TWI484431B - Multi - source heterogeneous network data community analysis method - Google Patents

Description

Multi-source heterogeneous network data community analysis method

The invention relates to a method for identifying a user through a plurality of characteristic attributes in different networks to generate a household probability, and analyzing a multi-source heterogeneous network data community analysis method for users with cross-network influence.

The development and application of social networking is a new application of new forms of development in recent years. It is still in the initial stage of the market. However, the development and application of social networks is not only a new business model, but also a replacement for the traditional The communication required for social activities, so in the future, only more industries will invest in the development of social networks and social network analysis applications.

US Patent No. 20120284384 A1 "Computer processing method and system for Network data" is a technical feature of generating a sub-community based on a screening relationship and reusing any grouping algorithm, but its data source is limited to single source network data analysis, and cannot be targeted. Multiple data source analysis, and the value of the patent in the relationship screening is only the number of links that users have in a single network, and it is also impossible to filter in a multi-network environment.

The main purpose of the present invention is to provide a multi-source heterogeneous network data community analysis method for identifying users in a heterogeneous network with the same user probability in a multi-source heterogeneous network.

A secondary object of the present invention is to provide an analysis that has a more frequent and tight connection A dense mashup community cluster and a multi-source heterogeneous network data community analysis method for users with cross-network influence

To achieve the above objective, the multi-source heterogeneous network data community analysis method of the present invention obtains the user's network behavior and characteristic attributes from heterogeneous network data, and identifies different networks by comparing network behaviors and feature attributes. Whether the users in the road are the same user, and the entrants are generated to generate the affiliation, and then the connection is confirmed as the same user, and the candidate relationship is established with the candidate relationship, and each of the individual networks is connected. After the user relationship connection is normalized, it is compared with the relationship connection to obtain the relationship weight, and then grouped by the social network grouping algorithm to generate a plurality of mixed community clusters, and the mixed community according to the probability of returning the household. In the cluster, the same user has the highest chance of returning to the household, and then the heterogeneous network number/total heterogeneous network number in the mixed community clusters is generated, and the heterogeneous weight is generated, and then the social network analysis method is used. (Because there are many evaluation indicators, the connection degree is only one of them, so delete) The evaluation indicators calculate the influence status within the community to identify important images within the community. Power users.

100‧‧‧ Calculate the probability of multiple sources of information

200‧‧‧Creating a virtual concept network

300‧‧‧Computational relationship weights

400‧‧‧ screening relationship

500‧‧‧Generating a mixed community cluster

600‧‧‧Home-to-house heterogeneous network relationship

700‧‧‧ Calculate heterogeneous weights

800‧‧‧Analysis of user influence

101‧‧‧ Obtain the network behavior and feature attributes of each user and identification function from multiple heterogeneous network data, and give the characteristics attribute weight

102‧‧‧ will obtain the feature attributes to compare the users with the same feature attributes, and will identify the unique users

103‧‧‧Deduct the unique users identified, set the unrecognized users as candidate homecoming combinations, and use the unidentified repetitive feature attributes or similar network behaviors to obtain the household probability

104‧‧‧The probability of returning to the household is less than the threshold of the household registration, and the household registration action is not carried out, and the household registration rate is higher than the household registration threshold, then the household registration action is performed.

1 is a simplified flow chart of the present invention; FIG. 2 is a flow chart of calculating the probability of multi-source source attribution according to the present invention.

Referring to FIG. 1 , the multi-source heterogeneous network data community analysis method of the present invention is an analysis method applied to a media program, and the step process includes the steps 1. calculating a multi-source source attribution probability 100; and step 2: establishing a virtual Concept network 200; step 3, computing relationship weight 300; step 4, screening relationship 400; step 5, generating mixed community cluster 500; step 6, home heterogeneous network relationship 600; step 7, calculating heterogeneous weight 700; Analyze user impact 800.

