CN108182235A

CN108182235A - A kind of method and system for being used to carry out user characteristics distributed coding

Info

Publication number: CN108182235A
Application number: CN201711446827.5A
Authority: CN
Inventors: 罗维
Original assignee: Beijing Qihoo Technology Co Ltd
Current assignee: Beijing Qihoo Technology Co Ltd
Priority date: 2017-12-27
Filing date: 2017-12-27
Publication date: 2018-06-19

Abstract

The invention discloses a kind of method and systems for being used to carry out user characteristics distributed coding, multiple subfiles corresponding to processing node are divided by the consumer profiles for forming all user characteristics, the number of the user characteristics quantity and subfile in each subfile is obtained, subfile is sent to corresponding processing node respectively and is handled.It is encoded using distributed nature, it can lifting scheme operational efficiency, reduction exploitation and the additional workload safeguarded.Client can obtain the audient's demographic data to match with oneself actual demand, and precision is high, can fully meet the different demands of client.

Description

A kind of method and system for being used to carry out user characteristics distributed coding

Technical field

It is used to be distributed user characteristics the present invention relates to Internet technical field, and more particularly, to one kind The method and system of formula coding.

Background technology

In Internet advertising field, for the businessman for launching advertisement, advertisement is launched to arbitrary crowd on a large scale, is deposited It is too high in cost, it is difficult to which that the problem of bearing how from a large amount of netizen, pointedly selects suitable crowd, further according to not Each determined property with crowd goes out to need the advertisement crowd launched, is Internet advertising market development urgent problem.

At present, it is in Internet advertising field to provide more valuable crowd to advertiser using crowd's orientation method One important link, crowd's orientation method are by the analysis to user characteristic data, are found out and seed crowd behaviour feature The joint act feature of similar potential target crowd using machine learning model, predicts target audience's demographic data, and side is wide Accuse the main target group for finding and oneself being look for.The scale of wherein involved seed demographic data is at most in millions of amounts Grade, and the scale of non-seed demographic data is in several hundred million magnitudes, the two ratio great disparity when machine learning model is trained, can cause Memory increases model training and the memory overhead and time overhead of prediction using larger waste is above had.

Meanwhile in some Machine learning tools, need to encode plaintext feature, then can just do model training and Prediction, for example currently have 10,000,000 different characteristics, it needs to encode them with 1 to 1,000 ten thousand, possible feature " accessed Sports.sina.com.cn " is encoded as 11, and feature " searching for tourism " is encoded as 999.

In traditional scheme, using unit feature coding, i.e., using single machine, the file of storage feature is traversed, It encodes successively simultaneously.There are following 2 points deficiencies for the program：

If 1) tag file is especially big, than the feature if any tens times, then scheme operation is slower；

2) if tag file is originally to be stored in HDFS, while the tag file after coding is required also to be stored in HDFS On, then it needs first to download lower data from HDFS, while upload on HDFS the tag file after coding in this way, and These can give exploitation and maintenance to bring extra work.

Invention content

To solve the above-mentioned problems, a kind of method and system for being used to carry out user characteristics distributed coding is provided.

According to an aspect of the invention, there is provided a kind of method for being used to carry out user characteristics distributed coding, packet It includes：

Obtain associated with the network behavior of users all in data network statistical data, and to the statistical data into Row feature extraction is with the total quantity of determining multiple user characteristics and determining the multiple user characteristics；

Structure includes the tag file of the multiple user characteristics, and is based on preset division rule by the feature File is divided into multiple subfiles；

Content scanning is carried out to the user characteristics in each subfile, to determine the number of user characteristics in each subfile Amount；

User characteristics in space encoder based on user characteristics, the total quantity of the multiple user characteristics and each subfile Subnumber amount determine the coding subspace of user characteristics subset in each subfile；And

According to preset processing rule, each subfile and corresponding coding subspace are sent to multiple processing and saved Node is handled in point accordingly to be encoded by corresponding processing node to the user characteristics in user characteristics subset.

Preferably, behavior, browsing webpage behavior are clicked according to the search of all users in the data network and/or passed through The behavior that third party's cooperation obtains obtains the associated statistical data of network behavior of all users in the data network.

Preferably, according to the host features of user, n-gram features, surf time section, online institute in the statistical data Possession domain and/or browsing commodity behavior, carry out feature extraction, to determine multiple user characteristics.

Preferably, the multiple user characteristics are subjected to classification rejecting, the user of counting user demand according to user demand The quantity of feature, the total quantity as user characteristics.

Preferably, by division rule of the tag file of the multiple user characteristics based on hash function, it is divided into correspondence In multiple subfiles of the processing number of nodes.

Preferably, all user characteristics are divided into N barrels；

Calculating the quantity Array [i] of user characteristics described in each bucket, the i is the number of bucket, i=0,1,2,3 ... N；

Accumulation and AccumulatedArray [i], the AccumulatedArray [i] are converted to the Array [i] =AccumulatedArray [i-1]+Array [i]；

Coding proceeded by from start_index [i]+1 to user characteristics in each bucket, the start_index [i]= AccumulatedArray[i-1]。

Preferably, each user characteristics are divided into any one bucket according to hash function.

Preferably, the quantity of the processing node is identical with the quantity of the subfile, each node that handles is to an institute The user characteristics stated in bucket are encoded.

Preferably, as the i=0, AccumulatedArray [0]=Array [0].

Preferably, as the i=0, start_index [0]=0.

Preferably, filter the apparent abnormal dirty sample data in the user characteristics.

Preferably, determine that the coding of the user characteristics is empty according to the maximum group/cording quantity that the coding method can allow for Between.

Preferably, the coding subspace of each subfile is determined according to the processing capacity of each processing node.

According to another aspect of the present invention, a kind of system for being used to carry out user characteristics distributed coding is provided, The system comprises：

User characteristics unit, for obtaining statistical data associated with the network behavior of users all in data network, And feature extraction is carried out to the statistical data to determine multiple user characteristics and determine the total quantity of the multiple user characteristics；

Tag file construction unit for building the tag file for including the multiple user characteristics, and is based on setting in advance The tag file is divided into multiple subfiles by fixed division rule；

Feature quantity confirmation unit, it is each to determine for carrying out content scanning to the user characteristics in each subfile The quantity of user characteristics in subfile；

Subspace confirmation unit is encoded, for the space encoder based on user characteristics, the sum of the multiple user characteristics The subnumber amount of user characteristics determines the coding subspace of user characteristics subset in each subfile in amount and each subfile；And

Node allocation unit is handled, for regular according to preset processing, by each subfile and corresponding coding Subspace be sent to it is multiple processing nodes in handle accordingly node with by it is corresponding processing node to user characteristics subset In user characteristics encoded.

Preferably, all user characteristics are divided into N barrels；

Preferably, as the i=0, AccumulatedArray [0]=Array [0].

Preferably, as the i=0, start_index [0]=0.

According to another aspect of the present invention, a kind of mobile terminal is provided, including or for perform as above any one The system.

