network hotspot data, recording the change data of the network portrait of each user, then analyzing the hotspot influence data of each user in combination with the network hotspot data, and then establishing a machine learning model according to the network hotspot data, the user difference data and the hotspot influence data, and outputting the real-time user total portrait and the predicted user total portrait of the celebrity by the machine learning model.

[0039] The brand matching module is used for obtaining order history data, and matching the historical purchase record of the corresponding user according to the order history data, then combining the machine learning model with the historical purchase record of each user to analyze the predicted purchase preference, and analyzing the correlation degree between the real-time user total portrait, the predicted user total portrait and the predicted purchase preference of the celebrity and the brand to be matched, and recommending the brand matching for the celebrity according to the correlation degree.

[0040] Compared with the prior art, the application has the following advantages:

[0041] 1. A method and system for precise matching of influencers and brands based on machine learning, which divides data into continuous time periods according to collection time to generate user sub-portraits for each time period, captures the evolution of user behavior over time, aggregates the portraits of each time period to form a user total portrait, and quantifies the dynamic trends of user groups through difference analysis to provide a dynamic map for brand matching. Then, by introducing network hotspot data, the influence factors of hotspots on user portrait changes are quantified to identify emerging trends, a dynamic model is constructed using an LSTM network, real-time features are input to generate real-time total portraits, and future trends are predicted based on historical sequences, so that brand matching can respond to current needs and predict long-term trends.

[0042] 2. A method and system for precise matching of influencers and brands based on machine learning, which collects data on audience, fans, and work interactions in the influencer's live room, constructs a three-dimensional data set covering user travel, social attributes, and content preferences, avoids one-sidedness of single-dimensional evaluation, and filters core audiences based on effective viewing time to avoid interference from transient audiences. High participation potential users are identified by setting thresholds for likes and comments, and priority rules are used to ensure that each user is assigned only one high-value label, providing a pure data basis for subsequent portrait generation. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The overall flowchart of the present application is shown in the figure;

[0044] Figure 2 The flowchart of the present application for obtaining an audience list based on the collection results is shown in the figure;

[0045] Figure 3 The flowchart of the present application for obtaining the fan list corresponding to the influencer is shown in the figure;

[0046] Figure 4 The flowchart of the present application for obtaining user difference data corresponding to the influencer is shown in the figure;

[0047] Figure 5 The flowchart of the present application for outputting real-time user total portraits and predicted user total portraits corresponding to the influencer after learning data by the machine learning model is shown in the figure;

[0048] Figure 6 The flowchart of the present application for outputting the predicted purchase preferences of each user by the machine learning model is shown in the figure. DETAILED DESCRIPTION

[0049] With reference to the drawings and embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0050] As shown in Figures 1-6 One of the purposes of the present application is to provide a method for accurately matching KOLs and brands based on machine learning, which comprises the following steps:

[0051] S1, collecting audience data of a KOL live broadcast room, and screening the audience data according to a valid audience list according to a viewing time length;

[0052] The step of S1 is as follows:

[0053] S1.1, when a KOL enters a live broadcast mode, collecting audience data of a live broadcast room in real time, and calculating a real-time viewing time length of each audience according to the audience data;

[0054] When the KOL starts the live broadcast mode, relying on a monitoring system of a live broadcast platform, data of all audiences entering the live broadcast room in a current live broadcast period are obtained, covering audience entering and exiting time stamps and login identifiers;

[0055] For each audience, the entering time collected and the current time monitored in real time are used to calculate a real-time viewing time length of the audience from entering to the current time;

[0056] When the user exits, the calculation is stopped, or when the user exits and then enters again, the timing is restarted until the one-time viewing time length of the user in the live broadcast room exceeds the valid viewing time length.