Please refer to FIG. 2, step 1. Calculate the multi-source source attribution probability 100, which means that the multi-source heterogeneous network data includes the same user using the same or different accounts to log in to the heterogeneous network, and the process steps include: 1-1. Obtaining the network behavior and feature attributes of each user and the identification function from the plurality of heterogeneous network data, and assigning the characteristic attribute proportion 101; through each user's characteristic attribute and network behavior in the heterogeneous network, Identify the probability that the user in the heterogeneous network is the same user, each feature attribute has a specific gravity, and the feature attribute is an attribute that can be used to describe the user feature, such as name, license number, e-mail address, mobile phone number, residence The place, the place of work, the graduate school, the interest, ..., etc., can be obtained by using the information provided by the user on each website, and the network behavior refers to the repeated common friends, the same content of the same place at the same time, etc., etc. Can be achieved by techniques such as web crawling; Steps 1-2, the feature attributes are acquired to be compared, and the users with the same feature attributes are compared, and The identified unique user is logged in to the household 102; Step 1-3, deducting the identified unique user, setting the unrecognized user as the candidate homing combination, using the unrecognized duplicated feature attribute or similar The network behavior is determined by the probability of returning home 103; Steps 1-4, setting a threshold value of the household account is used to determine whether the boundary of the home registration action is performed. The probability of returning home is between 0.0 and 1.0, and the threshold for returning home is between 0.5 and 1.0. If the user whose return rate is less than the threshold value of the household, the user who does not perform the household registration operation and the household registration rate is higher than the household registration threshold, the household registration action 104 is performed, and the household income rate table is generated, and the household income rate table is Its content describes the probability of user homing between heterogeneous networks.

Steps 1-2 and 1-3 generate multi-source source attribution probability through direct identification and indirect probability identification; direct identification is to identify a unique user in the common feature attribute between heterogeneous networks, and to return to the household. The user probability is 1, the feature attribute can be a single attribute or a composite attribute, and the indirect identification is to deduct the identified users, and to classify the unrecognized users into candidate homecoming combinations, using the repeated feature attributes of the users or The similar network behavior ratio is used to obtain the probability of returning to the household; finally, the degree of approximation between the individual characteristic attributes is calculated, and the user with the same characteristic attribute is the ratio of the household ownership rate of 1, and multiplied by the attribute weight of each characteristic attribute. Partial similarity can obtain the similar probability of household registration, thereby calculating the difference probability value, and finally summing the similar probability values of each attribute to obtain the attribution probability of each candidate homing combination.

The virtual concept network 200 is established, which calculates the probability of each user's attribution based on the calculation of the multi-source source attribution probability 100, and converts the multi-community network connection into an equivalent virtual concept, and the return-to-home ratio is expressed as one. 100% confirmed users, confirmed users need to use a unified registrar code to replace the relationship in the original network, and establish a relationship connection, and the probability of returning to the household is not one, the user expressed as a candidate relationship, can retain the original The registrar code, but the candidate relationship connection must be established. The relationship between the user and other users in the original network needs to be transferred to another candidate user to establish a heterogeneous network connection, and finally a relationship structure table is generated, and the content record is The connection between users.

The calculation relationship weight 300 is used to determine the strength of the relationship between users. The strength of the relationship between users is determined according to the degree of interaction of the users in the network. When multiple sources are used, the size of each network must be considered separately to avoid excessive network influence. Diluting the influence of small networks and avoiding excessive amplification of small network influences, so each user relationship must be normalized by first dividing the relationship strength of each user relationship in each network. With the maximum relationship strength in each network, the strength of each relationship is between 0 and 1, and then the relationship with the virtual concept network 200 is established. The line is compared. If the relationship is transferred, the relationship strength of the relationship is multiplied by the probability of returning to obtain the relationship weight. If the user connection relationship is repeated, the same home connection is added to the average. In order to obtain the relationship weight, then calculate all the relationship weights in order, and finally get a relationship weight table, the content of which records the relationship weight between each household and the related network.

The relationship 400 is selected according to the relationship weight table generated by the calculation relationship weight 300. Each relationship strength is between 0.0 and 1.0, and then a relationship threshold is used to set the relationship threshold value to determine whether to retain the relationship connection. The relationship threshold is between 0.5 and 1.0. The relationship between the relationship weight and the relationship threshold is deleted. The relationship weight is higher than the relationship threshold. The threshold value can be directly filtered by an absolute value, or all relationship weights can be used. Perform a power-down sort, taking the previous percentage value to make the threshold between 0.5 and 1.0.

