Data mining model construction method and system based on machine learning

Abstract

The invention relates to the technical field of big data delivery, in particular to a data mining model construction method and system based on machine learning. Comprising the following steps: firstly, generating an initial target user portrait by collecting personal data of a target user; secondly, establishing an initial friend recommendation sequence, and calculating three different types of recommendation users by the sequence, mainly comprising the following steps of: calculating a first recommendation value by a target user by using a contact person of a registered user on the mobile equipment, and judging through a threshold value; therefore, the relevance of recommendation can be improved; a recommendation expectation is calculated through the ratio and intimacy of the common friends, and judgment is carried out through a threshold value; therefore, the social relation between the target user and the potential recommended friends can be reflected more accurately; a strange user carries out recommendation by matching the initial user portrait; and finally, the initial friend recommendation sequence is adjusted by enhancing the user portrait and analyzing historical recommendation, so that the optimal friend recommendation sequence is obtained.

Description

Technical Field

[0001] The present invention relates to the field of big data delivery technology, and in particular to a method and system for constructing a data mining model based on machine learning. Background Art

[0002] With the rapid development of Internet technology, today's society is in an era of information explosion; the Internet makes it very convenient to obtain and share information, resulting in an exponential growth of information. The amount of data generated every day has reached a massive level, including data in various forms such as text, images and videos.

[0003] There are many forms of potential value behind each piece of data, which needs to be mined and realized through appropriate data management and analysis methods. As computer technology matures, better use of this data can create greater value and benefits for individuals, enterprises and society.

[0004] When analyzing data, data mining technology is a key method used to extract useful patterns, relationships or trends from large amounts of data. It combines theories and methods from multiple fields such as database technology, machine learning and statistics. In the data mining process, the following main steps are included: data cleaning, data integration, data selection, data transformation, data mining, pattern evaluation and knowledge representation. Data mining has a wide range of applications in various fields, such as marketing, finance, healthcare, e-commerce and social media. It can help companies discover customer behavior patterns, predict sales trends, and improve the accuracy of product recommendations, thereby optimizing operations and decision-making processes. Data mining has a wide and far-reaching application in the field of social media. It can help analyze large amounts of user-generated data, reveal user behaviors, trends and preferences, and provide valuable insights and strategies for companies and decision makers.

[0005] As an important service in social media applications, friend recommendation has become a research hotspot in recent years. Accurate friend recommendation can be achieved through data mining technology, which not only increases user participation, improves user retention rate, and improves user experience and satisfaction, but also enables users to develop and expand their social circles, save search time and energy, and enhance social satisfaction and sense of belonging.

[0006] In the existing literature, many methods mine user data based on machine learning technology to achieve accurate friend recommendations; these methods combine the powerful functions of machine learning technology and the ability to process big data, and can more accurately understand and predict user behavior and preferences, thereby achieving more effective friend recommendation services. The development of this method not only improves the user experience of social platforms, but also provides new possibilities for the innovation of business models. For example, the Chinese patent with publication number CN104731962B discloses a friend recommendation method and system based on similar communities in social networks; the invention finds social communities in social networks through user social information, proposes a rule for calculating the similarity between communities from the aspects of structure and attributes, classifies users by using the topological structure of user nodes in communities, and defines the role of nodes in the network by analyzing the role status of user nodes in their own communities and other communities, and provides users of the target community with corresponding similar or multiple categories of users in similar communities as friend recommendation sets. This method expands the traditional similarity calculation between individual users to the calculation of similarity between communities, and expands the traditional friend recommendation problem between users to classify users through role discovery in similar communities, and recommends the same or multiple categories of users to each category of users in the target community; for example, the Chinese patent with publication number CN111782963A discloses a social network data mining method, system, and service equipment based on SNS; the invention constructs a data model of the social network based on the interactive relationship between any two users in SNS; based on the data model, performs social network data mining operations. The above-mentioned prior art realizes a good friend recommendation service in the social network, but there are some defects, that is, it only considers the similarity between two users, but ignores that social relationships in real dating scenarios are also an important factor in becoming friends; such a method has some limitations, resulting in inaccurate friend recommendations.

[0007] To this end, the present invention proposes a method and system for constructing a data mining model based on machine learning. Summary of the invention

[0008] The purpose of the present invention is to propose a data mining model construction method and system based on machine learning for realizing friend recommendation in a social platform; the present invention mainly starts from the following points; first, the present invention generates an initial user portrait according to the basic information filled in by the user on the social platform, and generates an initial friend recommendation sequence according to the types of recommendations of three different friends; wherein, it includes calculating qualified candidate recommended friends according to the contacts of the social platform account registered by the user's mobile device, calculating candidate recommended friends according to users with common friends, and calculating candidate recommended friends according to strangers; in terms of contacts, the first recommendation value is mainly calculated by the intimacy of the user's notes on the contact combined with the time the contact is saved, and the first recommended user list is created by comparing the first recommendation threshold; in terms of having common friends, the friend recommendation expectation is calculated by the probability of having common friends between the recommended friend and the user combined with the intimacy between the friends, and the second user recommendation list is created by comparing the second recommendation threshold; in terms of strangers, the matching degree of the initial user portrait is compared, and the third user recommendation list is created by comparing the third recommendation threshold; wherein, in order to further improve the friend recommendation Reliability, the present invention collects historical behavior data of users on social platforms for analysis, enhances the initial user portrait, and obtains an enhanced user portrait; in the enhancement process, the historical behavior data is mainly characterized and enhanced in combination with the initial user portrait; finally, the present invention proposes a screening method for a friend recommendation list, which enhances the initial target user portrait by collecting the historical behavior data of the target user on the social platform, thereby performing initial screening on the generated initial friend recommendation list, which can help users find friends who meet their needs more quickly; in addition, the present invention analyzes the historical friend recommendation results of the target user, and establishes a target user-successful recommendation user positive correlation matrix, a successful recommendation user-successful recommendation user deviation matrix, a target user-failed recommendation user negative correlation matrix, and a failed recommendation user-failed recommendation user deviation matrix for successfully recommended users and failed users; these four matrices analyze the relationship between the target user and the historically recommended friends from the positive and reverse directions respectively, and finally calculate the friend recommendation adjustment matrix based on the above four matrices; the friend recommendation adjustment matrix can be used to further refine the friend recommendation scheme.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] Obtaining the personal information submitted by the target user on the social platform; generating an initial target user portrait; wherein the personal information includes: text information and image information;

[0011] The process of generating the initial target user portrait includes:

[0012] Further, the personal information is preprocessed to obtain standard personal information; wherein the preprocessing includes: text preprocessing and image preprocessing;

[0013] The text preprocessing includes: segmenting the text information to obtain multiple phrases; embedding the multiple phrases into word vectors to obtain multiple high-dimensional real number vectors; performing data cleaning on the high-dimensional real number vectors to obtain standard high-dimensional real number vectors; and performing data compression on the standard high-dimensional data vectors to obtain standard dimensional data vectors.

[0014] The image preprocessing includes: adjusting the size of the image information to obtain a uniform size image; enhancing the uniform size image to obtain an enhanced image; extracting features from the enhanced image to obtain image features; converting the image features into feature vectors; and performing dimensionality reduction processing on the feature vectors to obtain standard dimensional image feature vectors;

[0015] Further, the standard dimension data vector and the standard dimension image feature vector are input into a machine learning model to obtain an initial user portrait analysis value;

[0016] Furthermore, the initial target user portrait is obtained according to the initial user portrait analysis value.