In the scheme that each embodiment of the present invention is provided, drawn by the consumer profiles for forming all user characteristics It is divided into multiple subfiles corresponding to processing node, obtains the number of the user characteristics quantity and subfile in each subfile, Subfile is sent to corresponding processing node respectively to handle.Encoded using distributed nature, can lifting scheme operational efficiency, Reduce exploitation and the additional workload safeguarded.Client can obtain the audient's demographic data to match with oneself actual demand, precisely Degree is high, can fully meet the different demands of client.

Above description is only the general introduction of technical solution of the present invention, in order to better understand the technological means of the present invention, And it can be implemented in accordance with the contents of the specification, and in order to allow above and other objects of the present invention, feature and advantage can It is clearer and more comprehensible, below the special specific embodiment for lifting the present invention.

Description of the drawings

By reading the detailed description of hereafter preferred embodiment, it is various other the advantages of and benefit it is common for this field Technical staff will become clear.Attached drawing is only used for showing the purpose of preferred embodiment, and is not considered as to the present invention Limitation.And throughout the drawings, the same reference numbers will be used to refer to the same parts.In the accompanying drawings：

Fig. 1 is DMPLook-alike online system structure diagrams provided in an embodiment of the present invention.

Fig. 2 is the method flow diagram provided in an embodiment of the present invention that extending user is determined according to the statistical data degree of association.

Fig. 3 is the system structure signal provided in an embodiment of the present invention that extending user is determined according to the statistical data degree of association Figure.

Fig. 4 is the method flow diagram provided in an embodiment of the present invention that extending user is determined according to weighted calculation.

Fig. 5 is the system structure diagram provided in an embodiment of the present invention that extending user is determined according to weighted calculation.

Fig. 6 is the method flow diagram provided in an embodiment of the present invention that extending user is determined according to statistical data interest-degree.

Fig. 7 is the system structure signal provided in an embodiment of the present invention that extending user is determined according to statistical data interest-degree Figure.

Fig. 8 is the method flow diagram provided in an embodiment of the present invention for being used to carry out user characteristics distributed coding.

Fig. 9 is the system structure diagram provided in an embodiment of the present invention for being used to carry out user characteristics distributed coding.

Specific embodiment

Exemplary embodiments of the present invention are introduced referring now to attached drawing, however, the present invention can use many different shapes Formula is implemented, and be not limited to the embodiment described herein, and to provide these embodiments be to disclose at large and fully The present invention, and fully convey the scope of the present invention to person of ordinary skill in the field.Show for what is be illustrated in the accompanying drawings Term in example property embodiment is not limitation of the invention.In the accompanying drawings, identical cells/elements use identical attached Icon is remembered.

Unless otherwise indicated, term used herein has person of ordinary skill in the field (including scientific and technical terminology) It is common to understand meaning.Further it will be understood that with the term that usually used dictionary limits, should be understood as and its The linguistic context of related field has consistent meaning, and is not construed as Utopian or too formal meaning.

The each embodiment of the present invention be based on the DMPLook-alike online systems shown in Fig. 1, as shown in Figure 1, its In：

Offline flows：Based on Distributed Computing Platform (Hadoop+Spark), to the network behavior (ratio of the whole network user Such as：Search click behavior, passes through behavior for being obtained with third company cooperation etc. at the behavior of browsing webpage) extraction user characteristics (compare Such as：Host features, n-gram features, surf time section, region, browsing commodity etc. belonging to online).

Online flows：The whole network user characteristics and scheduler calculated based on offline flows are sent out by task scheduling The seed crowd come, using appropriate machine learning model (such as in disaggregated model in supervised learning, unsupervised learning Clustering Model etc.), by model training and prediction, find similar target audience crowd.

The demand of advertiser greatly with the limited always contradiction of computing resource.In order to ensure the profit of each advertiser from the overall situation Benefit, scheduler modules therein with regard to extremely important, it can consider factors (such as：What advertiser had had submitted Look-alike number of tasks, advertiser are DSP launches the consuming capacity at end, advertiser's special demand (for example double 11), is It is not no mined massively to same Ziren with multiple extension multiples) look-alike mission requirements are dispatched, once some look-alike Task is dispatched, then can start online flows and calculate it similar target audience crowd.

During similar crowd is excavated, crowd's extension, i.e., look-alike Main Basiss user essential attribute and its The behavioural information possessed, this just needs huge data storage as analysis source.Data management platform (DMP, Data- Management Platform) be crowd's growth data analysis method basis.Crowd's growth data analysis method development company Can be based on a large number of users itself covered, acquisition and depth excavate the big number of corresponding behavioral chain under the premise of individual privacy is protected According to, such as the search of user is clicked, browsing webpage network behavior data.In general, client, i.e. advertiser will be apparent that oneself The product of advertisement and behind is wanted to touch the user group reached, for example App advertiser can be accurately grasped in oneself App product Any active ues IMEI or IFA, electric business website advertiser have the cookie of interested user or cell-phone number, O2O to certain commodity Advertiser might have telephone number of client etc..Therefore, the first party data that client has by oneself can also be obtained, as official website visitor, It places an order, pay close attention to the data such as wechat, concern microblogging and installation mobile application client.It can also be obtained outer by cooperating with third party The labeling data of portion affiliate, for example, user access website, using APP, watch video, place an order shopping and connection hot spot etc. Network behavior data.Thus the network behavior off-line data of the user obtained becomes the integration of DMP emphasis after anonymous desensibilization Critical data.

In the network behavior off-line data for obtaining user, Distributed Computing Platform (Hadoop+Spark) can be also based on, According to the network behavior daily record of user, the network behavior off-line data of the whole network user is obtained, by the network behavior daily record of user Data good data for researchs such as user interest discovery, resource recommendations can be provided support, convenient for the user's according to acquisition Network behavior off-line data extracts user characteristic data.

By long-term data acquisition and subdivision, can user characteristics number be extracted according to the network behavior off-line data of user According to, such as the extraction host features of user, n-gram features, surf time section, region and/or browsing commodity behavior belonging to online User characteristics are waited, user are finely divided management, from latitudes such as the action trail of user, interest preference, consumer behavior, geographical locations Degree realizes seeing clearly and analyzing for all types of user, obtains the characteristic that all types of user matches, is stored, the use of daily output The disk storage overhead of family characteristic is more than TB magnitudes.Delineation mesh can be freely combined in the user characteristic data that analyze extraction Mark crowd, launch displaying, search, brand, using downloads ad when, can quickly and accurately be directed to a certain category feature people Group.

The demand of client greatly with the limited always contradiction of computing resource, in order to ensure the interests of each client from the overall situation, Agree with the personalized focal need of client, need to consider all multi-parameters in scheduling process to dispatch look-alike mission requirements, Specifically, default scheduling parameter includes look-alike number of tasks, client that client had submitted and launches end in DSP Whether consuming capacity, client's special demand and/or client mine massively to same Ziren extends multiples etc. with multiple, once some Look-alike tasks are dispatched, then can be according to seed demographic data and user characteristic data after obtaining seed demographic data To calculating the target audience crowd similar to seed demographic data, more similar target audiences are found, expands precision marketing and covers Lid range.In addition, may also be combined with the first party data that client has by oneself, such as official website visitor, place an order, pay close attention to wechat, concern microblogging and The data such as mobile application client are installed as seed demographic data, in the same of the accurately customer data grasped using client When, it can also meet the personalized customization demand of client.

