CN103577876A - Credible and incredible user recognizing method based on feedforward neural network - Google Patents

Abstract

The invention discloses a credible and incredible user recognizing method based on a feedforward neural network. The method aims to solve the problems that in the prior art, precision is low, recognition bases are not sufficient, flexibility is poor and social network analysis granularity is rough. The method includes the steps of (1) obtaining a special user and determining users contained in a training set, (2) analyzing and quantifying user characteristics so as to represent the users as user characteristic vectors, (3) building the feedforward neural network, (4) training the feedforward neural network, and (5) achieving credible and incredible user recognition through the trained feedforward neural network. The credible and incredible user recognition includes the steps of (1) obtaining user information in a social network, (2) quantifying the user information to generate the user characteristic vectors, and (3) inputting the user characteristic vectors into the feedforward neural network and recognizing credible users and incredible users according to output values of output nodes.

Description

Credible and insincere user identification method based on feedforward neural network

Technical field

The present invention relates to the credible and insincere user identification method in a kind of community network field, or rather, the present invention relates to a kind of credible and insincere user identification method based on feedforward neural network.

Background technology

Nowadays, community network provides an information interchange comparatively easily and efficiently and resource sharing platform for the network user, yet because community network has open and virtual property, a large amount of unreal information even deceptive information is full of cyberspace, cause non-sincere phenomenon day by day serious, society is caused to harmful effect, upset civil order, so recognition network deceptive information becomes to attach most importance to and hot research problem.In view of the network user is the medium of information issue and propagation, the primary work of the identifying information true and false is the identification to credible in community network and insincere user.

Existing credible and insincere user identifies work and mostly around user's credit worthiness, calculate to launch, and is mainly divided three classes: based on user behavior, topological structure based on community network with based on user behavior and network topology structure.But calculate according to comparatively simple:

1. front two class methods have all only been considered single credit worthiness influence factor.

Based on user behavior: quote the credit worthiness of user in behavior evaluation community network according to video sharing behavior and key word; Utilize sets theory to carry out credit worthiness assessment in conjunction with individual and social recognition grouping; Contact between the user who finds not know each other mutually according to the up-to-date buddy list of user in online community network and specific user's whole attitude; By number of labels, mark user's credit worthiness, the time dynamic of text credit worthiness relevant with term assessed user's credit worthiness jointly; According to the comment about once transaction, comment publisher credit worthiness, comment issuing time, transaction number and situational factor design credit worthiness valuation functions, assess credit worthiness.

Based on community network topological structure: utilize the assessment of the relational implementation credit worthiness between user; For online auction, the S-graph that sets up social bond between reflection buyer and buyer calculates evaluate parameter, by demonstration/implicit feedback of participant, infers the feedback of deliberately omitting, and with this, assesses its credit worthiness; Transitivity for trust and influence power is assessed credit worthiness.

2. although last class methods have considered user behavior and topological structure, do not introduce all sidedly user's individual factors, more do not relate to the interbehavior between user, ignored user's own characteristic.

Based on user behavior and community network topological structure: combine three kinds of data elements, social status, social adjacency and social similarity are carried out society and trusted and measure.

Although work on hand has realized the assessment of user's credit worthiness in community network to a certain extent, but its assessment is according to having limitation, the user profile that can not utilize all sidedly community network to provide, appraisal procedure dirigibility is lower, and assessment result lacks objectivity, accuracy and credibility.Therefore, in order to address the above problem between topological structure, user's individual factors and the user who proposes to using community network interbehavior as assessment foundation, utilize credible and insincere user in feedforward neural network identification community network, not only overcome the deficiency of only considering single influence factor, also consider all sidedly user's individual factors, improved precision and the objectivity of recognition methods.Wherein, the topological structure of community network and user's individual factors are mainly the segmental society's networks for user place, have proved that local trust tolerance has higher degree of accuracy than global reputation measurer.In addition, neural network can be found undefined assessment foundation in the process of study, has further improved the precision of dirigibility and the identification of appraisal procedure.

Summary of the invention

Technical matters to be solved by this invention is to overcome prior art to have the problems such as precision is inadequate, dirigibility is not enough, intelligent disappearance, and a kind of credible and insincere user identification method based on feedforward neural network is provided.

For solving the problems of the technologies described above, the present invention adopts following technical scheme to realize: the described credible and insincere user identification method based on feedforward neural network comprises the steps:

1. obtain special user and determine training set user:

(1) grade according to user in community network builds initial user set, initial user sets definition:

M＝{μ∈M|μ∈OSW _s，Φ _μ→Φ _tag}

Wherein: M represents initial user set, OSW _suser's set under the s of expression field, μ represents that user gathers OSW _sin user, Φ _μthe individual summary that represents user, Φ _tagrepresent the label that grade high user has;

(2) according to user time relevant information to initial user set delete, the inactive user of filtering, using remaining user as special user;

(3) according to the trusting relationship between other users in special user and community network, build initial community network;

(4) according to the trusting relationship between other users in special user and community network, obtain kind of child user and upgrade initial community network, planting child user is the user that training set comprises, and obtains condition:

ProUser＝{μ _p∈ProUser||μ _s→μ _p|≥2}

Wherein: ProUser represents kind of a child user set, μ _srepresent definite special user in (2), μ _prepresent a user in set ProUser, | μ _s→ μ _p| represent to trust user μ _pspecial user's number.

2. analyze and quantize user characteristics, subscriber's meter is shown to user characteristics vector:

(1) with three aspect information of interbehavior between user's community network topological structure, user's individual factors and user, user is carried out to signature analysis;

(2) quantize user characteristics, subscriber's meter is shown to the user characteristics vector being formed by a plurality of features.