[0017] Further, a first recommended user list is created; wherein the process of creating the first recommended user list includes: retrieving the contact list of the target user's mobile device; obtaining the contact information in the contact list; determining whether the contact information has registered an account on the social platform, and obtaining a set of registered account contacts; calculating a first recommendation value for each contact in the set of registered account contacts; wherein the calculation formula for the first recommendation value is: FRV i =α1*CNW i +α2*CRL i Among them, FRV i Represents the first recommended value of contact i; CNW i Represents the remark weight of contact i; CRL i represents the standardized storage time of contact i in the contact list; α1 represents the recommendation influence coefficient of the remark information; α2 represents the recommendation influence coefficient of the storage time of the contact information; and α1≥α2; when the first recommendation value is greater than the first recommendation threshold, the contact is added to the first recommended user list;

[0018] Further, creating a second recommended user list;

[0019] The process of creating the second recommended user list includes:

[0020] Obtaining the first friend list of the target user on the social platform;

[0021] Further, obtaining a second friend list for each friend object in the first friend list;

[0022] Furthermore, the friend recommendation expectation of the target user and each user in the second friend list is calculated to construct a friend recommendation expectation vector; wherein the calculation of the friend recommendation expectation includes two forms:

[0023] If there is no friendship between the users in the first friend list, and the second friend list contains common friends; wherein the calculation formula for the friend recommendation expectation is:

[0024]

[0025] Among them, E (T_user,R_user) N represents the friend recommendation expectation for the target user and the user in the second friend list; R_user N is represented by the number of friends in the second friend list and the first friend list; S_user It is represented by the number of users in the first friend list; (T_user，S_user) It is represented by the friend familiarity between the target user and the user in the first friend list; (S_user，R_user) represents the friend familiarity between the user in the first friend list and the user in the second friend list; b1 represents the friend recommendation coefficient of the proportion of common friends; b2 represents the friend recommendation coefficient when there is no friend relationship between the users in the first friend list; ∑ represents the symbol for calculating the total friend familiarity;

[0026] The calculation process of the friend familiarity includes: collecting friend familiarity analysis information; wherein the friend familiarity analysis information includes: friend remark information and friend interaction information; performing information processing on the friend familiarity analysis information to obtain standard friend familiarity analysis data; using a feature engineering algorithm to characterize the standard friend familiarity analysis data to obtain a friend familiarity analysis feature matrix; establishing a friend familiarity evaluation index; inputting the friend familiarity analysis feature matrix and the friend familiarity evaluation index into a friend familiarity evaluation model to obtain the friend familiarity; wherein the model function of the friend familiarity evaluation model is expressed as:

[0027] FFEF(X,FFEI)=OutputLayer(HiddenLayer(InputLayer(X,FFEI)));

[0028] Among them, FFEF() represents the model function of the friend familiarity evaluation model; X represents the friend familiarity analysis feature matrix; FFEI represents the friend familiarity evaluation index; InputLayer() represents the input layer of the friend familiarity evaluation model, which is used to receive the friend familiarity analysis feature matrix and the friend familiarity evaluation index; HiddenLayer() represents the hidden layer of the friend familiarity evaluation model, which is used to learn the mapping relationship between the friend familiarity analysis feature matrix and the friend familiarity evaluation index; OutputLayer() represents the output layer of the friend familiarity evaluation model, which is used to output friend familiarity;

[0029] If there is a friendship between the users in the first friend list, and the second friend list contains common friends;

[0030]

[0031] b3 represents the friend recommendation coefficient when there is a friend relationship between users in the first friend list; f1 represents the sum of the friend familiarity between users in the first friend list;

[0032] Furthermore, when the median of the friend recommendation expected vector is greater than the second recommendation threshold, the corresponding user in the second friend list is added to the second recommended user list.

[0033] Further, creating a third recommended user list; the third recommended user list is obtained by matching the initial target user portrait with the initial user portrait of the unfamiliar user;

[0034] Further, generating an initial friend recommendation sequence according to the first recommended user list, the second recommended user list and the third recommended user list;

[0035] Further, the initial target user portrait is enhanced to obtain an enhanced target user portrait;

[0036] The process of generating the enhanced target user portrait includes:

[0037] Obtaining the historical behavior data of the target user on the social platform; wherein the historical behavior data includes: historical browsing content, historical browsing frequency, historical browsing duration, historical like information, historical comment information, historical collection information, historical attention information, historical work release information, and historical friend addition information;

[0038] Further, the historical behavior data is processed to obtain standard historical behavior data;

[0039] Further, the standard historical behavior data is discretely distributed to obtain historical behavior discrete data;

[0040] Further, the historical behavior discrete data is mapped to a high-dimensional feature space to obtain a high-dimensional feature vector;

[0041] Further, the initial target user portrait is characterized, and the feature dimension is kept consistent with the high-dimensional feature vector to obtain an initial target user portrait feature vector;

[0042] Further, the initial target user portrait feature vector is concatenated with the high-dimensional feature vector, input into a machine learning model, and an enhanced value of the target user portrait is output;

[0043] Furthermore, an enhanced target user portrait is obtained according to the target user portrait enhancement value.

[0044] Further, the recommended friends are adjusted to obtain a final friend recommendation sequence;

[0045] The process of obtaining the final friend recommendation sequence includes: screening the initial friend recommendation sequence according to the enhanced target user portrait to obtain a first friend recommendation list; collecting the historical recommendation results of the target user; calculating the feedback deviation to obtain a friend recommendation adjustment matrix; wherein the calculation of the friend recommendation adjustment matrix includes:

[0046] Dividing historically recommended users into a recommendation success set and a recommendation failure set according to the historical recommendation results;

[0047] Establish a positive correlation matrix between target users and successful recommended users, a deviation matrix between successful recommended users and successful recommended users, a negative correlation matrix between target users and failed recommended users, and a deviation matrix between failed recommended users and failed recommended users;

[0048] The recommendation success adjustment matrix is ​​calculated based on the target user-recommended successful user positive correlation matrix and the recommended successful user-recommended successful user deviation matrix, which is expressed as:

[0049] M R+ (v,w)=M A+ (v,w)*M S+ (w,w);

[0050] Among them, M R+ is represented by the recommendation success adjustment matrix; v is represented by the target user; w is represented by the recommended successful user; M A+ It is represented as the positive correlation matrix of the target user and the recommended successful user; M S+ It is represented as the recommended successful user - recommended successful user deviation matrix;

[0051] The recommendation failure adjustment matrix is ​​calculated according to the target user-recommendation failure user negative correlation matrix and the recommendation failure user-recommendation failure user deviation matrix, which is expressed as:

[0052] M R- (v,y)=M A- (v,y)*M S- (y,y);

[0053] Among them, M R- M represents the recommendation failure adjustment matrix; y represents the recommendation failure user; A- It is represented as the negative correlation matrix of the target user and the failed recommendation user; M S- It is represented as the recommendation failure user-recommendation failure user deviation matrix;

[0054] The friend recommendation adjustment matrix is ​​obtained according to the recommendation success adjustment matrix and the recommendation failure adjustment matrix, which is expressed as:

[0055] M R =β*M R+ +(1-β)*M R- ;

[0056] Among them, M R It is represented as the friend recommendation adjustment matrix; β is represented as the weight factor;

[0057] The first friend recommendation list is adjusted according to the friend recommendation adjustment matrix to obtain the final friend recommendation sequence.

