METHOD FOR CLASSIFYING A USER PROFILE, CLASSIFICATION SYSTEM

Abstract

METHOD FOR CLASSIFYING A USER PROFILE, CLASSIFICATION SYSTEM System comprising: an electronic terminal (T1) including means of communication, said means of communication enabling the receipt of data from a plurality of data resources and being configured to: select a set of content exchanged (Dk) with a first user (U1) of the terminal (T1), said content being associated with the first user (U1), extract terms (Tki) from the set of digital content (Dk); generate messages containing said extracted data;receive data processed by said remote server, a remote server (SERV1) configured to receive data emitted by the electronic terminal and implement a neural network (CNN1) trained to produce outputs (KF) feeding different classifiers, said server (SERV1) being, in addition, configured to generate a list of user profiles that have exchanged data with the first user (U1). Figure for the abbreviation: Fig.2;

Description

Title of the invention: METHOD FOR CLASSIFYING A USER PROFILE, CLASSIFICATION SYSTEM Field of the invention

[0001] The invention relates to the field of methods and systems for automatically classifying a user profile using at least one resource of a data network based on the data produced by its interactions with said resources and said users. The field of the invention relates more particularly to the automatic detection of a user profile interacting uniquely within a community and more particularly with respect to at least one user. In particular, the field of the invention relates to methods and systems for detecting user profiles whose interactions within a community can be analyzed and detected from a numerical model of said interactions. Prior art

[0002] Methods exist for identifying suspicious or abnormal behavior of individuals towards other individuals within a community. Generally, these methods rely on user profiling based on intellectual techniques, implementing a psychological or psychic study of individuals that involves comparing them to archetypes of known typical profiles.

[0003] However, these methods are not very effective because they involve detecting the suspect after an action has occurred that could cause harm or injury to a third party within a community. Furthermore, these methods are ineffective within certain data networks that involve a large number of interacting individuals, making it difficult to detect suspicious behavior quickly.

[0004] However, there is a real need to identify at-risk user profiles very early on within digital platforms. Indeed, these platforms attract the interest of younger people, bring together numerous communities, and facilitate rapid digital contact and exchanges between users.

[0005] This need exists to detect harassers, pedophiles, terrorists or individuals with racist remarks or behavior or any other individual likely to harm an individual or present a risk within a community.

[0006] Regardless of the community and profile type, there is a need to automatically identify unique user profiles within a community based on the type of digital interactions used. More generally, there is a need to generate a list of individuals associated with a probability of actions, said individuals having a computer link with a given individual that can be characterized by tangible interactions, quantifiable and measurable by technical criteria.

[0007] An unsolved problem is how to automatically identify a harmful profile or list of profiles and categories of profiles exercising different types of harm. Summary of the invention

[0008] The present invention makes it possible to overcome the aforementioned drawbacks. In particular, the technical solution detailed below makes it possible to overcome the current limitations because it involves technical means and an automatic implementation of user profile detection based on the implementation of a neural network trained to classify characteristic vectors specific to data interactions of a plurality of users.

[0009] According to one aspect, the invention relates to a computer-implemented method for automatically classifying a user profile from a plurality of resources on a data network, said method comprising: • automatic selection of a set of initial digital content exchanged between at least two users, the digital content being associated with an identifier of an initial user; • calculation based on a first learning function to which input vectors generated from the first numerical contents of output characteristic vectors are applied, said first output characteristic vectors feeding: • a first user profile classifier comprising at least a first class associated with the age of said first individual and a second class associated with the gender of the first individual; • calculation from a second learning function to which input data or input vectors generated from the first digital content exchanged between at least two users is applied, of second characteristic output vectors, said second characteristic output vectors feeding: • a second semantic classifier comprising at least a third ontological class characterizing a predefined semantic domain based on a first predefined ontology and the occurrences of the extracted terms in each selected digital content and; • Calculation of third characteristic output vectors, from a third function whose inputs include: • at least one piece of data characterizing the number of exchanges between the first individual and at least one second user within a predefined digital resource, • at least one piece of data characterizing a temporal data point relating to the exchanges of the first individual with at least one second user within a predefined digital resource, • at least one piece of data characterizing a number of individuals associated with a data exchange with the first user within a resource;

[0010] said characteristic output vectors supplying: • a third classifier characterizing data exchanges between at least the two users; • generation by means of a calculator of a list of user profiles, said list of user profiles comprising at least one user profile whose characteristic vectors have been classified into a plurality of classes of each classifier.