[0057] S1.2, according to a set valid viewing time length (a valid viewing time length threshold is set in the system, which can be determined according to industry experience and past live broadcast data of the KOL, for example, set to 10 minutes), then the real-time viewing time length of each audience is compared with the valid viewing time length, when the real-time viewing time length exceeds the valid viewing time length, the audience is determined to be valid, otherwise, when the real-time viewing time length does not exceed the valid viewing time length, the monitoring is continued;

[0058] The audience determined to be valid in S1.2 is collected, and an audience list is obtained according to the collection result;

[0059] S1.3, collecting the audiences determined to be valid in S1.2, and obtaining an audience list according to the collection result;

[0060] The above determination is continuously performed during the live broadcast, and after the live broadcast ends, a temporary list of valid audiences is summarized, and the final audience list is obtained after deduplication.

[0061] S2, collect the fan data of the person, and collect the fan list according to the fan data, and obtain the like data and comment data of the works of the person, and collect the potential user list according to the like data and the comment data;

[0062] The steps of S2 are as follows:

[0063] S2.1, log in to the network account of the person, then collect the fan data in the network account, thereby obtaining the fan data of the person, and collect the fan list according to the fan data, and obtain the fan list corresponding to the person;

[0064] The API interface opened by the platform is used to obtain the structured data related to the fans from the logged-in network account of the person, which covers the unique identification of the fans (such as user ID), the time of following, the interactive relationship mark and other information;

[0065] The collected fan data is cleaned, and the repeated and invalid (such as unfollowed but not updated in time) fan records are removed, the unique identification information of the fans is extracted, and the fan list corresponding to the person is formed by arranging and summarizing according to a unified format (such as a list, a database table, etc.), which is used for subsequent user portrait, matching analysis and other links.

[0066] S2.2, use the network account of the person to collect the like data and comment data of the network works of the person, and then set the like number threshold and the comment number threshold according to the number of network works of the person (statistically obtain the total number of network works of the person, set the like number threshold and the comment number threshold according to the number of works and industry experience, for example, if the person has ten network works, set three likes or three comments), the formula is as follows:

[0067] ;

[0068] Wherein, T p is the like number threshold, n is the total number of network works of the person, p i is the like number of the i th work, and a p is the like adjustment coefficient;

[0069] ;

[0070] Wherein, T c is the comment number threshold, c i is the comment number of the i th work, and a c is the comment adjustment coefficient.

[0071] Through the authorization interface of the expert network account, all network works (such as videos, texts, etc.) published by the expert network account are batch acquired, and for each work, like data (containing a list of like user IDs) and comment data (containing comment user ID, content and timestamp) are extracted;

[0072] S2.3, according to the like data and the comment data, user attribution analysis is performed to obtain the like number and the comment number of each user, and then the like number and the comment number of the user are compared with a like number threshold and a comment number threshold;

[0073] When one of the like number and the comment number of the user exceeds the like number threshold and the comment number threshold, the user is determined to be a potential user, and the potential user list is recorded;

[0074] When the like number and the comment number of the user do not exceed the like number threshold and the comment number threshold, the user is continuously monitored.

[0075] S3, according to the audience list, the fan list and the potential user list in different time periods, a network portrait of each user is generated, and the network portraits of the users corresponding to the expert in different time periods are summarized, and user difference data analysis in different time periods is performed;

[0076] In the network portrait generation of S3, when the same user appears in the audience list, the fan list and the potential user list, the user is only assigned to one list, and the priority is as follows:

[0077] Fan list > audience list;

[0078] Audience list > potential user.

[0079] If the user appears in the fan list and other lists at the same time, the user is preferentially assigned to the fan list;

[0080] If the user does not appear in the fan list but appears in the audience list and the potential user list at the same time, the user is preferentially assigned to the audience list;

[0081] For each user, only the highest priority list is retained, and other low-priority list associations are discarded;

[0082] According to the priority allocation result, the fan list, the audience list and the potential user list are reorganized to ensure that each user only appears in one list.

