RU2469389C1 - Method of integrating user profiles of online social networks - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: method comprises steps of entering all possible pairs of profiles; constructing a Conditional Random Field model for all profiles and connections between them; for each pair of profiles, calculating the similarity value of their attributes using a string or graph similarity metric; constructing a feature vector from the obtained similarity metric values, which is sent to a machine learning algorithm which calculates unary energy or binary energy for each pair of profiles, wherein the profiles belong to different social graphs; calculating profile similarity; checking whether the obtained profile similarity value exceeds a given threshold value; if so, the pair of profiles is entered into a list of candidates; a priori true projections are selected from the obtained list of candidates; the Conditional Random Field model is broken into independent components; for each component of the model, the optimum configuration of projections is sought; the lists of the found projections are merged for all components of the model.

EFFECT: high efficiency of integrating user profiles of online social networks.

6 dwg

Description

The invention relates to the field of processing user data obtained from graphs of online social networks in order to integrate data of various profiles belonging to one user. It can be used to build databases of user information obtained from various sources, in particular, to build an extended social graph containing user data obtained from several different social graphs. Such an expanded social graph can be used to improve the quality of results in a number of tasks, such as searching for information on the Internet, online advertising of goods and services, building recommendations for goods and services to users, etc.

Consider the basic concepts necessary for understanding the presented invention:

1. Data integration includes combining data from various sources and providing data to users in a unified form. The invention provides the integration of data at a logical level in order to provide access to data contained in various sources in terms of a single global scheme that describes their joint presentation taking into account structural properties.

2. A social graph is a digital representation of the relationships of users of online social networks, which is clearly defined by various types of relationships between users (for example, friendship, following, etc.). The user data in the social graph is presented in the form of a profile, which is a combination of attributes (mainly string) in the form of name-value pairs that contain various information about the user (for example, name, gender, address, phone number, etc.). The invention is intended for combining various social graphs by comparing the attributes of user profiles and the intensive use of information hidden in the links between profiles.

3. Conditional Random Fields is a graphical probabilistic model in which dependencies between random variables are represented as an undirected graph. The nodes of the graph are divided into two disjoint sets - observable variables, which are set as input, and hidden variables. The edges of the graph correspond to probabilistic relationships between nodes. Vertices and edges can be assigned numerical values called energies. The calculation of the values of hidden variables is called inference from the model.

4. Machine learning - a section of artificial intelligence that studies the methods of constructing algorithms that can be learned. In the present invention, a kind of machine learning is used, called learning with a teacher: the algorithm generates a function that associates input with output in a certain way (classification task). As training data, examples of the relationship between input and output are used. Machine learning algorithms make extensive use of the concept of trait. Signs are the individual measurable properties of the observed phenomenon, which are used to create its numerical representation (for example, the value of string similarity metrics for pairs of profile attributes).

5. String similarity metrics return the numerical value of the similarity of a pair of strings, based on the arrangement of their constituent characters (for example, the Jaro-Winkler distance [17]).

6. Graph similarity metrics return the numerical value of the similarity of a pair of graph nodes based on the structure of the relationships between them (for example, the Dyce coefficient [18]).

The largest study on the integration of user profiles of social networks is the thesis Veldman [1]. It offers many heuristics that use both user profile data and existing relationships between them. The results of such studies are also presented in Motoyama et al [2], Gae-won et al [3], Raad et al [4] and Vozecky et al [5]. In [2], authors compare MySpace and Facebook user profiles. In [3], the authors do the same with profiles from Twitter and EntityCube. The authors [4] generate synthetic user profiles and apply various complex heuristics to them, trying to use any potentially useful source of data on a social network. In [5], the user profiles of Facebook and StudiVZ are presented in the form of n-dimensional vectors, which are subsequently compared using various techniques, including fuzzy comparison. The authors also examine the effect of various profile attributes on the accuracy of comparison results.

Of interest are also the projects Foaf-o-matic [6] and Okkam [7], the purpose of which is the integration of social profiles using the formal semantics of FOAF (Friend-of-a-friend). The Stanford Entity Resolution Framework [8] is also designed to solve problems like this one. In addition to the source code of the framework, there are many works on theoretical aspects of data integration, such as scalability, quality assessment, etc.

Despite the successes achieved by the authors of the above works, they use a too simple model for comparing profiles, based mainly on pairwise comparison using string similarity of individual attributes. In addition, existing connections between profiles are not taken into account sufficiently or are not taken into account at all, the same applies to the features of the compared social graphs.