A hybrid (Mashup) community cluster 500 is generated. According to the screening relationship 400, a relationship with a weight greater than 0.5 may cover multiple networks, and then grouping by social network grouping algorithm to find out that the contact is frequent and close. Mixed community clusters. Those skilled in the art can use the social network grouping algorithms such as n- clique, n- clan, and k- core.

The home heterogeneous network relationship 600, according to the mixed community (Mashup) cluster 500 generated in the mixed community cluster may contain the same user only to establish relationships in different networks, in order to avoid users in the sub-community are the same user To create a meaningless mix of mashups, it is necessary to categorize users who may be the same user in the same mashup community cluster according to the registrar probability table.

Calculate the heterogeneous weight of 700, after getting the community cluster after the home, because the user and connection relationship in each mixed (Mashup) community cluster may come from the hybrid network, across the network The impact is smaller than the same network. Therefore, each relationship weight must be recalculated for the Mashup community cluster. The calculation method is the number of heterogeneous networks/total heterogeneity in the mixed community clusters. The number of networks, you can generate heterogeneous weights, and get a heterogeneous weight table that records the user relationships and heterogeneous weights within each mashup community cluster.

The user influence 800 is analyzed, and the heterogeneous weight table is generated according to the homing result of the home heterogeneous network relationship 600 and the calculated heterogeneous weight 700, and the heterogeneous weight of each user relationship is used by using the evaluation index of the social network analysis method. After the addition, multiply the probability of returning to the household to generate an influence score to identify important influence users in the community. The evaluation indicators of the social network analysis method include the degree of association and the degree of continuity. , eigenvector, shortest path, etc.

In the following, an embodiment of the present invention provides a detailed description of the multi-source heterogeneous network data community analysis method: assuming that an e-commerce company wants to promote a new product, the company has obtained heterogeneous network data from three networks, so the company The present invention can be used in a heterogeneous network of materials to identify a mixed community with strong relationships and cross-network influences in a heterogeneous network, and then launch a series of word-of-mouth marketing activities.

Step 1: Calculate the probability of multiple data source attributions. Tables 1, 2, and 3 are the heterogeneous network data obtained by the company. First, the characteristic attributes in each network have an attribute weight, among which, unique identification The characteristic attributes of the function (such as e-mail address and mobile phone number) have a property weight of 1, and the attribute weight of other feature attributes is less than 1. Secondly, two heterogeneous networks are selected to calculate the household. First, the direct identification is used. It can identify the user's unique feature attribute and assign it to the household, that is, the attribute weight is one, and it is known that (N1, A) and (N2, M) are the same user, and the attribution rate is 1, and then the indirect probability identification is performed. The directly identified users, the remaining users are combined into a candidate homing combination, and the candidate homing combinations generated by the D users of N1 are [(N1.D)(N2.J)], [(N1.D)(N2 .K)], [(N1.D)(N2.I)], [(N1.D)(N2.L)], according to the characteristic attribute computer rate value in each candidate user, the N1 and N2 repetition attributes have 4 And each attribute has a different proportion. If you have the same attribute, you will get the attribute weight and then divide the number of duplicate attributes. After calculation, you get: [(N 1.D)(N2.I)] is the same user's chance (0.9)/4=0.225

[(N1.D)(N2.J)] is the same user's chance (0.9) / 4 = 0.225

[(N1.D)(N2.K)] is the same user's chance (0.9+0.9+0.8)/4=0.625

[(N1.D)(N2.L)] is the probability of the same user (0.9/4=0.224

The more feature attributes that can be used, the more accurate the resulting probabilities. In this embodiment, it is determined that the probability value is less than 0.5, and there is no homing value, so the homing action is not performed. Then select the next user to calculate, until all users in the network have completed the home calculation; two heterogeneous networks are finished After the home registration is calculated, the third network user is added to continue the calculation until all calculations are completed, and the home rate table as shown in Table 4 is obtained.