[0058] A data mining model construction system based on machine learning, the system comprising: a data collection unit, a user portrait generation unit, a recommended friend generation unit, a recommended friend optimization unit and a recommendation unit;

[0059] Wherein, the data collection unit is used to collect relevant data on friend recommendations;

[0060] The user portrait generation unit is used to generate a user portrait, including: an initial user portrait generation module and a user portrait enhancement module;

[0061] The recommended friend generation unit is used to generate initial recommended friends;

[0062] Wherein, the friend recommendation generating unit includes: a first recommended user generating module, a second recommended user generating module and a third recommended user generating module;

[0063] The first recommended user generation module includes: retrieving the contact list of the target user's mobile device; obtaining the contact information in the contact list; determining whether the contact information has registered an account on the social platform, and obtaining a set of registered account contacts; calculating the first recommendation value of each contact in the set of registered account contacts; wherein the calculation formula of the first recommendation value is: FRV i =α1*CNW i +α2*CRL i Among them, FRV i Represents the first recommended value of contact i; CNW i Represents the remark weight of contact i; CRL i represents the standardized storage time of contact i in the contact list; α1 represents the recommendation influence coefficient of the remark information; α2 represents the recommendation influence coefficient of the storage time of the contact information; and α1≥α2; when the first recommendation value is greater than the first recommendation threshold, the contact is added to the first recommended user list;

[0064] The second recommended user generation module includes: obtaining a first friend list of a target user on a social platform; obtaining a second friend list of each friend object in the first friend list; calculating the friend recommendation expectation of the target user and each user in the second friend list, and constructing a friend recommendation expectation vector; wherein the calculation of the friend recommendation expectation includes two forms:

[0065] If there is no friendship between the users in the first friend list, and the second friend list contains common friends, the calculation formula for the friend recommendation expectation is:

[0066]

[0067] Among them, E (T_user,R_user) N represents the friend recommendation expectation for the target user and the user in the second friend list; R_user N is represented by the number of friends in the second friend list and the first friend list; S_user It is represented by the number of users in the first friend list; (T_user，S_user) It is represented by the friend familiarity between the target user and the user in the first friend list; (S_user，R_user) represents the friend familiarity between the user in the first friend list and the user in the second friend list; b1 represents the friend recommendation coefficient of the proportion of common friends; b2 represents the friend recommendation coefficient when there is no friend relationship between the users in the first friend list; ∑ represents the symbol for calculating the total friend familiarity;

[0068] If there is a friendship between the users in the first friend list, and the second friend list contains common friends;

[0069]

[0070] b3 represents the friend recommendation coefficient when there is a friend relationship between users in the first friend list; f1 represents the sum of the friend familiarity between users in the first friend list;

[0071] When the median of the friend recommendation expected vector is greater than a second recommendation threshold, adding the corresponding user in the second friend list to the second recommended user list;

[0072] Furthermore, the third recommended user generation module matches the initial target user portrait with the initial user portrait of the unfamiliar user to obtain a third recommended user list;

[0073] Further, an initial friend recommendation sequence is generated according to the first recommended user list, the second recommended user list and the third recommended user list.

[0074] The friend recommendation optimization unit is used to optimize the recommendation of the initially recommended friends;

[0075] The recommendation unit is used to recommend friends to the target user.

[0076] Compared with the prior art, the present invention has the following beneficial effects:

[0077] 1. The present invention proposes an initial friend recommendation method, which is used to make recommendations to three different types of users; mainly including: recommendation of users with existing contact information, recommendation of users with common friends and recommendation of unfamiliar users; wherein, recommendation of users with existing contact information is achieved by retrieving the contacts of registered users on a mobile device, calculating a first recommendation value by the remark information and the time of saving the contact information, and screening by comparing with a first recommendation threshold; users with common friends calculate the recommendation expectation by the proportion of common friends between friends and the familiarity of friends, and screen by comparing with a second recommendation threshold; unfamiliar users are screened by matching according to the initial portrait of the target user; such a method can effectively improve the accuracy of recommendations.

[0078] 2. In the process of calculating the expected recommendation of common friends, a friend familiarity evaluation model is proposed to calculate the familiarity between friends; the model maps the characteristic data of friend notes and friend interaction information with the friend familiarity evaluation index in the hidden layer, and outputs the friend familiarity through the output layer; the model can more accurately measure and predict the familiarity between friends, thereby improving the accuracy of recommendations in the recommendation algorithm. The accuracy of familiarity evaluation helps to avoid recommending people who have common friends but are not actually familiar, thereby improving the quality of recommendations.

[0079] 3. The present invention proposes a screening method for a friend recommendation list. The method enhances the initial target user portrait by collecting the historical behavior data of the target user on the social platform, thereby performing initial screening on the generated initial friend recommendation list, which can help users find friends who meet their needs more quickly; in addition, the present invention analyzes the historical friend recommendation results of the target user, and establishes a target user-successful recommendation user positive correlation matrix, a successful recommendation user-successful recommendation user deviation matrix, a target user-failed recommendation user negative correlation matrix, and a failed recommendation user-failed recommendation user deviation matrix for the successfully recommended users and failed users; these four matrices analyze the relationship between the target user and the historically recommended friends from the positive and reverse directions respectively, and finally calculate the friend recommendation adjustment matrix based on the above four matrices; the friend recommendation adjustment matrix can be used to further refine the friend recommendation scheme. This method not only improves the accuracy of the recommendation, but also optimizes the operating efficiency and intelligence of the recommendation system. Through the data-driven friend recommendation adjustment matrix, the system can learn and adapt to the changing needs of users more quickly, realize instant and accurate recommendation feedback, and provide users with more valuable social connections. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 A flowchart of a method for constructing a data mining model based on machine learning provided by an embodiment of the present invention;

[0081] Figure 2 A structural diagram of a data mining model construction system based on machine learning provided by an embodiment of the present invention;

[0082] Figure 3 A structural diagram of a friend recommendation generating unit provided in an embodiment of the present invention;

[0083] Figure 4 A schematic diagram of a situation in which there is no friend in the first friend list in a friend relationship provided by an embodiment of the present invention;

[0084] Figure 5 A schematic diagram of a friend existing in a first friend list in a friend relationship provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0085] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0086] With the rapid development of Internet technology, the acquisition and sharing of information data has become very convenient and is growing at an exponential rate; the data comes in various forms, including text, images, audio, and video.

[0087] As a form of information dissemination, data contains various forms of potential value; this value needs to be analyzed through appropriate data mining methods. Many researchers combine machine learning technology with data mining technology to achieve data analysis; for example, the above technology is used to implement an important service such as friend recommendation on major social platforms.

[0088] In recent years, most of the friend recommendation methods in existing literature use data mining of user interests to achieve matching with the most similar users; however, this method only considers the suitability of recommendation from one side and has limitations; therefore, the present invention proposes a data mining model construction method and system based on machine learning to achieve more accurate friend recommendation and improve the success rate of user matching. The present invention will be described in detail from two embodiments.

[0089] Embodiment 1:

[0090] In the embodiment of the present application, the process of recommending friends on social platform A is described; wherein social platform A is a dating application software; its platform includes many functions, such as friend recommendations, video browsing, posting updates, and community discussions.

[0091] A brief introduction is given to the social platform A in the embodiment of the present application, and some of its functions are explained; the above functions are only for introduction and do not include all the functions of the social platform A.

[0092] Further, recommend friends to users registered on social platform A; for specific steps, refer to Figure 1 and Figure 2 Content, the embodiment of the present application combines the method and system proposed by the present invention; Figure 1The specific steps of friend recommendation are given in the text, including: S10. Obtaining the personal information submitted by the target user on the social platform; S20. Generating an initial target user portrait; S30. Generating an initial friend recommendation sequence; S40. Enhancing the initial target user portrait; S50. Adjusting the initial friend recommendation sequence to obtain a final friend recommendation sequence; S60. Recommending friends. Figure 2 The structure of the system includes: a data collection unit, a user portrait generation unit, a recommended friend generation unit, a recommended friend optimization unit and a recommendation unit. Figure 1 and Figure 2 The content is introduced in detail.