[0011] The method of the invention can also be described as a method for automatically predicting a type of data exchange likely to occur between two individuals or as a method for identifying a user profile exchanging an actor with a first predefined user.

[0012] An advantage of the invention is the correlation of different characteristic vectors, making it possible to determine, with greater reliability, user profiles of interest according to a defined semantic field and a profile of a first defined user. The automatic determination of characteristic vectors corresponding to exchanges between two users allows for better identification of user profiles presenting a risk within a community or for an individual. The exchanges can be characteristic of a given semantic field, which makes it possible to better distinguish false positives from such identification.

[0013] According to one embodiment, • The first learning function includes the implementation of a first neural network and / or a function implementing a support vector machine, said first learning function being trained from a dataset comprising digital content exchanged between at least two users; and • the second learning function includes the implementation of a learning function integrating statistical learning and / or a second neural network, the input of which includes descriptors extracted from the set of digital content exchanged between at least the two users, the second neural network being trained from a dataset including digital content exchanged between at least two users: • The third function includes the implementation of an expert system and / or a third neural network.

[0014] One advantage is to configure an artificial intelligence that is adapted to the nature of the system's input data for each data model processed. Another advantage is to combine, on the one hand, an artificial intelligence to process the data of the given semantic field and, on the other hand, an expert system to process the nature of the exchanges between two individuals. Such a configuration makes it possible to obtain the best benefit from the two systems, which are combined according to the data model.

[0015] According to one embodiment, where the third function includes the implementation of an expert system, said expert system comprising a fact base, a rule base and an inference engine, the rule base having configured thresholds and durations, said configuration being able to be automatically updated according to the classification of output vectors.

[0016] According to one embodiment, at least one input of the third function is an input vector comprising a set of time-stamped events, said events corresponding to data exchanges between at least the first user and at least one other user.

[0017] One advantage is that the system handles all events related to an exchange between at least two individuals, regardless of the data resource, the number of individuals, the number of exchanges, etc. This configuration improves the quality of artificial intelligence learning.

[0018] According to one embodiment, the first learning function, the second learning function and the third function are the same function.

[0019] One advantage of this configuration is the availability of a unique artificial intelligence in which input data is formatted to define input vectors containing all semantic descriptors, data qualifying the data exchanges and the exchanged content, and metadata. Advantageously, such a solution is useful when each formatted data element is sufficiently well-defined. Thus, the solution can initially be implemented with an expert system to classify the characteristic vectors. corresponding to data exchanges, this allows for the definition of threshold parameters, the coefficients of certain functions, and more generally, the configuration of the variable validity domains. Subsequently, artificial intelligence can be configured to replace the expert system or complement it through a learning function, for example, by developing coefficients and thresholds, etc., to qualify each data exchange.

[0020] According to one embodiment, the extracted descriptors include textual terms extracted from textual digital content, metadata from digital content, from multimedia digital files.

[0021] One advantage is having a semantic function that allows the extraction of terms from a text or, more generally, from multimedia content, and, using thesauri, dictionaries, and ontologies, the definition of formatted semantic descriptors to be processed by a learning function. The learning function is implemented using computational means, and the specially formatted data can populate a database within the system.

[0022] According to one embodiment, the set of initial digital content exchanged between at least two users includes at least one piece of data corresponding to: • A message of numerical data; • A digital image; • A digital video, • An interaction with a digital publication containing at least one digital annotation.

[0023] According to one embodiment, the semantic analysis includes a statistical analysis of the terms extracted from each of the digital contents exchanged between the two individuals.

[0024] According to one embodiment, said characteristic vectors classified in the set of classifiers are associated with a likelihood score, the set of likelihood scores allowing the calculation of a quantified indicator determining the inclusion or not in the generated list of a user who has exchanged data with the first user.

[0025] One advantage of using a likelihood score is the ability to leverage outputs from different classifiers of different artificial intelligences or expert systems. For example, the likelihood score allows an individual to be excluded from the generated list when a vector is not classified by one of the classifiers or when its classification is not relevant to a given class, while the output vectors from other artificial intelligences are. Or, to give another example, this solution allows individuals to be considered for which the vector classifications in each of the classifiers corroborate by crossing likelihood thresholds.

[0026] According to one embodiment, the process includes a timestamping of said first selected digital content.

[0027] One advantage is to construct entries that take into account a sequencing of exchanges. These can be characteristic of a type of exchange between two individuals according to the semantic field considered.

[0028] According to one embodiment, the semantic analysis includes the implementation of a neural network by processing input images and providing output characteristic vectors classified in the first classifier.

