CN104463601A - Method for detecting users who score maliciously in online social media system - Google Patents

Abstract

The invention discloses a method for detecting users who score maliciously in an online social media system. The method for detecting users who score maliciously in the online social media system arms at scoring feedback. Firstly, clustering is conducted according to scores for products by users, and the normalized user confidence degree is calculated; secondly, the reliability degree of user scoring is calculated according to the user confidence degree to obtain a candidate list of the users who score maliciously; finally, candidate users who score maliciously are sorted in combination with the deviation degree of user scoring and product quality to obtain the final list of the users who score maliciously. The method has advantages in the aspects of calculation accuracy and efficiency and can be applied to large-scale online social media websites.

Description

A method for detecting malicious rating users in an online social media system

technical field

The invention relates to a method for detecting malicious evaluation users in an online social media system, in particular to a method for detecting malicious evaluation users in a social media system aimed at rating feedback.

Background technique

As the carrier of business, the Internet has become an indispensable tool for information collection, transmission and exchange. The advent of the information age has injected new vitality into the IT service industry based on the Internet. Among them, social media has attracted much attention and has been recognized as a new economic model and catalyst of the 21st world pattern, known as "sunrise industry and green industry". Social media is a new type of networked economic activity, which is developing rapidly at an unprecedented speed. It has become an effective means for the country to enhance its economic competitiveness and win the advantages of global resource allocation. Through social media, people are no longer face-to-face, looking at real goods, and relying on paper documents (including cash) for buying and selling transactions, but present a wide range of commodity information, perfect logistics distribution systems and convenient and safe fund settlement through the Internet. system to trade. There are tens of thousands of e-commerce companies and hundreds of millions of consumers in social media. How to establish an effective reputation evaluation mechanism, create an environment of orderly competition, and guide consumers reasonably is particularly important.

Most of the current reputation evaluation systems are based on information such as user reviews or ratings on products. Users comment on purchased products or perform satisfaction ratings to express users' opinions and satisfaction with a certain product. These reviews provide valuable information resources for manufacturers and potential consumers. By analyzing this information, manufacturers can keep abreast of the current market situation and consumer feedback, and potential consumers can also use this as an important reference for purchasing products. Whether a potential consumer decides to buy a product or not, the most direct and important references are often the ratings obtained by the product itself and the quality of the review content. For large social media trading platforms, most of the recommendation systems for recommending products to potential users are based on users' historical rating data and review content for products. If most of the reviews for an item are positive, then the user has a very high probability of buying the product; if most of the reviews are negative, the item will hardly be purchased. In reality, in order to increase their own interests, some unscrupulous merchants hire a group of people to make malicious comments on certain products, and the content of the reviews does not match the actual value of the products, or maliciously tout or maliciously slander. Malicious ratings and review information affect the reference value of review information, seriously mislead consumers' choices, weaken the significance of normal users' ratings and review information, and make consumers gradually lose trust in social media product evaluation systems, thereby endangering And ultimately hurt the entire social media industry. It can be seen that the authenticity and validity of the scoring data and comment information in the reputation evaluation system are of great significance to the benign competition of social media, and the importance of how to identify malicious scoring users in the reputation evaluation system is self-evident.

In order to detect users with cheating comments or malicious ratings, there are currently two main methods:

The first method is manual labeling. By artificially observing the user's ratings, comment content and other comment behaviors, it is judged whether the user is a cheating comment user. However, this detection method is highly subjective, and due to the large amount of data that needs to be processed, it is difficult for manual methods to be truly applied to the detection of malicious evaluation users in large-scale social media systems.

The second method is to use computer automatic identification. First mark the typical cheating comment users, and then classify the unmarked users through machine learning algorithms. There are two typical approaches. One is to judge the similarity of user comments in reviews with text comments, and the other is to calculate the degree of deviation between user ratings and product inherent quality.