[0083] The steps of S3 are as follows:

[0084] S3.1, according to the collection time, the audience list, the fan list and the potential user list are time period divided, the network portrait of each user in the audience list, the fan list and the potential user list in different time period is generated, and the network portrait corresponding to each user is obtained according to the generation result, the specific steps are as follows:

[0085] Time period division: according to the data collection time range (one month), the time axis is divided into several continuous and non-overlapping time periods, and each user record in the audience list, the fan list and the potential user list is distributed to the sub list of the corresponding time period according to the first appearance or interaction time (such as entering the live room time, attention time, first comment time);

[0086] Portrait feature extraction: for each user sub list in each time period, the user features are extracted from multiple dimensions:

[0087] Behavioral characteristics (watching time, interaction frequency, content preference, etc.);

[0088] Social characteristics (attention relationship, interaction object, social group attribute, etc.);

[0089] Network portrait generation: the features extracted for each user in each time period are integrated to generate the network portrait corresponding to the time period.

[0090] S3.2, the network portrait of the user is summarized and time period divided, so as to obtain the portrait corresponding to each user in different time period and the total portrait of the user corresponding to the influencer in different time period;

[0091] The network portrait generated by each user in different time period is associated according to the user ID, forming the portrait time sequence of each user, and for each time period, the portraits of all users in the time period are summarized to construct the total portrait of the user of the influencer in the time period (statistical mean).

[0092] S3.3, the total network portrait of all users corresponding to the influencer in different time period is analyzed, the user difference data corresponding to the influencer is obtained, and the user difference data includes the following:

[0093] Characteristic value change (average watching time, increase and decrease of interaction frequency);

[0094] Distribution shift (age layer distribution, transfer of interest preference);

[0095] User loss or addition (analysis of the intersection and complement set of user set in different time period);

[0096] The analysis result is arranged as the user difference data (change trend chart) corresponding to the influencer.

[0097] S4, obtain network hotspot data, record the change data of each user network portrait at the same time, then analyze the hotspot influence data of each user combined with the network hotspot data, and then establish a machine learning model according to the network hotspot data, user difference data and hotspot influence data, and output the real-time user total portrait and the predicted user total portrait of the expert from the machine learning model;

[0098] The steps of S4 are as follows:

[0099] S4.1, obtain network hotspot data at different time periods, and analyze the changes of each user's network portrait at different time periods to obtain the change data corresponding to the network portrait of the user at different time periods, and analyze the hotspot influence data by combining the change data at different time periods with the network hotspot data to obtain the corresponding hotspot influence data of the user, the specific steps are as follows:

[0100] Network hotspot data acquisition: Collect network hotspot topics, keywords and their heat values (such as search volume, discussion volume) at different time periods from authoritative information platforms, social media and other channels, and clean and classify the hotspot data (such as by industry, event type) to construct a hotspot feature vector;

[0101] User portrait change analysis: Compare the network portraits of the same user at adjacent or specific time periods, calculate the change value of each feature dimension (such as interest label weight change, consumption preference shift), and arrange the change value into a user portrait change vector to reflect the dynamic adjustment of user behavior and preference;

[0102] Hotspot influence data analysis: Correlate the user portrait change vector with the network hotspot feature vector at the same period to identify the influence of the hotspot topic on the user behavior (calculate the correlation between the hotspot topic and the user portrait change, and quantify the degree of hotspot influence), and generate the hotspot influence data corresponding to each user, the formula is as follows:

[0103] ;

[0104] Where, I j，k is the influence factor of the user to the hotspot, Corr is the correlation function, which measures the correlation between the hotspot and the user portrait change, h k is the heat value of the kth hotspot topic at time period t, Amp(h k ) is the influence magnifier of the hotspot, ΔP j is the portrait change vector of the user from time period t to t+1.