[0029] According to one embodiment, a first time input data for the third function includes the frequency of the data emitted by the second user.

[0030] According to one embodiment, a second time input to the third function includes a daily time variability of the data emitted by the second user.

[0031] According to one embodiment, a third time input to the third function includes an average response time between two successive exchanges between two users exchanging data.

[0032] According to one embodiment, at least one neural network is a convolutional neural network and the data classification is performed from a regression operation.

[0033] According to one embodiment, the semantic corpora are updated from a classification carried out from a neural network on a corpus of predefined digital documents.

[0034] According to another aspect, the invention relates to a system comprising: • an electronic terminal comprising a computer, at least one memory and a display, and means of communication for exchanging data with a data network, said means of communication enabling the reception of data from a plurality of data resources, such as web platforms and social networks, said computer being configured for; • select a set of content exchanged with a first user of the terminal, said content being associated with the first user, • extract from a set of descriptors the entire set of digital content exchanged between at least two users; • generate messages containing the extracted data to a remote server; • receive data processed by said remote server, • at least one remote server configured to receive data transmitted by the electronic terminal and implement • first learning function to which input vectors generated from the first digital contents of output characteristic vectors are applied, said first output characteristic vectors feeding: • a first user profile classifier comprising at least a first class associated with the age of said first individual and a second class associated with the gender of the first individual; • a second learning function to which input data or input vectors generated from the first digital content exchanged between at least two users of second output characteristic vectors are applied, said second output characteristic vectors feeding: • a second semantic classifier comprising at least a third ontological class characterizing a predefined semantic domain based on a first predefined ontology and the occurrences of the extracted terms in each selected digital content and; • a third function delivering third characteristic output vectors and whose inputs include: • at least one piece of data characterizing the number of exchanges between the first individual and at least one second user within a predefined digital resource, • at least one piece of data characterizing a temporal data point relating to the exchanges of the first individual with at least one second user within a predefined digital resource, • at least one piece of data characterizing a number of individuals associated with a data exchange with the first user within a resource;

[0035] said characteristic output vectors supplying: • a third classifier characterizing data exchanges between at least the two users;

[0036] said server being, moreover, configured to generate a list of user profiles associated with a classification of characteristic vectors in each classifier, said list including at least one second user who has exchanged data with the first user, said server being configured to send an alert message to a second electronic terminal when said second user profile is identified in the generated list.

[0037] According to one embodiment, the set of digital content is collected from a plurality of digital resources. Brief description of the figures

[0038] Other features and advantages of the invention will become apparent from the following detailed description, with reference to the accompanying figures, which illustrate:

[0039] [Fig-1]: an example of the implementation of a process of the invention;

[0040] [Fig.2]: an example of an implementation of the system of the invention,

[0041] [Fig.3]: an example of a classifier allowing the generation of a list of individuals classified according to a given parameter. Detailed description

[0042] Figure 1 shows a diagram of the main steps of the process of the invention.

[0043] A first step, SEL, comprises the selection of a set of digital content Dk from at least one data resource R;. A data resource R; can be a digital platform for exchanging data between different interconnected users, such as a digital social network, a website, a web application, etc.

[0044] According to one embodiment, the method includes a preliminary step of selecting a set of R resources installed on an electronic device such as a tablet, smartphone, or PC. The selection is preferably configured during setup and then performed automatically during the execution of the method. In a particular embodiment, R data resources can be added or removed.

[0045] A data resource R is associated with a first user identifier Ub. This identifier may be associated with a password. The first user is preferably the user of an electronic terminal Ti from which the data resources R are accessible with a first user identifier Up.

[0046] According to one embodiment, the digital content Dk can be of different types of format such as: - digital textual content formatted in a computer language such as HTML, XML, PDF, PostScript, XHTML, TXT, PHP, etc.; - digital image content encoded in a format such as PNG, TIFF, JPEG, etc.; - digital video content encoded in a format such as MPEG, OGM, AVI, FLV, boomerang, live, etc.

[0047] Digital content Dk can correspond to digital messages exchanged on a messaging service, digital comments edited on a data sharing platform and attached to existing digital content such as a data forum, digital statuses associated with a data page or user profile, videos downloaded from or to a remote site or videos streamed from a platform, photos downloaded from or to a remote site and possibly shared with other users.

[0048] The method of the invention includes an automatic and continuous extraction of the exchanged digital content Dk. This content may include metadata such as data labels, keywords, description fields, dates, positions, URLs or more generally pointers to data resources.