For example, an article published on EPL in 2011 (A robust ranking algorithm to spamming. EPL, 94(2011), 48002.) proposed a correlation-based user reputation ranking algorithm to detect malicious scoring users. The algorithm mainly calculates the user reputation value and the product mean value through an iterative strategy, and finally detects malicious scoring users according to the ranking of user reputation. The essence of this algorithm is to use the user reputation to weight the product ratings to calculate the product quality. In essence, it detects the deviation between the user rating value and the inherent quality of the product. The greater the deviation, the greater the possibility of the user becoming a malicious rating user. . Although this method is simple, the inherent quality of the product itself is an immeasurable value, and different users' satisfaction with the same product varies from person to person. In general, the product quality is represented by the mean value of all the scores obtained by the product, which objectively has certain errors, which will lead to low detection accuracy. In addition, the algorithm shows good robustness when the proportion of malicious scoring users is particularly large, but it is not effective for the real scoring system where the proportion of malicious scoring users and the proportion of cheating users are small.

As another example, the 2012 WWW conference paper (Spotting Fake Reviewer Groups in Consumer Reviews. WWW'12, 2012, pp , 191-200.) proposed a method for detecting malicious scoring users based on the similarity of user review content. This method detects cheating comment users by analyzing the similarity of user comment text content. If the similarity between two comments is high, the user who posted comments in the past two days is more likely to become a cheating comment user. Although this method can effectively detect cheating commenters, it needs textual analysis of the comment content in the entire social media system, which requires a large amount of data and low processing efficiency; on the other hand, users in many social media systems do not actively participate Comments, and even participating comments are only short text, which makes analysis based on comment content unusable in many systems. The system based on scoring is currently available in most systems. Since the cost of user evaluation is not high, there are more users involved, and the method of discrimination based on comment text cannot be used in this type of system.

With the continuous development of social networks, the US patent US8176057 authorized on August 5, 2012 discloses a user reputation detection method based on social networks, which transmits the reputation value through the feedback of high reputation users, thereby detecting low reputation User. Although this method can effectively calculate the user reputation value, it is mainly used to identify users with high reputation, and the accuracy of user detection for malicious scoring is not high.

To sum up, the existing methods cannot meet the actual needs of most social media sites, or have biases in recognition accuracy, or cannot be efficiently applied to actual detection, or are not suitable for some evaluation systems.

Contents of the invention

The purpose of the present invention is to provide an effective method suitable for malicious scoring user detection in online social media systems. The present invention is aimed at a social media system with score feedback, and detects malicious score users by analyzing the user's score value, avoiding the large amount of calculation caused by analyzing and processing user comment text content, improving detection efficiency and high accuracy.

The technical solution adopted by the present invention to solve its technical problems is a method for detecting malicious scoring users in an online social media system, including the following steps:

Step 1: Extract the user rating data in the system, preprocess the data, and obtain standardized user rating data including user ID, product ID, and user ratings for products, and divide these three types of data into triplets (u,p , v) in the form of storage;

Step 2: clustering of user ratings, and calculating the confidence vector of user ratings;

Step 2-1: Clustering users who give the same rating for the same product into one group;

Step 2-2: Calculate the confidence vector of each user. Each component of the confidence vector represents the reputation value of the user for a product. The ratio of numbers, which is defined as the herd ratio;

Step 3: According to the user confidence vector calculated in step 2, calculate the reliability of user ratings, regard the least reliable N users as malicious rating users, and generate a list of malicious rating user candidates, where N is based on the user ratings of the actual system Factors such as ratio and detection accuracy are set;

Step 4: Reorder the malicious rating user candidate list according to the degree of deviation between the user rating and the inherent quality of the product in the malicious rating user candidate list, and select M users with the largest deviation to obtain the final malicious rating user, where M is based on the actual system Factors such as the proportion of user ratings and detection accuracy are set.

Among them, the specific steps of step 1 are:

Step 1-1: Remove users whose scoring times are lower than the threshold K, where the threshold K can be adjusted according to the system scoring situation and the fineness of specific detection;

Step 1-2: According to the principle of rounding, carry out integer discretization for scores whose scores are not integers;

Step 1-3: Store the user ID, product ID, and user rating data on the product in the form of triples (u, p, v).

In the step 1, usually the value of K is 8.

In the step 2, the dimensions of the confidence vectors of each user are inconsistent, and are stored in an xml file.

The concrete steps of described step 3 are:

Step 3-1: Calculate the average value and variance of each user's confidence vector, and divide the calculated average value by the size of the variance to obtain the user reliability;

Step 3-2: Arrange all users in ascending order of reliability, select the top N users, and generate a list of malicious scoring user candidates.