Client can be based on its own demand, autonomous to determine look-alike number of tasks, its consumption energy at DSP dispensings end Power, special needs, such as the scheduling parameters such as double 11, Holiday Sale can also independently determine user's magnitude after extension, be No to be mined massively to same Ziren with multiple extension multiples, the specific multiple that extends is how many, passes through and the customer demand of input is carried out Scheduling, obtains seed demographic data, carries out automation extension convenient for the subsequent shared attribute according to seed demographic data, can fill Divide and meet different advertisers to the different demands precisely with covering.

DSP (Demand Side Platform) described above is party in request's platform, is responsible for receiving dispensing demand, It looks for demographic data, realizes a central management control platform for launching the functions such as bid, from being mainly characterized by precise positioning mesh Mark crowd.For example, advertiser when launching advertisement, dispensing demand is inputted on DSP, target audience is drawn a circle to approve according to dispensing demand Description, such as age, gender, occupation and hobby etc. can also set fixed condition, as user is using PC to click every time extensively The unit price of announcement is no more than 2 points of, and then these conditions are sent in dsp system, and dsp system is linked up with DMP, according to DMP systems The user characteristic data of middle analysis extraction, is found out the crowd of condition coupling, is carried out by using actual environments such as media resources Advertisement is launched.

After seed demographic data is obtained, seed crowd need to be analyzed from multiple dimensions, therefrom filtered out most Representational common characteristic according to these feature combination user characteristic datas, filters out another batch from a large amount of any active ues The user similar to seed crowd.Specifically：Can computation model be established according to machine learning, which can be used supervision and learn Seed demographic data and user characteristic data are substituted into and calculate mould by the Clustering Model in disaggregated model or unsupervised learning in habit Type carries out calculating analysis, obtains the user similar to seed crowd, and the user scope that can orient advertisement is from extensive characteristic According to more accurate user is contracted to, advertiser is met to the different demands precisely with covering, improves crowd's expansion efficiency.

In fact, a simple examples of the above-mentioned only embodiment of the present invention, particular content of the invention will be by following Each embodiment describes one by one.

As shown in Fig. 2, for the method provided in an embodiment of the present invention that extending user is determined according to the statistical data degree of association, In,

Step 201, statistical data associated with the network behavior of users all in data network is obtained, and to the system It counts and carries out feature extraction to determine the user characteristics of all users.

Data network can be general internet data or various dedicated networks.Wherein, it needs to obtain data The network behavior data of all users in network.The network behavior of user can pass through various operations of the user during online Which which which obtain, for example, website can be logged in including user, browse content or watch video to determine.User Network behavior it is varied, can be the operation content of user or the behavior trace content of user.

Obtaining user network behavior can be carried out by recording the network behavior of user, can also be soft by various networks Hardware obtains.In fact, the acquisition of user network behavior, is more sorted out and is recorded based on the analysis to user network behavior.

After the network behavior of user is recorded, need it is for statistical analysis, can be by user network by various statistical analyses Network behavior is sorted out.User network behavior includes a plurality of types of behaviors and plurality of kinds of contents, needs classification storage.It is deposited in classification On the basis of storage, it is counted, so as to obtain statistical data.Include in statistical data all user network behaviors with And the various possible network behaviors by user network behavior sorted generalization.The network behavior of user can include：Search is clicked Behavior, browsing webpage behavior and/or the behavior obtained by third party's cooperation.

In fact, the analysis to user network behavior, can extract user characteristics.User characteristics are that user registration exists Various actions feature on network.User characteristics are the motion characteristics of user, the operation behavior including user on network and can The action behavior of energy.User characteristics have characterized each action details of user's operation, so as to therefrom determine user's debarkation net The behavioural habits of network and corresponding prediction is made to its behavior.

User characteristics can include the host features of user, n-gram features, surf time section, region belonging to online and/ Or browsing commodity behavior.User characteristics are different from user property.User property is usually the fixed attribute of user, including with User identifier, age, the IP at family etc. are attached to the attribute information of user itself.And user characteristics are dynamic during user's online Make behavioural information, be the operability breath of user in a network, be dynamic.

Thus, in the environment of mass users, user property is attached to number of users and magnanimity.And user characteristics, Due to being related to a variety of behaviors of mass users, thus, data volume is even more more huger than user attribute data.

Step 202, the extended requests that fellow users extension is carried out to basic user are received, the extended requests are solved Analysis is with the setting quantity of determining extending user and the positive sample collection including multiple basic users.

Basic user is the user that client is provided as seed.Client can be that advertiser etc. has user is expanded The client of exhibition.Basic user is put forward by client, is set according to the demand of client.Generally, the generation of basic user, The set for the multiple user properties that can be selected according to client and the basic user quantity of client's input, determine the multiple base Plinth user.Namely the client user property paid close attention to according to itself sets basic user.

Client after above-mentioned look-alike systems are logged in, can select key user's attribute of itself concern, system According to user property corresponding basic user dynamic listing and quantity are provided for client.According to the group for the user property that client inputs Variation is closed, dynamic adjusts basic user dynamic listing, until the quality and quantity of basic user meet the demand of client.

Client also needs to set simultaneously the scale of required user's extension namely client sets extension demand, according to expansion Exhibition demand determines the quantity of extending user.For example, client can be used by obtaining 1,000,000 bases to the adjustment of user property Family, it is 10,000,000 then to input overall extending user scale.At this point, the scale of user's extension is 10 times.

After generating basic user, basic user can be incorporated into positive sample collection.Positive sample is grouped as basic user, passes through These basic users can extract the overall user feature of customer demand, so as to obtain the actual demand of client.

Step 203, it determines to include multiple instructions according to all users in the data network and the multiple basic user Practice the negative sample collection of user, wherein the ratio of the basic user and the quantity of training user is less than or equal to predetermined threshold.

After obtaining positive sample collection, it is also necessary to which training obtains negative sample collection.Negative sample collection is all users out of data network User characteristics in extract user composition.Wherein, the user that negative sample is concentrated is the basic user phase concentrated with positive sample The minimum user's set of closing property.That is, negative sample concentrates user substantially to concentrate user completely unrelated with positive sample.

The acquisition of negative sample collection needs the user characteristics of basic user to be concentrated to carry out analysis extraction positive sample, finds it The general character of user characteristics therefrom extracts the elemental user feature of basic user, then compares one by one in all users of whole network again To user characteristics, so as to find user corresponding with the user characteristics of elemental user feature correlation minimum.These users are put Enter negative sample collection.

The user that negative sample is concentrated also need to carry out it is relevant be further processed, to dispose wherein unrelated data. For example, it is desired to filter out the user data of apparent exception.