[0049] The extracted Dk content is stored in memory. This can be the memory of the terminal Ti and / or the memory of a remote server SERVi when the latter content is transmitted to a remote server SERVi. In one embodiment, the extracted Dk content includes text terms or fields formatted in a predefined data language, metadata associated with a given digital content, and media such as an image or video of a given digital content.

[0050] In one embodiment, the extracted contents Dk are processed to define inputs for one or more functions. For example, three types of functions, Fb, F2, and F3, are used in the invention. In one embodiment, functions F1 and F2 are learning functions, whose learning is performed by artificial intelligence, and function F3 includes the implementation of an expert system based on a set of rules, a fact base, and an inference engine. In another embodiment, function F3 is implemented by an artificial intelligence technique that can incorporate an expert system. By way of example, the values ​​of the constants and parameters defining rules, thresholds, and the parameterization of the expert system can be deduced from the implementation of artificial intelligence configured to generate coefficients or parameterization of the expert system.

[0051] The first function Fi allows for the automatic determination of the user profile of terminal Tb. The second function F2 allows for the automatic determination of a set of semantic content likely to be in a predefined ontological domain from which characteristic vectors are to be extracted. of this domain. The third function F3 allows the determination of user profiles interacting with the first user likely to correspond to a risk associated with the predefined ontological domain.

[0052] According to an example, the third function F3 allows itself to construct the reasoning adapted to the particular case it has to deal with by repeating a cycle of detection and selection of relevant rules to use and trigger.

[0053] According to one embodiment, the first and second functions FB F2 are implemented by neural networks or a statistical learning method implemented by a machine learning type algorithm, better known as "Machine Learning" or "automatic learning".

[0054] The input vectors can be of type Lexicon, CharLSYM, Fasttext, LIWC, Surface, Meta with respect to textual content. The input vectors can be of type image.

[0055] The machine learning model can be a neural network such as a CNN (Convolutional Neural Network), an SVM (Support Vector Machine), a BERT (Bidirectional Encoder Representations from Transformers), or an LSTM (Long Short-Term Memory). Other artificial intelligence technologies can be implemented within the scope of the invention. When different artificial intelligence technologies are used to classify semantic data, multimedia data (photos, videos, etc.), metadata, or data characterizing a type of exchange, these technologies can be implemented by different computers or even different servers.The classifiers can then be grouped within the same application on the same server in order to calculate indicators that generate a list of profiles identified with a risk.

[0056] Thus, the step of calculating the characteristic vectors KF in [Fig.1] can be carried out on one or more machines.

[0057] According to one embodiment, the extracted data is transmitted to a remote data server SERVi. The remote data server SERVi is configured to implement at least one neural network or a machine learning method such as a statistical learning method.

[0058] In one embodiment, each extracted content item is annotated with context data, namely the R data resource from which it was extracted, a date, a format, or a content type. In some variations, other data may be associated with the extracted content, such as social data indicating the type of exchange and the number of participants involved. 1st Classifier

[0059] According to one embodiment, the invention implements a CLP classifier configured to automatically classify the gender and sex of the first user Up. Recall that the first user Ui is the user of an electronic terminal Ti on which the process executes the steps, possibly with a remote server SERVi. The terminal Tl is configured to retrieve the extracted content and send it to at least one remote server SERVi. This CLP user profile classifier comprises at least a first class Ci associated with the age of the first individual Ui and a second class C2 associated with the gender of the first individual Up. Each class may include subclasses allowing the classification of different age groups or different subcategories of a class.One advantage of the first CLP classifier is that it provides a solution where it is not necessary to define a user profile for the first user (Up). Specifically, it is not necessary to define the age and gender of the first user since the profile is automatically generated by the first CLP classifier. The data classified by the first CLP classifier allows, in particular, the automatic determination of the age and gender data of the first user (Up). To this end, the interactions of the first user (Up) within at least one data resource (R) allow the collection of digital content that is exchanged, commented on, or downloaded within said data resource (R) by the first user (Up). The first remote server (SERVi) then collects the extracted data (Dk) from the various interactions of the first user (Ui) and defines input vectors for a neural network.A neural network, or any other technology implementing artificial intelligence such as a statistical learning method, can generate output vectors that feed into a CLP classifier. The first CLP classifier is then able to discriminate the age and gender of the first user, Up. 2nd Classifier

[0060] A second CLS classifier is configured to classify output vectors from a neural network and / or artificial intelligence implementing statistical learning according to a predefined ontology. The second CLS classifier is configured according to a given semantic domain. The domain could be that of harassment, terrorism, sexual predators, or individuals with racist remarks or behavior, etc. To this end, the neural network or a system that has learned automatically from a semantic corpus can be trained on a set of data constituting a training corpus.