Closest to the presented invention is a method proposed by Singla et al [9] for identifying duplicates in a citation network of scientific papers using the model of Conditional Random Fields. The authors formulate the problem in terms of Markov random logic and construct a model from facts and statements about the objects being compared, and then calculate their probabilities. To optimize the speed of the algorithm, the authors split the model into intersecting components. However, the presented approach has the following disadvantages that impede its use for solving the problem under consideration:

- the model provides for the presence of only one data source (graph of the citation network of scientific works);

- the nodes of the model are facts and statements about the compared objects, and not the objects themselves. In particular, two types of nodes are defined: record nodes and attribute nodes. The first type of nodes is designed to store the question "Is this object identical to another object?", While the second type stores information about the similarity of the attributes of objects;

- to compare objects, only metrics of string similarity of their attributes are used, while metrics of graph similarity are not used.

The present invention has the following advantages compared to previously proposed approaches:

- allows you to simplify the representation of a social graph in the memory of a computer, since the model of Conditional Random Fields is built on the basis of one of the compared graphs, while the nodes of the model are directly profiles, and the edges are the connections between them. Then, the link energies of the model are calculated based on the data on the string and graph similarity of the profiles, after which the search for the optimal solution in the form of a configuration projection of the profiles of one social network on another is performed;

- using the Conditional Random Fields model, all available information about the relationships between the profiles is taken into account, which allows you to use the latent information hidden in these relationships to refine the results. Thus, the method allows the integration of profiles whose attributes contain only a small amount of useful information, which greatly complicates the application of the generally accepted approach based on string proximity of attributes;

- the adequacy and effectiveness of various metrics of string and graph proximity, as well as the relative importance of attributes are estimated using machine learning techniques on a pre-compiled set of real experimental data, which allows you to take into account the features of the selected social graphs and optimize the algorithm parameters to achieve better results.

The technical result of the use of the invention is that the invention allows a previously unknown method to obtain a list of pairs of user profiles of online social networks, in which the profiles in each pair contain information about the same user, based only on the information contained in the profile attributes and connections between them. A universal approach is also proposed that allows you to take into account the features of any pair of social graphs and optimize the parameters of the method to obtain better results.

For a better understanding of the claimed invention the following is a detailed description with the corresponding drawings.

Figure 1 is a block diagram of the algorithm of the invention

Figure 2 is a diagram for calculating the similarity values of profile attributes and the construction of a feature vector

Figure 3 is a Model of Conditional Random Fields

Consider two social graphs A and B. For a vertex ν∈A, the corresponding profile from graph B will be called the projection pr (ν) of the vertex ν∈A onto graph B. If a projection from B is not defined for the vertex ν∈A, then the neutral projection is assigned to it : pr (ν) = N.

The task of integrating user profiles is to determine the maximum possible number of correct projections (ν, u), ν∈A, u∈A. in terms of the model of Conditional Random Projection Fields pr (ν) for each ν∈A are hidden variables whose value must be set.

The presented invention is based on the following main statements:

- the problem of determining the correct projection for a node from graph A is related to the problems of determining the correct projections for all adjacent nodes from graph A;

- if two nodes in column A are connected, then their projections in column B should have the highest value of graph similarity.

The proposed method for integrating user profiles of online social networks contains an algorithm that includes entering all possible pairs of profiles, 19 basic steps and displaying the result in the form of a list of pairs of profiles, in which each pair contains information about the same user, while the profiles that make up the projection belong to different social graphs.

Consider in more detail the steps of the algorithm (figure 1).

Step 100. Entering all possible pairs of profiles.

All possible pairs of profiles are involved in the comparison, since the algorithm for deriving Conditional Random Fields from the model used in the invention gives the best results when there is information about the energies of all possible pair combinations of nodes. A set of profiles from columns A and B with all known relationships between them form a model of Conditional Random Fields.

Step 101. Select the next pair of profiles.

Step 102. Calculation of the similarity values of profile attributes and construction of a feature vector.

Comparison of the profile attributes from columns A and B is performed using the correspondence scheme, which sets the order of attribute comparison and the similarity metrics used (examples of correspondence schemes for calculating the similarity of attributes between nodes from different graphs and between two nodes of ν are given in Table 1 and Table. 2, respectively).