3. build feedforward neural network:

(1) according to the dimension of user characteristics vector, determine the input node number of feedforward neural network;

(2) according to the complexity of trusted users identification, determine the number of plies of feedforward neural network and the nodes that each layer comprises;

(3) according to structure and the performance requirement of feedforward neural network, determine the type that hidden layer and output layer comprise node.

4. train feedforward neural network:

(1) the k value according to k-fold cross validation algorithm is divided into k subset by training set, and the common factor of any two subsets is empty;

(2) feedforward neural network is carried out to k training, choose a different subset as training set at every turn, k-1 remaining subset is as test set;

(3) by k the determined frequency of training of training, according to corresponding accuracy of identification, give different weights, according to the corresponding weight value of k frequency of training, try to achieve weighted sum as final frequency of training;

(4) the final frequency of training of determining according to step (3) is trained feedforward neural network on complete training set.

5. by the feedforward neural network after training, realizing credible and insincere user identifies:

(1) obtain three aspect information of interbehavior quantification between user's community network topological structure, user's individual factors and user, subscriber's meter is shown to user characteristics vector;

(2) user characteristics vector step (1) being obtained is input to feedforward neural network and carries out credible and insincere user's identification, obtains being identified user's output valve;

(3) the output valve identification user according to feedforward neural network is credible or insincere.

The special user that obtains described in technical scheme comprises the steps:

(1) user in initial user set is pressed to descending sort according to user time relevant information;

(2) take be positioned at centre position user as separatrix is divided into two parts by the user after sequence, calculate respectively two parts user's time related information average, obtain non-special user's initial cluster center and special user's initial cluster center, computing formula:

${\mathrm{\Ω}}_c=\frac{\underset{i=1}{\overset{\left|\mathrm{\Θ}\right|/2}{\mathrm{\Σ}}}{\mathrm{\Γ}}_i}{\left|\mathrm{\Θ}\right|/2},{\mathrm{\Ω}}_s=\frac{\underset{i=\left|\mathrm{\Θ}\right|/2}{\overset{\left|\mathrm{\Θ}\right|}{\mathrm{\Σ}}}{\mathrm{\Γ}}_i}{\left|\mathrm{\Θ}\right|/2}$

Wherein: set Θ represents initial user set, | Θ | represent the quantity of user in initial user set, Γ _irepresent user time relevant information, Ω _cand Ω _srepresent respectively non-special user's initial cluster center and special user's initial cluster center;

(3) calculate the distance of each user to two cluster centre in initial user set, computing formula:

$\mathrm{dist}({\mathrm{\Γ}}_i,{\mathrm{\Ω}}_c)={({|{\mathrm{\Γ}}_{i1}-{\mathrm{\Ω}}_{c1}|}^h+\·\·\·+{|{\mathrm{\Γ}}_{\mathrm{ir}}-{\mathrm{\Ω}}_{\mathrm{cr}}|}^h)}^{\frac{1}{h}}$

$\mathrm{dist}({\mathrm{\Γ}}_i,{\mathrm{\Ω}}_s)={({|{\mathrm{\Γ}}_{i1}-{\mathrm{\Ω}}_{s1}|}^h+\·\·\·+{|{\mathrm{\Γ}}_{\mathrm{ir}}-{\mathrm{\Ω}}_{\mathrm{sr}}|}^h)}^{\frac{1}{h}}$

Wherein: Γ _irepresent user time relevant information, Ω _cand Ω _srepresent respectively non-special user's cluster centre and special user's cluster centre, Γ _ir, Ω _crand Ω _srrepresent respectively vectorial component;

(4) by user assignment in the shorter cluster of cluster centre distance;

(5) calculate non-special user's cluster centre and special user's cluster centre, computing formula:

${\mathrm{\Ω}}_c^{\′}=\frac{\underset{i=1}{\overset{\left|{\mathrm{\Θ}}_c^{\′}\right|}{\mathrm{\Σ}}}{\mathrm{\Γ}}_i}{\left|{\mathrm{\Θ}}_c\right|},{\mathrm{\Ω}}_s^{\′}=\frac{\underset{i=1}{\overset{\left|{\mathrm{\Θ}}_s^{\′}\right|}{\mathrm{\Σ}}}{\mathrm{\Γ}}_i}{\left|{\mathrm{\Θ}}_s\right|}$

Wherein: Ω _c' and Ω _s' represent respectively domestic consumer's cluster centre and special user's cluster centre, gather Θ _cand Θ _srepresent respectively domestic consumer's set and special user's set that cluster obtains, | Θ _c| and | Θ _s| represent respectively the quantity of user in domestic consumer's set and special user's set, Γ _irepresent user time relevant information;

(6) check in two clusters, whether user changes.If change, jump to step (3), otherwise, finish.

Quantification user characteristics described in technical scheme comprises the steps:

(1) quantize the topological structure of community network: core degree, user kernel degree is divided into out-degree (Out-Link) and two parts of in-degree (In-Link), and the trusting relationship according to user in community network quantizes out-degree and in-degree, adopts following formula:

Out-Link＝｜Trustee|/(|Trustee|+|Trustor|)

In-Link＝|Trustor|/(|Trustee|+|Trustor|)

Wherein: | Trustee| represents user's that user trusts number, | Trustor| represents to trust user's number of users;

(2) quantize user's individual factors: liveness and influence power, with user, deliver the percentage that Review number and all users of choosing deliver the summation of Review number and recently quantize liveness, adopt following formula:

$\mathrm{Activity}={\mathrm{\μ}}_{\mathrm{RW}}/\underset{\mathrm{\μ}\∈\mathrm{ProUser}}{\mathrm{\Σ}{\mathrm{\μ}}_{\mathrm{RW}}}$

Wherein: μ _rWrepresent that user delivers the number of Review, ProUser represents the user that training set comprises.