[0061] To this end, specific dictionaries, thesauri and / or ontologies may be used to extract and analyze the textual content of digital documents excerpts. According to an example implementation, different ontologies can be used. These could be, for example, grammatical or syntactic ontologies. 3rd classifier

[0062] The invention makes it possible to take into account the data characterizing data exchanges between individuals within different data resources R;. This data characterizing the exchanges is automatically deduced from the context in which the digital content Di is exchanged between individuals Ui, UP within the different data resources.

[0063] These data define inputs to a third function F3 implementing an expert system and / or artificial intelligence so that output characteristic vectors are classified within a third classifier CLD. This classification is performed, for example, using a neural network when the third function is a learning function, or by a function implementing an expert system. The latter can also be implemented by a learning function.

[0064] Among the data characterizing data exchanges, the method and system of the invention make it possible to take into account the number of messages exchanged between two individuals, including the first individual Ui and at least one other individual UP. This can then be a quantitative classification of data emitted by certain individuals.

[0065] This feature can be obtained by regularly and automatically recording the dates and identifiers of individuals exchanging with the first individual Up. The data characterizing the number of exchanges within the same resource R; are input data for the neural network and can be formatted by a vector containing the date of the exchanged data, the identifier associated with the author of an exchanged data, the resource R; from which the data is exchanged, etc.

[0066] According to one example, a set of events from a set of resources comprises a plurality of time-stamped events which correspond to actions of the first user, including messages sent, digital interactions within data resources, messages received or notifications received, or data generated by other users and which are associated with an identifier of the first user on a data resource of the network.

[0067] In this configuration, a class C4 allows for the classification of exchanges between individuals based on the quantity and possibly the density or frequency of exchanges, particularly between individuals Ui and UP. This classification makes it possible, in particular, to distinguish between cases in which numerous exchanges between two individuals are observed and, conversely, sporadic or isolated exchanges between individuals. The number of messages exchanged can be related to a time interval in order to process message density over a time scale.

[0068] The method and system of the invention can be configured according to an embodiment such that it includes the definition of a fifth class C5 associated with a characterization of a temporal data of the exchanges of the first individual Ui with at least one second user UP within a predefined data resource R.

[0069] According to one embodiment, the characterization of temporal data can correspond to a quantification of the variability within the hours of a day of the sending of messages by a user Ui or by users such as the user UP with whom the first user Ui exchanges data. This classification makes it possible to define classes of messages sent by individuals that can be grouped according to certain times of the day or evening, or conversely, classes of messages distributed within the same day at very different times. Advantageously, a subclass C5i of the class C5 corresponding to this characterization can be defined.

[0070] According to one embodiment, the characterization of a temporal data can correspond to the average time interval between two successive messages sent by a user UP to the first user Up. This characterization can allow the definition, for example, of a subclass C5i of the class C5.

[0071] According to another embodiment, the characterization of a temporal data point can correspond to the average time interval between the transmission of a message by the first user and the response of a message transmitted by a user UP. This exchange characterization can be classified, for example, in a third subclass C53 of class C5.

[0072] According to another example, a subclass of C5 can include the characterization of the duration of a set of messages exchanged between Ui and another user UP. Indeed, the exchanges can continue over a timescale of a day, a week, a month, or even years. Thus, this characterization makes it possible to define a class of exchange between users. This class can be a subclass of class C5, denoted C54 for example.

[0073] According to one embodiment, the classifier comprises different C5 classes or subclasses of C5 characterizing temporal data. The classification of this data can be performed for a plurality of users interacting with the first user Ui

[0074] According to one embodiment, the CLD classifier includes a class C6 characterizing the number of individuals associated with a data exchange with the first user Ui within a data resource R;. This class allows, in particular, discrimination Whether the exchange is unilateral or involves multiple users interacting with the initial user (Ub), as can occur with digital comments associated with an individual's status posted on a digital platform, one interest lies in determining the probability of detecting a given individual profile or the probability that an individual presents a specific risk to an initial user (Up). This profile identification can correspond to a risk associated with a particular type of individual profile. Thus, data exchanges are correlated with the context of the exchange.For example, if the characterization of a data exchange containing abusive language with the first user (Ui) is conducted by a single user (UP), their interactions may be weighted upwards in the classification. If their approach corroborates those of other users, their interactions may be weighted downwards, or vice versa, depending on the other classifications of the characteristic vectors. The classification of an individual (UP) is performed based on the analysis of a set of vector classifications characterizing a plurality of exchange types involving said individual (UP). In other cases, the method and system of the invention make it possible to detect a case of harassment between N users against a single user. The repetition of a number of exchanges between a plurality of users, said users being associated with a data exchange with a first user (Ui), can then be identified as a risk.