[0105] S4.2, establish a machine learning model according to real-time network hotspot data, change data, user difference data and hotspot influence data, then output the real-time user total portrait and the predicted user total portrait of the expert from the machine learning model after learning the data, the formula is as follows:

[0106] ;

[0107] wherein, X m is the feature vector of the mth input sample, Diff j is the user difference data vector;

[0108] ;

[0109] wherein, G output is the user total portrait output by the model, Model is a machine learning model function (LSTM network), and θ is a set of model parameters;

[0110] ;

[0111] ;

[0112] wherein, G real-time is the real-time user total portrait, X current is the real-time feature vector of the current time period, G prediction is the predicted user total portrait, X history is a historical multi-time period feature sequence.

[0113] The user total portrait of S3.2 and the real-time user total portrait of S4.2 are different, the user total portrait of S3.2 is defined as a basic user total portrait, and the real-time user total portrait of S4.2 is defined after the basic user total portrait is corrected.

[0114] S5, obtaining order history data, and matching historical purchase records of corresponding users according to the order history data, and then combining a machine learning model with the historical purchase records of each user to perform prediction and purchase preference analysis;

[0115] The steps of S5 are as follows:

[0116] S5.1, obtaining order history data from a network account of a person with outstanding skills and a brand sales end, matching historical purchase records of corresponding users according to the order history data, and obtaining historical purchase records corresponding to each user;

[0117] Obtaining user order data through a network account of a person with outstanding skills and a brand sales platform (such as an e-commerce platform and a brand official website), including order ID, user ID, purchase product information (such as product category, brand, and price), purchase time, and other fields, taking the user ID as an index, associating and integrating all order records of the same user, and forming a historical purchase sequence of the user.

[0118] S5.2, Calculate the chain value between each product purchase, then input the historical purchase records of each user and the chain data into the machine learning model, and output the predicted purchase preference of each user from the machine learning model;

[0119] Analyze the correlation of adjacent products in the user purchase sequence:

[0120] Statistical product combination of the same user multiple purchases (such as "shampoo + conditioner"), calculate its co-occurrence frequency, and use association rule algorithm (Apriori) to quantify the chain strength between products;

[0121] Convert the historical purchase records (product ID sequence) and chain values (such as association rule weight) of each user into a feature vector recognizable by the model, input the feature vector to train the model, and the target is to predict the product category or brand that the user may purchase in the future;

[0122] Input the latest purchase records and chain data of the user into the trained model, and output the preference probability or ranking of the user for each product / brand, forming a predicted purchase preference list.

[0123] S6, According to the real-time user total portrait of the influencer, the predicted user total portrait, and the predicted purchase preference, perform correlation analysis with the brand to be matched, and recommend brand matching for the influencer according to the correlation degree, the specific steps are as follows:

[0124] Brand portrait construction: Collect the basic information of the brand to be matched, including target audience characteristics (such as age, gender, regional distribution), product categories, marketing tone (such as high-end, youth, environmental protection), and recent promotion keywords;

[0125] Real-time user total portrait and brand correlation degree calculation: Match the real-time user total portrait of the influencer with the brand portrait, compare the coincidence degree of user characteristics and brand target audience, and analyze the relevance of user recent interaction hotspots and brand marketing theme;

[0126] Predicted user total portrait and brand long-term matching degree analysis: Based on the predicted user total portrait, evaluate the long-term fit degree of the brand and the influencer audience, such as whether the predicted user demand change is consistent with the brand product line upgrade direction (such as the user's future demand for smart home rising, the brand plans to launch related products), and compare the user's predicted purchase preference with the brand's core product categories and SKU to calculate the category matching degree;

[0127] Comprehensive correlation degree score and recommendation: Set the weight of each dimension to calculate the comprehensive correlation degree score of the influencer and the brand, generate a brand recommendation list in descending order of score, and preferentially recommend high-correlation-degree brands, the formula is as follows:

[0128] ;

[0129] wherein R real is the cosine similarity between the real-time user total portrait and the brand, and B is the brand portrait vector;

[0130] ;

[0131] wherein R pred is the trend correlation coefficient between the predicted user total portrait and the brand future positioning, and B futur is the brand future strategy feature vector;

[0132] ;

[0133] wherein R purchase is the purchase preference matching degree, E is the product category set, w e is the weight of the category e, Match (p e , b e ) is the matching function of the user preference category p e and the brand main category b e , and the matching is 1 and the mismatching is 0;

[0134] ;

[0135] wherein A is the comprehensive score, β1, β2 and β3 are weight coefficients of each dimension, and β1+β2+β3=1.

