WO2018148894A1

WO2018148894A1 - Time-dependent reputation evaluation algorithm based on scoring network

Info

Publication number: WO2018148894A1
Application number: PCT/CN2017/073722
Authority: WO
Inventors: 廖好; 黄泽成; 毛一帆; 陆克中
Original assignee: 深圳大学
Priority date: 2017-02-16
Filing date: 2017-02-16
Publication date: 2018-08-23

Abstract

Disclosed is a time-dependent reputation evaluation algorithm based on a scoring network, wherein the following steps are comprised: S1: respectively initialising a first behaviour weight factor, a second behaviour weight factor, a user reputation and an object quality; S2: determining whether a quality variation is less than a set value, if so, obtaining an object quality Qi; and if not, proceeding to step S3; S3: successively obtaining a reputation accumulated value, a quality accumulated value, and a temporary reputation value of a user; and S4: after a reputation value of the user and the object quality are updated according to the reputation accumulated value, the quality accumulated value and the temporary reputation value of the user, returning to step S2. In the present invention, the two behaviour weight factors and two component accumulation processes are designed from the point of view of time and degree, a specific transaction process can be reflected to a certain degree, malicious behaviour can be effectively restrained, and the processes by which each object quality and user reputation change are directly refined. Compared with other algorithms, the algorithm achieves a significantly improved effect and is valuable for practical applications.

Description

A Time-based Reputation Evaluation Algorithm Based on Scoring Network

Technical field

The invention belongs to the field of Internet, and more specifically relates to a time-based reputation evaluation algorithm based on a scoring network.

Background technique

The prior art proposes a dynamic reputation calculation model that combines subjective and objective elements. The model mainly models the data provider's behavior, motivation, hobbies, attitudes, etc. into subjective elements of reputation calculation, and provides data provided by the data provider. Quality is modeled as an objective element of reputation calculation, and subjective and objective elements are integrated in the time domain. Adali et al. established a corresponding behavioral reputation calculation model based on the communication behavior of users in social networks: comprehensive session trust and communication trust to calculate behavior credibility. The premise of this model is that regular communication behavior has more random communication behavior than random communication behavior. Higher credibility. The session trust is used to indicate the session time and session frequency between users. The higher the frequency and the longer the time, the higher the trust between them. The communication trust refers to the degree of information dissemination, and the information is transmitted from one user to other users. The more you indicate the higher the trust in the information source user.

The prior art also proposes a priority-based reputation calculation model, which mainly constructs the service provider's reputation from four aspects: service requester, service provider's service experience and service provider's priority distribution of service quality attributes. The similarity, the suitability of the candidate service to the service demand, and the evaluation information provided by the third party, the four attributes together determine the final reputation value.

The prior art also proposes a dynamic trust computing model and reputation system for network transactions based on multi-agent system. The model includes the trust degree of the user's time domain, the weighted average calculation of the reputation feedback score, and the weighting of the community contribution. It establishes a pre-existing prevention, coordination in the matter, and a penalty-binding mechanism that punishes the Trinity afterwards.

The prior art also proposes to calculate user attitudes by interactive experience and to calculate using interactive time series. The method of user behavior pattern change, and then use the three kinds of information of attitude, interaction experience and behavior pattern to build a trust evaluation model. High quality and personalized recommendations are a key feature of many online systems in the current Internet environment. To help users find the right products or services from a large number of products or services, avoiding false or disappointing services is a fundamental research issue in the design of network service systems. In order to achieve this goal, researchers have proposed many methods.

One of the most representative methods in the prior art is called iterative refinement (abbreviated as IR) algorithm. In the IR algorithm, the user's reputation is inversely proportional to the difference between his scoring vector and the estimated quality vector of the object. Estimating the quality of the object and evaluating the user's reputation is iteratively updated until they become stable. Based on the IR algorithm, optimizing this iterative algorithm by creating a reputation for each individual score forms a new algorithm. An improved iterative algorithm (for CR) is mainly to calculate the user's reputation by the Pearson correlation formula between the user's score and the estimated quality of the object. This method is said to be very effective against spam malicious behavior. The prior art also proposes an iterative algorithm (referred to as IAAR) to deal with malicious behavior of users, mainly by adopting a reputation redistribution process to improve the influence of well-known users and two penalty factors to maliciously malicious users.

In the existing iterative algorithm, whether it is newly proposed or optimized based on the existing algorithm, most of the problems are caused by the application context information mining, which leads to the inaccurate evaluation of the user reputation, which leads to the failure of the related quality ranking algorithm.