[0093] The personal data of the target user of social platform A is obtained according to the data collection unit of the system, corresponding to the above-mentioned step S10; the target user refers to the user who needs to recommend friends; the personal data refers to the basic information, avatar and cover, etc. filled in by the user when registering on social platform A; in the implementation of this application, the user personal data of social platform A includes: basic information, personal profile, cover, avatar, personal signature, personal label, geographic location, education and career information and interests and hobbies; refer to Table 1, which gives some personal data about the target user.

[0094] Table 1. Partial personal information of target users

[0095] Nick name Free fish ID PY12138 age 25 gender male Personalized signature "30% is determined by fate, 70% depends on hard work" Location City A Education Undergraduate Profession programmer Hobbies Singing, reading, basketball

[0096] Furthermore, the user portrait generation unit of the system generates an initial target user portrait according to the personal information of the target user, corresponding to the above step S20; the user portrait generation unit includes two modules, namely: an initial user portrait generation module and a user portrait enhancement module; the initial target user portrait is generated by the initial user portrait generation module, and the specific process includes:

[0097] Preprocessing the personal data to obtain standard personal data;

[0098] Wherein, the preprocessing includes: text preprocessing and image preprocessing;

[0099] The text preprocessing includes: segmenting the text information to obtain multiple phrases; embedding the multiple phrases into word vectors to obtain multiple high-dimensional real number vectors; performing data cleaning on the high-dimensional real number vectors to obtain standard high-dimensional real number vectors; and performing data compression on the standard high-dimensional data vectors to obtain standard dimensional data vectors.

[0100] The image preprocessing includes: adjusting the size of the image information to obtain a uniform size image; enhancing the uniform size image to obtain an enhanced image; extracting features from the enhanced image to obtain image features; converting the image features into feature vectors; and performing dimensionality reduction processing on the feature vectors to obtain standard dimensional image feature vectors;

[0101] Further, the standard dimension data vector and the standard dimension image feature vector are input into a machine learning model to obtain an initial user portrait analysis value;

[0102] Among them, in the application embodiment, the machine learning model adopts a pre-trained clustering algorithm model, automatically identifies and organizes data points by inputting the feature vector of the target user, and classifies the target user and similar training users into the same category; the machine learning model outputs a pre-defined cluster label of the target user.

[0103] Furthermore, the initial target user portrait is obtained according to the initial user portrait analysis value.

[0104] According to the analysis of the partial personal information provided in Table 1 by the machine learning model, the initial user portrait of the target user can be defined as follows: If the target user is classified as a "work-oriented user (Type A)", the target user is a programmer located in City A. If the target user is classified as a "hobby-oriented user (Type B)", the target user likes singing, reading or basketball. If the target user is classified as "other types of users (Type C)", it may be some target users without specific feature matching.

[0105] The above only explains the process of generating the initial user portrait based on part of the personal data in Table 1, and does not represent all the content of the target user. The initial user portrait can be further analyzed in combination with other personal data of the target user, and cluster labels about personality and other aspects can be defined in the machine learning model mentioned in the embodiment of the present application to increase the analysis content.

[0106] In the embodiment of the present application, an initial target user portrait is generated by collecting the target user's personal information. During the generation process, the text information features and image information features in the personal information are extracted, and a machine learning model is used to perform user portrait analysis. This process provides more effective support for subsequent friend recommendations.

[0107] Furthermore, the recommended friend generation unit of the system generates an initial friend recommendation sequence, corresponding to the above step S30; the recommended friend generation unit includes three modules, see Figure 3; They are: a first recommended user generation module, a second recommended user generation module and a third recommended user generation module; wherein the first recommended user generation module is used to recommend contacts on a mobile device who have registered on social platform A; the second recommended user generation module is used to recommend users who have common friends on social platform A; the third recommended user generation module is used to recommend unfamiliar users on social platform A.

[0108] The first recommendation generation module can be executed only when the target user has enabled the social platform A to access the contact directory on the mobile device. When the current permission is enabled, the specific implementation process of the first recommendation generation module includes:

[0109] Retrieving the contact list of the target user's mobile device;

[0110] The mobile devices involved in the embodiments of this application include not only electronic devices such as mobile phones, tablets, and computers, but also other electronic devices that can be installed on the social platform A, which can be used as mobile devices mentioned in the examples of this application. The contact list mentioned in the implementation of this application refers not only to the contact list, but also to the contact list in other ways such as user registration and login with the social platform A;

[0111] Further, the contact information in the contact list is obtained; wherein the contact information includes: remarks information about the contact and the time when the contact information is saved;

[0112] Further, determining whether the contact information has registered an account on the social platform, and obtaining a set of registered account contacts;

[0113] Further, the first recommendation value of each contact in the registered account contact set is calculated; wherein the calculation formula of the first recommendation value is: FRV i =α1*CNW i +α2*CRL i Among them, FRV i Represents the first recommended value of contact i; CNW i Represents the remark weight of contact i; CRL i represents the standardized storage time of contact i in the contact list; α1 represents the recommendation influence coefficient of the remark information; α2 represents the recommendation influence coefficient of the storage time of the contact information; and α1≥α2; when the first recommendation value is greater than the first recommendation threshold, the contact is added to the first recommended user list;

[0114] In the embodiment of the present application, the first recommendation value is calculated by retrieving the contact list on the target user's mobile device, and the recommendation candidate is obtained by comparing the first recommendation threshold; wherein, the note name of the contact list and the time of saving the contact information are used as recommendation factors in the calculation process; the two recommendation influence coefficients involved in the calculation formula of the first recommendation value are α1 and α2, respectively, and α1 can be set to 0.8 and α2 can be set to 0.2 according to the success rate of recommending friends by historical contacts; for the first recommendation threshold, A / B testing is used to help verify different threshold settings, based on monitoring the feedback of each target user recommending friends from the contact list. The note name can interact with the relationship between the target user and the contact; the note name can be divided into: relative notes, friend notes, contact name notes and no notes, etc.; the degree of relationship for different note names is also different, generally: relative notes>friend nickname notes>name notes>no notes. The time for saving the contact information of the contact is also used as a recommendation factor, which can show the persistence and importance of the relationship between the target user and the specific contact.

[0115] In addition, in the embodiment of the present application, friend recommendations are made from the perspective of contacts, and the first recommendation value can also be calculated based on the grouping of some contact lists. Different groups reflect different degrees of relationship between the target user and the contacts.