According to Member Visits and two attributes of Total Visits, influence power size is quantized, adopts following formula:

$\mathrm{MVP}={\mathrm{\μ}}_{\mathrm{MV}}/\underset{\mathrm{\μ}\∈\mathrm{ProUser}}{\mathrm{\Σ}{\mathrm{\μ}}_{\mathrm{MV}}},\mathrm{TVP}={\mathrm{\μ}}_{\mathrm{TV}}/\underset{\mathrm{\μ}\∈\mathrm{ProUser}}{\mathrm{\Σ}{\mathrm{\μ}}_{\mathrm{TV}}}$

Wherein: MVP represents the number percent of the summation of user Member Visits value and all user of choosing Member Visits values, μ _mVthe Member Visits value that represents user, TVP represents the number percent of the summation of user Total Visits value and all user of choosing Total Visits values, μ _tVthe Total Visits value that represents user, ProUser represents the user that training set comprises;

(3) quantize interbehavior between user: supporting rate, opposition rate, the Review Rating that user's supporting rate and opposition rate are delivered Rview according to user quantizes, and adopts following formula:

$\mathrm{Support}-\mathrm{rate}=\underset{s\∈\left\{\begin{array}{c}\mathrm{VeryHelpful},\\ \mathrm{MostHelpful}\end{array}\right\}}{\mathrm{\Σ}{r}_s}/\left|R_w\right|$

$\mathrm{Oppose}-\mathrm{rate}=\underset{o\∈\left\{\begin{array}{c}\mathrm{OffTopic},\mathrm{NotHelpful},\\ \mathrm{SomewhatHelpful},\mathrm{Helpful}\end{array}\right\}}{\mathrm{\Σ}{r}_o}/\left|R_w\right|$

Wherein: r _srepresent that Review Rating is VeryHelpful, the Review of MostHelpful, r _orepresent that Review Rating is OffTopic, NotHelpful, SomewhatHelpful, the Review of Helpful, | R _w| represent user's Review number.

Structure feedforward neural network described in technical scheme comprises the steps:

(1) according to the dimension of user characteristics vector, determine the input node number of feedforward neural network, have seven input node: Out-Link, In-Link, Activity, MVP, TVP, Support-rate and Oppose-rate;

(2) according to the complexity of trusted users identification, determine the number of plies of feedforward neural network and the nodes that each layer comprises.Have three layers: input layer, seven input nodes; Hidden layer, two concealed nodes; Output layer, an output node;

(3) according to structure and the performance requirement of feedforward neural network, determine that the type that hidden layer and output layer comprise node is sigmoid threshold cell.

Training feedforward neural network described in technical scheme comprises the steps:

(1) the k value according to k-fold cross validation algorithm (in this method, k gets 5) is divided into k subset by training set, and the common factor of any two subsets is empty;

(2) feedforward neural network is carried out to k training, choose a different subset as training set at every turn, k-1 remaining subset, as test set, will be noted " concussion " of accuracy of identification, deconditioning when accuracy of identification is stablized in training process;

(3) recording accuracy of identification and the frequency of training of k training, is that corresponding frequency of training is given weights according to accuracy of identification, calculates the weighted sum of k frequency of training as final frequency of training;

(4) the final frequency of training of determining according to step (3) is trained feedforward neural network on complete training set, obtains the weights of threshold value and the corresponding edge of respective nodes.

Credible and insincere user's identification described in technical scheme comprises the steps:

(1) in community network, obtain between user's to be identified community network topological structure, user's individual factors and user aspect three of interbehaviors information and be quantified as user characteristics vector, the form of user characteristics vector is as follows: (Out-Link, In-Link, Activity, MVP, TVP, Support-rate, Oppose-rate);

(2) user's to be identified proper vector is input to feedforward neural network, obtains user's to be identified output valve;

(3) if the difference of user's to be identified output valve and 1 is less than predetermined threshold value, user is identified as trusted users, otherwise, be identified as insincere user.

Compared with prior art the invention has the beneficial effects as follows:

1. the credible and insincere user identification method based on feedforward neural network of the present invention has provided about clear, clear and definite, the computable information of user and has represented from three aspects of interbehavior between community network topological structure, user's individual factors and user.The nonumeric information that user in community network is discrete is converted into numerical information, enriched the foundation of trusted users and insincere user identification, and with user characteristics vector, represent user, for feedforward neural network provides, be convenient to input data that calculate, various dimensions.Community network has been carried out to stratification parsing: interbehavior between community network topological structure, user's individual factors and user, further clear and definite parsing object, refinement parsing granularity, for user feature analysis and quantification are laid a good foundation.Each level of community network after stratification is resolved has independence, user characteristics in this level is quantized can not exert an influence to other levels, there is high cohesion, the feature of low coupling, be convenient to the improvement of quantization method, made up the deficiencies such as existing method basis of characterization is single, community network parsing granularity is thicker.

2. the construction method that has improved feedforward neural network training set based on the credible and insincere user identification method of feedforward neural network of the present invention.Construction method in the past is only distinguished credible and insincere user according to the existing mark of user, it is ageing that but the problems such as the temporal correlation of community network, user's dynamic change and community network renewal delay cause mark not have, therefore, for the problems referred to above, need to introduce time related information user is screened, with this, solve mark and do not there is ageing problem.When building training set, introduced definite user method, it is ageing that binding time relevant information and clustering technique are that mark adds, and the user that delete flag was lost efficacy, for feedforward neural network provides more standard, training set accurately.Because definite user method be take clustering technique as basis, the distribution of user in training set after wherethrough reason in user characteristics space is more tight, reduce the interstitial content that is positioned at positive example and negative routine boundary in training set, avoided " vibration " problem of discrimination in cross-training process.