Figure 2 shows an example of a comparison of profiles from Twitter and Facebook. T _i and F _i correspond to two compared attributes from the profiles T and F, where i is the serial number of the records containing the given attributes in the correspondence scheme. A sim _i similarity metric is applied to each pair of attributes. The similarity metric values for all attribute pairs make up the feature vector, which is passed to the next step.

Step 103. Calculation of the energy of a pair of profiles using the machine learning algorithm.

We define the concepts of energies by first defining the necessary variables.

Graph A is used to construct a model of Conditional Random Fields. Let ν and u be users from graph A, and pr (ν) and pr (u) their projections from graph B.

Then the node energy (unary energy) is defined as

Ф (pr (ν) | ν) = α (ν) · (1-profile _{similarity (ν, pr (ν)} ),

binding energy (binary energy) is defined as

ψ (pr (ν), pr (u) | ν, u) = β (pr (ν), pr (u)) · (1-network _{similarity (pr (ν), pr (u)} ),

and the total energy is defined as

α (ν) and β (pr (ν), pr (u)) are the coefficients that govern the influence of each type of energy on the final model and the results of the work. The profile _similarity and network _similarity functions are profile _similarity functions that are normalized and increase with increasing similarity of the compared profiles.

To calculate unary and binary energies in order to maximize the accuracy of the results by analyzing real data, a machine learning technique is used that receives a set of similarity metric values for a given pair of profiles as a feature vector and returns the value of the binding energy of this pair of profiles. As a machine learning technique, one of the following methods is provided:

- A weighted linear combination of features, where weights are selected using linear regression [15], on the basis that the unary energy for the correct projections should be 0, and for the wrong ones - 1;

- Machine learning algorithm MultiBoostAB [13] over decision trees C 4.5 [14];

- LogitBoost machine learning algorithm over M5P decision trees [16].

The listed algorithms pass the training procedure before use, i.e. receive a set of feature vectors, obtained by calculating the similarity metrics for the profile attributes included in the data set, in which the correct projections are predefined manually. The feature vector used for training includes all values of similarity metrics for the pair of profiles being compared and an additional dimension containing a Boolean type value and indicating whether these profiles contain information about the same user. Such a data set can be compiled for any pair of social graphs.

Step 104. Is this pair of profiles the last? If NO, go to step 101; if YES, go to step 105.

Step 105. Selecting the next pair of profiles.

A sequential search of all possible pairs of profiles is carried out, in which one of the profiles belongs to column A and the second to column B.

Step 106. Calculation of the similarity of the pair of profiles.

The attribute of the selected pair of profiles is compared, which most likely uniquely identifies the profile using the string similarity metric.

Step 107. Does the similarity of the pair of profiles exceed the specified threshold value? If YES, go to step 108; if NO, go to step 109.

Step 108. Add pairs of profiles to the list of candidates.

Step 109. Is this pair of profiles the last? If NO, go to step 105; if YES, go to step 110.

Step 110. Selecting the best pairs of profiles from the list of candidates and compiling a list of a priori correct projections.

To select the best pairs of profiles, the Kuna-Mancresa algorithm [19] is applied to the list of candidates, which sequentially searches all pairs from the list of candidates and gives as a result some of them selected so that each profile appears in only one pair, and the profiles constituting a pair mutually maximized similarity to each other. The result is a set of a priori correct projections (pairs of profiles) that are entered into the model in order to improve the quality of the results.

Step 111. Partitioning the model into independent components by removing a priori true projections.

In order to reduce the computational complexity of the algorithm, the original problem is divided into sub-tasks by splitting the original model. This is achieved by removing the found a priori correct projections.

Step 112. Selecting the next model component.

Step 113. Finding the optimal projection configuration for the selected component.

To solve this problem, the total energy of the model should be minimized. For this, a conclusion is made from the constructed model of Conditional Random Fields. First, quadratic relaxation [10] is applied to the inference problem, then the problem is approximated using the Power Iteration [11] and Singular Value Decomposition [12] methods, and then it is solved as a quadratic programming problem. The result of this step is the configuration of the model (set of projections), which minimizes the total energy of the model and contains the maximum number of correct projections of the profiles.

Figure 3 schematically shows a model with a set of different projections. The nodes of graph A are shown by a continuous line, while the nodes of graph B by a dash-dotted line. Nodes in the form of squares correspond to a priori correct projections, nodes in the form of triangles are components of the projections found as a result of the derivation, while for all other nodes suitable projections were not found. The pair of nodes making up the projection is connected using a dashed line.