[0116] The specific implementation process of the second recommended user generation module of the recommended friend generation unit includes: obtaining the first friend list of the target user on the social platform;

[0117] Further, obtaining a second friend list for each friend object in the first friend list;

[0118] Further, calculating the friend recommendation expectation of the target user and each user in the second friend list, and constructing a friend recommendation expectation vector;

[0119] The calculation of friend recommendation expectation includes two forms:

[0120] If there is no friendship between the users in the first friend list, and the second friend list contains common friends, the calculation formula for the friend recommendation expectation is:

[0121]

[0122] Among them, E (T_user,R_user) N represents the friend recommendation expectation for the target user and the user in the second friend list; R_user N is represented by the number of users in the second friend list and the number of friends in the first friend list; S_user It is represented by the number of users in the first friend list; (T_user，S_user)It is represented by the friend familiarity between the target user and the user in the first friend list; (S_user，R_user) represents the friend familiarity between the user in the first friend list and the user in the second friend list; b1 represents the friend recommendation coefficient of the proportion of common friends; b2 represents the friend recommendation coefficient when there is no friend relationship between the users in the first friend list; ∑ represents the symbol for calculating the total friend familiarity;

[0123] The calculation process of the friend familiarity includes: collecting friend familiarity analysis information; wherein the friend familiarity analysis information includes: friend remark information and friend interaction information; performing information processing on the friend familiarity analysis information to obtain standard friend familiarity analysis data; using a feature engineering algorithm to characterize the standard friend familiarity analysis data to obtain a friend familiarity analysis feature matrix; establishing a friend familiarity evaluation index; inputting the friend familiarity analysis feature matrix and the friend familiarity evaluation index into a friend familiarity evaluation model to obtain the friend familiarity; wherein the model function of the friend familiarity evaluation model is expressed as:

[0124] FFEF(X,FFEI)=OutputLayer(HiddenLayer(InputLayer(X,FFEI)));

[0125] Among them, FFEF() represents the model function of the friend familiarity evaluation model; X represents the friend familiarity analysis feature matrix; FFEI represents the friend familiarity evaluation index; InputLayer() represents the input layer of the friend familiarity evaluation model, which is used to receive the friend familiarity analysis feature matrix and the friend familiarity evaluation index; HiddenLayer() represents the hidden layer of the friend familiarity evaluation model, which is used to learn the mapping relationship between the friend familiarity analysis feature matrix and the friend familiarity evaluation index; OutputLayer() represents the output layer of the friend familiarity evaluation model, which is used to output friend familiarity;

[0126] In the embodiment of the present application, a friend familiarity evaluation model is used to calculate the familiarity between friends; the model is mainly based on the analysis of the characteristics of the friend's note information and the characteristics of the friend's interaction information; the friend interaction information is mainly collected by collecting a series of information such as the frequency of mutual communication, likes, reposts, comments and the relevance of shared content between friends; the analysis data is characterized by feature engineering and input into the value model to obtain the evaluation result; in this process, in order to achieve accurate evaluation results; the present invention also quantifies the evaluation content and establishes a friend familiarity evaluation index; the index can be accurately mapped in the model according to the feature data, thereby improving the evaluation accuracy.

[0127] If there is a friendship between the users in the first friend list, and the second friend list contains common friends;

[0128]

[0129] b3 represents the friend recommendation coefficient when there is a friend relationship between users in the first friend list; f1 represents the sum of the friend familiarity between users in the first friend list;

[0130] When the median of the friend recommendation expected vector is greater than a second recommendation threshold, adding the corresponding user in the second friend list to the second recommended user list;

[0131] In the embodiment of the present application, a second recommended user list is generated for the target user when the target user has common friends. According to the above description, the specific common friends are as follows: Figure 4 and Figure 5 ;exist Figure 4 The first column shows that there is no friend relationship in the friend list when there are common friends; Figure 5 The first column of the friend list shows that there is a friend relationship in the case of having common friends. According to these two cases, the description of the embodiment of the present application is given in Table 2 and Table 3.

[0132] Table 2. When there are common friends, the first column does not have a friend relationship in the friend list

[0133]

[0134] According to the friend correspondence relationship, the friend recommendation expectation value can be calculated according to Table 2; for example, the friend recommendation expectation value of ID PY14128 is: The friend recommendation expectations for ID PY71425 are: The remaining users can be calculated based on the values ​​in Table 2;

[0135] Table 3. When there are common friends, the first column of the friend list has a friend relationship When there is a friend relationship in the first column of the friend list, the calculation method of the friend recommendation expectation is changed; for example, the calculation formula for the friend recommendation expectation of ID PY71425 is:

[0136]

[0137] In the embodiment of the present application, two different friend recommendation expectation calculation formulas are given in the case of having common friends; and are explained in detail in combination with the data in Table 2 and Table 3; in order to achieve the calculation of the friend recommendation expectation, the embodiment of the present application sets b1 to 0.4, b2 to 0.6, and b3 to 0.7; the second recommendation threshold can be set according to the feedback results of the social platform A's previous recommendations with common friends; this is not because the final result does not conform to the actual situation due to the excessive difference in the recommendation coefficient; it is obtained from the calculation results of Table 2 that when the first friend When PY25714 and PY36782 are not friends in the list, the expectation of user PY71425 recommending target user PY12138 is 0.4595; in Table 3, when PY25714 and PY36782 are friends in the first friend list, the expectation of user PY71425 recommending target user PY12138 is 0.5183; it can be seen that the closer the social graph between friends on the social platform, the greater the possibility of users becoming friends.

[0138] Furthermore, the third recommended user generation module is used to recommend stranger users on social platform A, mainly by matching the initial user portrait of the target user with the initial user portrait of the stranger user; comparing the matching result with the third recommendation threshold to obtain candidate recommended users; the third recommendation threshold can be obtained based on relevant data analysis of successful stranger user recommendations in the history of social platform A; the results obtained by the above first recommended user generation module, second recommended user generation module and third recommended user generation module can constitute the generation of an initial friend recommendation sequence.

[0139] Furthermore, the user portrait enhancement module of the user portrait generation unit enhances the portrait of the target user, corresponding to the above step S40; wherein the enhancement process of the initial portrait of the target user includes:

[0140] Obtain the historical behavior data of the target user from social platform A; wherein the historical behavior data includes: historical browsing content, historical browsing frequency, historical browsing duration, historical like information, historical comment information, historical collection information, historical attention information, historical work release information and historical friend addition information; in order to further refine the portrait of the target user, the historical behavior data is described in Table 4;

[0141] Table 4. Comprehensive information of historical behavior data of target users

[0142]

[0143] Further, the historical behavior data is processed to obtain standard historical behavior data; wherein the processing of the historical behavior data mainly includes: data cleaning, data conversion and integration, data dimension reduction and coding classification;

[0144] Furthermore, the standard historical behavior data is discretely distributed to obtain historical behavior discrete data; wherein the discrete distribution process is based on the value range [min, max] of the standard historical behavior data; min represents the minimum value in the standard historical behavior data; max represents the maximum value in the standard historical behavior data; the standard historical behavior data is discretized into k boxes, and the width of each box is w, which can be calculated as According to the specific value of the standard historical behavior data, the box to which it belongs is determined by the formula: bin() represents a binning function; x represents a specific value of the standard historical behavior data;

[0145] Further, the historical behavior discrete data is mapped to a high-dimensional feature space to obtain a high-dimensional feature vector;

[0146] Further, the initial target user portrait feature vector is concatenated with the high-dimensional feature vector, input into a machine learning model, and an enhanced value of the target user portrait is output;

[0147] Furthermore, an enhanced target user portrait is obtained according to the target user portrait enhancement value.

[0148] In order to improve the accuracy of friend recommendations in the embodiment of the present application, the initial user portrait is enhanced by collecting and analyzing the user's historical behavior data on the social platform to obtain an enhanced user portrait; in the enhancement process, the historical behavior data is mainly characterized and enhanced in combination with the initial user portrait to generate a richer and more accurate enhanced user portrait.