The concrete steps of described step 4 are:

Step 4-1: Calculate the average value of product ratings, which is regarded as the inherent quality of the product;

Step 4-2: Calculate the inherent quality deviation degree of each user for each product in the maliciously rated user candidate list obtained in step 3, that is, the difference between the user's rating of the product and the product's inherent quality;

Step 4-3: Calculate the absolute value of the inherent quality deviation degree of each user for each product, and then average them to obtain the user's rating deviation degree;

Step 4-4: Arrange the users in descending order according to the score deviation, select the top M users as the final malicious scoring users, and generate a list of malicious scoring users.

The invention detects based on user ratings. On the one hand, it saves the complicated process of processing texts, improves the detection efficiency, and is applicable to almost all evaluation systems. Performing secondary detection, this operation greatly improves the recognition accuracy of the present invention, especially in real evaluation systems where the ratio of user ratings to the total number of products is small and the number of malicious rating users is not large compared to the number of all users It works brilliantly.

Description of drawings

FIG. 1 is a flowchart of a method suitable for detecting malicious scoring users in a large-scale social media system provided by the present invention.

Fig. 2 is a flow chart of the process of generating user confidence vectors provided by the present invention.

Fig. 3 is a flowchart of user credibility calculation and malicious scoring user candidate list generation provided by the present invention.

Fig. 4 is a flow chart of reordering the malicious scoring user candidate list according to the degree of deviation between user ratings and product inherent quality in the malicious scoring user candidate list to obtain the final malicious scoring user provided by the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be described in further detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

The present invention will be described in detail below in conjunction with the accompanying drawings. It should be noted that the described examples are only intended to facilitate the understanding of the present invention, and have no limiting effect on it.

The overall process of the method for detecting malicious scoring users in social media based on scoring behavior clustering proposed by the present invention is shown in FIG. 1 .

Step 1 is the data preprocessing module. This module will preprocess the original data input by the system, filter the noise data, and perform discrete integerization on the scoring data. The preprocessed data is the input of the feature extraction operation in step S2.

Step 2 is the user confidence calculation module. This module performs score clustering on the data preprocessed in step S1, and calculates the user confidence vector according to the conformity ratio of the cluster size, and the user confidence will be used as the input data for the secondary feature extraction in step S3.

Step 3 is to calculate the reliability of user ratings and generate a malicious rating user candidate list module. Based on the user confidence vector, the module extracts the mean and variance of each user's confidence, calculates the ratio of the mean and variance as the user reliability, sorts the user reliability, and finally generates a list of malicious scoring user candidates.

Step 4 reorders the candidate list of malicious scoring users, and the final malicious scoring user generation module. This module will calculate the inherent quality of the product, and on the basis of the candidate list of malicious scoring users, use the degree of deviation between the user ratings and the inherent quality of the product to perform a second-stage detection on the preliminary detection results to generate the final malicious scoring user detection results.

The main steps are described in detail below:

1. Input the original evaluation data of the system, perform data preprocessing on the input data, and store the preprocessed results. (step 1).

The preprocessing work includes two main parts: noise data filtering and score value integerization. It is preferred to separate the user comment data from the input raw data, and filter the users whose comments are less than 8 times and the corresponding rating information. If the user's rating is not an integer, the user's rating will be rounded up based on the rounding principle. Since the noise data is the users with less ratings and the rating information, it has little effect on the whole system after being removed, but it can effectively improve the calculation efficiency. Through the discretization of scoring integers, the complexity of clustering calculations is reduced, and it is easier to apply to practical systems.

2. Clustering of user ratings to calculate the confidence vector of user ratings.

Step 2 mainly completes the calculation of the user confidence vector. The workflow is shown in Figure 2, including clustering of scoring behaviors, calculation of conformity ratio, and generation and storage of user confidence vectors.

The clustering of the rating behavior in step 2-1 is based on clustering the users who rated the same product into a group if the rating scores are the same. For each product in the system, user rating behavior clustering is performed. If the user has rated N products, then the user confidence is an N-dimensional vector, and each component is the reputation value obtained by the user after each rating. Since the scoring scores are discrete after preprocessing, a fixed number of groups are formed after clustering.

