The presented work deals with mathematical models' development for determining the significance degree of the social information. It contained in unstructured raw data from public web-sources. The significance criteria of social environment descriptions based on fuzzy logic and probability theory methods. Also, the set theory apparatus used to represent the extracted social objects. Results implementation allows to improve the data relevance and reliability determining in the social profiling system based on Big Data.
Modeling and analysis of social processes is currently based both on the classical techniques of sociology and various
theories of computer science [
The social profile (SP) [
Therefore, it is critical to identify discrepancies in the collected information and determine the data relevance degree. Results should be divided by importance in a particular research task. The initial SP information is taken from various sources and has a significant size (up to several terabytes). There is a problem of determining its reliability in a finite time. It is also impossible to exclude the human factor influence from the manual process of data`s social values. It is necessary to use automated tools to solve these problems. The instruments should be based on formalized mathematical criteria for the importance of social data.
Studies of textual data have been conducted in order to automatically determine their relevance and consistency by the scientific community for a long time. Currently, Text Mining methods are used for this purpose. Over the past five years, a sufficient number of scientific papers have been published, which served as sources of ideas for developing a criterion for the significance of the initial social data. We give a description of quite interesting works.
Abdul Ghaffar Shoro and Tariq Rahim Soomro note the possibilities of Big Data analysis and recognize meaningful
information from popular social networks [
A number of articles are devoted to the use fuzzy logic methods in the semantic analysis of unstructured text and natural
language. Farman Ali provides [
There are works on determining the records tonality from the Internet social media. Han Liu and Mihaela Cocea review
the concepts and techniques of granular computing in general, and focus on [
Also, there are researches on measuring the quantitative and bibliometric characteristics of information sources and
contained content. Mehdi Jafari and others simulate human-like methods by integrating fuzzy logic with traditional
statistical approaches to improve [
Works “Where the Truth Lies: Explaining the Credibility of Emerging Claims on the Web and Social Media” of
Kashyap Popat [
We take as a basis the scientific principles and analytical technologies from the above works. The popularity of using
fuzzy logic methods in analysis of natural language texts was discovered. A methods review was conducted for assessing
the accuracy of textual data from the world wide web public sources. The most interpreted bibliometric parameters for
digital sources of textual data are determined. However, we note that the developed significance criterion must be adaptable
to use in the system of a social profile building [
The developed mathematical model [
A personal SP contains many of the names and pseudonyms of the concerned person. Also, it consists key dates of life and related locations, personal characteristics and the most significant social phenomena, as well as links to profiles of other people and social groups, including a list of the person`s social roles and group relationships types. All this have links to sources in dynamic content. A conclusion is made about the mood of the person, both as a whole and in relation to the surrounding objects of the social environment. It based on the analysis of the source data context and tonality containing the elements presented above. Such content may include fairly subjective evaluation information, which is nonetheless important for drawing up a complete picture of the social profile.
In a group SP, its own unique parameters are added to the information from the profiles of its members. They can change drastically over time. Therefore, the included personal profiles have timestamps in the group relation descriptions to organize the analysis results. The complexity of defining the social groups boundaries also depends on the fact that they may be informal, conditional and implicit. In view of this, we restrict ourselves to the description of the personal social profile model only.
Social phenomena are information about the objects and events of the social environment. They are connected with a
person or group in the task of a social profile building. Phenomena can be described directly or indirectly - when existence
of one indicated by other phenomena. What is more, they have a number of properties [
All the above information should have the following qualities to achieve the most truthful and complete results: - Reliability - the analyst must know whether the collected initial information is true or false, and to what extent. This directly affects the veracity of the social profiling results.
- Relevance - social profile data from public sources can be updated at different intervals, ranging from a few changes per hour to just one change per week or even longer. The less time passes from the origin of a certain tangible event to its recording into the sources of the initial SP data and from its gathering to the completion of the analytical processing, the more accurately the results of the analysis will reflect reality.
- Detail - also affects the final results of social profiling. This quality allows for a more in-depth analysis of the social environment. The more properties an object has, the easier it is to compare one with other social environment elements.
The initial data for a social profile construction is taken from a variety of public Internet sources. It is impossible to state unequivocally that the sample will contain information that fully satisfies the indicated qualities. There may be a contradictory or completely false information in the samples. Moreover, even if the SP consist of only true information, it is not always fully used in further research due to the insignificant value of its individual components. Semantic value is determined primarily by the number of links of the selected element with the surrounding social environment objects. Therefore, it is required to introduce additional criteria related to information sources. The reliability and moderability show resources that have a high reputation. The attendance provides information how much the resource is known, readability — how popular the published content is. The appearance speed and the number of new unique publications give an idea how active the authors of the content are.