[0136] The second object of the present application is to provide a machine learning-based influencer and brand precise matching system, which comprises the machine learning-based influencer and brand precise matching method of any one of the above, and comprises a list collection module, a list analysis module and a brand matching module.

[0137] The list collection module is used for collecting audience data for effective screening of an audience list, collecting fan data for collection of a fan list, collecting like data and comment data for collection of a potential user list, generating a network portrait for each user according to the audience list, the fan list and the potential user list in different time periods, and performing user difference data analysis in different time periods.

[0138] The list analysis module is used for obtaining network hotspot data, recording change data of the network portrait of each user, then performing hotspot influence data analysis on each user in combination with the network hotspot data, and then establishing a machine learning model according to the network hotspot data, the user difference data and the hotspot influence data, and outputting the real-time user total portrait and the predicted user total portrait of the influencer by the machine learning model.

[0139] The brand matching module is used for obtaining order history data, and simultaneously performs historical purchase record matching for corresponding users according to the order history data, then combines a machine learning model with historical purchase records of each user to perform prediction purchase preference analysis, performs correlation degree analysis on the real-time user total image of the influencer, the predicted user total image, the predicted purchase preference and the brand to be matched, and performs brand matching recommendation for the influencer according to the correlation degree.

[0140] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited by the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for precise matching between experts and brands based on machine learning, characterized in that: It comprises the following steps: S1, collecting audience data of the live room of the influencer, and screening the audience data according to the viewing time; S2, collecting the fan data of the influencer, collecting the fan list according to the fan data, and obtaining the like data and comment data of the works of the influencer, and collecting the potential user list according to the like data and comment data; S3, generating a network portrait for each user according to the audience list, the fan list and the potential user list in different time periods, and summarizing the network portraits of the users corresponding to the influencer in different time periods, and analyzing the user difference data in different time periods; S4, obtaining network hotspot data, recording the change data of each user network portrait, and then analyzing the hotspot influence data of each user combined with the network hotspot data, and then establishing a machine learning model according to the network hotspot data, the user difference data and the hotspot influence data, and outputting the real-time total user portrait and the predicted total user portrait of the influencer from the machine learning model; S5, obtaining the historical order data, and then matching the historical purchase records of the corresponding users according to the historical order data, and then analyzing the predicted purchase preference of each user combined with the machine learning model; S6, analyzing the correlation degree between the real-time total user portrait, the predicted total user portrait and the predicted purchase preference of the influencer and the brand to be matched, and recommending the brand matching for the influencer according to the correlation degree. 2.The method of claim 1, wherein: The steps of S1 are as follows: S1.1, when the influencer enters the live mode, the audience data of the live room is collected in real time, and the real-time viewing time of each audience is calculated according to the audience data; S1.2, set the effective viewing time, then compare the real-time viewing time of each audience with the effective viewing time, if the real-time viewing time exceeds the effective viewing time, the audience is determined to be effective, otherwise, if the real-time viewing time does not exceed the effective viewing time, the audience is kept monitoring; S1.3, collect the audiences determined to be effective in S1.2, and obtain the audience list according to the collection result. 3.The method of claim 1, wherein: The steps of S2 are as follows: S2.1, log in to the network account of the influencer, and then collect the fan data in the network account to obtain the fan data of the influencer, and collect the fan list according to the fan data to obtain the fan list corresponding to the influencer; S2.2, use the network account of the influencer to collect the like data and comment data of the network works of the influencer, and then set the like times threshold and the comment times threshold according to the number of network works of the influencer; S2.3, analyze the user attribution according to the like data and comment data, obtain the like times and comment times of each user, and then compare the like times and comment times of the user with the like times threshold and the comment times threshold; if one of the like times and comment times of the user exceeds the like times threshold and the comment times threshold, the user is determined to be a potential user, and the potential user list is recorded; if the like times and comment times of the user do not exceed the like times threshold and the comment times threshold, the user is kept monitoring. 