EP3948579A1 - System and method for enriching data - Google Patents

Abstract

The present invention concerns a method for enriching data, comprising: a) receiving (S1) a plurality of datasets each comprising a fundamental datum and metadata, b) grouping (S3) the datasets according to fundamental data according to a similarity function (F), c) enriching (S4) each dataset with a label characterising the group to which the dataset belongs, d) searching (S5) for each enriched dataset, in a database (DB1, DB2) storing datasets each comprising metadata and a label, for a combination of at least a part of the metadata and the label of the enriched dataset, and e) if the combination of at least a part of the metadata and the label of the enriched dataset is missing from the database, removing (S6) the label from the enriched dataset.

Description

Description

Title: System and process for data enrichment

The field of the invention relates to the enrichment of data, in particular when the latter present the risk of including inaccuracies or errors due to the conditions of transmission and reception of these data. One of the main applications of the invention relates in particular to mobile banking, better known by the English term "mobile banking".

When data is transmitted, there is a risk that the data will be compromised and errors or inaccuracies often result from the transfer of the data. These errors or inaccuracies in the data may be present at the source but may also be introduced on transmission or reception.

The need to make the data received reliable is an important issue in all systems in which the transfer of data, sometimes repeatedly, is inevitable. In particular, the field of mobile banking, which designates all types of financial services accessible from mobile equipment connected to a wide area network, such as a mobile phone for example, is a field in which data transfers are numerous and restitution of this data is a necessary condition for the implementation of the services. In this specific area, the data transferred may include information such as the name of a merchant, his activity code, his location, his name, etc. It is therefore necessary to ensure the reliability of this data when it presents the risk of containing errors for the proper functioning of the services.

The present invention improves the situation.

As such, the present invention relates to a data enrichment method implemented by computer means and comprising:

a) receive several sets of data, a set of data comprising fundamental data and one or more metadata relating to the fundamental data, ...

b) grouping the data sets according to the fundamental data respectively associated with the data sets according to a similarity function, ...

c) enrich each set of data with additional data known as a label characterizing the group to which the set of data belongs,

d) searching for each enriched data set, in at least one database storing data sets each comprising metadata and a label, a combination of at least part of the metadata and the label of the enriched data set , and

e) if the combination of at least part of the metadata and the label of the dataset enriched is missing from the at least one database, remove the label from the enriched data set.

For example, the fundamental data is a series of characters, or a sound signal or a digital image.

According to one embodiment, the method further comprises, following the reception of the data sets: generating for each data set, by applying a processing for reducing a noise level to the fundamental datum, a processed datum associated with the dataset. In such an embodiment, the grouping of the data sets is implemented as a function of the processed data respectively associated with the data sets.

For example, the fundamental data is a sequence of characters and the processed data is generated by deleting the sequence of characters of one or more characters from a list of predetermined characters.

According to one embodiment, the grouping of the datasets uses an unsupervised learning algorithm.

According to one embodiment, each set of data stored in the at least one database further comprises fundamental data and, if the combination of at least part of the metadata and of the label of an enriched data set is present in at least one database in a corresponding data set, the fundamental data of the enriched data set is replaced if necessary by the fundamental data of the corresponding data set.

According to one embodiment, each set of data stored in the at least one database further comprises fundamental data, and the search is carried out on a plurality of databases, each database being characterized by a coefficient of reliability, and, if the combination of at least part of the metadata and the label of the enriched data set is present in corresponding data sets respectively stored in separate databases of the plurality of databases, the label of the enriched data set is removed if the fundamental data of the enriched data set is distinct from the fundamental data of the corresponding data set stored in the database characterized by the greatest reliability coefficient.

According to one embodiment, each set of data stored in the at least one database further comprises fundamental data, and the search is carried out on a plurality of databases, each database being characterized by a coefficient of reliability, and, if the combination of at least part of the metadata and the label of the enriched dataset is present in corresponding data sets respectively stored in separate databases of the plurality of databases, each fundamental data present in at least one of the corresponding data sets is associated with a likelihood factor determined as a function of the reliability coefficient of each database storing a corresponding data set comprising the fundamental data in question, and the label of the enriched data set is removed if the fundamental data of the enriched data set is distinct from the fundamental data associated with the factor highest likelihood.

According to one embodiment, each metadata of an enriched data set being associated with a weight, the combination of at least part of the metadata and of the label is present in a database if and only if a value of one. presence function, calculated as a function of the respective weights of the metadata of the combination present in the database, is greater than or equal to a predetermined threshold.

According to one embodiment, if, after the search in the at least one database, an enriched data set retains its label, the enriched data set is again enriched by data representative of the similarity function and / or at least one database within which the combination of at least part of the metadata and of the label of the enriched data set has been found.

According to one embodiment,: steps b) to e) are repeated for the data sets from which the label has been removed with a new similarity function, so that a data set cannot be enriched by a label already aggregated then removed previously.

For example, the repetition of steps b) to e) is limited to a predetermined maximum number of iterations.