For ease of the training of subsequent computation model, the positive sample collection in the user characteristic data of said extracted can be demarcated It is negative class sample data by the negative sample collection data scaling in the user characteristic data of said extracted for positive class sample data.Cause For data characteristics, the scale of positive class sample is at most in millions of magnitudes, and the scale of negative class sample is in several hundred million magnitudes, class positive in this way The ratio of sample and negative class sample reaches 1:10 or even 1:100, this is unfavorable for machine learning model, particularly disaggregated model, learns Practise effective model.For this purpose, the processing of following step can be made to above-mentioned sample.

Over-sampling is carried out to the positive class sample data, negative sampling is carried out to the negative class sample data.It specifically, can be with According to positive class sample data and the ratio of negative class sample data, the sample rate of adjustment setting over-sampling and time sampling, by multigroup Experiment, determines a feasible ratio.Preferably, before being sampled to positive class sample data and negative class sample data, also The user characteristic data of above-mentioned acquisition can be analyzed and filter its apoplexy involving the solid organs sample data, avoid influencing the accurate of subsequent disaggregated model Property.

Step 204, the user characteristics of the multiple training users concentrated to the negative sample carry out signature analysis, to determine to use In the computation rule calculated each user's degree of being associated.

The sampled data obtained to the over-sampling and negative sampling trains computation model.Obtain positive sample collection data it Afterwards, user characteristics need to be analyzed from multiple dimensions, therefrom filters out most representational common characteristic, according to these spies Sign combines user characteristic data, and another crowd of user similar to seed crowd is filtered out from a large amount of any active ues.Specifically：It is first First need select computation model, computation model may include logistic regression (logistic regression algorithm model) and/or The models such as linear SVM (supporting vector machine model), by the sampled data obtained through above-mentioned over-sampling and negative sampling using calculating Model carries out model training, obtains effective computation model.

The user characteristics of basic user that multiple training users concentrate with the positive sample is concentrated to distinguish to the negative sample It extracts, compares the relevance of the two, extract the computation rule.

Here computation model is computation rule, and complete computation rule is extracted by way of model training.

Step 205, the degree of association score value of each user in all users is calculated based on the computation rule, according to described The descending order of degree of association score value is ranked up all users to generate user list.

The computation rule obtained by above-mentioned model training can do model prediction to the whole network user, be sorted based on prediction Go out user of the prediction point more than certain threshold value as extension crowd, i.e., similar target audience crowd, the use that advertisement can be oriented Family range is contracted to more accurate user from extensive characteristic, meet advertiser to precisely and covering different demands, Improve crowd's expansion efficiency.

According to the computation rule, calculating is compared in the user characteristics of users multiple in the data network one by one, Each user and the degree of association score value of the basic user are assigned according to contrast conting result.It will be more in the data network A user is ranked up according to its degree of association score value, and the result of sequence is adjusted according to user property.

One degree of association score value generates each user by computation rule respectively, this degree of association score value characterizes user With the degree of association of basic user.By the arrangement of the size of all users degree of association score value from big to small, user list is obtained, In include putting in order for all users and respective degree of association score value.

Step 206, by the highest setting quantity of degree of association score value in the user list for eliminating the multiple basic user User be determined as extending user.

After obtaining the relevant user's arrangement of specific degree of association score value, can it be chosen according to the size of degree of association score value The higher certain customers of middle degree of association score value are as extending user.Specifically quantity is determined according to the setting of client, Ke Yishi The extending user scale amounts of client's setting.

The basic user initially selected due to including client in the whole network user, and these basic users not necessarily degree of association Score value is higher, therefore, it is possible to the selection according to client, it is determined whether needs to delete in final extending user recommendation list Basic user.

Delete basic user time can be before calculating correlation score value, can also calculating correlation score value it Afterwards.It alternatively, can be before or after extending user be recommended.

In the present embodiment, the customer demand of input is scheduled, obtains the positive sample collection and and base for including basic user The negative sample collection of the completely unrelated user's composition of plinth user, the user characteristics for concentrating user by positive sample collection and negative sample are instructed Practice model, obtain computation rule, calculated one by one for each user of the whole network according to computation rule and concentrate being associated with for user with positive sample Score value is spent, is expanded user according to degree of association score value.Client can obtain the audient crowd's number to match with oneself actual demand According to precision is high, can fully meet the different demands of client.

Fig. 3 shows that the present invention provides a kind of system that extending user is determined according to the statistical data degree of association, the systems System includes：

User characteristics unit 301, for obtaining statistical number associated with the network behavior of users all in data network According to, and feature extraction is carried out to the statistical data to determine the user characteristics of all users；

Positive sample collection unit 302, for receiving the extended requests that fellow users extension is carried out to basic user, to the expansion Exhibition request is parsed the setting quantity to determine extending user and the positive sample collection including multiple basic users；

Negative sample collection unit 303, for true according to all users in the data network and the multiple basic user Surely include the negative sample collection of multiple training users, wherein the ratio of the basic user and the quantity of training user is less than or equal to Predetermined threshold；

Computation rule unit 304, the user characteristics of multiple training users for being concentrated to the negative sample carry out feature Analysis, to determine the computation rule for calculating each user's degree of being associated；

Calculation of relationship degree unit 305, for calculating the association of each user in all users based on the computation rule Score value is spent, all users are ranked up according to the descending order of the degree of association score value to generate user list；

Extending user unit 306, for degree of association score value will to be eliminated in the user list of the multiple basic user most The user of high setting quantity is determined as extending user.

Preferably, by the highest setting quantity of degree of association score value in the user list for not removing the multiple basic user User is determined as extending user.

Preferably, according to the statistics of the network behavior off-line data of all users of data network, all users are extracted User characteristics.

Preferably, the network behavior of the user includes：Search click behavior browses webpage behavior and/or passes through third The behavior that Fang Hezuo is obtained.

Preferably, the user characteristics include：The host features of user, n-gram features, surf time section, belonging to online Region and/or browsing commodity behavior.

Preferably, the basic user quantity that the set for the multiple user properties selected according to client and client input, really Determine the positive sample collection of basic user.

It preferably, will be in all users in the data network and the user characteristics degree of association of the multiple basic user The smaller user of score value, classification become the negative sample collection for including multiple training users.

Preferably, filtering the apparent abnormal dirty sample data of user characteristics in all users in the data network, obtain To negative sample collection.

Preferably, the user characteristics of all users in the data network are carried out with negative sampling, according to the setting threshold Value and basic user quantity, obtain the quantity of the negative sample concentration training user.

Preferably, concentrate the user of basic user that multiple training users concentrate with the positive sample special to the negative sample Sign extracts respectively, compares the relevance of the two, extracts the computation rule.

Preferably, according to the computation rule, the user characteristics of users multiple in the data network are compared one by one To calculating, each user and the degree of association score value of the basic user are assigned according to contrast conting result.

Preferably, users multiple in the data network are ranked up, and according to its degree of association score value to the knot of sequence Fruit is adjusted according to user property.