Step 114. Is this component of the model the last? If NO, go to step 112; if YES, go to step 115.

Step 115. Combining the results for all components of the model.

The lists of found projections for all components of the model are combined.

Step 116. Select the next projection.

Step 117. Building a feature vector for the selected projection and redirecting it to the classifier as input.

For the selected projection, a vector is constructed consisting of the following features:

- unary peak energy;

- average binary binding energy with a priori true projections;

- the proportion of a priori correct projections in the list of vertices associated with this one;

- the quality of a set of a priori true projections.

The quality of the set of a priori faithful projections is calculated as the sum of the largest N weights assigned to the a priori faithful projections associated with the vertex in question, where the vertex weight is calculated as the average binary binding energy between this vertex and other weighted vertices.

Step 118. Is the selected projection correct?

To clarify the results, a classifier algorithm for machine learning is used (MultiBoostAB boost [13] over decision trees C 4.5 [14]). The classification algorithm takes the constructed feature vector as input and returns a decision on whether the given vertex is correctly projected. If the projection is incorrect (the solution of the classifier does not coincide with the solution of the algorithm), then go to step 119, otherwise, go to step 120.

Step 119. Removing the selected projection from the results.

Step 120. Is this projection the last? If NO, go to step 116; if YES, go to step 121.

Step 121. Displaying a list of projections in which each projection contains information about the same user, while the profiles that make up the projection belong to different social graphs.

Work example

Figure 3 shows models of two social graphs, where profiles are represented by nodes, and the connections between them are edges. We assume that nodes 1-8 belong to column A (Twitter), and nodes 9-14 belong to column B (Facebook). All connections between nodes within each of the graphs are given initially, while connections between nodes of different graphs (pairing nodes) are the result of work. A Twitter graph with a full set of nodes and edges is used to construct a model of Conditional Random Fields, and nodes from the Facebook graph are considered hidden model variables. At the same time, the relationship in the Twitter column is the reciprocal relationship (each pair profile follows a different profile, i.e. receives notifications of new information in the profile; this relationship is established by unilateral activation of notifications by the owner of the profile that follows the other profile), and in the Facebook column - the friendship relationship (each profile of the couple is friends with another profile; this relationship is established by sending the owner of one of the profiles a request for establishing a relationship and receiving an explicit confirmation Nia request).

At step 102, the similarity values of the profile attributes are calculated and a feature vector is constructed. Comparison of the profile attributes from columns A and B is performed using the correspondence scheme, which sets the order of attribute comparison and the similarity metrics used (examples of correspondence schemes for calculating the similarity of attributes between Twitter and Facebook profiles and between Facebook profiles are given in Table 1 and Table 2 respectively).

Table 1 Compliance scheme for calculating the similarity of profile attributes between nodes 2 and 11 Twitter Profile Attribute (Node 2) Twitter profile attribute value Facebook Profile Attribute (Node 11) Facebook profile attribute value Similarity metric Metric value Name John smith Name J smith VMN 0.66 User place New York, US Current city New york Jaro 0.45 URL www.my.site Website www.no.site URL measure 0

Table 2. Compliance scheme for calculating the similarity of profile attributes by nodes 11 and 14 Facebook profile attribute First Facebook Profile Attribute Value (Node 11) Facebook Second Profile Attribute Value (Node 14) Similarity metric Metric value Contact list 9, 10, 12, 13, 14 11, 12,13 Bidirectional contact score one Contact list 9, 10, 12, 13, 14 11,12,13 Weighted dice 0.3

At step 103, the values of unary and binary energies are calculated.

To calculate unary and binary energies, a machine learning technique is used, which receives a vector of attributes and returns the value of the binding energy of a given pair of profiles. For example, for profiles from Table 1, the feature vector is as follows:

[0.66; 0.45; 0].

The machine learning algorithm returns the value of unary energy for a given pair of profiles equal to 0.52.

At step 106, a pairwise comparison of the "Name" attribute of all pairs of Twitter-Facebook profiles is performed using the metric similarity string VMN. The result is a set of triples of the form “profile number - profile number - similarity value”. The similarity threshold value is assumed to be 0.51.

At step 107, the string similarity metric is compared with a predetermined threshold.

At step 108, all pairs for which the value of the string similarity metric has exceeded a predetermined threshold are entered in the candidate list.