${\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; \{</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; r</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; Rate</span>}<span\; class="notranslate">\; \}</span>$

in The group formed by clustering the ratings of product O _j , r _{i, j} represents the rating of user U _i on product O _j . Rate is the integer discrete rating range of the product.

Step 2-2 is to calculate the ratio of the size of the group to which each user belongs to the total number of users who evaluate the product. The larger the ratio, the stronger the conformity. This ratio value reflects the deviation degree of user evaluation behavior from most people. If the user belongs to a small group, then the proportion value is small, and the user evaluation behavior deviates greatly from the public evaluation. On the contrary, if the user belongs to a larger group, it means that the comments are consistent with those of the vast majority of people, and the degree of deviation is small, which is trustworthy. The system uses the method of cluster size normalization to calculate the herd strength. Generate and store a user confidence vector. The user confidence is assigned according to the group to which the user belongs and the conformity ratio assigned by the group, and the degree of confidence represents the conformity strength of the group to which the user belongs. By giving the same confidence to users who rate the same product and are clustered into a group, the confidence of each user for each product that has been rated is obtained, and the user's confidence vector is generated. Finally, the confidence vector corresponding to the user is stored.

3. Calculate the reliability of user ratings and generate a candidate list of users with malicious ratings. (step S3)

Step 3 is to calculate user reliability based on the user confidence vector generated in step S2, and sort according to the size of user reliability, and take the top K percent users and add them to the malicious scoring user candidate list. The flow chart of step S3 is shown in FIG. 3 , including calculation of confidence average and variance (step S31 ), calculation of user reliability (step S32 ), generation and storage of user candidate sets with malicious scores.

In step 3-1, extract the mean and variance of all confidences for each user. The average value of the confidence reflects the average level of the user's reliability, and the variance of the confidence reflects the fluctuation degree of the user's reliability. The user credibility is the final credibility obtained by further calculation on the basis of the user confidence. Use the user's average reliability and reliability fluctuation to comprehensively generate the user's reliability. The specific calculation method is shown in the following formula:

${\mathrm{Score}}_i=\frac{{\mathrm{Rs.}}_i}{{\mathrm{PS}}_i},$ in ${\mathrm{Rs.}}_i=\frac{{\mathrm{\&Sigma;}}_{j\&Element;o_{i\&Element;u}}{\mathrm{rp}}_{i,j}}{\dim {\stackrel{\&Right\; Arrow;}{\mathrm{rp}}}_i},{\mathrm{PS}}_i=\frac{{\mathrm{\&Sigma;}}_{j\&Element;o_{i\&Element;u}}{({\mathrm{rp}}_{i,j}-{\mathrm{Rs.}}_i)}^2}{\dim {\mathrm{rp}}_i};$

Among them, Score _i is the reliability of user U _i , that is, the final trustworthiness of the user; Rs _i is the average value of user confidence, representing the average level of user reliability; Ps _i is the variance of user confidence, representing user reliability degree of volatility. When the average value is large and the variance is small, the user can stably obtain a high-score rating reputation every time. This type of user has high reliability and is trustworthy.

In step 3-2, generating the user candidate list with malicious rating is performed on the basis of the user reliability calculated in step S32. Sort the reliability of users in ascending order, and take the top K percent users in the list to join the user candidate set for malicious scoring to complete the preliminary detection. The sorting algorithm adopts a mature quick sorting algorithm, which does not belong to the content emphasized in the present invention. When the amount of data is large, the sorting algorithm can be well distributed and improve the sorting efficiency.

4. According to the degree of deviation between user ratings and product inherent quality in the malicious scoring user candidate list, the malicious scoring user candidate list is reordered to obtain the final malicious scoring user.

The overall process of step 4 is shown in Figure 4, which mainly includes extracting the inherent quality of the product (step S41), calculating the deviation degree of the rating (step S42), sorting the deviation between the user rating and the inherent quality of the product (step S43), and finally generating Malicious scoring user detection results.

In step 4-1, the inherent quality of the product is measured by the mean value of all ratings of the product. The inherent quality of the product itself is an immeasurable quantity, and some algorithms are usually used to estimate the inherent quality of the product. In the present invention, the arithmetic average of the scores obtained by calculating each product is used as the inherent quality of the product. The larger the average value, the better the quality of the product itself, otherwise, the worse the quality of the product itself.