[0149] Furthermore, the friend recommendation optimization generation unit of the system optimizes the initial friend recommendation sequence to obtain a final friend recommendation sequence, which corresponds to the above step S50. The specific process includes:

[0150] Performing friend recommendation screening on the initial friend recommendation sequence according to the enhanced target user portrait to obtain a first friend recommendation list; wherein the screening process of the first friend recommendation list is based on the social relationship density and matching degree between the target user and the users in the initial friend recommendation sequence;

[0151] Collect the historical recommendation results of the target user; perform feedback deviation calculation to obtain a friend recommendation adjustment matrix; wherein the calculation of the friend recommendation adjustment matrix includes: dividing the historical recommendation users into a recommendation success set and a recommendation failure set according to the historical recommendation results;

[0152] Establish a positive correlation matrix between target users and successful recommended users, a deviation matrix between successful recommended users and successful recommended users, a negative correlation matrix between target users and failed recommended users, and a deviation matrix between failed recommended users and failed recommended users;

[0153] The recommendation success adjustment matrix is ​​calculated based on the target user-recommended successful user positive correlation matrix and the recommended successful user-recommended successful user deviation matrix, which is expressed as:

[0154] M R+ (v,w)=M A+ (v,w)*M S+ (w,w);

[0155] Among them, M R+ is represented by the recommendation success adjustment matrix; v is represented by the target user; w is represented by the recommended successful user; M A+ It is represented as the positive correlation matrix of the target user and the recommended successful user; M S+ It is represented as the recommended successful user - recommended successful user deviation matrix;

[0156] The recommendation failure adjustment matrix is ​​calculated according to the target user-recommendation failure user negative correlation matrix and the recommendation failure user-recommendation failure user deviation matrix, which is expressed as:

[0157] M R- (v,y)=M A- (v,y)*M S- (y,y);

[0158] Among them, M R- M represents the recommendation failure adjustment matrix; y represents the recommendation failure user; A- It is represented as the negative correlation matrix of the target user and the failed recommendation user; M S- It is represented as the recommendation failure user-recommendation failure user deviation matrix;

[0159] The friend recommendation adjustment matrix is ​​obtained according to the recommendation success adjustment matrix and the recommendation failure adjustment matrix, which is expressed as:

[0160] M R =β*M R+ +(1-β)*M R- ;

[0161] Among them, M R It is represented as the friend recommendation adjustment matrix; β is represented as the weight factor;

[0162] Furthermore, the first friend recommendation list is adjusted according to the friend recommendation adjustment matrix to obtain the final friend recommendation sequence.

[0163] In the embodiment of the present application, a friend recommendation adjustment method is used to adjust the generated initial friend recommendation sequence; the method first performs a screening based on the enhanced target user portrait to retain users with high relationship degrees; then, the historical friend addition records of the target user are analyzed to distinguish between friends who are successfully added and those who are failed to be added, and a target user-successfully recommended user positive correlation matrix, a successfully recommended user-successfully recommended user deviation matrix, a target user-failed recommended user negative correlation matrix, and a failed recommended user-failed recommended user deviation matrix are established; these four matrices analyze the relationship between the target user and the historically recommended friends from a positive and reverse perspective respectively; based on the above four matrices, a friend recommendation adjustment matrix is ​​finally calculated, and the friend recommendation scheme can be further refined through the friend recommendation adjustment matrix.

[0164] Furthermore, the system recommendation unit performs friend recommendation according to the final friend recommendation sequence, corresponding to the above step S60.

[0165] In the application embodiment, the method of the present invention is combined with the system to realize the recommendation of user friends of social platform A; the present invention mainly starts from the following aspects; first, an initial target user portrait is generated according to the personal information of the target user; then an initial friend recommendation sequence is generated, and the sequence is generated according to three different types of user types; namely, contacts of registered users of the mobile device used by the target user, users with common friends on the social platform, and strangers; in terms of contacts, the first recommendation value is calculated by obtaining the remark information and the time of saving the contact information, and a first recommended user list is created by comparing the first recommended threshold; in terms of common friends, the recommendation expectation is calculated by analyzing the intimacy and friend ownership rate of the common friends of the target user and the recommended user, and a second recommended user list is created by comparing the second threshold; in terms of stranger users, the match between the initial user portraits of the target user and the stranger user is calculated, and a third recommended user list is created by comparing the third threshold; the above three methods can well screen out users recommended by different channels; secondly, in order to further analyze the target user, the present invention enhances the user portrait by analyzing the historical behavior data of the target user on the basis of the initial target user portrait, so that these portraits can be enriched to include more details and depth. This more detailed user portrait can improve the accuracy of the recommendation system and reduce the recommendation of irrelevant or inappropriate friends. Finally, a friend recommendation method is adopted, which matches the enhanced target user portrait with the enhanced portrait of the recommended user, conducts preliminary screening, and retains users with high relationship degrees. Then, the historical friend addition records of the target user are analyzed to distinguish between successful and failed friends, and establish a target user-successfully recommended user positive correlation matrix, a successfully recommended user-successfully recommended user deviation matrix, a target user-failed recommended user negative correlation matrix, and a failed recommended user-failed recommended user deviation matrix. These four matrices analyze the relationship between the target user and the historically recommended friends from the positive and reverse directions respectively. Based on the above four matrices, the friend recommendation adjustment matrix is ​​finally calculated, and the friend recommendation scheme can be further refined through the friend recommendation adjustment matrix.

[0166] Embodiment 2:

[0167] In Example 1, the method of the present invention is combined with the system to implement the friend recommendation task on social platform A. In order to further illustrate the effectiveness of the present invention, in the embodiment of the present application, friend recommendations are made to other target users on social platform A. It is known that the profile of the target user in the embodiment of the present application shows that the gender is "female", and the interests and hobbies include: mountain climbing, dancing, and playing badminton. The specific friend recommendation process of the present application is basically the same as that of Example 1, and mainly includes:

[0168] Obtain the personal information submitted by the target user on the social platform; generate the initial target user portrait;

[0169] Further, a first recommended user list is created; wherein the process of creating the first recommended user list includes: retrieving the contact list of the target user's mobile device; obtaining the contact information in the contact list; determining whether the contact information has registered an account on the social platform, and obtaining a set of registered account contacts; calculating a first recommendation value for each contact in the set of registered account contacts; wherein the calculation formula for the first recommendation value is: FRV i =α1*CNW i +α2*CRL i Among them, FRV i Represents the first recommended value of contact i; CNW i Represents the remark weight of contact i; CRL i represents the standardized storage time of contact i in the contact list; α1 represents the recommendation influence coefficient of the remark information; α2 represents the recommendation influence coefficient of the storage time of the contact information; and α1≥α2; when the first recommendation value is greater than the first recommendation threshold, the contact is added to the first recommended user list;

[0170] Further, creating a second recommended user list;

[0171] Further, creating a third recommended user list; the third recommended user list is obtained by matching the initial target user portrait with the initial user portrait of the unfamiliar user;

[0172] Further, generating an initial friend recommendation sequence according to the first recommended user list, the second recommended user list and the third recommended user list;

[0173] Furthermore, the initial target user portrait is enhanced to obtain an enhanced target user portrait; wherein the generation process of the enhanced target user portrait includes:

[0174] Obtaining the historical behavior data of the target user on the social platform; wherein the historical behavior data includes: historical browsing content, historical browsing frequency, historical browsing duration, historical like information, historical comment information, historical collection information, historical attention information, historical work release information, and historical friend addition information;

[0175] Further, the historical behavior data is processed to obtain standard historical behavior data;

[0176] Further, the standard historical behavior data is discretely distributed to obtain historical behavior discrete data;

[0177] Further, the historical behavior discrete data is mapped to a high-dimensional feature space to obtain a high-dimensional feature vector;

[0178] Further, the initial target user portrait is characterized, and the feature dimension is kept consistent with the high-dimensional feature vector to obtain an initial target user portrait feature vector;

[0179] Further, the initial target user portrait feature vector is concatenated with the high-dimensional feature vector, input into a machine learning model, and an enhanced value of the target user portrait is output;

[0180] Furthermore, an enhanced target user portrait is obtained according to the target user portrait enhancement value.