According to one embodiment, the fundamental data relates to an individual or of an entity, and the metadata comprises at least contact data of the individual or of the entity, and in which the enriched data set is transmitted, using contact data, to the individual or entity for verification of the aggregated label.

For example, contact data is a postal address, phone number, email address and / or an address of an application user account.

The present invention also relates to a computer program comprising instructions for implementing the method described above, when the instructions are executed by at least one processor.

Finally, the present invention relates to a data enrichment system comprising:

- a communication module designed to receive several sets of data, a set of data comprising fundamental data and one or more metadata relating to the fundamental data,

- a processing unit designed for:

group the data sets according to the fundamental data respectively associated with the data sets according to a similarity function,

enrich each set of data with additional data known as a label characterizing the group to which said set of data belongs, and

- at least one database configured to store sets of data each comprising metadata and a label,

the processing unit being furthermore arranged for:

search for each enriched data set, in at least one database, a combination of at least part of the metadata and of the label of said enriched data set, and whether the combination of at least part of the metadata and of the label of the enriched data set is absent from the at least one database, removing the label from the enriched data set.

Other characteristics, details and advantages of the invention will become apparent on reading the detailed description below, and on analyzing the accompanying drawings, in which:

- [Fig. 1] illustrates a data enrichment system according to the invention; and

- [Fig. 2] illustrates a data enrichment method according to the invention.

The [Fig. 1] illustrates a data enrichment system, hereinafter SYS system, according to the invention.

The SYS system is designed to receive data presenting the risk of including errors or inaccuracies and to enrich these data despite these potential errors or inaccuracies.

In the context of the invention, the data received by the SYS system is indeed likely to include a certain level of noise. For example, when the data corresponds to a series of characters, this data is liable to contain erroneous characters or inaccuracies. It is therefore understood here that noise typically designates any error introduced into a character sequence at the source, on transmission or on reception or during the transmission of data.

Typically, in the context of mobile banking (also known by the English term “mobile banking”), the data received at the entrance of the SYS system are data allowing a user to access financial services from a mobile device. , for example a cell phone. The data transferred then makes it possible to consult an online account or to make a transfer. The data can correspond to the name of a merchant, to his activity code, to his location, therefore his city, his address and his postal code, or even to his name. In such a field of application, for example, certain information is limited to a maximum number of characters. The transfer of data representative of such information is therefore necessarily imprecise and incomplete since not all the characters could be entered. In the present invention, noise refers, for example, to this type of imprecision.

The SYS system is designed to allow, even when these data contain errors, sometimes introduced at the source but also at the sending or receiving of the data, the provision of the service.

These data can also be a sound signal comprising noise or a digital image comprising digital noise. The SYS system is designed to enrich the data despite this potential noise

As illustrated in [Fig. 1], the SYS system comprises a processing unit UNT and at least one database, here two databases DB1, DB2.

The UNT processing unit is designed to, upon receipt of several data sets, enrich each data set and verify the relevance of this data enrichment using databases DB1, DB2. More specifically, the processing unit UNT is designed to generate, for each set of data received, an additional piece of data called a label and to aggregate or append the label generated to the associated set of data. In the literature, we will also talk about label to designate the label.

Furthermore, the processing unit UNT is also arranged to apply processing to at least part of the data received in order to reduce a level of noise that the data is likely to contain.

The processing unit UNT is furthermore arranged for, once a data set has been enriched, forwarding this data set to an address in order to allow a user to take cognizance of the enriched data and to verify that this data has been correctly enriched. .

In the example illustrated in [Fig. 1], three sets of data DAT1, DAT2, DAT3 are transmitted to the processing unit UNT.

Each set of data comprises a fundamental datum Di, D ₂ , D ₃ and one or more metadata relating to this fundamental datum. Metadata accompanying fundamental data is descriptive data used to describe or define fundamental data. For example, in the field of mobile banking, the fundamental data is the description of a merchant while the metadata characterizes his activity code, his location or any other information concerning the merchant in question. In the example of [Fig. 1], the first set of data DAT1 comprises the fundamental data item Di and further comprises metadata MDi ¹ , MDi ^m . The second set of data DAT2 comprises the fundamental data item D ₂ and further comprises metadata MD ₂ \ MD ₂ ″. Finally, the third set of data DAT3 comprises the fundamental data item D ₃ and further comprises metadata MD ₃ \ MD ₃ ^P. In the previous notations, m, n and p are natural numbers designating the respective number of metadata of the first, second and third data sets DAT1, DAT2, DAT3.

In the context of the invention, the fundamental data of each set of data is likely to present a certain level of noise and therefore to include errors or inaccuracies.

Of course, metadata is also likely to present a certain level of noise. It is understood that the processing applied by the UNT to fundamental data to reduce noise can also be applied to metadata. In the remainder of the description, we focus in particular on the case in which the fundamental data may be noisy. As explained below, the processing unit UNT is designed in particular to generate a piece of data processed by applying a processing for reducing a noise level to the fundamental datum of a set of data.

However, it is understood that the metadata can also be noisy and that the processing unit UNT can also be arranged to generate new metadata by applying a noise reduction processing to the received metadata.