4.The method of claim 1, wherein: The S3 is in the process of network portrait generation, when the same user appears in the audience list, the fan list and the potential user list, only one list is assigned to the user, and the priority is as follows: Fan list > audience list; Audience list > potential user. 5.The method of claim 1, wherein: The steps of the S3 are as follows: S3.1, according to the collection time, the audience list, the fan list and the potential user list are divided into time periods, the network portrait of each user in the audience list, the fan list and the potential user list in different time periods is generated, and the network portrait corresponding to each user is obtained according to the generation result; S3.2, the network portraits of users are summarized and divided into time periods, so as to obtain the portrait corresponding to each user in different time periods and the total portrait of users corresponding to the influencer in different time periods; S3.3, the total network portrait of all users corresponding to the influencer in different time periods is analyzed, and the user difference data corresponding to the influencer is obtained. 6.The method of claim 1, wherein: The steps of the S4 are as follows: S4.1, the network hotspot data in different time periods is obtained, the network portrait of each user in different time periods is analyzed, the change data corresponding to the network portrait of the user in different time periods is obtained, and the change data in different time periods is combined with the network hotspot data to analyze the hotspot influence data, and the hotspot influence data corresponding to the user is obtained; S4.2, a machine learning model is established according to the real-time network hotspot data, the change data, the user difference data and the hotspot influence data, and then the real-time total portrait of users corresponding to the influencer and the predicted total portrait of users are outputted after the machine learning model learns the data. 7.The method of claim 5, wherein: The total portrait of users of S3.2 and the real-time total portrait of users of S4.2 are different, the total portrait of users of S3.2 is defined as the basic total portrait of users, and the real-time total portrait of users of S4.2 is defined as the modification of the basic total portrait of users. 8.The method of claim 1, wherein: The steps of the S5 are as follows: S5.1, the order history data of the influencer network account and the brand sales end is obtained, the historical purchase record of the corresponding user is matched according to the order history data, and the historical purchase record corresponding to each user is obtained; S5.2, the chain value between the purchase of each product is calculated, and then the historical purchase record of each user and the chain data are inputted into the machine learning model, and the predicted purchase preference of each user is outputted by the machine learning model.

9. A machine learning based influencer and brand precision matching system for implementing the machine learning based influencer and brand precision matching method of any one of claims 1-8, characterized in that: It comprises a list collection module, a list analysis module and a brand matching module; The list collection module is used for collecting audience data to effectively screen the audience list, collecting fan data to collect the fan list, collecting like data and comment data to collect the potential user list, generating the network portrait of each user according to the audience list, the fan list and the potential user list in different time periods, and analyzing the user difference data in different time periods; The list analysis module is used for obtaining network hotspot data, recording the change data of the network portrait of each user, then analyzing the hotspot influence data of each user combined with the network hotspot data, and then establishing a machine learning model according to the network hotspot data, the user difference data and the hotspot influence data, and outputting the real-time total portrait of users of the influencer and the predicted total portrait of users by the machine learning model; The brand matching module is used for obtaining order history data, and simultaneously matches historical purchase records of corresponding users according to the order history data. Then, the machine learning model is combined with the historical purchase records of each user to perform prediction purchase preference analysis. According to the real-time user total image of the influencer, the predicted user total image, and the predicted purchase preference, correlation degree analysis is performed on the to-be-matched brand. According to the correlation degree, brand matching recommendation is performed on the influencer.

Images