Accounting for these qualities leads to conduct activities to study the significance of the collected data variety used in the social profiling process. As a result, each element from the social environment description must be assigned a certain weighting factor. Weights show the conformity of the element to these qualities. Assessment of the quality indicators carried out not only by computer means, but also handled by experts. They can determine arbitrary values from Absolute Truth to Absolute Falsehood. Therefore, the apparatus of the fuzzy logic theory is the most suitable for modeling the significance criterion of the initial social data.
Then the mathematical model of personal social profile will take the form [
PSP P X , v , R Y , u , Q m n , (1) where P P1,, P – the social objects array of the person in question (themes, events, persons, etc.); m R R ,, R – the social connections between social objects array of the person in question;
1 n m, n - the number of social objects and connections in the personal SP, respectively, and m 1 n ; Q(m n) {Q1,, Qmn} – the fuzzy set of “intonations” of social characteristics, fundamental for a person`s complex mood assessment and depending on the number of social characteristics; it may take values in the interval: Q 1; 1 ; X, Y – property sets of social objects and relationships, respectively; v, u – weights of social objects and connections, respectively, they are fuzzy values.
Approaches to the processing [
A SP multimedia content is analyzed as follows: 1. Content is divided by nature (audio, images); 2. Content is divided into author's content and information about the selected person or group; 3. For author's content, multimedia objects are compared with existing samples from the Internet and the social profiles; also, the service information is extracted to find out information about the studied person`s activities and preferences; 4. For information about the studied person, a detailed content-analysis is carried out using the available tools of multimedia analytics (such as Machine Learning) in order to isolate essential information about the social environment from the multimedia object itself;
5. For all types of data, tonality analysis is carried out on the basis of the content expert assessment itself and the phrases extracted from it; 6. The results of the analysis are recorded to the information card with links to the processed content.
In this way, textualized information from multimedia content can also be evaluated. The most efficient way is to check the data significance at the stage of secondary lexical analysis. The SP information card is filled at this step. It also shows the semantic properties of the future social profile elements.
We pay attention to the belief functions and plausible reasoning definitions from Dempster-Shafer evidence theory [
Initial data characteristics were selected empirically. The weight coefficients u, v are calculated on its basis. There are verifiable difficulties, subjectivity of expert assessments and the prevalence of the human factor in data interpreting. It is not possible to give an accurate importance assessment of an object or phenomenon. Formalized criteria can only be considered ratings Vdeg, presented in numerical or percentage form. The fuzzy logic methods are proposed to solve this problem. We defined the fuzzy variables that affect the value of the significance coefficient for SP entities: relevance (Vvol), meaning (Vval), originality (Vunq), validity (Varg), credibility (Vver) of data and source dependability (Vauth). Their characteristic functions µ are:
– Relevance - the freshness of initial information. It actually represents the information appearance time into the raw sample. We will consider the interval of updating SP information as one day. This value provides the convenience of carrying out relevance calculations. The choice of a shorter period is not advisable, since the changes are mostly insignificant during this time. A longer period may affect that some gathered information will be irrelevant at the collection time.
The fewer days that have passed since the publication until the moment of data collection or processing the more relevant it is. We suppose the published social information will lose half its relevance a month later. Further reduction of the relevance follows a power law. The 31 days interval was chosen because the people’s activities associated to calendar cycles. For example, payouts and vacations are proportional to the months. The choice of the relevance weakness degree = 2 is justified by the fact that a person is usually socially active in the long term, but shortly he can have frequent intervals of 'silence'. We also determine that the information is most relevant during day of publication. Author may make changes to the published message during this period, and readers may not be able to immediately get acquainted with it (e.g. when they are in different time zones). This number will decrease in the future because the pace of life acceleration and the spread of mobile telecommunications grow.
However, sometimes information is not relevant yet at the collection time. This is usually represented by forecasts and conditional assumptions. Past events information always relevant at the time of publication, as well as facts are not tied to time.
Vvol = {not actual yet, actual, not actual anymore} – possible values of relevance criteria; vol = {no, yes} – shows the availability of more recent information; time = {past, present, future} – shows the time of statements, which contain the social essence, to determine their relevance; xvol = [0; ∞) – elapsed time since publication; 0, vol 1 time "future "; Vvol 1, xvol 1 time "past"; (2) 1 , xvol 1, 1 ( xvol 1)2 30 – The meaning of information - shows the power of connectedness with other entities SP. This indicator depends on their weights. We modify the class S membership function so that information is considered secondary if objects with the connections no more than half of the all considered entity links, otherwise - primary.