3. the credible and insincere user identification method based on feedforward neural network of the present invention combines trusted users identification with machine learning method.Trusted users identification is a kind of thinking activities of people, by the nervous system that is positioned at brain, completed, therefore adopt feedforward neural network to carry out simulated nervous system, neuron in the corresponding nervous system of node in feedforward neural network, the corresponding neuronic electric signal of the input value of node and output valve receives and transmits.In addition, the input value of input node comes from the stratification of community network resolves, and belongs to respectively different classifications, the corresponding input the most original, diversification that in brain, nervous system receives.Trusted users identification problem, with " attribute-value " formal definition, is applicable to using feedforward neural network to solve.In addition, with existing recognition methods, compare the actual settling mode that more approaches problem, and feedforward neural network hidden layer can be found prior undefined basis of characterization in training process, has not only enriched basis of characterization, and the dirigibility of method and the accuracy of identification have been improved.

In sum, the present invention be directed to the Biological characteristics of the level, diversity of community network, the feature such as ageing and trusted users identification, by analogy, the nervous system on biological significance is mapped to feedforward neural network, in conjunction with take structure and credible and the insincere user that clustering technique realizes training set as basic algorithm and corresponding machine learning algorithm, identifies.

Accompanying drawing explanation

Below in conjunction with accompanying drawing, the present invention is further illustrated:

Fig. 1 is the function that respectively forms module, the annexation schematic block diagram of implementing the computer program of the credible and insincere user identification method based on feedforward neural network of the present invention.

Fig. 2 is the FB(flow block) of the definite user method based on cluster of the present invention.

Fig. 3 is feedforward neural network schematic block diagram of the present invention.

Fig. 4 is initial user set of the present invention, and in figure, each node represents an initial user, and the label on node is node representative user's user name, and the distribution of these nodes is without any rule.

Fig. 5 is definite user method result that set is deleted to initial user of the present invention, and Lead node and UnLead node are dummy nodes, represent cluster centre; The node being connected with Lead node has represented special user, and the node being connected with UnLead node has represented disallowable non-special user.

Fig. 6 is initial community network schematic block diagram of the present invention, by special user freak369, the common user who trusts of bryan_carey and popsrocks is divided into four classes: the common user who trusts of special user freak369 and special user bryan_carey, the common user who trusts of special user freak369 and special user popsrocks, common user and the special user freak369 trusting of special user bryan_carey and special user popsrocks, the common user who trusts of bryan_carey and popsrocks, four user's set of this four class user corresponding diagram 6 cores.

Fig. 7 is community network of the present invention (part) schematic block diagram.

Fig. 8 is feedforward neural network cross-validation method training schematic diagram of the present invention.

Fig. 9 is the credible and insincere user's result schematic diagram of feedforward neural network identification of the present invention, exists the node on limit to represent respectively credible and insincere user with node Cre and node Ucr; Exist the node on limit to represent respectively by feedforward neural network with node ANNCre and node ANNUCre and be identified as credible and incredible user; Exist the node representative on limit by feedforward neural network, to be identified as insincere user's trusted users with node Cre and node ANNUCre simultaneously.

Embodiment

Below in conjunction with accompanying drawing, the present invention is explained in detail:

Credible and insincere user identification method technical matters to be solved based on feedforward neural network of the present invention is to overcome the deficiencies in the prior art, based on feedforward neural network, proposing training set builds, obtain special user, determine training set user, quantize user characteristics, build feedforward neural network, training feedforward neural network, key issue in credible and insincere user's identifications such as application feedforward neural network, propose and realize a series of new technology and methods towards trusted users identification, effectively solved basis of characterization not enough, the problems such as community network parsing granularity is thicker, improved the accuracy of identification, for trusted users identification and the combination of machine learning techniques provide technical support.Wherein feedforward neural network is to be subject to biological inspiration, by analogy, biological nervous system principle of work is mapped to Computer Science and Technology field, with perceptron or other types unit, simulate the neuron of biological nervous system, with reception and the transmission of neuron electric signal in the input value of perceptron or other types unit and output valve simulated nervous system.Through carefully analyzing, the feedforward neural network technology that this method adopts can comparatively relevantly be simulated the mode of biological nervous system solving practical problems, and corresponding, the hidden layer that is mainly reflected in information value size in weight and the practical problems of corresponding, unit of the different types of input value of input block and information diversification in practical problems found in new basis of characterization and practical problems the corresponding of relation between information automatically.

Consult Fig. 1, for the credible and insincere user identification method this purpose realizing based on feedforward neural network has been worked out computer program voluntarily, it includes five functional modules, comprise the line module that obtains special user and definite training set and comprise, quantize user characteristics module, build feedforward neural network module, training feedforward neural network module and credible and insincere subscriber identification module, modules function:

1. obtain the line module that special user and definite training set comprise

Described obtain line module that special user and definite training set comprise and be divided into and determine user method and obtain two major parts of kind of child user.Determine that user method comprises the set of structure initial user, obtains user gradation and obtain user time relevant information etc., obtains kind of a child user and comprises the users to trust relation of obtaining, according to users to trust relation, builds initial community network etc.First according to user gradation, build initial user set, then, by determining that user method obtains the user who builds initial community network, finally according to users to trust relation, build initial community network, by the structure of initial community network, obtain kind of a child user.

(1) obtain user gradation

Building more standard, before training set, needing existing subscriber to screen for the first time accurately, screening is according to being the Social Grading of user in community network.Obtain user gradation and mainly adopted Web Crawler, in conjunction with matching regular expressions, go out the class information of user in community network.According to the user gradation information of obtaining, screen for the first time, using the user who remains as initial user set.

(2) determine user method

In method in the past, the use that initial user set middle grade is higher is regarded as trusted users per family.But community network has ageing, some grades compared with high user because its credibility being reduced in disabled state for a long time, do not meet the condition of trusted users in training set, therefore after obtaining initial user set, need by determining that user method retains any active ues, separated inactive user.Determine that user method is to take clustering technique as basis, in conjunction with user time relevant information, guarantee the method that user is ageing, be intended to remove noise node in training set.