Furthermore, still with reference to the example illustrated in [Fig. 1], three enriched data sets DAT1 *, DAT2 *, DAT3 * are generated by the UNT processing unit. As explained previously, the processing unit UNT is more particularly designed to generate, for each set of data received, additional data also called label or label and to enrich each set of data by aggregating or adding to it the label generated. In the example described here, the first and second data sets DAT1, DAT2 are enriched by a single label label (Ci) while the third data set DAT3 is enriched by a label label (C ₃ ).

The UNT processing unit comprises a COM communication module, a MEM memory and a PROC processor.

The COM communication module is designed to receive several sets of data. In the example illustrated in [Fig. 1], the communication module COM is arranged to receive the first, second and third sets of data DAT1, DAT2, DAT3. Furthermore, the communication module COM is furthermore designed to send several enriched data sets. In the example illustrated in [Fig. 1], the communication module COM is arranged to send the first, second and third enriched data sets DAT1 *, DAT2 *, DAT3 *.

It is known to those skilled in the art that there are many different types of data communication networks, for example radio communication networks, cellular or non-cellular, and that depending on the embodiment, the communication module COM can integrate one or more communication modules, for example radiofrequency communication and be configured for the transmission and reception of radiofrequency signals, according to one or more technologies, such as TDMA, FDMA, OFDMA, CDMA, or one or more standards communications, such as GSM, EDGE, CDMA, UMTS, HSPA, LTE, LTE-A, WiFi (IEEE 802.11) and WiMAX (IEEE 802.16), or their variants or evolutions, currently known or developed later.

In other words, the COM communication module is arranged to communicate with a wide area network (also known by the English acronym WAN for "Wide Area Newtork"), a local network (also known by the English acronym LAN for " Local Area Network ”) or any other type of network.

Data sets are, for example, sent to the COM communication module of the UNT processing unit following the use of an application. Such an application is typically implemented on a terminal, for example a mobile terminal of smartphone type (common English term for a smart phone), and is for example intended to be used by a user. For example, the user makes a payment via this application and this payment generates the generation of at least part of the data of a set of data, whether it is the fundamental data and / or the metadata. It is typically in such a case that noise can be introduced with errors or inaccuracies. For example, this information is a series of characters. The memory MEM is arranged to store instructions in the form of a computer program whose execution by the processor PROC results in the operation of the processing unit UNT.

The operation of the PROC processor and therefore of the processing unit UNT will be described in more detail in the remainder of the description with reference to [Fig. 2].

As explained previously, the SYS system also includes at least one database. In the example illustrated in [Fig. 1], the SYS system includes two databases DB1, DB2. Nevertheless, those skilled in the art understand here that the SYS system can only include a single database. Each DB1, DB2 database is configured to store data sets each including metadata and a label. Advantageously, one or more data sets stored in a database DB 1, DB2 also include, in addition to metadata and a label, fundamental data.

Additionally, each DB1, DB2 database is configured to be accessible to the UNT processing unit within the SYS system. As explained in the remainder of the description, this accessibility results in the possibility for the processing unit UNT to perform a search within each database DB1, DB2 to establish, if possible, a correspondence between a set of enriched data and data sets stored in databases. This research aims in particular to verify the relevance of the enrichment of the set of data produced and its conformity with known databases.

As explained above, one of the fields of application of the present invention is mobile banking. In such a context, the databases sent by the processing unit UNT to verify that a set of data received has been correctly enriched is for example a database of the SIREN type (for "Identification system of the directory of companies ”), SIRET (for“ Identification system of the establishment directory ”) or even Infogreffe. These public databases provide access to data relating to the identification of a company, a company, an establishment, an organization or an association with activities in France. Of course, those skilled in the art will understand that DB1 and DB2 databases can refer to any database of this type and not only for France. Of course, databases DB1, DB2 can also refer to other types of databases accessible by programming interfaces (also known by the acronym API for "Application programming interface").

A method for enriching data according to the invention will now be described with reference to

[Fig. 2]

In the context of the implementation of the method described below, several data sets are sent to the SYS system as described previously with reference to [FIG. 1].

These data are transmitted to the SYS system for example via a wide area network of the Internet type or via a local network. This method is typically implemented for the purpose of providing a service within the framework of mobile banking. A set of data then relates, for example, to a company, a company or a merchant and the metadata included in the set of data are informative or descriptive data of a fundamental data concerning the name of the company, the company or the merchant. This fundamental data is, due to the transfer of the data set, likely to include errors or inaccuracies and therefore to be corrupted by a certain noise level. This noise may have been introduced at the source, on transmission or even on reception.

During an SI step, the SYS system receives several sets of data. More specifically, the data sets are received by the COM communication module of the SYS system UNT processing unit.

As explained previously, in the field of mobile banking, the generation of such data sets is for example triggered by an application on a user's mobile terminal.

In the example illustrated in [Fig. 1], the communication module COM receives a first set of data DAT1, a second set of data DAT2 and a third set of data DAT3. Of course, this example is purely illustrative and the SYS system may have to process a much larger number of data sets.