The calculation of score deviation in step 4-2 is a process that can be calculated offline. Calculate the absolute value of the difference between the user's rating and the inherent quality of the product as the user's rating deviation.

In step 4-3, do the same process for all the products evaluated by the user to obtain the user's rating deviation vector, and finally calculate the average value of the rating deviation vector as the final user rating deviation. Similarly, do the same processing above for all users in the user candidate set with malicious ratings generated in step 3. The "user rating deviation" mentioned below refers to the value after the absolute value averaging of the user rating deviation.

In step 4-4, based on the degree of deviation of user ratings in step 4-2, the users in the user candidate set with malicious ratings are sorted in descending order. The higher the ranking, the greater the deviation of ratings, and the greater the possibility of becoming a user with malicious ratings , take the top K users to generate the final malicious scoring user list. The K value can be adjusted according to factors such as the user rating ratio and detection accuracy of the actual system. The final sorting result thus obtained is the malicious scoring user detection result.

The process of the execution of the present invention will be described below with a practical example

To simplify the description, in this example, there are a total of 10 users' original evaluations on 5 products in the social media website system, and the ratings range from 1 to 5, with 5 rating levels in total. As shown in Table 1, the row in Table 1 represents the user (U), the column represents the product (O), and the value in the corresponding cell is the user's rating of the product. If the cell is empty (-), it means that the user has not purchased passed this product, the rating is empty. In this way, the user-product rating matrix R in Table 1 is formed.

O1 O2 O3 O4 O5 U1 4 5 3 4 - U2 - 4 4 2 5 U3 3 4 - 5 3 U4 5 - - 4 3 U5 3 4 5 - 3 U6 2 4 3 5 3 U7 - 3 1 5 3

U8 1 - 3 3 4 U9 5 2 2 5 - U10 5 - 2 1 4

Table 1

To simplify the description, here only a specific implementation of clustering based on rating behavior is used as an example for illustration, wherein the clustering method is performed according to the specific description in step S21 to obtain clustered rating groups. Rows correspond to the 5 products in Table 1, and columns correspond to rating ranges from 1 to 5. The corresponding cells are groups formed by clustering based on rating behavior. A cell that is empty (-) indicates that no user has rated the corresponding score. The value indicates the size of the group after clustering, as shown in Table 2.

Table 2

The user clusters obtained for each product are normalized according to the cluster size to obtain user confidence, and the same confidence is given to users who score the same product and are clustered into a group. As shown in Table 3, the row represents 10 users, the column represents 5 products, and the cell is the confidence level of the user's evaluation of the product. If the cell is empty, it means that the user has not rated the product.

O1 O2 O3 O4 O5 U1 0.125 0.143 0.375 0.222 - U2 - 0.571 0.125 0.111 0.125

U3 0.250 0.571 - 0.444 0.625 U4 0.375 - - 0.222 0.625 U5 0.250 0.571 0.125 - 0.625 U6 0.125 0.571 0.375 0.444 0.625 U7 - 0.143 0.125 0.444 0.625 U8 0.125 - 0.375 0.111 0.250 U9 0.375 0.143 0.250 0.444 - U10 0.375 - 0.250 0.111 0.250

table 3

Calculate the Gaussian distribution statistical features of each user's confidence vector according to step S31 to obtain the mean value and variance. The final reliability of the user is obtained by using the average value ratio, as shown in Table 4, the row represents the user, and the column represents the average value, variance, and final reliability of the user's confidence.

average value variance Reliability u1 0.2163 0.1139 1.899 u2 0.2330 0.2254 1.034 u3 0.4725 0.1666 2.836 u4 0.4073 0.2034 2.002 u5 0.3927 0.2434 1.613 u6 0.4280 0.1963 2.180 u7 0.3342 0.2429 1.376 u8 0.2153 0.1235 1.743 u9 0.3030 0.1335 2.269 u10 0.2465 0.1079 2.285

Table 4

Sort in ascending order according to the user reliability calculated in Table 4, and the result is: u2, u7, u5, u8, u1, u4, u6, u9, u10, u3. The top 40% users in the list are added to the user candidate set with malicious rating, and the user candidate set with malicious rating is obtained: {u2,u7,u5,u8}.