Vval = {meaningless, minor, major, critical} – information may not relate to the actual problem, have little or great importance in solving it, and also be central by providing a task solution or opening previously unknown areas for research;
xval = [0; nlink] – the integer number of connections between social entities whose weight is greater than a certain threshold value umin; nlink – total number of unique connections with the entity under investigation; (3) (4)
Vval 0, xval 0;
2 2xval , 0 xval nlink 2 1 nlink 2 1, xval nlink , 2( xval nlink )2 nlink 2
; , nlink 2
xval nlink; 0, xauth 0; Vauth 2xauth2 , 0 xauth 0.5; 1 2( xauth 1)2 , 0.5 xauth 1, – The source dependability is determined empirically by an expert. There are no exact criteria. A zero value of the indicator means that the source authority cannot be determined.
Vauth = {not defined, low, high} – values showing how dependable the source is; xauth = [0; 1] – average expert scale of authority; – Data validity – defined by the existence of arguments, evidence, references of considered social element. The resulting assessment is influenced by the fact that the sample contains the most reliable sources of information and the average level of sources dependability. This is done in order to identify information noise: some social media resources copy facts from the outside hastily. The assessment of this indicator more than 0.5 says that the considered fact has weighty evidence of truth.
Varg = {not enough, enough} – states of validity criteria; xarg = [0; ∞) – number of possible arguments; 0, xarg 0; 1 Varg xarg , xarg 0, (5)
1 e2(2maxxarg Vauth xarVgauth ) – Data originality - the number of information mentions. This index used to determine the less common facts. This value shows primary sources of information. Like validity, this indicator depends on the source dependability. It allows you to separate the facts from the unsubstantiated rumors and assumptions. This simplifies the process of identifying verified primary sources.
Vunq = {no mention, few mention, widely known} – list of uniqueness values; xunq = [0; ns] – the number of sources containing the analyzed statement, with a publication date less than in the original; ns = [1; ∞) – the total number of sources containing the subject statement from the SP; 1, xunq 0; Vunq 1 2( xnunqs2Vauth )2 , 0 xunq n2s ; (6) (ns 1 )2 ns2xunq Vauth , n2s xunq ns, – Data credibility - shows the different points of view (PoV) prevalence. It depends on the validity and originality. If there is a single displayed opinion, we will consider it true until alternatives appear. Otherwise, the credibility of this PoV will be depend on confidence indicator of others sights.
Vver = {false, undefined, true} – possible states of credibility criteria; xver = [0; ∞) – number of different points of view in initial dataset;
Vver
1 maxixve1r veri , xver 1;
1, xver 1, where ver – the confidence indicator of proceeded data. It defined as the most authoritative opinion, which must either be supported by many facts or be novel.
ver max Varg Vauth , Vunq Vauth , nsmin , (8)
Thus, the summary mathematical representation of the weighting coefficient for data significance definition will be as follows:
u Vvol Vval umin Vver ver , The coefficient v is depicted similarly. The difference is that it focuses directly on the definition of actants. (7) (9)
The introduction of the developed significance criterion for the initial text data allowed to detect and exclude from consideration extraneous and fake information sources. Testing was conducted on a sample of unstructured social texts from 30 public sources. These results were issued by Google search engine for the query “famous blockchain developers biography”. Some selected resources in the sample contained a meaningless compilation of sentences and phrases borrowed from trusted sources. Thereby they mimicking popular information resources. Some data sources were randomly sampled and did not contains the required information.
Also, the usage of a fuzzy significance criterion (relevance indicator) has simplified the sorting of links between social objects in chronological order. It is a part of the ongoing research devoted to methods creation for social profiling task.
In the future, it is planned to introduce the developed criteria into text-analytical module of the social profile building system based on Big Data technologies. Detailed performance testing of the weight criterion will be carried out on a sample of 1500 unique public web-pages with socialized data. If necessary, the characteristic functions of existing quantitative indicators will be corrected to process bibliometric attributes.
The result of this work is a fuzzy criterion model. It determines the significance coefficient of social profile elements in a comprehensive analysis. The criterion based on the such qualitative information characteristics as: connectivity, reasoning, relevance, reliability, uniqueness of data and expert assessment of the source verification level. These parameters are primary and necessary to identify outdated and unreliable information in the SP, as well as finding unreliable sources of social media. The resulted fuzzy model can be implemented in the developed concept of the social environment states description. It allows you to increase the consistency degree of the aggregated social profile elements. Also developed significance criterion does simple the tonality analysis procedure of social media data by identifying the semantic content of the particular textual elements. Next stage of our research work devoted to the development of modular implementation for the software and instrumental complex of social profiling.
Acknowledgments This work is carried out with the support of the Russian Foundation for Basic Research grant №18-07-00408 in a research project named “Fundamental theoretical bases development for self-adaptation of applied software systems”.