Each data set comprises a fundamental data and one or more metadata relating to the fundamental data. In other words, metadata makes it possible to define, describe or provide additional information about the fundamental data.

In the example described here, the first set of data DAT1 comprises metadata MDi ¹ , ..., MDi ^m describing the fundamental data item Di. The second data set DAT2 comprises metadata MD ₂ \ ..., Ml) ₂ "describing the fundamental data item D _2. Finally, the third data set DAT3 comprises metadata MD3 ¹ , ..., MD ₃ ^P describing the fundamental data D ₃ .

Optionally, during a step S2, the processing unit UNT of the system SYS generates for each set of data, by applying a processing for reducing a noise level to the fundamental datum, a processed datum associated with the 'dataset.

Indeed, as explained previously, due to the transfer of the data set, noise can be introduced at the source, on the transmission or on the reception in the data set and more specifically in the fundamental data. The implementation of the service requiring the correct routing of the data set is then compromised by such errors or inaccuracies.

During this step, the UNT processing unit applies any type of data processing allowing the noise level level of the fundamental data to be reduced. Those skilled in the art are familiar with the techniques usually employed to decrease the noise level or completely remove it from one or more data.

For example, when the fundamental data is a sequence of characters, the processed data is generated by deleting the sequence of characters of one or more characters from a list of characters predetermined. This list of characters is for example stored in the memory MEM of the processing unit UNT so that, when the processing unit detects a character from this list in a fundamental datum taking the form of a series of characters, this character is deleted to generate the processed data.

Of course, the fundamental data can also be a sound signal or a digital image. Again, the various techniques for reducing or eliminating noise in a sound signal or a digital image are widely known to those skilled in the art so that the UNT processing unit can be configured to be able to apply such techniques on the fundamental data of each set of data received by the SYS system.

In the present case, Di ′ denotes the processed data item generated by the first set of data DAT1 by reducing the noise level of the fundamental data item Di. Likewise, D ₂ 'denotes the processed data item generated by the second set of data DAT2 by reducing the noise level of the fundamental datum D ₂ , and D ₃ ' denotes the processed datum generated by the third set of data DAT3 by reduction of the noise level of the fundamental datum D ₃ .

At this stage of the process, the data processed for a data set can be aggregated or appended to the data set in addition to or in place of the fundamental data, and in the company of the corresponding metadata.

In the remainder of the description of the method, it is considered that this step S2 has been implemented and that the processed data item replaces the fundamental data. Those skilled in the art understand here that in the embodiment in which this step is not implemented, this processed data item is therefore not generated and what is carried out subsequently using the processed data item l 'is using the fundamental data. Moreover, in the case where the processed data is generated, this processed data can be identical to the fundamental data. In particular, if the fundamental datum does not include any noise, the processed datum is identical to the fundamental datum.

With reference to [Fig. 2], from now on, a counter i, initialized to 1, is incremented and a similarity function I j is selected. Advantageously, the memory MEM stores a set of similarity functions.

During a step S3, the processing unit UNT groups the data sets according to the processed data respectively associated with the data sets according to the similarity function. The grouping of data implemented by the processing unit UNT is better known under the English term “data clustering” or more simply “clustering”. We can also speak here of partitioning or clustering of data. Here again, the grouping techniques used by the processing unit UNT are techniques known to those skilled in the art. Advantageously, the grouping implemented by the processing unit UNT makes it possible to obtain a great intra-group similarity, namely a high homogeneity between the elements, here data sets, of the same group, and a low similarity inter-group, in order to have well-differentiated groups.

The techniques of grouping or “clustering” used within the framework of the implementation of the method are varied. Advantageously, the grouping implemented by the UNT processing unit comprises a partitioning algorithm, a hierarchical algorithm, a density-based algorithm, a grid algorithm or even a model algorithm.

Advantageously, the grouping of data sets uses an unsupervised learning algorithm. Such algorithms are known to those skilled in the art.

At the end of this step, therefore, the data sets are grouped into groups, better known under the English term "clusters", according to the similarity function used.

Typically, the similarity function is a distance function defined on a space of M + l dimensions, where M is the number of metadata (M + l therefore corresponding to the cardinality of a set of data received with M metadata and a fundamental data ). For example, if the data is numeric data, the similarity function can be a Euclidean distance. If the data is characters, typically letters, the similarity function can be a Levenshtein distance. Of course, when some data in a dataset is numeric while others are letters, the similarity function can be a combination of a Euclidean distance and a Levenshtein distance. We can then define a threshold so that, when the distance between two sets of data is less than or equal to this threshold, then the two sets of data are grouped together, thus forming part of the same cluster.

Referring again to the example illustrated in [Fig. 1], the first set of data DAT1 and the second set of data DAT2 are grouped together in the same group or “cluster” Ci. The third set of data DAT3 is for its part placed in a group C ₂ . The first, second and third data sets DAT1, DAT2, DAT3 have been grouped together according to their respective processed data Di ', D ₂ ', D ₃ '.