For example, the initial list of pairs has the form:

one 9 0 one 10 0.65 one eleven 0.55 2 9 0.5 2 10 0.6 2 eleven 0.66 3 9 0.35 3 10 0.7 3 eleven 0

Then the list of candidate pairs will be as follows:

one 10 0.65 one eleven 0.55 2 10 0.6 2 eleven 0.66 3 10 0.7

At step 110, the Kuhn-Mancres algorithm is applied to the candidate list to select the best projection profiles that mutually maximize similarity to each other. The result is a set of a priori true projections that are entered into the model:

2 eleven 3 10

In Fig. 3, a priori correct projections correspond to pairs of nodes in the form of squares connected by a dashed line (pairs 2-11.3-10 and 5-13). Each such projection introduces additional useful information into the model and, thus, not only helps to reduce the number of calculations necessary to obtain the optimal configuration, but also reduces the likelihood of incorrect answers in the work results.

At step 111, the original model is partitioned into independent components by removing a priori true projections. In the example under consideration, after deleting nodes 2, 3, and 5, 3 independent subgraphs (1, 7), (8), and (4) remain. The corresponding nodes 10, 11, and 13 from the Facebook graph are also not further considered and do not participate in the selection of projections for the nodes remaining in the model.

At step 113, for each of the resulting independent components of the model, a conclusion is made by applying an iterative algorithm that sortes out the possible configurations of the projections of the model and minimizes its total energy.

At step 115, the results for all components of the model are combined.

An example of the results of this step is a set of projections 1-9, 1-12 and 6-14.

The following steps refine the results.

In step 117, for each projection found in step 115, a feature vector is compiled, which is fed to the input of the classifier.

At step 118, the classifier decides on the fidelity of each of the projections. Let the vector for the projection 1-9 be [0.85; 0.5; 0.4; 0.5], and for the projection 1-12 [0.7; 0.6; 0.6; 0.35]. Then for the projection 1-9 the classifier gives the answer "true", and for the projection 1-12 - the answer is "false".

At step 119, the projection 1-12 is excluded from the results of the work, since the classifier answer for this projection does not coincide with the solution of the algorithm.

At step 121, a list of projections is displayed in which each projection contains information about the same user, while the profiles that make up the projection belong to different social graphs. The final result of the invention on the example in question is shown in figure 3 and contains 2 projections: 1-9 and 6-14. This means that peaks 1 and 9 belong to different social graphs, but at the same time contain information about the same user. The same goes for peaks 6 and 14.

Claims (1)

A way to integrate user profiles of online social networks, characterized in that they enter all possible pairs of profiles, build a model of Conditional Random Fields from all profiles and the relationships between them, after which for each pair of profiles they calculate the similarity values of their attributes using string and graph similarity metrics, based on correspondence schemes for profiles from various social graphs, after which a feature vector is constructed from the obtained similarity metrics, which is transmitted to the machine learning algorithm which calculates unary energy by the formula
Ф (pr (ν) | ν) = α (ν) · (1-profile _{similarity (ν, pr (ν))} )
for a pair of profiles ν and pr (ν) belonging to different social graphs, or binary energy according to the formula
Ψ (pr (ν), pr (u) | ν, u) = β (pr (ν), pr (u)) · (1-network _{similarity (pr (ν), pr (u))} )
for a pair of profiles pr (ν) and pr (u) belonging to the same social graph, after which for each pair of profiles in which profiles belong to different social graphs, the similarity of profiles is calculated by applying the string similarity metric to the attribute with the highest the degree of probability uniquely identifies the profile, then it is checked whether the obtained profile similarity value exceeds a predetermined threshold value, in case of a positive answer, a pair of profiles is entered in the candidate list, then from the received the candidate list, a priori correct projections are selected by applying an algorithm to the candidate list that sequentially searches all the pairs from the candidate list and gives as a result some of them selected so that each profile appears in only one pair, and the profiles that make up the pair mutually maximized similarity to each other, then the Model of Conditional Random Fields is divided into independent components by removing a priori true projections, after which for each component of the model the search for the optimal projection configuration by applying an iterative algorithm minimizing the total energy of the model

in order to obtain the maximum number of correct projections of profiles, after which the lists of found projections are combined for all components of the model, then a feature vector is constructed for each projection found, which is transmitted to the classifier, after which the classifier determines whether this projection is correct, in case of a negative answer, the projection is excluded from the results, after which a list of projections (pairs of profiles) is displayed, each of which contains information about the same user, while Projection profiles refer to various social graphs.

Images