Preferably, all users in the data network are carried out with negative sampling obtains the negative sample collection, to the basis User carries out positive sampling and obtains positive sample collection；The negative sampling and the downsampling factor just sampled are set as needed.

Preferably, the positive sample of the negative downsampling factor for sampling and just sampling according to actual needs concentrates basic user number Amount and the setting of negative sample concentration training number of users.

As shown in figure 4, for a kind of method that extending user is determined according to weighted calculation provided in an embodiment of the present invention, In,

Step 401, statistical data associated with the network behavior of users all in data network is obtained, and to the system It counts and carries out feature extraction to determine the user characteristics of all users.

Step 402, the extended requests that fellow users extension is carried out to basic user are received, the extended requests are solved Analysis is with the setting quantity of determining extending user and multiple basic users.

Step 403, each corresponding sample set determining respectively being directed in preset multiple training rules, and root Signature analysis is carried out according to the user characteristics in each sample set to determine the computation rule calculated each user's degree of being associated.

Preset training rules can include a variety of, it can be common that classification supervised training to the user characteristics, Cluster training to the user characteristics and/or the semi-supervised training to the user characteristics.

Supervised learning training of classifying is instructed by existing training sample (i.e. given data and its corresponding output) Practice, so as to obtain an optimal models, recycle this model by all new data samples be mapped as accordingly export as a result, Output result is carried out simply to judge the purpose so as to fulfill classification, then this optimal models is also just provided with to unknown number According to the ability classified.

The unsupervised learning training of cluster is to need directly to build data without any training data sample in advance Mould.It is generally necessary to train cluster centre by clustering algorithm, exercised supervision study using cluster centre as classifying rules.

In addition, also semi-supervised learning training pattern, i.e., with reference to supervised learning training pattern and unsupervised learning training mould The prioritization scheme of type.For example, it may be the model that is clustered on basis of classification calculates or enterprising on cluster basis One step thinks that the model of interference classification calculates.

Specific training rules can select according to actual needs, be a variety of different training rule of selection in the present embodiment Then, the whole network user is trained respectively according to a variety of training rules, so that it is determined that sample set corresponding with a variety of training rules. Each sample set be as obtained from corresponding training rules, it is uncorrelated mutually.

User characteristics in each sample set further carry out signature analysis, it may be determined that go out and all users are associated Spend the computation rule of analysis.Equally, each computation rule is to be directed to different training rules, orthogonal.

Sample set is respectively trained by multiple training rules in the present embodiment, and computation rule is being extracted by sample set.It calculates The extraction of rule is typically by the way of model training.User characteristics are analyzed from multiple dimensions, are therefrom filtered out most Representative common characteristic according to these feature combination user characteristic datas, filters out another from a large amount of any active ues Criticize the user similar to seed crowd.Specifically：Firstly the need of selection computation model, computation model may include logistic The models such as regression (logistic regression algorithm model) and/or linear SVM (supporting vector machine model), will be through above-mentioned sample The sampled data of this concentration carries out model training using computation model, obtains effective computation model.

According to the multiple training rules, corresponding sample set is determined respectively；According to the user in each sample set Feature analyzes all users, determines the degree of association of each user, and obtains the computation rule of calculation of relationship degree.

According to the computation rule, all user's degree of being associated are calculated respectively, obtain the degree of association point of each user Value；The user is ranked up according to the degree of association score value of each user.

Step 404, the pass of each user in all users is calculated based on each computation rule in multiple computation rules Connection degree score value is ranked up all users according to the descending order of the degree of association score value to generate multiple user lists.

A variety of computation rules can calculate the degree of association score value of all users of a set of the whole network respectively.If for example, by three kinds Computation rule, then each user of the whole network 3 degree of association score values can be calculated.According to the degree of association score value of each user, difference The whole network user is arranged, multiple user lists is obtained namely each computation rule corresponds to a user list, including All user and respective degree of association score value put in order.

Step 405, weighted value is set for each user list according to the accuracy of each training rules, according to each user The degree of association score value of each user is weighted in the weighted value of list, to determine each to use according to the result of weighted calculation The output score value at family.

According to the degree of association score value of each user, the accuracy of corresponding computation rule is calculated；According to described accurate Degree sets weighted value to corresponding computation rule.Since the accuracy of different computation rules is different, it is then desired to according to Its accuracy for each user list set a weighted value, this weighted value be according to the accuracy of corresponding computation rule come Setting.

According to the weighted value of each computation rule, the degree of association of each user obtained to each computation rule Score value is weighted processing.

Degree of association score value by each user respectively in multiple user lists is multiplied with its weighted value, then weighted calculation is most Whole output score value.Extending user is chosen according to final output score value.

In the present embodiment, it is the degree of association score value that aggregative weighted calculates each user under a variety of computation rules, then weights Calculate final score value.By the weighting that each computation rule of each user obtains handle as a result, being multiplied or phase Add, obtain the output score value of each user.The user is ranked up according to the output score value of each user, output setting number The user of amount is as extending user.

Users multiple in data network are exported score value according to it to be ranked up, and to the result of sequence according to user Attribute is adjusted.The user of the highest setting quantity of score value will be exported in the user list for eliminating the multiple basic user It is determined as extending user.

In the present embodiment, multiple sample sets are trained by preset multiple training rules, and then determine multiple Computation rule；Respectively all users are carried out each user is calculated according to multiple computation rules and be directed to each computation rule Degree of association score value, then calculate the weighted value of each computation rule, the degree of association point of each computation rule corresponded to reference to user The weighted value of value and corresponding computation rule, each user's final output score value of weighted calculation determine to set according to output score value The extending user of quantity.Client can obtain the audient's demographic data to match with oneself actual demand, and precision is high, can fully expire The different demands of sufficient client.

Fig. 5 shows a kind of system that extending user is determined according to weighted calculation provided in an embodiment of the present invention, described System includes：

User characteristics unit 501, for obtaining statistical number associated with the network behavior of users all in data network According to, and feature extraction is carried out to the statistical data to determine the user characteristics of all users；

Basic user unit 502, for receiving the extended requests that fellow users extension is carried out to basic user, to the expansion Exhibition request is parsed setting quantity and the multiple basic users to determine extending user；

Computation rule unit 503, it is each determining corresponding respectively in preset multiple training rules for being directed to Sample set, and user characteristics in each sample set carry out signature analysis to determine in terms of to each user's degree of being associated The computation rule of calculation；

Degree of association score value computing unit 504, for all to calculate based on each computation rule in multiple computation rules The degree of association score value of each user, is ranked up with life all users according to the descending order of the degree of association score value in user Into multiple user lists；And

Score value computing unit 505 is exported, for being each user list setting power according to the accuracy of each training rules Weight values are weighted the degree of association score value of each user according to the weighted value of each user list, in terms of according to weighting The result of calculation determines the output score value of each user.

Preferably, the system also includes：By degree of association score value in the user list for not removing the multiple basic user The user of highest setting quantity is determined as extending user.

Preferably, the basic user quantity that the set for the multiple user properties selected according to client and client input, really Fixed the multiple basic user.