As explained previously, the generation of the processed data item is optional. Thus, when no processed data has been generated, the grouping of the datasets is implemented according to the respective fundamental data of the datasets.

During a step S4, the processing unit UNT enriches each set of data with an additional piece of data called a label characterizing the group to which the set of data considered belongs. In other words, a set of data receives, at the end of the grouping, a data additional characterization of the group into which the data set in question has been classified. We can consider that this additional data, also called label or tag, is aggregated or appended to the data set.

Referring again to the example illustrated in [Fig. 1], the first and second data sets DAT1, DAT2 have been classified in the same group or “cluster” Ci. These two data sets DAT1, DAT2 are therefore enriched by the same additional data item referenced label (Ci). Likewise, the third set of data DAT3 having been classified in the group or "cluster" C ₂ , it is enriched by the additional data item label (C ₂ ).

During a step S5, the processing unit UNT searches, for each enriched data set, in at least one database storing data sets each comprising metadata and a label, a combination of at least part metadata and the label of the enriched data set considered.

In the case of illustration developed here, for example, the first enriched set DAT1 comprises the fundamental data Di, metadata MDi ¹ , ..., MDi ^m , a label label (Ci) and, optionally, the processed data Di ' . The search performed by the processing unit UNT in at least one of the databases DB1, DB2 therefore aims to determine whether the combination of at least part of the metadata MDi ¹ , ..., MDi ^m and the label label ( Ci) is present in a dataset among the datasets stored in the database DB1, DB2. Advantageously, such a search is carried out in all the databases, therefore here the database DB 1 and the database DB2.

In the remainder of the description, we will speak of a "corresponding data set" to designate a data set stored in a database and comprising the desired combination. We say that this dataset is a corresponding dataset of the enriched dataset from which the sought combination is derived.

With reference again to the method illustrated in [Fig. 2], there are then two possibilities for an enriched data set:

- either such a combination of metadata and label is present in at least one database;

- or such a combination is absent.

Advantageously, in one or more embodiments, each metadata of an enriched data set is associated with a weight. This weight makes it possible to characterize the importance of a metadata within a set of data. The combination of at least part of the metadata and the label is then considered to be present in a database if and only if a value of a presence function, calculated according to the respective weights of the metadata of the combination present in the database in question is greater than or equal to a predetermined threshold.

In other words, in this specific embodiment, an additional criterion is applied to determine whether a dataset stored in a database can be considered a "corresponding dataset". This criterion consists of verifying whether a potential corresponding dataset is sufficiently meaningful, according to the metadata it contains and shares in common with an enriched dataset. Of course, the need for the label of this potential corresponding data set to be the same as the enriched data set considered remains in this specific embodiment.

As an example, we consider the first enriched data set DAT1 *. It is assumed that the metadata MDi ¹ , ..., MDi ^m are all respectively associated with a weight Pi ¹ , ..., Pi ^m . It is also assumed that, in the database DB1, a set of data includes the metadata MDi ¹ , ..., MDi ^k and the label label (Ci), where k is a natural integer strictly less than m. In other words, this data set found in the database DB1 does indeed include at least part of the metadata of the first enriched data set DAT1 * as well as the label label (Ci). So this is a potential matching dataset.

However, in this embodiment, in order to determine whether this potential corresponding data set is relevant, the criterion explained above is applied in addition by calculating the value Vi taken by the presence function, denoted G below, for the metadata of the first enriched data set DAT1 *, and more exactly for the weights respectively associated with these metadata. In other words:

Vf = G (Pi \ ..., Pi ^k )

This value Vi is then compared with a predetermined threshold and, if this value is greater than or equal to the predetermined threshold, then the dataset found in the database DB1 is relevant and is retained as a corresponding dataset.

For example, the presence function G is an addition or a multiplication.

Still by way of example, it is also assumed that, within the DB2 database, another combination of at least part of the metadata and of the label of the first enriched data set DAT1 * is found in a set of data, so potentially a corresponding data set. The metadata common to this data set and to the first enriched data set DAT1 * can of course be different from those found in the database DB1. For example, the dataset found in the DB2 database includes the metadata MD, MDi ^m , where j is a natural number less than m, and the label label (Ci).

As for the previous example, the processing unit UNT then calculates the value V ₂ taken by the function G taken for this combination found. In other words:

V ₂ = G (P ₁ ^j , ..., P ₁ ^k )

This value V ₂ is then compared with the predetermined threshold and, if this value is greater than or equal to the predetermined threshold, then the data set found in the database DB2 is relevant and is retained as a corresponding data set.

For example, the dataset found in the DB1 database is retained according to this criterion while the one found in the DB2 database is not.

In the rest of the description of the method, it is understood that, in this embodiment, a corresponding data set is not only a data set stored in a database comprising the combination of at least part of the metadata and the label of an enriched data set but also a data set verifying the criterion described above concerning the respective weights of the metadata that it shares with the enriched data set on the basis of which the search is carried out by the UNT processing.