(3) build initial community network

By determining that user method can obtain special user.Adopt Web Crawler to obtain special user's trusting relationship, according to trusting relationship, build initial community network.The initial community network building is a directed networks, and direction is by trusting and being trusted decision.Because community network is larger, if adopt the adjacency matrix storage can the more space of waste, so this method be with the form storage community network of file, user's trustor and being stored in the file with user's filename by name by trustor.

(4) obtain kind of a child user

In initial community network, comprised special user, the user of special user's trust and trust special user's user, this three classes user is by trusting and being connected each other by trusting relationship.In initial community network, this method is only paid close attention to the user of special user and special user's trust, by differentiating the relation between user, select the user that special user trusts, again, through determining user method processing, the user who retains is offered to training set as kind of a child user.

2. quantize user characteristics module

Described quantification user characteristics module is divided into quantification community network topological structure, quantizes user's individual factors and quantizes interbehavior three parts between user, this three some work is all that the user who comprises for training set launches, be intended to user's categorical data to be quantified as numeric data, merge various species basis of characterization, obtain user characteristics vector, subscriber's meter is shown to the pattern that feedforward neural network can identifying processing.

(1) quantize community network topological structure

Adopt Web Crawler to obtain the trusting relationship that training set comprises user, according to trusting relationship, build community network, by user kernel metrization community network topological structure.Core kilsyth basalt show a user and other users contact number, to contact its credibility of more user also higher with other users.In this method, core degree comprises out-degree (Out-Link) and in-degree (In-Link) two parts, Trustee represents the user of users to trust, Trustor represents to trust user's user, ratio with the same Trustee of Trustee, Trustor sum quantizes out-degree (Out-Link), with the ratio quantification in-degree (In-Link) of the same Trustee of Trustor, Trustor sum.

(2) quantize user's individual factors

User's individual factors comprises liveness and influence power two parts, embodied respectively user to the sensitivity of certain realm information and user's personal view the influence degree to other users.Percentage with the summation of user Review value and all user of choosing Review values in this method recently quantizes liveness (Activity), with user Member Visits value and Total Visits value, carrys out quantization influence power.Influence power comprises two parts content: MVP and TVP.MVP is the number percent of summation of user Member Visits value and all users' of choosing Member Visits value, and TVP is the number percent of summation of user Total Visits value and all users' of choosing Total Visits value.

(3) quantize interbehavior between user

Between user, interbehavior comprises supporting rate (support-rate) and opposition rate (oppose-rate) two parts.Its credibility of user that supporting rate is higher is higher, and its credibility of user that opposition rate is higher is lower.The Review that user delivers can receive different gradings---Review Rating, to be rated the Review number of support and the number percent of Review sum, quantize supporting rate (support-rate), to be rated the Review number of opposition and the number percent of Review sum, quantize opposition rate (oppose-rate).

3. build feedforward neural network module

Described structure feedforward neural network module is mainly responsible for determining hierarchical structure and node number and the type comprising at all levels of feedforward neural network.

(1) determine input node number

Feedforward neural network in this method comprises seven input nodes, seven components of respective user proper vector.

(2) determine that the number of plies and each layer comprise nodes

In this method, comprise three layers: input layer, hidden layer and output layer.Input layer is as described in step (1), and hidden layer comprises two concealed nodes, and output layer comprises an output node.

(3) determine node type

Input node adopts perceptron unit, is only responsible for receiving the user characteristics vector of input.Concealed nodes and output node adopt sigmoid unit, and output valve is arrived to interval (0,1) by extruding Function Mapping.Extruding function: f (x)=1/ (1+e ^-x), wherein x is the output valve that sigmoid unit is not processed through extruding function.

4. train feedforward neural network module

Described training feedforward neural network module adopts cross validation method to train feedforward neural network.

(1) k-fold cross-training feedforward neural network

Training set is divided into five subsets, and the common factor of any two subsets is empty.Each select a subset as training set, all the other four subsets are as test set, frequency of training when best identified precision obtained in record.After five training, obtain five frequency of training, according to five accuracy of identification, give five weights that frequency of training is certain, calculate the weighted sum of five frequency of training as final frequency of training.

(2) feedforward neural network after training

The final frequency of training of determining according to step (1) is trained feedforward neural network on complete training set, obtains the threshold value of each node of feedforward neural network and the weights on each limit.

5. credible and insincere subscriber identification module

Described credible and insincere subscriber identification module mainly comprises user and is converted to the credible and insincere identification two parts of user characteristics vector sum user.

(1) user is converted to user characteristics vector

Adopt Web Crawler in community network, to obtain user's to be identified information, then the input information obtaining is quantized to user characteristics module and obtain the user characteristics vector by interbehavior three aspects: basis of characterization forms between community network topological structure, user's individual factors and user.

(2) the credible and insincere identification of user

User characteristics vector is input to the feedforward neural network after training, according to the output valve of feedforward neural network output node, user is carried out to credible and insincere identification.

Consult Fig. 2, the step of definite user method of the present invention is as follows:

(1) setting value in k(the present invention is 2) individual cluster centre;

(2) user is pressed to time related information descending sort;

(3) user after sequence is divided into k decile;

(4) calculate each user to the distance of each cluster centre;

(5) by user assignment in cluster corresponding to bee-line;

(6) recalculate k cluster centre;

(7) check whether the user in k cluster changes;

(8) if change, return to (4); If unchanged, finish;

Consult Fig. 3, Architecture of Feed-forward Neural Network of the present invention is as follows:

The feedforward neural network adopting in this method consists of input layer, hidden layer and output layer three parts.