Preferably, the multiple training rules, including to the user characteristics classification supervised training, to user spy The cluster training of sign and/or the semi-supervised training to the user characteristics.

Preferably, according to the multiple training rules, corresponding sample set is determined respectively；According in each sample set User characteristics all users are analyzed, determine the degree of association of each user, and obtain the computation rule of calculation of relationship degree.

Preferably, according to the computation rule, all user's degree of being associated are calculated respectively, obtain the pass of each user Connection degree score value；The user is ranked up according to the degree of association score value of each user.

Preferably, according to the degree of association score value of each user, the accuracy of corresponding computation rule is calculated；According to institute It states accuracy and weighted value is set to corresponding computation rule.

Preferably, according to the weighted value of each computation rule, each user for being obtained to each computation rule Degree of association score value be weighted processing.

Preferably, by the weighting that each computation rule of each user obtains handle as a result, being multiplied or phase Add, obtain the output score value of each user.

Preferably, the user is ranked up by the output score value according to each user, the use of output setting quantity Family is as extending user.

It is ranked up, and to the result of sequence preferably, users multiple in data network are exported score value according to it It is adjusted according to user property.

Preferably, the use for the highest setting quantity of score value being exported in the user list for eliminating the multiple basic user Family is determined as extending user.

Fig. 6 shows a kind of method that extending user is determined according to statistical data interest-degree, the method includes：

Step 601, statistical data associated with the network behavior of users all in data network is obtained, and to the system It counts and carries out feature extraction to determine the user characteristics of all users.

Step 602, the extended requests that fellow users extension is carried out to basic user are received, the extended requests are solved Analysis is with the setting quantity of determining extending user and multiple basic users.

Step 603, user characteristics in set sample set carry out signature analysis with determine to each user into The computation rule of row calculation of relationship degree, and it is each in all users to calculate based on each computation rule in multiple computation rules The initial association degree score value of user.

Training rules are a variety of to include, it can be common that classification supervised training to the user characteristics, to user spy The cluster training of sign and/or the semi-supervised training to the user characteristics.

The degree of association score value of each user in all users is calculated based on each computation rule in multiple computation rules, All users are ranked up according to the descending order of the degree of association score value to generate multiple user lists.

Weighted value is set for each user list according to the accuracy of each training rules, according to the power of each user list The degree of association score value of each user is weighted in weight values, to determine that each user's is initial according to the result of weighted calculation Degree of association score value.

According to the computation rule, all user's degree of being associated are calculated respectively, obtain the initial association of each user Spend score value；The user is ranked up according to the initial association degree score value of each user.

Step 604, interest-degree extraction is carried out to the statistical data to determine the interest-degree score value of each user, and base Initial association degree score value is adjusted in interest-degree score value and corrects degree of association score value to generate.

Each user has unique interest during surfing the Internet again, can be extracted in the statistical data of mass users corresponding Interest-degree, and then determine each user interest-degree score value.This interest-degree score value characterize each user relative to it is specific certain The interest-degree of one user characteristics.For example, for the interest-degree of browse advertisements or particular advertisement, each use is different per family, has User interest degree score value it is high, some user interest degree score values are low.

User characteristics are extracted according to the statistical data of the network behavior of the user, according to the statistical number of the user characteristics It is extracted according to the interest-degree of extraction user.Corresponding interest-degree score value, the user are calculated according to the interest-degree of user statistics There is different interest-degree score values relative to different interest-degrees.

Multiple basic users are obtained according to the fellow users that the client inputs, are determined according to the basic user relevant Interest-degree carries out all users according to the interest-degree interest-degree extraction and calculates the interest-degree point for corresponding to the interest-degree Value.

According to this interest-degree score value, the initial association degree score value of user can be corrected, obtain the correction degree of association Score value.The interest-degree score value is multiplied or is added with initial association degree score value, generates corrected correction degree of association score value. For example, user is zero for the interest-degree score value of browse advertisements, user shields all advertisements, no matter the then user Initial association degree score value is how high, and after weighted correction interest-degree score value, correction degree of association score value is zero.

Step 605, all users are ranked up to generate user according to the descending order of the correction degree of association score value List, the user of the highest setting quantity of the user list lieutenant colonel's positive association degree score value for eliminating the multiple basic user is true It is set to extending user.

By the correction degree of association score value descending arrangement, the correction degree of association of setting quantity is chosen according to described put in order The highest user of score value is as extending user.By users all in the data network according to the correction degree of association score value sequence Afterwards, the highest user of correction degree of association score value for choosing setting quantity is determined as extending user.

It, can be according to the size of correction degree of association score value, choosing after obtaining the relevant user's arrangement of specific degree of association score value The certain customers that wherein correction degree of association score value is higher are taken as extending user.Specific quantity is true according to the setting of client It is fixed, can be the extending user scale amounts of client's setting.

Users multiple in data network are corrected degree of association score value according to it to be ranked up, and to the result root of sequence It is adjusted according to user property.The highest setting quantity of score value will be exported in the user list for eliminating the multiple basic user User be determined as extending user.

The basic user initially selected due to including client in the whole network user, and these basic users not necessarily correct pass Connection degree score value is higher, therefore, it is possible to the selection according to client, it is determined whether needs in final extending user recommendation list Delete basic user.

In the present embodiment, multiple computation rules are determined by preset multiple training rules and sample set；According to Multiple computation rules carry out each user is calculated to all users is directed to the degree of association score value of each computation rule, then count The weighted value of each computation rule is calculated, corresponds to the degree of association score value of each computation rule and corresponding computation rule with reference to user Weighted value, each user's initial association degree score value of weighted calculation；In conjunction with the interest-degree score value of user, weighted calculation is final Degree of association score value is corrected, the extending user of setting quantity is determined according to correction degree of association score value.Client can obtain practical with oneself Audient's demographic data that demand matches, precision is high, can fully meet the different demands of client.

Fig. 7 shows a kind of system that extending user is determined according to statistical data interest-degree provided in an embodiment of the present invention, The system comprises：

User characteristics unit 701, for obtaining statistical number associated with the network behavior of users all in data network According to, and feature extraction is carried out to the statistical data to determine the user characteristics of all users；

Basic user unit 702, for receiving the extended requests that fellow users extension is carried out to basic user, to the expansion Exhibition request is parsed setting quantity and the multiple basic users to determine extending user；

Initial association degree computing unit 703 carries out signature analysis for the user characteristics in set sample set To determine the computation rule that calculates each user's degree of being associated, and based on each computation rule in multiple computation rules come Calculate the initial association degree score value of each user in all users；

Calculation of relationship degree unit 704 is corrected, for carrying out interest-degree extraction to the statistical data with determining each user Interest-degree score value, and based on interest-degree score value initial association degree score value is adjusted with generate correct degree of association score value；

Extending user unit 705, for being arranged according to the descending order of the correction degree of association score value all users Sequence will eliminate the highest setting of user list lieutenant colonel's positive association degree score value of the multiple basic user to generate user list The user of quantity is determined as extending user.

Preferably, user characteristics are extracted according to the statistical data of the network behavior of the user, according to the user characteristics Statistical data extraction user interest-degree extraction.