During a step S6, implemented in particular in the case where a combination of at least part of the metadata and of the label of an enriched data set is absent from the at least one database, the label previously assigned is removed from the enriched dataset. In other words, since there is no trace in any database of a combination of at least part of the metadata and the label, it is considered that it was by mistake that it was assigned during the grouping of step S3 to the data set considered. This previously enriched is therefore removed from the additional data or label which has been aggregated or added to it.

With reference again to the case illustrated in [Fig. 1], we consider for example the second group of data DAT2. At the end of step S4, this has been enriched by the additional data item label (Ci). The processing unit UNT therefore then searched, during step S5, in at least one of the databases DB 1, DB2 if a set of data stored in one of these databases DB 1, DB2, comprises both at least part of the metadata M1) ₂ ', ..., MD ₂ "and the additional data label (Ci). If no data set stored in the databases DB1, DB2 does not include such a combination, the label label (Ci) is therefore removed from the second enriched data set DAT2 *. The latter then no longer comprises, at the end of this step S5, only the fundamental data D2, the metadata MD ₂ \ ..., MD ₂ "and, optionally, the processed data item D ₂ '. Alternatively, with reference to [Fig. 2], if the combination of at least part of the metadata and the label of an enriched data set is indeed present in at least one database, it is determined whether this combination has been determined in a single database or in several databases. Of course, in an embodiment in which a single database is integrated into the SYS system and is accessible to the processing unit UNT, a set of data corresponding to the combination of at least part of the metadata and label searched can only be found in this database alone.

Nevertheless, in the case illustrated for example in [Fig. 1], such a search can be performed on a plurality of databases, here two databases DB1, DB2 and a corresponding set of data can be found in several different databases. For example, concerning the third enriched data set DAT3 *, the processing unit UNT searched in the database DB1 but also in the database DB2 for a data set comprising the combination of at least part of the metadata MD ₃ ¹ , ... MD ₃ ^P and the label label (C ₂ ). It is quite possible that a matching dataset was found in the DB1 database, while another matching dataset was found in the DB2 database. In other words, the processing unit UNT has found a data set stored in the database DB1 comprising the combination of at least part of the metadata and the label of the third enriched data set DAT3 * but also found a dataset stored in the DB2 database including this same combination of metadata and label.

In such a case, during a step S7, the processing unit UNT applies a predefined criterion to determine whether the result of this search which resulted in finding a corresponding set of data in more than one database of the system SYS makes it possible to conclude on the relevance of the label attributed or not.

During this step S7, there are two embodiments:

In a first embodiment, each database is characterized by a reliability coefficient. In addition, each set of data stored in a database further comprises fundamental data.

We then compare the processed data of the considered data set and the fundamental data of the corresponding data set stored in the database characterized by the highest reliability coefficient.

As explained previously, we are here in the particular case where, for each set of data, a piece of data processed has been generated during the optional step S2 and replaces the fundamental data for the implementation of the method. It is therefore quite clear here that, if the processed data item has not been generated, during this step S7, the fundamental data item of the data set is compared. considered and the fundamental data of the corresponding data set stored in the database characterized by the greatest reliability coefficient.

For example, still referring to [Fig. 1], in which the SYS system includes two databases DB1, DB2. Since there are several databases, each is assigned a reliability coefficient to quantify its relevance or reliability. Advantageously, the respective reliability coefficients of two distinct databases are distinct. Thus, the database DB1 is characterized by a reliability coefficient CFI while the database DB2 is characterized by a reliability coefficient CF2. It is also considered that the database DB 1 being more reliable than the database DB2, we have: CF1> CF2.

It is then assumed that the combination of at least part of the metadata MDi ¹ , ..., MDi ^m and of the additional data item label (Ci) of the first enriched data set DAT1 * has been found in a corresponding stored data set in the DB 1 database but also in another corresponding data set stored in the DB2 database. These two corresponding sets therefore each comprise a fundamental datum. These respective fundamental data may be similar or different.

Since, among the databases storing a corresponding data set, the database with the highest reliability coefficient is the database DB 1 characterized by the reliability coefficient CFI. During this step S7, in this first embodiment, the processed data Di ’of the first enriched data set DAT1 * is therefore compared with the fundamental data of the corresponding data set found in the database DB1.

With reference to the data enrichment process illustrated in [Fig. 2], if the combination of at least part of the metadata and the label of an enriched data set is present in corresponding data sets respectively stored in databases distinct from the plurality of databases, the label of the enriched data set is removed during step S6 then implemented by the processing unit UNT if the processed data of the enriched data set is distinct from the fundamental data of the corresponding data set stored in the database characterized by the highest coefficient of reliability.

In a second embodiment, each database is again characterized by a reliability coefficient. In addition, each set of data stored in a database further comprises fundamental data.

This time, in this alternative embodiment, we do not consider only the database having the highest reliability coefficient among all the databases comprising a set of data corresponding to the sought combination. In this embodiment in particular, the processing unit UNT takes into account all the databases comprising a corresponding set of data.