Summary of the invention

Aiming at the defects of the prior art, the object of the present invention is to provide a time-based reputation evaluation algorithm based on a scoring network, which aims to solve the problem that the relevant quality ranking algorithm fails due to inaccurate evaluation of user reputation due to uncertain factors in the social network. .

The invention provides a time-based reputation evaluation method based on a scoring network, comprising the following steps:

S1: initializing the first behavior weighting factor, the second behavior weighting factor, the user reputation, and the object quality respectively;

S2: determining whether the amount of quality change is less than a set value, and if so, obtaining an object quality Qi; if not, Then proceeds to step S3;

S3: sequentially obtaining the credit accumulation value, the quality accumulation value, and the temporary credit value of the user;

S4: Update the user reputation value and the object quality according to the reputation accumulation value, the quality accumulation value, and the temporary credit value of the user, and then return to step S2.

Further, in step S1, the formula is passed

Initialize the user behavior weighting factor; pass the formula

Initialize the object behavior weighting factor; pass the formula

Initialize user reputation; pass the formula

Initializing the quality of the object; wherein wu _ij is a user behavior weighting factor for a certain period of time, i is a serial number identifier, t is a specific time period identifier, and UU _j is a collection of users in a certain period of time, T _j is the set of objects in a certain period of time in a total number of records, V _t is a collection of objects of a single user in a certain period of time, UT _ix is a collection of users in a certain period of time; wo _ij is an object at a certain time Within the weight, i is a serial number identifier, t is a specific time period identifier, OT _ix is the set of objects in a certain period of time, OO _j is the set of objects in a certain period of time, T _j is the total number of records The set of objects in a certain period of time, V _t is the set of objects of a single user in a certain period of time, OT _ix is the set of objects in a certain period of time; R _i is the reputation value of a certain user, x is a The serial number identifier, O _ij is the set of objects selected by the user in a certain period of time, wu _ij is the weight of the user at a certain time, r _ix is the score given by the user to an object, and ku _i is the degree of a certain user; Q _i is the mass of an object, i is a serial number identifier U _ij for the selection of an object in a certain period of time the user set, wo _ij rights object in a certain period of weight, R _i is a value of a user's credibility, r _ix to an object for the user rating.

Further, in step S2, the mass change amount

Where |QQ'| is a condition variable, which is used to end the algorithm operation, Q ₁ is the quality of an object, Q _l ' is the quality of other objects, and acco _i is the mass accumulation process component value of an object, l is A serial number identifier, O _total is the total number of objects, and o is the object.

Further, the set value Δ = 10 ^-4 .

Further, in step S3, the two component accumulation processes

with

Reputation value

Among them, accu _i is the reputation value of a certain user credit accumulation process, ku _i is the degree of a certain user, ko _k is the degree of an object, k is a serial number identifier, and O _ij is the user within a certain period of time. The selected object set, r _ix is the user's rating for an object; Q _i is the quality of an object, R _i is the reputation value of a certain user; acco _i is the value of an object quality accumulation process, ku _i is The degree of a user, ko _k is the degree of an object, k is a serial number identifier, U _ij is the set of users who select an object in a certain period of time, Q _i is the quality of an object, R _i For a user's reputation value; accu _i is the reputation value of a user's reputation accumulation process, Q _i is the quality of an object, x is a serial number identifier, and O _ij is the object selected by the user within a certain period of time. The set, wo _ij is the weight of the object in a certain period of time, r _ix is the score given by the user to an object, and wu _ij is the weight of the user at a certain time.

The average of the scoring vectors for a user, Dri is the standard deviation of a user's scoring vector, and D _Qi is the standard deviation of an object's mass vector.

The average of the quality of an object.

Further, in step S4, the formula is passed

Update reputation; pass formula

Make the object quality better.

According to the above technical solution conceived by the present invention, compared with the prior art, since the reputation of the user can be effectively predicted, and the malicious user behavior can be effectively restricted, and an iterative algorithm of two scoring time factors can be combined, The degree has curbed the user's malicious scoring behavior and greatly enhanced the robustness. The two behavior weighting factors of the time segmentation can quantify the user's weight value in a time manner, and the comprehensive calculation of the weighting values of the multiple time periods is more in line with the actual situation. By designing two accumulation processes from both degrees of time and time, the quality of the object and the reputation of the user can be effectively improved.