Preferably, calculating corresponding interest-degree score value according to the interest-degree of user statistics, the user is not relative to Same interest-degree has different interest-degree score values.

Preferably, multiple basic users are obtained according to the fellow users that the client inputs, it is true according to the basic user Fixed relevant interest-degree carries out interest-degree extraction to all users according to the interest-degree and calculates corresponding to the interest-degree Interest-degree score value.

Preferably, according to the computation rule, all user's degree of being associated are calculated respectively, obtain the first of each user Beginning degree of association score value；The user is ranked up according to the initial association degree score value of each user.

Preferably, the interest-degree score value to be multiplied or be added with initial association degree score value, corrected correction is generated Degree of association score value.

Preferably, by the correction degree of association score value descending arrangement, the school of setting quantity is chosen according to described put in order The highest user of positive association degree score value is as extending user.

Preferably, after users all in the data network are sorted according to the correction degree of association score value, setting is chosen The highest user of correction degree of association score value of quantity is determined as extending user.

Fig. 8 shows a kind of method for being used to carry out user characteristics distributed coding provided in an embodiment of the present invention, institute The method of stating includes：

Step 801, statistical data associated with the network behavior of users all in data network is obtained, and to the system It counts and carries out feature extraction to determine multiple user characteristics and determine the total quantity of the multiple user characteristics.

In above-mentioned each embodiment, actually it is required to handle mass users feature, and in the mode handled, by In the limitation of hardware condition, the calculation amount of mass users characteristic is excessively huge, it is difficult to completed in specific computer, because And it needs after carrying out distributed coding to user characteristics, then carry out relevant calculating and processing.

In the user characteristic data of magnanimity, it is necessary first to determine the total quantity of user characteristics.The sum of this user characteristics Amount is variation, is adjusted according to the demand of client, thus, for each user demand, it is required to redefine user The total quantity of feature.

In fact, after client accesses look-alike systems, mission requirements are submitted.System recalls need according to mission requirements The user characteristics and the processing quantity of user characteristics wanted.

The multiple user characteristics are subjected to classification rejecting, the number of the user characteristics of counting user demand according to user demand Amount, the total quantity as user characteristics.

Step 802, structure includes the tag file of the multiple user characteristics, and will based on preset division rule The tag file is divided into multiple subfiles.

For all user characteristics, it is necessary first to build consumer profiles, include this in consumer profiles All user characteristic datas of mission requirements.Then, it needs to be drawn all user characteristics according to preset division rule It is divided into multiple subfiles.User characteristics quantity in each subfile is roughly equal.Alternatively, the user in each subfile is special Sign quantity can be determined according to the load capacity of specific processing equipment.

In general, being respectively divided multiple user characteristics into different subfiles, need using clustering algorithm.It will be the multiple Division rule of the tag file of user characteristics based on hash function is divided into multiple sons corresponding to the processing number of nodes File.Each user characteristics are divided into any one bucket according to hash function.

One of which clustering algorithm is referred to following proposal：

All user characteristics to be encoded are divided into N barrels；

This method may also include：It analyzes the user characteristics to be encoded and filters its apoplexy involving the solid organs sample data, after avoiding influence The accuracy of continuous disaggregated model.

Whole user characteristics can be divided into N barrels according to user characteristics scale and calculate node regulation setting N, it optionally, can Each user characteristics are divided into any one bucket using Hash hash functions, treat that whole user characteristics calculate affiliated bucket Afterwards, the user characteristics quantity of each bucket is calculated, this user characteristics quantity can be recorded with array, is denoted as Array [i], wherein, I is the number of bucket, and i=0,1,2,3 ... N, Array [0] represent the user characteristics quantity that the 0th bucket contains, Array [i-1] Represent the user characteristics quantity that (i-1)-th bucket contains.

The above-mentioned Array [i] calculated is converted into accumulation and AccumulatedArray [i], calculation formula is AccumulatedArray [i]=AccumulatedArray [i-1]+Array [j], wherein AccumulatedArray [0]= Array[0]；Coding, the start_index are proceeded by from start_index [i]+1 to user characteristics in each bucket again [i]=AccumulatedArray [i-1].

N number of calculate node can be called, each calculate node encodes the user characteristics in a bucket, specifically, first It first handles i-th, if i=0, remembers start_index [0]=0, otherwise remember start_index=AccumulatedArray [i-1]；Coding is proceeded by from start_index [i]+1 to bucket interior element again.By the above process, you can complete distributed User characteristics coding, and it can guarantee that the coding between each bucket does not have conflict.

Step 803, content scanning is carried out to the user characteristics in each subfile, to determine that user is special in each subfile The quantity of sign.

Each subfile includes a certain number of user characteristics, in order to accurately encode, needs to each subfile In user characteristics carry out content scanning, determine in each subfile determine user characteristics quantity.

Meanwhile between each subfile and have associated order, it is specific to need by encoding progress.The quantity of subfile It can be determined according to the quantity of the processing unit of actual treatment task.It also needs to filter apparent abnormal in the user characteristics Dirty sample data.

Step 804, in the total quantity and each subfile of the space encoder based on user characteristics, the multiple user characteristics The subnumber amount of user characteristics determines the coding subspace of user characteristics subset in each subfile.

The space encoder of the user characteristics is determined according to the maximum group/cording quantity that the coding method can allow for.According to The processing capacity of each processing node determines the coding subspace of each subfile.Thus, the volume of each subfile Numeral space is actually that the user characteristics subset of corresponding each subfile determines.

Step 805, according to preset processing rule, each subfile and corresponding coding subspace are sent to more Handled accordingly in a processing node node with by it is corresponding processing node to the user characteristics in user characteristics subset into Row coding.

Preset processing rule as combines the processing capacity of each processing node, select the subfile adapted into Row processing.When encoding user characteristics, need the corresponding subfile of corresponding character subset sending corresponding processing section Point carries out.

In the present embodiment, it is divided by the consumer profiles for forming all user characteristics corresponding to processing node Multiple subfiles obtain the number of the user characteristics quantity and subfile in each subfile, subfile are sent respectively corresponding Processing node handled.Encoded using distributed nature, can lifting scheme operational efficiency, reduce exploitation and safeguard it is additional Workload.Client can obtain the audient's demographic data to match with oneself actual demand, and precision is high, can fully meet client's Different demands.

Fig. 9 shows a kind of system for being used to carry out user characteristics distributed coding provided in an embodiment of the present invention, packet It includes：

User characteristics unit 901, for obtaining statistical number associated with the network behavior of users all in data network According to, and feature extraction is carried out to the statistical data to determine multiple user characteristics and determine the sum of the multiple user characteristics Amount；

Tag file construction unit 902, for building the tag file for including the multiple user characteristics, and based on advance The tag file is divided into multiple subfiles by the division rule of setting；

Feature quantity confirmation unit 903, it is every to determine for carrying out content scanning to the user characteristics in each subfile The quantity of user characteristics in a subfile；

Encode subspace confirmation unit 904, for the space encoder based on user characteristics, the multiple user characteristics it is total The subnumber amount of user characteristics determines the coding subspace of user characteristics subset in each subfile in quantity and each subfile；

Node allocation unit 905 is handled, for according to preset processing rule, each subfile to be compiled with corresponding Numeral space is sent in multiple processing nodes handles node with sub to user characteristics by corresponding processing node accordingly The user characteristics of concentration are encoded.