Each fundamental data item present in at least one of the corresponding data sets is associated with a likelihood factor determined as a function of the reliability coefficient of each database storing a corresponding data set comprising the considered fundamental data.

To illustrate this embodiment, reference is again made to the example illustrated in [Fig. 1]. It is further assumed that a third database (not shown here) is included in the SYS system and is searched by the processing unit UNT in addition to the databases DB1, DB2. Finally, it is assumed that this third database is characterized by a reliability coefficient CF3. It is also assumed for this embodiment that the database DB1 is characterized by a reliability coefficient CFI while the database DB2 is characterized by a reliability coefficient CF2. Advantageously here too, the respective reliability coefficients of two distinct databases are distinct.

It is then assumed that the combination of at least part of the metadata MD3 ¹ , ..., MD ₃ ^P and the additional data item label (C ₂ ) of the third enriched data set DAT3 * has been found in a data set correspondent stored in the database DB 1 but also in a corresponding dataset stored in the DB2 database and in a corresponding dataset stored in the third database. These three corresponding sets therefore each comprise a fundamental datum. These respective fundamental data may be similar or different.

It is assumed by way of example that the corresponding data set stored in the database DB 1 and the corresponding data set stored in the database DB2 comprise the same fundamental data, denoted hereinafter DFi ₂ . On the other hand, the corresponding data set stored in the third database comprises a different fundamental datum, denoted hereinafter DF ₃ . The processing unit UNT then determines a likelihood factor FV (DF _{I 2} ) associated with the fundamental datum DFi ₂ . This likelihood factor FV (DF _{I 2} ) is calculated as a function of the reliability coefficients of the database DB1 and of the database DB2, namely CFI and CF2. Likewise, the processing unit UNT determines a likelihood factor FV (DF ₃ ) associated with the fundamental datum DF ₃ . This likelihood factor FV (DF ₃ ) is calculated as a function of the reliability coefficient of the third database, namely CF3.

For example, a likelihood factor is determined by adding the reliability coefficients. We then have:

FV (DF _U ) = CF1 + CF2 FV (DF ₃ ) = CF3

Another possibility can also be to calculate the average of the reliability coefficients to calculate the likelihood factor. We then have:

FV (DF) = (CFl + CF2) / 2

FV (DF ₃ ) = CF3

Those skilled in the art understand here that several calculations are possible to determine the likelihood factor of a fundamental datum as a function of the reliability coefficients.

In the example developed here, the processed data D ₃ 'of the third enriched data set DAT3 * is then compared with the fundamental data associated with the highest likelihood factor.

As explained above, we are here in the particular case where, for each set of data, a piece of processed data was generated during the optional step S2 and replaces the fundamental data for the implementation of the method. It is therefore clear here that, if the processed data item has not been generated, during this step S7, the fundamental data of the set of data considered and the fundamental data associated with the highest likelihood factor are compared. .

With reference to the data enrichment process illustrated in [Fig. 2], if the combination of at least part of the metadata and the label of the enriched data set is present in corresponding data sets respectively stored in separate databases of the plurality of databases, each data item fundamental present in at least one of the corresponding data sets is associated with a likelihood factor determined as a function of the reliability coefficient of each database storing a corresponding data set comprising the fundamental data considered, and the label of the set of enriched data is removed during step S6 then implemented by the processing unit UNT if the processed data of the enriched data set is distinct from the fundamental data associated with the highest likelihood factor.

Step S6, as specified previously, is implemented for an enriched data set either following step S5 if it turns out that the combination of at least part of the metadata and the label of this enriched data set is not present in any database, either following step S7 if this combination has been found in several databases and it turns out that the assigned label is incorrect. During this step S6, therefore, the label of the enriched data set is removed. Then, as illustrated in [Fig. 2], it is determined whether the counter i characterizing the number of iterations of the method is less than or equal to a predetermined maximum number of iterations N. If this maximum number of iterations has not yet been reached, the counter is incremented . Advantageously, when the counter is incremented, a new similarity function, for example stored in the memory MEM of the processing unit UNT, is selected. Steps S3 and following are then repeated for the data sets whose label has been removed with the new similarity function, so that a data set cannot be enriched by a label already aggregated and then removed previously.

As explained previously, a similarity function calculates a distance between two sets of data so that two sets of data are grouped together in the same group or cluster when the distance between these two sets of data is less than or equal to a certain threshold. . Also, when a new similarity function is selected, it is also possible to modify this threshold, for example by increasing it. Furthermore, it is also possible to keep the same similarity function and only change the threshold.

On the other hand, if the predetermined maximum number of iterations N of the method has been reached, the processing unit UNT interrupts the loop and proceeds to step S8 even though some data sets are found without an assigned label.

Step S8 is implemented at the end of step S7 if it is determined that the label attributed to a set of data during the enrichment of the latter is correct in view of the search carried out on T at least one database, it is then considered that this data set has been correctly enriched.

Step S8 can also be implemented if the maximum number of predetermined iterations N of the method has been reached.