According to step S41, the average value of all ratings obtained by the product is calculated as the inherent quality of the product, and the user candidate set with malicious ratings is reordered according to the degree of deviation between the user ratings and the inherent quality of the product. As shown in Table 5, rows represent users and columns represent products. There are a total of 4 users and 5 products, and the corresponding cell is the degree of deviation between the user rating and the inherent quality of the product.

table 5

Arrange the average deviation degree calculated in Table 5 in descending order, and the obtained list of users with malicious ratings is U2, U7, U8, and U5. Compared with the scoring list in the malicious scoring candidate set, the U8 score deviates more than U5, and it is easier to become a malicious scoring user. At this point, the detection is completed, and the list of malicious scoring users is obtained. The higher the ranking, the greater the possibility of becoming a malicious scoring user.

Although the illustrative embodiments of the present invention have been described above, it is for those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes It is obvious that all inventions and creations using the concept of the present invention are included in the protection.

Claims (6)

1. in online Social Media system detection of malicious scoring user a method, the method comprises:

Step 1: the user's score data in extraction system, pre-service is carried out to data, obtains normalized user's score data and comprise by user ID, product IDs, user to the scoring of product, by these three classes data according to tlv triple (u, p, v) form store;

Step 2: user marks cluster, calculates the degree of confidence vector of user's scoring;

Step 2-1: be one group by the user clustering giving identical scoring for same product;

Step 2-2: the degree of confidence vector calculating every user, this user of each representation in components of this degree of confidence vector is to a kind of credit value of product, and this credit value is for user is for the ratio value of comforming of phylogenetic group size belonging to this product and all evaluation numbers of users;

Step 3: according to the user's degree of confidence vector calculated in step 2, calculate the fiduciary level of user's scoring, be considered as least reliable N number of user maliciously to mark user, generate malice scoring user candidate list, wherein N to mark ratio and detect the factors such as degree of accuracy and set according to the user of real system;

Step 4: with the departure degree of product proper mass, user's candidate list of maliciously marking is resequenced according to user's scoring in malice scoring user candidate list, choose the maximum M of a departure degree user, obtain final malice scoring user, wherein M to mark ratio and detect the factors such as degree of accuracy and set according to the user of real system.

2. in a kind of online Social Media system as claimed in claim 1 detection of malicious scoring user method, it is characterized in that the concrete steps of step 1 are:

Step 1-1: remove the user of scoring number of times lower than threshold k, wherein threshold k can regulate according to the situation of system scoring and the concrete fine degree detected;

Step 1-2: according to the principle rounded up, to mark not for integer discretize is carried out in the scoring of integer;

Step 1-3: by user ID, product IDs, user store the form of the score data of product according to tlv triple (u, p, v).

3. in a kind of online Social Media system as claimed in claim 2 detection of malicious scoring user method, it is characterized in that in described step 1-1, usual K value is 8.

4. the method for detection of malicious scoring user in a kind of online Social Media system as claimed in claim 1, is characterized in that the degree of confidence vector dimension of each user in described step 2 is inconsistent, adopts xml file to store.

5. in a kind of online Social Media system as claimed in claim 1 detection of malicious scoring user method, it is characterized in that the concrete steps of described step 3 are:

Step 3-1: the mean value and the variance that calculate every user's degree of confidence vector, at calculating mean value divided by square extent, obtain user's fiduciary level;

Step 3-2: by all users according to the arrangement of fiduciary level size ascending order, choose top n user, generates malice scoring user candidate list.

6. in a kind of online Social Media system as claimed in claim 1 detection of malicious scoring user method, it is characterized in that the concrete steps of described step 4 are:

Step 4-1: calculate a mean value of product scoring, this mean value is considered as the proper mass of product;

Step 4-2: in malice that calculation procedure 3 obtains scoring user candidate list, each user is for the proper mass irrelevance of each product, namely user is to the difference of the scoring of product and this product proper mass;

Step 4-3: calculate the proper mass irrelevance absolute value of each user to each product, then it is averaging, obtain the scoring irrelevance of this user;

Step 4-4: each user is carried out descending sort according to irrelevance of marking, chooses a front M user for the final user that maliciously marks, generates user list of maliciously marking.