(1) the input data of input layer are user characteristics vectors, every one dimension corresponding input block: Out-Link, an In-Link of user characteristics vector, Activity, MVP, TVP, Support-rate and corresponding seven input blocks of Oppose-rate.

(2) hidden layer comprises two hidden units altogether, and each hidden unit has a threshold value H-Thresholdi(i=1,2), this threshold value has determined the output valve of hidden unit.There are a weights ω ij(i={1,2,3,4,5,6,7}, j={1,2} in every the limit that connects input block and hidden unit), ω ij represents the contribution rate of i input block to j hidden unit output.

(3) output layer comprises an output unit, and threshold value O-Threshold has determined the output valve of output unit.There are weights ω hi(i={1, a 2} in every the limit that connects hidden unit and output unit), ω hi represents the contribution rate of i hidden unit to output unit output.

The output valve of hidden unit and output unit is all processed through sigmoid function.The codomain of sigmoid is 0 to 1, monotone increasing, and by sigmoid function, input codomain can be mapped to scope is 0 to 1 output codomain.

Embodiment:

Consult Fig. 1, the step of the credible and insincere user identification method based on feedforward neural network of the present invention is as follows:

1. obtain the user that special user and training set comprise

(1) adopt Web Crawler to crawl user gradation information, by the set of user gradation information architecture initial user, result is consulted Fig. 4;

(2) by determine user method to initial user set delete, obtain special user, result is consulted Fig. 5;

(3) adopt Web Crawler to crawl users to trust relation, according to trusting relationship, build initial community network, result is consulted Fig. 6;

(4) according to topological structure and definite user method of initial community network, obtain kind of a child user, adopt Web Crawler to crawl the trusting relationship of kind of child user, build community network, result is consulted Fig. 7.

2. quantize user characteristics

Adopt Web Crawler crawl between user's community network topological structure, user's individual factors and user interbehavior three aspects: information and quantize, then, subscriber's meter is shown to user characteristics vector.User characteristics vector is as follows:

<42english>

<Out-Link>0.0030065141139134794</Out-Link>

<In-Link>0.00454078021288246</In-Link>

<Activity>0.002798286587597133</Activity>

<MVP>0.002927235461531413</MVP>

<TVP>0.0034997794019903895</TVP>

<Support-rate>0.03653846153846154</Support-rate>

<Oppose-rate>0.9634615384615385</Oppose-rate>

</42english>

3. build feedforward neural network

Consult Fig. 3, in initialize routine, partly build feedforward neural network.

The input layer of feedforward neural network, hidden layer and output layer are by following code construction:

Input=new?double[inputSize]；

Hidden=new?double[hiddenSize]；

Output=new?double[outputSize]；

Feedforward neural network concealed nodes threshold value and output node threshold value are stated by following code:

HidThresHold=new?double[hiddenSize]；

OptThresHold=new?double[outputSize]；

Between feedforward neural network input node and concealed nodes, between the weights on limit and concealed nodes and output node, the weights on limit are stated by following code:

IptHidWeights=new?double[inputSize][hiddenSize]；

HidOptWeights=new?double[hiddenSize][outputSize]；

By random function, be that above-mentioned threshold value and weights are composed initial value, scope is [0.05,0.05], and code is as follows:

CreatRandomWeight(IptHidWeights);

CreatRandomWeight(HidOptWeights);

4. train feedforward neural network

Adopt cross-validation method training feedforward neural network, training process is consulted Fig. 8.

After training, the threshold value of each node of feedforward neural network and the weights on each limit are as follows:

Concealed nodes threshold value :-0.9997232652135897 ,-0.9997855324086498;

Output node threshold value :-0.9985211755028933;

The weights on limit between input node and concealed nodes:

0.0897245066322621，0.030246991857638347，0.08625975560191126，

0.05646571427266416，0.12009404029133104，0.030833920608579368，

0.07035358302802967，0.13070006844338009，0.07024489700918604，

0.09999458520853907，0.4235673648091106，0.4004164285673208，

2.1125328654365805，2.113744324370964

The weights on limit between concealed nodes and output node:

1.0038892999564673，0.9847503518050365

5. credible and insincere user's identification

(1) in community network, obtain user's to be identified information, quantize and be expressed as user characteristics vector;

(2) the input data using user characteristics vector as feedforward neural network, according to the credible and insincere user of output valve identification of output node, result is consulted Fig. 9, and the vector of user characteristics shown in step 2 of take is example, and identifying is as follows:

The output valve of concealed nodes 1 (HN1OutPut):

Out-Link*W11+In-Link*W21+Activity*W31+MVP*W41+TVP*W51+Support-rate*W61+Oppose-rate*W71+H-Threshold1

The output valve of concealed nodes 2 (HN2OutPut):

Out-Link*W11+In-Link*W21+Activity*W31+MVP*W41+TVP*W51+Support-rate*W61+Oppose-rate*W72+H-Threshold2

The output valve of output node (ONOutput):

HN1OutPut*Wh1+HN2OutPut*Wh2+O-Threshold

Bring respective value into above-mentioned computing formula and obtain following result:

The output valve of concealed nodes 1 (HN1OutPut): 1.051599979598784;

The output valve of concealed nodes 2 (HN2OutPut): 1.052814435066734;

The output valve of output node (ONOutput): 1.09336101901679;

The difference of ONOutput and target output value 1 is less than 0.2, and this user is judged as trusted users.