[0181] Further, the recommended friends are adjusted to obtain a final friend recommendation sequence;

[0182] The process of obtaining the final friend recommendation sequence includes: screening the initial friend recommendation sequence according to the enhanced target user portrait to obtain a first friend recommendation list; collecting the historical recommendation results of the target user; calculating the feedback deviation to obtain a friend recommendation adjustment matrix; wherein the calculation of the friend recommendation adjustment matrix includes:

[0183] According to the historical recommendation results, the historical recommended users are divided into a recommendation success set and a recommendation failure set;

[0184] Establish a positive correlation matrix between target users and successful recommended users, a deviation matrix between successful recommended users and successful recommended users, a negative correlation matrix between target users and failed recommended users, and a deviation matrix between failed recommended users and failed recommended users;

[0185] The recommendation success adjustment matrix is ​​calculated based on the target user-recommended successful user positive correlation matrix and the recommended successful user-recommended successful user deviation matrix, which is expressed as:

[0186] M R+ (v,w)=M A+ (v,w)*M S+ (w,w);

[0187] Among them, M R+ is represented by the recommendation success adjustment matrix; v is represented by the target user; w is represented by the recommended successful user; M A+ It is represented as the positive correlation matrix of the target user and the recommended successful user; M S+ It is represented as the recommended successful user - recommended successful user deviation matrix;

[0188] The recommendation failure adjustment matrix is ​​calculated according to the target user-recommendation failure user negative correlation matrix and the recommendation failure user-recommendation failure user deviation matrix, which is expressed as:

[0189] MR- (v,y)=M A- (v,y)*M S- (y,y);

[0190] Among them, M R- M represents the recommendation failure adjustment matrix; y represents the recommendation failure user; A- It is represented as the negative correlation matrix of the target user and the failed recommendation user; M S- It is represented as the recommendation failure user-recommendation failure user deviation matrix;

[0191] The friend recommendation adjustment matrix is ​​obtained according to the recommendation success adjustment matrix and the recommendation failure adjustment matrix, which is expressed as:

[0192] M R =β*M R+ +(1-β)*M R- ;

[0193] Among them, M R It is represented as the friend recommendation adjustment matrix; β is represented as the weight factor;

[0194] The first friend recommendation list is adjusted according to the friend recommendation adjustment matrix to obtain the final friend recommendation sequence.

[0195] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a data mining model based on machine learning, characterized in that: include: Obtain the personal information submitted by the target user on the social platform; The personal data includes: text information and image information; generating an initial target user portrait; creating a first recommended user list; the process of creating the first recommended user list includes: retrieving the contact list of the target user's mobile device; obtaining the contact information in the contact list; determining whether the contact information has registered an account on the social platform, and obtaining a set of registered account contacts; calculating the first recommendation value of each contact in the set of registered account contacts; wherein the calculation formula of the first recommendation value is: FRV i =α1*CNW i +α2*CRL i Among them, FRV i Represents the first recommended value of contact i; CNW i Represents the remark weight of contact i; CRL i represents the standardized storage time of contact i in the contact list; α1 represents the recommendation influence coefficient of the remark information; α2 represents the recommendation influence coefficient of the storage time of the contact information; and α1≥α2; when the first recommendation value is greater than the first recommendation threshold, the contact is added to the first recommended user list; a second recommended user list is created; a third recommended user list is created; the third recommended user list is obtained by matching the initial target user portrait with the initial user portrait of the stranger user; an initial friend recommendation sequence is generated according to the first recommended user list, the second recommended user list and the third recommended user list; the initial target user portrait is enhanced to obtain an enhanced target user portrait; the recommended friends are adjusted to obtain a final friend recommendation sequence.

2. The method for constructing a data mining model based on machine learning according to claim 1, characterized in that: The generation process of the initial target user portrait includes: preprocessing the personal information to obtain a standard personal information; wherein the preprocessing includes: text preprocessing and image preprocessing; the text preprocessing includes: segmenting the text information to obtain multiple phrases; embedding the multiple phrases into word vectors to obtain multiple high-dimensional real number vectors; performing data cleaning on the high-dimensional real number vectors to obtain standard high-dimensional real number vectors; performing data compression on the standard high-dimensional data vectors to obtain standard dimensional data vectors; the image preprocessing includes: adjusting the size of the image information to obtain a uniform-sized image; enhancing the uniform-sized image to obtain an enhanced image; extracting features from the enhanced image to obtain image features; converting the image features into feature vectors; performing dimensionality reduction processing on the feature vectors to obtain standard dimensional image feature vectors; inputting the standard dimensional data vectors and the standard dimensional image feature vectors into a machine learning model to obtain an initial user portrait analysis value; and obtaining the initial target user portrait according to the initial user portrait analysis value.

3. The method for constructing a data mining model based on machine learning according to claim 1, characterized in that: The creation process of the second recommended user list includes: obtaining the first friend list of the target user on the social platform; obtaining the second friend list of each friend object in the first friend list; calculating the friend recommendation expectation of the target user and each user in the second friend list, and constructing a friend recommendation expectation vector; wherein the calculation of the friend recommendation expectation includes two forms, namely: If there is no friendship between the users in the first friend list, and the second friend list contains common friends, the calculation formula for the friend recommendation expectation is: Among them, E (T_user,R_user) N represents the friend recommendation expectation for the target user and the user in the second friend list; R_user N is represented by the number of users in the second friend list and the number of friends in the first friend list; S_user It is represented by the number of users in the first friend list; (T_user，S_user) It is represented by the friend familiarity between the target user and the user in the first friend list; (S_user，R_user) represents the friend familiarity between the user in the first friend list and the user in the second friend list; b1 represents the friend recommendation coefficient of the proportion of common friends; b2 represents the friend recommendation coefficient when there is no friend relationship between the users in the first friend list; ∑ represents the symbol for calculating the total friend familiarity; If there is a friendship between the users in the first friend list, and the second friend list contains common friends; b3 represents the friend recommendation coefficient when there is a friend relationship between users in the first friend list; f1 represents the sum of the friend familiarity between users in the first friend list; When the median of the friend recommendation expected vector is greater than a second recommendation threshold, the corresponding user in the second friend list is added to the second recommended user list.

4. The method for constructing a data mining model based on machine learning according to claim 3, characterized in that: The calculation process of the friend familiarity includes: collecting friend familiarity analysis information; wherein the friend familiarity analysis information includes: friend remark information and friend interaction information; performing information processing on the friend familiarity analysis information to obtain standard friend familiarity analysis data; using a feature engineering algorithm to characterize the standard friend familiarity analysis data to obtain a friend familiarity analysis feature matrix; establishing a friend familiarity evaluation index; inputting the friend familiarity analysis feature matrix and the friend familiarity evaluation index into a friend familiarity evaluation model to obtain the friend familiarity; wherein the model function of the friend familiarity evaluation model is expressed as: FFEF(X,FFEI)=OutputLayer(HiddenLayer(InputLayer(X,FFEI))); Among them, FFEF() represents the model function of the friend familiarity evaluation model; X represents the friend familiarity analysis feature matrix; FFEI represents the friend familiarity evaluation index; InputLayer() represents the input layer of the friend familiarity evaluation model, which is used to receive the friend familiarity analysis feature matrix and the friend familiarity evaluation index; HiddenLayer() represents the hidden layer of the friend familiarity evaluation model, which is used to learn the mapping relationship between the friend familiarity analysis feature matrix and the friend familiarity evaluation index; OutputLayer() represents the output layer of the friend familiarity evaluation model, which is used to output friend familiarity.