[0075] Figure 1 represents a step in calculating the characteristic output vectors of the neural network; this step is denoted KF. Each neural network or artificial intelligence is trained with a labeled dataset.

[0076] At the output of neural networks, or more generally artificial intelligence engines, a set of CLP, CLS, and CLD classifiers are represented. These classifiers allow the output vectors to be classified. The invention makes it possible to leverage a set of classifiers to create a list of user profiles comprising a set of classified characteristics that can be collectively associated with a risk. The classified users have at least exchanged or posted data within an R data resource accessible to the first user, Ui.

[0077] Data is continuously collected from the various resources and sent to a server called SERVi, as shown in [Fig. 2]. To this end, [Fig. 2] represents a simplified example in which different terminals Ti and TP, respectively, belong to the first user Ui and at least one other user UP. A resource Ri is shown. This resource R can be a website or an exchange platform on which the exchanges between the first user Ui and the user UP take place.

[0078] Figure 2 further represents a terminal U2 corresponding to terminal T2, on which the process has been installed, or at least a part of the installation. Terminal T2 receives notifications issued based on the results of the profile classification of those with whom user Ui is communicating. Thus, parents, represented by U2, can be alerted if a child Ui is in conversation with a suspicious individual UP.

[0079] Figure 1 represents a GEN step for generating an indicator of a profile that has been classified with a percentage of similarity to a predefined threshold in each classifier. One advantage is the ability to define criteria for discriminating between a set of conditions that must be met in order to select a profile or not.

[0080] Figure 3 illustrates an embodiment of the method of the invention applied to the classification of four user profiles (U3, U4, U5, U6) that have interacted with user U1. User U2 receives notifications when a suspicious user, such as user U5 in this example, is detected. Detection is performed by analyzing the classifications of each user profile into a plurality of classes or subclasses.

[0081] In the example in [Fig. 3], the first CLP classifier identifies, for example, that the user (Ui) is a girl between 14 and 15 years old. These are the age classes Ci and the gender class C2. The classification of the user's profile (Ui) is performed automatically using a neural network or artificial intelligence implementing a statistical learning method to analyze all the image and text content of the digital content with which the first user (Ui) has interacted. These interactions may include commenting on content, downloading content, exchanging content with a third party, or viewing a data page containing said content. In the example in [Fig. 3], likelihood thresholds of 90% are defined in each of the classes Ci and C2 or a subclass.These thresholds are used to reduce error or uncertainty and to draw conclusions with a given level of confidence regarding the gender and age of the individual. According to one embodiment, the likelihood thresholds are parameterizable for each of the classes or subclasses. The method and system of the invention make it possible to define a class or subclass based on a threshold value to be reached for the value of the characteristic vector being classified. According to another embodiment, this value is defined from an average of the values ​​of the characteristic vectors of a class or subclass.

[0082] The second CLS classifier is an ontological classifier in which a set of ontological classes is configured to collect semantic characteristic vectors of a set of individuals interacting with the first user Ui One advantage of this classifier is that it limits the scope or selection of all users who share an interest or have a given relationship with the first user, Up. Ontologies can be defined to reflect a center of interest, such as horseback riding. They can also be defined to characterize a way of addressing someone, such as racism, pedophilia, or harassment. In other examples, the ontology can define a semantic register specific to terrorism. Other examples can be implemented using the method of the invention. The execution of the method of the invention is independent of the ontological class(es) used. Thus, the method can be adapted to any given semantic field.The invention allows within a given semantic field the definition of classes enabling the identification of user profiles having interactions with a first user Ui showing characteristic vectors specific to temporal data or a number of users exchanging with the first user Up.

[0083] In the case of [Fig. 3], a class C3 corresponds, for example, to the semantic field of harassment. To this end, training data for neural networks or artificial intelligence incorporating statistical learning may include dictionaries, pre-identified labels or tags, ontological databases, thesauri, etc.

[0084] In the example shown in [Fig. 3], two users, U5 and U4, out of the four users in the example, are classified in class C3 by the classifier CLs. Class C3, in this example, relates to the semantic domain of harassment. Class C3 is defined for a similarity of 70% in the lexical field extracted from the content exchanged between, on the one hand, these users U5 and U4, and, on the other hand, the first user, Up.