Claims (6)

1. the credible and insincere user identification method based on feedforward neural network, is characterized in that, the described credible and insincere user identification method based on feedforward neural network comprises the steps:

(1) obtain special user and determine the user that training set comprises:

1) grade according to user in community network builds initial user set, initial user sets definition:

M＝｛μ∈Ｍ|μ∈OSW _s，Φ _μ→Φ _tag}

Wherein: M represents initial user set, OSW _suser's set under the s of expression field, μ represents that user gathers OSW _sin user, Φ _μthe individual summary that represents user, Φ _tagrepresent the label that grade high user has;

2) according to user time relevant information to initial user set delete, the inactive user of filtering, using remaining user as special user;

3) according to the trusting relationship between other users in special user and community network, build initial community network;

4) according to the trusting relationship between other users in special user and community network, obtain kind of child user and upgrade initial community network, planting child user is the user that training set comprises, and obtains condition:

ProUser＝{μ _p∈ProUser||μ _s→μ _p|≥2}

Wherein: ProUser represents kind of a child user set, μ _sexpression 2) definite special user in, μ _prepresent a user in set ProUser, | μ _s→ μ _p| represent to trust user μ _pspecial user's number;

(2) analyze and quantize user characteristics, subscriber's meter is shown to user characteristics vector:

1) with three aspect information of interbehavior between user's community network topological structure, user's individual factors and user, user is carried out to signature analysis;

2) quantize user characteristics, subscriber's meter is shown to the user characteristics vector being formed by a plurality of features;

(3) build feedforward neural network:

1) according to the dimension of user characteristics vector, determine the input node number of feedforward neural network;

2) according to the complexity of trusted users identification, determine the number of plies of feedforward neural network and the nodes that each layer comprises;

3) according to structure and the performance requirement of feedforward neural network, determine the type that hidden layer and output layer comprise node;

(4) training feedforward neural network:

1) the k value according to k-fold cross validation algorithm is divided into k subset by training set, and the common factor of any two subsets is empty;

2) feedforward neural network is carried out to k training, choose a different subset as training set at every turn, k-1 remaining subset is as test set;

3) by k the determined frequency of training of training, according to corresponding accuracy of identification, give different weights, according to the corresponding weight value of k frequency of training, try to achieve weighted sum as final frequency of training;

4) the final frequency of training of determining according to step 3) is trained feedforward neural network on complete training set;

(5) by the feedforward neural network after training, realizing credible and insincere user identifies:

1) obtain three aspect information of interbehavior quantification between user's community network topological structure, user's individual factors and user, subscriber's meter is shown to user characteristics vector;

2) user characteristics vector step 1) being obtained is input to feedforward neural network and carries out credible and insincere user's identification, obtains being identified user's output valve;

3) the output valve identification user according to feedforward neural network is credible or insincere.

2. according to the credible and insincere user identification method based on feedforward neural network claimed in claim 1, it is characterized in that, the described special user that obtains comprises the steps:

(1) according to user time relevant information, the user in initial user set is pressed to descending sort;

(2) take be positioned at centre position user as separatrix is divided into two parts by the user after sequence, calculate respectively two parts user's time related information average, obtain non-special user's initial cluster center and special user's initial cluster center, computing formula:

${\mathrm{\&Omega;}}_c=\frac{\underset{i=1}{\overset{\left|\mathrm{\&Theta;}\right|/2}{\mathrm{\&Sigma;}}}{\mathrm{\&Gamma;}}_i}{\left|\mathrm{\&Theta;}\right|/2},{\mathrm{\&Omega;}}_s=\frac{\underset{i=\left|\mathrm{\&Theta;}\right|/2}{\overset{\left|\mathrm{\&Theta;}\right|}{\mathrm{\&Sigma;}}}{\mathrm{\&Gamma;}}_i}{\left|\mathrm{\&Theta;}\right|/2}$

Wherein: set Θ represents initial user set, | Θ | represent the quantity of user in initial user set, Γ _irepresent user time relevant information, Ω _cand Ω _srepresent respectively non-special user's initial cluster center and special user's initial cluster center;

(3) calculate the distance of each user to two cluster centre in initial user set, computing formula:

$\mathrm{dist}({\mathrm{\&Gamma;}}_i,{\mathrm{\&Omega;}}_c)={({|{\mathrm{\&Gamma;}}_{i1}-{\mathrm{\&Omega;}}_{c1}|}^h+\&CenterDot;\&CenterDot;\&CenterDot;+{|{\mathrm{\&Gamma;}}_{\mathrm{ir}}-{\mathrm{\&Omega;}}_{\mathrm{cr}}|}^h)}^{\frac{1}{h}}$

$\mathrm{dist}({\mathrm{\&Gamma;}}_i,{\mathrm{\&Omega;}}_s)={({|{\mathrm{\&Gamma;}}_{i1}-{\mathrm{\&Omega;}}_{s1}|}^h+\&CenterDot;\&CenterDot;\&CenterDot;+{|{\mathrm{\&Gamma;}}_{\mathrm{ir}}-{\mathrm{\&Omega;}}_{\mathrm{sr}}|}^h)}^{\frac{1}{h}}$

Wherein: Γ _irepresent user time relevant information, Ω _cand Ω _srepresent respectively non-special user's cluster centre and special user's cluster centre, Γ _ir, Ω _crand Ω _srrepresent respectively vectorial component;

(4) by user assignment in the shorter cluster of cluster centre distance;

(5) calculate non-special user's cluster centre and special user's cluster centre, computing formula:

${\mathrm{\&Omega;}}_c^{\&prime;}=\frac{\underset{i=1}{\overset{\left|{\mathrm{\&Theta;}}_c^{\&prime;}\right|}{\mathrm{\&Sigma;}}}{\mathrm{\&Gamma;}}_i}{\left|{\mathrm{\&Theta;}}_c\right|},{\mathrm{\&Omega;}}_s^{\&prime;}=\frac{\underset{i=1}{\overset{\left|{\mathrm{\&Theta;}}_s^{\&prime;}\right|}{\mathrm{\&Sigma;}}}{\mathrm{\&Gamma;}}_i}{\left|{\mathrm{\&Theta;}}_s\right|}$

Wherein: Ω _c' and Ω _s' represent respectively domestic consumer's cluster centre and special user's cluster centre, gather Θ _cand Θ _srepresent respectively domestic consumer's set and special user's set that cluster obtains, | Θ _c| and | Θ _s| represent respectively the quantity of user in domestic consumer's set and special user's set, Γ _irepresent user time relevant information;

(6) check in two clusters, whether user changes, if change, jump to step (3), otherwise, finish.