DRAWINGS

FIG. 1 is a flowchart of implementing a time-based reputation evaluation algorithm based on a scoring network according to an embodiment of the present invention.

detailed description

The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

The time-based reputation evaluation algorithm based on the scoring network provided by the embodiment of the present invention can be applied to multiple fields, for example, applied to the scientific research field, and can be used as a comparison algorithm, and the application mode is program implementation and importing the pseudo-detection data set operation. Applied to the Internet field, it can be applied to national products, thesis system, e-commerce, etc. The application method is to implement and import the data to be tested. It can be applied to industrial fields and can be applied to network intrusion detection, credit card fraud detection, traffic flow detection, etc. The application method is to implement and import the data to be tested. Applied to the field of life, it can be applied to personal reputation assessment, competition ranking, etc. The application method is to programmatically implement and import the data to be tested.

The time-based reputation evaluation algorithm based on the scoring network provided by the embodiment of the present invention mainly includes: the two behavior weight factors of the time segmentation can quantize the weight value of the user in a time manner, and the weight values of the plurality of time periods are The comprehensive calculation is more in line with the actual situation; designing two accumulation processes from the perspective of time and time can effectively improve the quality of the object and the credibility of the user.

details as follows:

Two different behavioral weighting factors:

A user will behave differently at different times. To reflect the weight of this behavior under time, you can use the following formula:

An object will reflect different qualities at different times. In order to reflect the weight value of this quality under the time factor, it can be set by the following formula:

The reputation of user i is represented by R _i , and the initial value of each user reputation is set by the following formula:

The quality of the object a is represented by Q _a , and the initial value of each object quality can be set by the following formula:

From equations (1)(2)(3)(4), the magnitude of a user's behavioral weight value factor depends on the number of objects he sees over a period of time. In this article, the number of movies viewed is referred to. The behavior weighting factor of an object depends on the frequency it is involved in by the user for a period of time. The higher the frequency, the greater the weight value during this time. The reputation value of a user is no longer simply initialized by the average value, but depends on the user's reputation, the weight of the user's behavior, and the degree of the user. This can better reflect the influence of multiple factors, and the aspect is also consistent with the actual situation.

In the process of updating, in order to calculate the reputation value of user i, we designed a credit accumulation process based on the perspective of time and degree, as follows:

Since the credibility of a user is not established once, but is gradually accumulated, we analyze and calculate the information related to this user. From the formula, we can see the credit component of a user and the quality of the object that the user is watching, the degree of the object, the score of the object, the overall reputation of the user, and the degree of the user. Present a certain proportional relationship. In addition, this accumulation process can effectively deal with malicious scoring, because a malicious comment has little effect on the overall reputation. As the user involves more and more objects, the value of the reputation component will become larger and larger, which means it The credibility is getting better and better.

Similarly, in the process of updating, in order to calculate the quality value of the object, we also designed a quality accumulation process from the perspective of time and degree, as follows:

The quality value of an object can not be tested at one time, but through many user behaviors, so we comprehensively analyze the relevant information and do the calculation. From the above formula, we can know the mass component of an object. It involves a certain relationship between the user reputation, the quality of the object, the degree of the object, and the degree of the user. In fact, this quality accumulation process can also effectively deal with poor reputation users and poor quality objects, because in this accumulation process, a single value does not have a great impact on the quality of the entire object. As an object is more and more acted by different users, its corresponding mass component value will become higher and higher, which means that its quality has been recognized by everyone, and it is indeed a high-quality object.

In addition, R _i is updated during the iteration process, so in a certain update process, in order to calculate the reputation value of user i, the improved Pearson correlation between the user score vector and the corresponding object quality vector is calculated by integrating the two behavior weighting factors. The coefficient is used as the temporary credit value of the user as follows:

If TR _{i is} less than or equal to 0, the temporary credit value of user i is set to 1 so that the interval of the temporary credit value TR _i is [0, 1]. In addition, the user's reputation is inversely proportional to the average squared error of his scoring vector and the weighted average scoring vector of the corresponding object. Reputation based on Pearson correlation analysis has proven to be more powerful in responding to spam scores than in the above methods, and is able to more accurately assess the quality of objects.

Considering that the user's reputation is affected by the user's degree, the reputation value is balanced. In a certain iteration process, the ratio of the temporary credit value to the user credit accumulation process component value is taken as the final reputation of the user after this iteration, as follows:

Through the above formula, we know that the user reputation accumulation process is similar to the well-known K-nearest neighbor algorithm, completely eliminating the noise outside the user's degree, and this effect is accumulated in each update iteration process, and finally in the evaluation target quality accuracy. Lieutenant will get a big improvement, although the method of modification seems to be small, but the improvement is huge.