[0085] The CLD classifier allows the application of artificial intelligence trained on data technically characterizing the interactions between the first user Ui and all the users studied.

[0086] Class C4 characterizes a quantitative classification of exchanges between two individuals. Thus, users U4 and U6 exchanged a large volume of data, either directly, for example by messaging, or through comments on digital content on a platform.

[0087] In this example, the fact that user U4 belongs to classes C4 and C6 shows that the large volume of exchanges took place among a large community of users, since class C6 represents a classification in which exchanges occur within a large number of users. Thus, it is possible to identify an individual engaging in a certain type of group harassment.

[0088] In this same example, user U5 is classified according to criteria based on the fact that their exchanges, for example, are characterized by a high variability in the timing of exchanges throughout the day. This characterization thus makes it possible to identify a given user profile and to corroborate this user profile with its classification in another class, such as a class (not shown) or subclass C54, which designates messages sent over long periods of time, for example, over several months.

[0089] The invention thus makes it possible to identify at least one list of user profiles that interact with a user UI, and each profile of which can be associated with a risk to that user. For example, it is possible to identify interactions specific to extortion, psychological harassment, sexual harassment, harassment leading to the isolation of an individual, etc., not by analyzing only the semantic field, but by analyzing the types of data exchanges and their technical characterization, taking into account temporal or contextual indicators of interaction depending on the data resource.

Claims

1. Claims Computer-implemented method for automatically classifying a user profile (UP) of a plurality of resources (Si, S2, S3) of a data network (NETi), characterized in that it comprises: • automatic selection (SEL) of a set of first digital contents (Dk) exchanged between at least two users (Ui, UP), the digital contents being associated with an identifier of a first user (Ui); • calculation (AS) from a first learning function (Fi) to which are applied input vectors generated from the first digital contents (Dk) of output characteristic vectors (KF), said first output characteristic vectors (KFi) supplying: • a first user profile classifier (CLP ) comprising at least a first class (Ci) associated with an age of said first individual (Ui) and a second class (C2) associated with the gender of the first individual (Ui); • calculation (AS) from a second learning function (F2) to which are applied input data or input vectors generated from the first digital contents (Dk) exchanged between at least the two users (Ui, UP) of second output characteristic vectors (KF2), said second output characteristic vectors (KF2) supplying: • a second semantic classifier (CLS) comprising at least a third ontological class (C3) characterizing a predefined semantic domain based on a first predefined ontology and the occurrences of the terms extracted in each selected digital content and; • calculation (AS) of third characteristic output vectors (KF3), from a third function (F3) whose inputs include:

2. • at least one item of data characterizing a number of exchanges between the first individual (Ui) and at least one second user (UP) within a predefined digital resource, • at least one piece of data characterizing temporal data relating to the exchanges of the first individual (Ui) with at least one second user (UP) within a predefined digital resource, • at least one piece of data characterizing a number of individuals associated with a data exchange with the first user (Ui) within a resource; said characteristic output vectors (KF 3) supplying: • a third classifier characterizing data exchanges (CLD) between at least the two users (Ui, UP); • generation (GEN) by means of a calculator of a list of user profiles, said list of user profiles comprising at least one user profile (UP) whose characteristic vectors have been classified into a plurality of classes of each classifier (CLP, CLS, CLD), • sending an alert message to an electronic terminal (T2) when at least one user profile (Up) is classified in one of the classes of the classifier (CLP, CLS, CLd). Method for automatic classification of a user profile (UP) according to claim 1, characterized in that: • The first learning function (FJ) comprises the implementation of a first neural network (CNNi) and / or a function implementing a support vector machine (SVMi), said first learning function (Fi) being trained from a set of data comprising digital content (Dk) exchanged between at least two users; and • The second learning function (F2) includes the implementation of a learning function integrating statistical learning and / or a second neural network (CNN2 ), said input of which comprises descriptors (Tki) extracted (EXTi) from all the digital contents (Di) exchanged between at least the two users (Ub UP ), the second neural network (CNN2) being trained from a set of data comprising digital contents (Dk) exchanged between at least two users: • The third function (F3) comprises the implementation of an expert system and / or a third neural network (CNN3).

3. Method for automatically classifying a user profile (UF) according to claim 2, characterized in that when the third function (F3) comprises the implementation of an expert system, said expert system comprising a fact base, a rule base and an inference engine, the rule base comprising configured thresholds and durations, said configuration being able to be automatically updated according to the classification of the output vectors (KFb (KF2, KF3).