Step 2: Establish a virtual concept network 200. According to the probability of each user's registrar in Table 4, the registrar rate is one, and the unified registrar code is used to replace the user code in the original network, [(N1.A), (N2, M )] is the same user, the home code is A, the M user code in the original N2 needs to be replaced with A, so the original (M->K), (M->J) connection relationship needs to be replaced with (A- >K)(A->J); the probability of returning to the household is not one, which is expressed as a candidate household combination, and the original number can be retained, but the candidate relationship connection must be established, and the users in the candidate household combination are in the original network. Relationships with other users must be transferred to another person to establish a heterogeneous network connection. [(N2.L), (N3, Y)] candidate homing households have a return rate of 0.8, the original N2 (L- >I) There is a connection relationship, which needs to be transferred to (N3.Y) to form (I->Y), N3 (Y->X) has a connection relationship, and must also be transferred to (N2.L) to form ( L->X), calculated in order, until all users have established a heterogeneous network relationship connection, and finally generate a relationship connection table (Table 5), the content records the connection relationship between each user, and the duplicate network Number of roads.

Step 3: Calculate the relationship weight 300. First, each user relationship connection in the respective network must be normalized, so that each relationship weight is between 0 and 1. In this embodiment, the strength of each network relationship is excluded from the Internet. The maximum relationship strength within the network, then the strength of all relationships in the network is between 0 and 1. Then, in step 200, a relationship connection table is obtained for comparison. If the relationship of the transfer is connected, the probability of user attribution in Table 4 needs to be taken into consideration, and the original (I->Y) relationship strength is 0.6, but because of the transfer relationship The probability of confirming the relationship strength process is 0.6*0.8=0.48; if the user connection relationship is repeated, the same household connection must be aggregated to average, and the (A->B) relationship strength is 0.6, 1. Then (A->B) the final relationship weight is 0.8. Calculate all the relationship weights in order, and finally get a relationship weight table (Table 6), and record the contents of each household. Relationship weights and related networks.

Step 4: Filter the relationship 400, and delete the relationship weight less than 0.5.

Step 5: Generate a mashup community cluster 500, select a closer relationship according to the screening relationship 400, and use a k-core method of the social network grouping algorithm to perform a k-value equal to 2 in the group to obtain a user. Codes A, B, C, I, J, and L are the same community.

Step 6, the home heterogeneous network relationship 600, according to the user affiliation table generated in step 100, the C and I in the community may be the same user probability of 0.625, need to be vested, the method is the same as establishing the virtual concept network 200 The household registration rate is one, the attribution code is C, and the original I user code needs to be replaced with C. Therefore, the original (I->J) and (I->L) connection relationships need to be replaced with (C). ->J), (C->L), the new community clusters are A, B, C, J, L, resulting in a new heterogeneous relationship table (Table 7).

Step 7. Calculate the heterogeneous weight 700. Step 600 obtains a heterogeneous relationship table. It can be seen that the inter-user connection relationship in each cluster may come from multiple networks, and the network that each side can influence is calculated. When each side can cover the network. The more the road, the heavier the weight should be. In this example, the number of networks covered by each relationship connection is divided by the total number of networks, and finally a heterogeneous weight table (Table 8) is obtained, and the content records each community cluster. User relationships and heterogeneous weights within.

Step 8. Analyze the user influence 800. In this example, the degree of influence in the community is calculated by using the degree of influence evaluation in the social network analysis method to find the important influence characters in the group, but each user All are generated by the household registration process, so each user’s step is required. In step 100, the probability of generating the user's household value is included in the calculation. The heterogeneous weights associated with the household code A are respectively added, 2/3, 1/3, 1/3, and then multiplied by the household code in Table 4. The relative attribution rates of A are 1 and 1, respectively, and the impact score is 1.67, which is calculated in order, and finally the user influence table of each sub-community cluster is generated (Table 9). The content records each community cluster. The impact scores of the users can be used to find out the characters with cross-network influence in the heterogeneous network relationship. In this embodiment, they are the home code A and B.