3. according to the credible and insincere user identification method based on feedforward neural network claimed in claim 1, it is characterized in that, described quantification user characteristics comprises the steps:

(1) quantize the topological structure of community network: core degree, user kernel degree is divided into out-degree (Out-Link) and in-degree (In-Link) two parts, and the trusting relationship according to user in community network quantizes out-degree and in-degree, adopts following formula:

Out-Link＝|Trustee|/(|Trustee|+|Trustor|)

In-Link＝|Trustor|/(|Trustee|+|Trustor|)

Wherein: | Trustee| represents user's that user trusts number, | Trustor| represents to trust user's number of users;

(2) quantize user's individual factors: liveness and influence power, with user, deliver the number percent quantification liveness that Review number and all users of choosing deliver the summation of Review number, adopt following formula:

$\mathrm{Activity}={\mathrm{\&mu;}}_{\mathrm{RW}}/\underset{\mathrm{\&mu;}\&Element;\mathrm{ProUser}}{\mathrm{\&Sigma;}{\mathrm{\&mu;}}_{\mathrm{RW}}}$

Wherein: μ _rWrepresent that user delivers the number of Review, ProUser represents the user that training set comprises;

According to Member Visits and two attributes of Total Visits, influence power is quantized, adopts following formula:

$\mathrm{MVP}={\mathrm{\&mu;}}_{\mathrm{MV}}/\underset{\mathrm{\&mu;}\&Element;\mathrm{ProUser}}{\mathrm{\&Sigma;}{\mathrm{\&mu;}}_{\mathrm{MV}}},\mathrm{TVP}={\mathrm{\&mu;}}_{\mathrm{TV}}/\underset{\mathrm{\&mu;}\&Element;\mathrm{ProUser}}{\mathrm{\&Sigma;}{\mathrm{\&mu;}}_{\mathrm{TV}}}$

Wherein: MVP represents the number percent of summation of user Member Visits value and all users' of choosing Member Visits value, μ _mVthe Member Visits value that represents user, TVP represents the number percent of summation of user Total Visits value and all users' of choosing Total Visits value, μ _tVthe Total Visits value that represents user, ProUser represents the user that training set comprises;

(3) quantize interbehavior between user: supporting rate, opposition rate, the Review Rating that user's supporting rate and opposition rate are delivered Rview according to user quantizes, and adopts following formula:

$\mathrm{Support}-\mathrm{rate}=\underset{s\&Element;\left\{\begin{array}{c}\mathrm{VeryHelpful},\\ \mathrm{MostHelpful}\end{array}\right\}}{\mathrm{\&Sigma;}{r}_s}/\left|R_w\right|$

$\mathrm{Oppose}-\mathrm{rate}=\underset{o\&Element;\left\{\begin{array}{c}\mathrm{OffTopic},\mathrm{NotHelpful},\\ \mathrm{SomewhatHelpful},\mathrm{Helpful}\end{array}\right\}}{\mathrm{\&Sigma;}{r}_o}/\left|R_w\right|$

Wherein: r _srepresent that Review Rating is VeryHelpful, the Review of MostHelpful, r _orepresent that Review Rating is OffTopic, NotHelpful, SomewhatHelpful, the Review of Helpful, | R _w| represent user's Review number.

4. according to the credible and insincere user identification method based on feedforward neural network claimed in claim 1, it is characterized in that, described structure feedforward neural network comprises the steps:

(1) according to the dimension of user characteristics vector, determine the input node number of feedforward neural network, have seven input node: Out-Link, In-Link, Activity, MVP, TVP, Support-rate and Oppose-rate;

(2) according to the complexity of trusted users identification, determine the number of plies of feedforward neural network and the nodes that each layer comprises.Have three layers: input layer, seven input nodes; Hidden layer, two concealed nodes; Output layer, an output node;

(3) according to structure and the performance requirement of feedforward neural network, determine that the type that hidden layer and output layer comprise node is sigmoid threshold cell.

5. according to the credible and insincere user identification method based on feedforward neural network claimed in claim 1, it is characterized in that, described training feedforward neural network comprises the steps:

(1) the k value according to k-fold cross validation algorithm (in this method, k gets 5) is divided into k subset by training set, and the common factor of any two subsets is empty;

(2) feedforward neural network is carried out to k training, choose a different subset as training set at every turn, k-1 remaining subset, as test set, will be noted " concussion " of accuracy of identification, deconditioning when accuracy of identification is stablized in training process;

(3) recording accuracy of identification and the frequency of training of k training, is that corresponding frequency of training is given weights according to accuracy of identification, calculates the weighted sum of k frequency of training as final frequency of training;

(4) the final frequency of training of determining according to step (3) is trained feedforward neural network on complete training set, obtains the weights of threshold value and the corresponding edge of respective nodes.

6. according to the credible and insincere user identification method based on feedforward neural network claimed in claim 1, it is characterized in that, the described feedforward neural network by after training is realized credible and insincere user's identification and is comprised the steps:

(1) in community network, obtain between user's to be identified community network topological structure, user's individual factors and user aspect three of interbehaviors information and be quantified as user characteristics vector, the form of user characteristics vector is as follows: (Out-Link, In-Link, Activity, MVP, TVP, Support-rate, Oppose-rate);

(2) user's to be identified proper vector is input to feedforward neural network, obtains user's to be identified output valve;

(3) if the difference of user's to be identified output valve and 1 is less than predetermined threshold value, user is identified as trusted users, otherwise, be insincere user.

Images