4. Method for automatically classifying a user profile (UF) according to any one of claims 1 to 3, characterized in that at least one input of the third function (F3) is an input vector comprising a set of time-stamped events, said events corresponding to data exchanges between at least the first user (Ui) and at least one other user (UP).

5. Method for automatically classifying a user profile (UP) according to claim 1, characterized in that the first learning function, the second learning function and the third function are the same function.

6. Method for automatically classifying a user profile (UP) according to claim 1, characterized in that the extracted descriptors comprise textual terms extracted from textual digital content, digital content metadata, multimedia digital file.

7. Method for automatically classifying a user profile (UP) according to any one of claims 1 to 6, characterized in that all of the first digital contents (Dk) exchanged between at least two users (Ub UP) comprise at least one piece of data corresponding to: • A digital data message; • A digital image; • A digital video, • An interaction with a digital publication including at least one digital annotation.

8. Method for automatically classifying a user profile (UP) according to any one of claims 1 to 7, characterized in that the semantic analysis comprises a statistical analysis of the terms extracted from each of the digital contents (Dk) exchanged between the two individuals.

9. Method for automatically classifying a user profile (UP) according to any one of claims 1 to 8, characterized in that said characteristic vectors classified in the set of classifiers (CLP, CLS, CLD) are associated with a likelihood score, the set of likelihood scores making it possible to calculate a quantified indicator determining the inclusion or not in the generated list of a user (UP) having exchanged data with the first user (Ui).

10. Method for automatically classifying a user profile (UP) according to any one of claims 1 to 9, characterized in that it comprises a timestamp of said first selected digital contents (Dk).

11. Method for automatically classifying a user profile (UP) according to claim 2, characterized in that a first temporal data item at the input of the third function (F3) comprises the frequency of the data transmitted by the second user (UP).

12. Method for automatically classifying a user profile (UP) according to claim 2, characterized in that a second temporal data item at the input of the third function (F3) comprises a daily temporal variability of the data emitted by the second user (UP).

13. Method for automatically classifying a user profile (UP) according to claim 2, characterized in that a third temporal data item at the input of the third function (F3) comprises an average response time between two successive exchanges between two users (Ub UP) exchanging data.

14. System comprising: an electronic terminal (Ti) comprising a computer (KJ, at least one memory and a display, characterized in that it comprises communication means for exchanging data with a data network, said communication means making it possible to receive data from a plurality of data resources, such as web platforms and social networks, said computer being configured for; • select a set of contents exchanged (Dk) with a first user (Ui) of the terminal (Ti), said contents being associated with the first user (Ui), • extract from a set of descriptors all digital content (Dk) exchanged between at least two users (Ub UP); • generate messages containing said extracted data to a remote server; • receive data processed by said remote server, at least one remote server (SERVi) configured to receive the data transmitted by the electronic terminal and implement • first learning function (Fi) to which input vectors generated from the first digital contents (Dk) of output characteristic vectors (KF) are applied, said first output characteristic vectors (KFi) supplying: • a first user profile classifier (CLP) comprising at least a first class (Ci) associated with the age of said first individual (Ui) and a second class (C2) associated with the gender of the first individual (Ui); • a second learning function (F2) to which input data or vectors are applied input generated from the first digital contents (Dk) exchanged between at least the two users (Ui, UP) of second output characteristic vectors (KF2), said second output characteristic vectors (KF2) supplying: • a second semantic classifier (CL s) comprising at least a third ontological class (C3) characterizing a predefined semantic domain based on a first predefined ontology and the occurrences of the terms extracted in each selected digital content and; • a third function (F3) delivering third output characteristic vectors (KF3) and whose inputs include: • at least one item of data characterizing a number of exchanges between the first individual (Ui) and at least one second user (UP) within a predefined digital resource, • at least one piece of data characterizing temporal data relating to the exchanges of the first individual (Ui) with at least one second user (UP) within a predefined digital resource, • at least one piece of data characterizing a number of individuals associated with a data exchange with the first user (Ui) within a resource; said output characteristic vectors (KF3) supplying: • a third classifier characterizing data exchanges (CLD) between at least the two users (Ui, UP); said server (SERVi) being, in addition, configured to generate a list of user profiles associated with a classification of characteristic vectors in each classifier (CLP, CLS, CLD), said list comprising at least one second user (UP) having exchanged data with the first user (Ui), said server being configured to send to a second electronic terminal (T2) an alert message when said second user profile is identified in the generated list.

Images