Preferably, all user characteristics are divided into N barrels；

Preferably, as the i=0, AccumulatedArray [0]=Array [0].

Preferably, as the i=0, start_index [0]=0.

Further, the present embodiment provides a kind of mobile terminal, including or for perform as above any one embodiment institute The system stated.

In the specification provided in this place, numerous specific details are set forth.It is to be appreciated, however, that the implementation of the present invention Example can be put into practice without these specific details.In some instances, well known method, structure is not been shown in detail And technology, so as not to obscure the understanding of this description.

Similarly, it should be understood that in order to simplify the disclosure and help to understand one or more of each inventive aspect, Above in the description of exemplary embodiment of the present invention, each feature of the invention is grouped together into single implementation sometimes In example, figure or descriptions thereof.However, the method for the disclosure should be construed to reflect following intention：I.e. required guarantor Shield the present invention claims the more features of feature than being expressly recited in each claim.More precisely, as following Claims reflect as, inventive aspect is all features less than single embodiment disclosed above.Therefore, Thus the claims for following specific embodiment are expressly incorporated in the specific embodiment, wherein each claim is in itself Separate embodiments all as the present invention.

Those skilled in the art, which are appreciated that, to carry out adaptively the module in the equipment in embodiment Change and they are arranged in one or more equipment different from the embodiment.It can be the module or list in embodiment Member or component be combined into a module or unit or component and can be divided into addition multiple submodule or subelement or Sub-component.Other than such feature and/or at least some of process or unit exclude each other, it may be used any Combination is disclosed to all features disclosed in this specification (including adjoint claim, abstract and attached drawing) and so to appoint Where all processes or unit of method or equipment are combined.Unless expressly stated otherwise, this specification is (including adjoint power Profit requirement, abstract and attached drawing) disclosed in each feature can be by providing identical, equivalent or similar purpose alternative features come generation It replaces.

In addition, it will be appreciated by those of skill in the art that although some embodiments described herein include other embodiments In included certain features rather than other feature, but the combination of the feature of different embodiments means in of the invention Within the scope of and form different embodiments.For example, embodiment claimed in detail in the claims is one of arbitrary It mode can use in any combination.

The all parts embodiment of the present invention can be with hardware realization or to be run on one or more processor Software module realize or realized with combination thereof.The present invention is also implemented as performing side as described herein The some or all equipment or program of device (for example, computer program and computer program product) of method.It is such Realizing the program of the present invention can may be stored on the computer-readable medium or can have the shape of one or more signal Formula.Such signal can be downloaded from internet website to be obtained either providing or with any other shape on carrier signal Formula provides.

It should be noted that the present invention will be described rather than limits the invention, and ability for above-described embodiment Field technique personnel can design alternative embodiment without departing from the scope of the appended claims.Word "comprising" is not arranged Except there are element or steps not listed in the claims.Word "a" or "an" before element does not exclude the presence of more A such element.The present invention can be by means of including the hardware of several different elements and by means of properly programmed calculating Machine is realized.If in the unit claim for listing equipment for drying, several in these devices can be by same Hardware branch embodies.

The above is only the specific embodiment of the present invention, it is noted that for the ordinary skill people of this field Member for, without departing from the spirit of the invention, can make it is several improve, modification and deformation, these improve, modification, It is regarded as falling in the protection domain of the application with deformation.

Claims

1. a kind of method for being used to carry out user characteristics distributed coding, including：

Statistical data associated with the network behavior of users all in data network is obtained, and the statistical data is carried out special Sign extraction is with the total quantity of determining multiple user characteristics and determining the multiple user characteristics；

Structure includes the tag file of the multiple user characteristics, and is based on preset division rule by the tag file It is divided into multiple subfiles；

Content scanning is carried out to the user characteristics in each subfile, to determine the quantity of user characteristics in each subfile；

The son of user characteristics in space encoder based on user characteristics, the total quantity of the multiple user characteristics and each subfile Quantity determines the coding subspace of user characteristics subset in each subfile；And

According to preset processing rule, each subfile and corresponding coding subspace are sent in multiple processing nodes It is corresponding to handle node to be encoded by corresponding processing node to the user characteristics in user characteristics subset.

2. the method as described in claim 1 clicks behavior, browsing webpage according to the search of all users in the data network Behavior and/or the behavior obtained by third party's cooperation, obtaining the network behavior of all users in the data network is associated Statistical data.

3. method as claimed in claim 2, according to the host features of user, n-gram features, online in the statistical data Region belonging to period, online and/or browsing commodity behavior, carry out feature extraction, to determine multiple user characteristics.

4. the multiple user characteristics are carried out classification rejecting by the method as described in claim 1 according to user demand, statistics is used The quantity of the user characteristics of family demand, the total quantity as user characteristics.

5. the method as described in claim 1 advises division of the tag file of the multiple user characteristics based on hash function Then, multiple subfiles corresponding to the processing number of nodes are divided into.

6. a kind of system for being used to carry out user characteristics distributed coding, the system comprises：

User characteristics unit, for obtaining statistical data associated with the network behavior of users all in data network, and it is right The statistical data carries out feature extraction to determine multiple user characteristics and determine the total quantity of the multiple user characteristics；

Tag file construction unit, for building the tag file for including the multiple user characteristics, and based on preset The tag file is divided into multiple subfiles by division rule；

Feature quantity confirmation unit, for carrying out content scanning to the user characteristics in each subfile, to determine each Ziwen The quantity of user characteristics in part；

Encode subspace confirmation unit, for the space encoder based on user characteristics, the multiple user characteristics total quantity and The subnumber amount of user characteristics determines the coding subspace of user characteristics subset in each subfile in each subfile；And

Node allocation unit is handled, for regular according to preset processing, by each subfile and corresponding coding sky Between be sent to it is multiple processing nodes in handle accordingly node with by it is corresponding processing node in user characteristics subset User characteristics are encoded.

7. system as claimed in claim 6 clicks behavior, browsing webpage according to the search of all users in the data network Behavior and/or the behavior obtained by third party's cooperation, obtaining the network behavior of all users in the data network is associated Statistical data.

8. system as claimed in claim 7, according to the host features of user, n-gram features, online in the statistical data Region belonging to period, online and/or browsing commodity behavior, carry out feature extraction, to determine multiple user characteristics.

9. the multiple user characteristics are carried out classification rejecting by system as claimed in claim 6 according to user demand, statistics is used The quantity of the user characteristics of family demand, the total quantity as user characteristics.

10. system as claimed in claim 6 advises division of the tag file of the multiple user characteristics based on hash function Then, multiple subfiles corresponding to the processing number of nodes are divided into.