The detailed description above is a detailed description of one of the possible embodiments of the present invention, and is not intended to limit the scope of the invention, and the equivalents and modifications may be included in the present invention. The patent scope of this case.

100‧‧‧ Calculate the probability of multiple sources of information

200‧‧‧Creating a virtual concept network

300‧‧‧Computational relationship weights

400‧‧‧ screening relationship

500‧‧‧Generating a mixed community cluster

600‧‧‧Home-to-house heterogeneous network relationship

700‧‧‧ Calculate heterogeneous weights

800‧‧‧Analysis of user influence

Claims (7)

A multi-source heterogeneous network data analysis method, in particular, an analysis method applied to a media program, the steps of which include: Step 1: Obtaining network behaviors and characteristics of each user and identification function from a plurality of heterogeneous network data Attribute, and assign the weight of the feature attribute; Step 2: Through the comparison of the obtained feature attributes, if the unique user with the same attribute attribute is compared, the chance of returning the household is 1, and the process proceeds to step 4; otherwise, the process proceeds to step 3; Deducting the identified unique users, setting the unrecognized users as candidate homecoming combinations, and using the unrecognized repeated feature attributes of users or similar network behaviors to obtain the attribution probability; Step 4, returning The user whose occupancy rate is greater than the threshold of the household account is the household registration action, and the user whose household rate is less than the threshold of the household registration does not perform the household registration action, and generates the household registration probability table; Step 5: the household operation will be performed The user's home code in the heterogeneous network data of probability 1 is set to be the same, and the relationship connection is established, and the heterogeneous network with the probability of belonging to the home is not 1 The user's home code is reserved, and the candidate relationship is established. Step 6. Divide the relationship strength of each user relationship in each network by the maximum relationship strength in each network, so that each The relationship strength of the relationship connection is between 0 and 1, and is compared with the relationship connection. If the comparison result is a connection relationship, the relationship strength of the relationship connection is multiplied by the probability of returning the home. Relationship weight, if the comparison result is a duplicate relationship, the relationship strength of the relationship connection is summed with the probability of returning the household, and the relationship weight is obtained; Step 7: The threshold value is less than the relationship threshold by the threshold value screening method The relationship of the values is deleted, and the relationship weight is greater than the relationship threshold value. The social network is grouped. The algorithm is grouped to form a plurality of mixed community clusters; step eight, according to the household rate table, the mixed community clusters, and the same user may be assigned to the home operation; step IX, the mixed community clusters are connected The number of heterogeneous networks appearing on the line divided by the total number of heterogeneous networks, resulting in heterogeneous weights; Step 10, using the evaluation indicators of the social network analysis method, adding the heterogeneous weights of each user relationship connection, and the household The odds are multiplied to generate an impact score to identify important influential users within the community. The multi-source heterogeneous network data analysis method according to claim 1, wherein the attribution rate is 0.0 to 1.0, the household threshold is between 0.5 and 1.0, and the relationship threshold is between 0.5 and 1.0. The multi-source heterogeneous network data analysis method according to claim 1, wherein the characteristic attribute includes a name, a license number, an e-mail address, a mobile phone number, a place of residence, a work place, a graduate school, an interest, or use provided on another website. Information . The multi-source heterogeneous network data analysis method according to claim 1, wherein the network behavior is a mode of operation behavior of the user on the network, including repeated common friends or the same content at the same time and place. The multi-source heterogeneous network data analysis method according to claim 1, wherein the unique user having the same characteristic attribute is the same as the characteristic attribute of the user having the single characteristic of the license number, the mobile phone number or the electronic mail box. User. The multi-source heterogeneous network data analysis method according to claim 1, wherein the repeated characteristic attribute or similar network behavior includes any two or more of a duplicate name, a place of residence, a work place, and a graduate school. , with the same feature attribute that can be identified as the same user The user, or a similar network behavior that has a common friend or the same content at the same time or place. The multi-source heterogeneous network data analysis method according to claim 1, wherein the obtained attribution probability is to compare the attribute weights of the same feature attribute to the number of duplicate feature attributes in the total/heterogeneous network data.