Further in an iterative process, the object Q _i will change, to update the value of the mass of the object during each iteration by the following equation:

The user's reputation and the quality of the object will be updated at every step. When the mass change is as small as a minimum Δ (Δ=10 ^{-4 in} this paper), the algorithm stops;

In the embodiment of the present invention, the entire algorithm can be developed based on a large number of languages, such as spark, hadoop, and the like.

The invention designs two behavior weighting factors and two component accumulation processes from the perspective of time and degree to reflect the specific transaction process to a certain extent, can effectively curb malicious behavior, and directly refine the changes of each object quality and user reputation. Process, compared with other algorithms, the effect achieved by this algorithm has been greatly improved, which can bring value to practical applications.

Those skilled in the art will appreciate that the above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and scope of the present invention, All should be included in the scope of protection of the present invention.

Claims

A time-based reputation evaluation algorithm based on a scoring network, comprising the following steps:

S1: initializing the first behavior weighting factor, the second behavior weighting factor, the user reputation, and the object quality respectively;

S2: determining whether the mass change amount is less than the set value, and if so, obtaining the object quality Q i ; if not, proceeding to step S3;

S3: sequentially obtaining the credit accumulation value, the quality accumulation value, and the temporary credit value of the user;

S4: Update the user reputation value and the object quality according to the reputation accumulation value, the quality accumulation value, and the temporary credit value of the user, and then return to step S2.
The method of claim 1 wherein in step S1, the formula is passed
Initialize the user behavior weighting factor; pass the formula
Initialize the object behavior weighting factor; pass the formula
Initialize user reputation; pass the formula
Initializing the quality of the object;

Where wu ij is the user behavior weighting factor for a certain period of time, i is a serial number identifier, t is a specific time period identifier, UU j is a set of users in a certain period of time, and T j is the total number of records. The set of objects in a certain period of time, V t is the set of objects of a single user in a certain period of time, UT ix is the set of the user in a certain period of time; wo ij is the weight of the object in a certain period of time, i is a serial number The identifier, t is a specific time period identifier, OT ix is a collection of objects in a certain period of time, OO j is a collection of objects in a certain period of time, and T j is an object in a total number of records in a certain period of time Set, V t is a collection of objects of a single user in a certain period of time, OT ix is a collection of objects in a certain period of time; R i is a reputation value of a certain user, x is a serial number identifier, and O ij is a user The set of objects selected in a certain period of time, wu ij is the weight of the user at a certain time, r ix is the score given by the user to an object, ku i is the degree of a certain user; Q i is the quality of an object , i is a serial number identifier, U ij is to select a certain time period The user collection of objects, wo ij is the weight of the object in a certain period of time, R i is the reputation value of a certain user, and r ix is the rating of the user to an object.
The method according to claim 1 or 2, wherein in step S2, said mass change amount

Where |QQ'| is a condition variable, which is used to end the operation of the algorithm, Q 1 is the quality of an object, Q l ' is the quality of other objects, and acco i is the mass accumulation process component value of an object, l is A serial number identifier, O total is the total number of objects, and o is the total number of objects.
The method of claim 3 wherein said set value Δ = 10 -4 .
The method according to claim 1 or 2, wherein in step S3, the two component accumulation processes
with
Reputation value

Where ku i is the degree of a certain user, ko k is the degree of an object, k is a serial number identifier, O ij is the set of objects selected by the user in a certain period of time, and r ix is the user giving an object The score is Q i is the quality of an object, R i is the reputation value of a certain user; acco i is the value of the process of accumulating the quality of an object, ku i is the degree of a certain user, and ko k is the object Degree, k is a serial number identifier, U ij is the set of users who select an object in a certain time period, Q i is the quality of an object, R i is the reputation value of a certain user; accu i is some The reputation value of the user reputation accumulation process, Q i is the quality of an object, x is a serial number identifier, O ij is the set of objects selected by the user in a certain period of time, and wo ij is the weight of the object in a certain period of time. , r ix is the user's rating for an object, wu ij is the user's weight for a certain period of time,
The average of the scoring vectors for a user, D ri is the standard deviation of a user's scoring vector, and D Qi is the standard deviation of an object's mass vector.
The average of the quality of an object.
The method according to claim 1 or 2, wherein in step S4, the formula is passed
Update reputation; pass formula
Make the object quality better.