5. The method for constructing a data mining model based on machine learning according to claim 1, characterized in that: The generation process of the enhanced target user portrait includes: obtaining the historical behavior data of the target user on the social platform; wherein the historical behavior data includes: historical browsing content, historical browsing frequency, historical browsing time, historical like information, historical comment information, historical collection information, historical attention information, historical work release information and historical friend addition information; processing the historical behavior data to obtain standard historical behavior data; performing discrete distribution on the standard historical behavior data to obtain historical behavior discrete data; mapping the historical behavior discrete data to a high-dimensional feature space to obtain a high-dimensional feature vector; characterizing the initial target user portrait, and keeping the feature dimension consistent with the high-dimensional feature vector to obtain an initial target user portrait feature vector; splicing the initial target user portrait feature vector with the high-dimensional feature vector, inputting the result into a machine learning model, and outputting a target user portrait enhancement value; and obtaining the enhanced target user portrait according to the target user portrait enhancement value.

6. The method for constructing a data mining model based on machine learning according to claim 1, characterized in that: The process of obtaining the final friend recommendation sequence includes: screening the initial friend recommendation sequence according to the enhanced target user portrait to obtain a first friend recommendation list; collecting the historical recommendation results of the target user; calculating the feedback deviation to obtain a friend recommendation adjustment matrix; wherein the calculation of the friend recommendation adjustment matrix includes: Dividing historically recommended users into a recommendation success set and a recommendation failure set according to the historical recommendation results; Establish a positive correlation matrix between target users and successful recommended users, a deviation matrix between successful recommended users and successful recommended users, a negative correlation matrix between target users and failed recommended users, and a deviation matrix between failed recommended users and failed recommended users; The recommendation success adjustment matrix is ​​calculated based on the target user-recommended successful user positive correlation matrix and the recommended successful user-recommended successful user deviation matrix, which is expressed as: M R+ (v,w)=M A+ (v,w)*M S+ (w,w); Among them, M R+ is represented by the recommendation success adjustment matrix; v is represented by the target user; w is represented by the recommended successful user; M A+ It is represented as the positive correlation matrix of the target user and the recommended successful user; M S+ It is represented as the recommended successful user - recommended successful user deviation matrix; The recommendation failure adjustment matrix is ​​calculated according to the target user-recommendation failure user negative correlation matrix and the recommendation failure user-recommendation failure user deviation matrix, which is expressed as: M R- (v,y)=M A- (v,y)*M S- (y,y); Among them, M R- M represents the recommendation failure adjustment matrix; y represents the recommendation failure user; A- It is represented as the negative correlation matrix of the target user and the failed recommendation user; M S- It is represented as the recommendation failure user-recommendation failure user deviation matrix; The friend recommendation adjustment matrix is ​​obtained according to the recommendation success adjustment matrix and the recommendation failure adjustment matrix, which is expressed as: M R =β*M R+ +(1-β)*M R- ; Among them, M R It is represented as the friend recommendation adjustment matrix; β is represented as the weight factor; The first friend recommendation list is adjusted according to the friend recommendation adjustment matrix to obtain the final friend recommendation sequence.

7. A data mining model construction system based on machine learning, characterized in that: The system includes: a data collection unit, a user portrait generation unit, a recommended friend generation unit, a recommended friend optimization unit and a recommendation unit; wherein the data collection unit is used to collect relevant data on friend recommendations; the user portrait generation unit is used to generate user portraits, including: an initial user portrait generation module and a user portrait enhancement module; the recommended friend generation unit is used to generate initial recommended friends; the recommended friend optimization unit is used to optimize the recommendation of the initially recommended friends; the recommendation unit is used to recommend friends to target users.

8. The data mining model construction system based on machine learning according to claim 7, characterized in that: The initial user portrait generation module includes: preprocessing the data acquired by the data collection unit to obtain a standard personal data; wherein the preprocessing includes: text preprocessing and image preprocessing; the text preprocessing includes: segmenting the text information to obtain multiple phrases; embedding the multiple phrases into word vectors to obtain multiple high-dimensional real number vectors; performing data cleaning on the high-dimensional real number vectors to obtain standard high-dimensional real number vectors; performing data compression on the standard high-dimensional data vectors to obtain standard dimensional data vectors; the image preprocessing includes: adjusting the size of the image information to obtain a uniform-sized image; enhancing the uniform-sized image to obtain an enhanced image; extracting features from the enhanced image to obtain image features; converting the image features into feature vectors; performing dimensionality reduction processing on the feature vectors to obtain standard dimensional image feature vectors; inputting the standard dimensional data vectors and the standard dimensional image feature vectors into a machine learning model to obtain an initial user portrait analysis value; and obtaining an initial user portrait based on the initial user portrait analysis value.

9. The data mining model construction system based on machine learning according to claim 7, characterized in that: The recommended friend generation unit includes: a first recommended user generation module, a second recommended user generation module and a third recommended user generation module; The first recommended user generation module includes: retrieving the contact list of the target user's mobile device; obtaining the contact information in the contact list; determining whether the contact information has registered an account on the social platform, and obtaining a set of registered account contacts; calculating the first recommendation value of each contact in the set of registered account contacts; wherein the calculation formula of the first recommendation value is: FRV i =α1*CNW i +α2*CRL i Among them, FRV i Represents the first recommended value of contact i; CNW i Represents the remark weight of contact i; CRL i It is represented as the standardized storage time of contact i in the contact list; α1 is represented as the recommendation influence coefficient of the remark information; α2 is represented as the recommendation influence coefficient of the storage time of the contact information; and α1≥α2; when the first recommendation value is greater than the first recommendation threshold, the contact is added to the first recommended user list; the third recommended user generation module matches the initial target user portrait with the initial user portrait of the unfamiliar user to obtain the third recommended user list; an initial friend recommendation sequence is generated according to the first recommended user list, the second recommended user list and the third recommended user list.

10. The data mining model construction system based on machine learning according to claim 9, characterized in that: The second recommended user generation module includes: obtaining a first friend list of a target user on a social platform; obtaining a second friend list of each friend object in the first friend list; calculating the friend recommendation expectation of the target user and each user in the second friend list, and constructing a friend recommendation expectation vector; wherein the calculation of the friend recommendation expectation includes two forms: If there is no friendship between the users in the first friend list, and the second friend list contains common friends, the calculation formula for the friend recommendation expectation is: Among them, E (T_user,R_user) N represents the friend recommendation expectation for the target user and the user in the second friend list; R_user N is represented by the number of users in the second friend list and the number of friends in the first friend list; S_user It is represented by the number of users in the first friend list; (T_user，S_user) It is represented by the friend familiarity between the target user and the user in the first friend list; (S_user，R_user) represents the friend familiarity between the user in the first friend list and the user in the second friend list; b1 represents the friend recommendation coefficient of the proportion of common friends; b2 represents the friend recommendation coefficient when there is no friend relationship between the users in the first friend list; ∑ represents the symbol for calculating the total friend familiarity; If there is a friendship between the users in the first friend list, and the second friend list contains common friends; b3 represents the friend recommendation coefficient when there is a friend relationship between users in the first friend list; f1 represents the sum of the friend familiarity between users in the first friend list; When the median of the friend recommendation expected vector is greater than a second recommendation threshold, the corresponding user in the second friend list is added to the second recommended user list.

Images