Furthermore, with reference to [Fig. 2], step S8 can also be implemented in the case where, for an enriched data set, the combination of at least part of the metadata and of the label of this enriched data set has only been found. in a single database at the end of step S5. In such a case, during step S8 then implemented by the processing unit UNT of the system SYS, the fundamental data of the enriched data set is replaced if necessary by the fundamental data of the data set corresponding. By "necessary" is meant here that the fundamental data of the corresponding data set is distinct from the enriched data set. Moreover, this fundamental data present in the corresponding dataset may correspond to the processed data.

We understand here the enriched data set at the output of the system comprises at least either the original fundamental data, the processed data or the fundamental data found in the corresponding data set. With reference to [Fig. 1], it is assumed for example that the combination of at least part of the metadata MD ₂ \ MD ₂ "and of the label label (Ci) from the second enriched data set DAT2 * has been found only in a database, for example the database DB1, therefore in a single corresponding data set during step S5. During step S8 immediately implemented thereafter, the processing unit UNT then compares the fundamental datum D ₂ of the second enriched data set DAT2 * is compared with the fundamental data of the corresponding data set stored in the database DB1. If the fundamental data of the corresponding data set is distinct from the fundamental data of the second data set enriched, the latter is then replaced in the second enriched data set by the fundamental data of the corresponding data set.

At the end of step S8, therefore, each set of data has benefited from at most N iterations of steps S3 and following to be assigned a consistent label in view of the search carried out on one or more databases of the system. SYS. Some datasets are, at the end of this step, still enriched with an additional data or label while other datasets may remain unlabeled.

Moreover, concerning the enriched data sets which, at the end of step S8, have retained their label because of the positive result of the search in the database or databases DB1, DB2, or because the combination d '' at least part of the metadata and the label were found in a single database, either because this combination was found in several databases and the label finally appeared correct in the light of the search, such enriched sets may also be supplemented by new metadata from databases.

Indeed, as explained previously, the test consists in determining whether, for a set of data, the combination of at least part of the metadata and of the label generated are included in at least one set of data, called the corresponding set of data, at least one database. But such corresponding sets can of course include other data in addition to the desired combination. This additional metadata can then be retrieved by the UNT processing unit to advantageously complement the enriched data sets.

In the example illustrated in [Fig. 1], the enriched datasets DAT1 *, DAT2 *, DAT3 * do not include additional metadata compared to the datasets DAT1, DAT2, DAT3 received by the system. Nevertheless, those skilled in the art understand here that the enriched data can comprise additional metadata originating from databases DB1, DB2. Still during step S8, advantageously, the enriched data sets can be enriched again so as to keep, for the sake of traceability, a history of the enrichment of the data and of the search within the databases. For example, an enriched data set can be completed by a piece of data representative of the similarity function used to implement the grouping in step S2.

Still advantageously, an enriched data set can also or alternatively be supplemented by data representative of the database within which the most relevant corresponding set has been found. Of course, in the more complex cases explained previously in which the combination of at least part of the metadata and the label of an enriched set has been found in several databases, the data added to the enriched data set in the during step S8 may be representative of at least part of the databases within which these corresponding data sets are stored.

Thus, a set of enriched data at the output of the SYS system can include, in addition to the label and possibly the data processed with or in place of the original fundamental data, data making it possible to characterize the different steps of the process that led to generation and verification of enriched data sets. This additional enrichment of a set of data typically comprises data representative of the similarity function used and / or one or more data representative of databases in which corresponding sets are stored.

In other words, if, after the search in the database (s), an enriched data set retains its label, the enriched data set in question is again enriched by data representative of the similarity function and / or at least one database within which the combination of at least part of the metadata and the label of this enriched data set has been found.

During a step S9, optionally implemented at the end of step S8, the metadata of the enriched data sets are used in order to carry out a verification of the assigned label. Such a check can also make it possible to correct the fundamental data if necessary.

According to one embodiment, for example in the specific field of mobile banking, the fundamental data relates to an individual or of an entity, and the metadata comprises at least contact data of the individual or of the entity. . The enriched data set is transmitted, using contact data, for verification of the aggregated label. An entity can designate here a company, a company, an organization or an establishment.

As explained previously, such contact data may already be present in the data set received and then enriched, but may also be retrieved from one of the databases. DB1 data, DB2 if the search result is satisfactory. Typically, when the metadata of an enriched data set at the end of step S4 does not include contact data allowing subsequent transmission of the enriched data set for verification, such contact data is sought. in the corresponding set or sets within the database or databases.

With reference to [Fig. 1], the enriched data sets are processed before transmission in order to keep either the fundamental data as received by the communication module COM or the data processed or the fundamental data recovered from a database. For example, in the first enriched data set DAT1 *, the fundamental data item Di received is kept alone. In the second enriched data set DAT2 *, the processed data item D ₂ 'generated is kept alone. Likewise, in the third enriched data set DAT3 *, the processed data item D ₃ 'generated is kept alone.

The contact data can be for example a postal address, a telephone number and / or an e-mail address. In the example described here, it is considered for example that the metadata of the first enriched data set comprises contact data relating to an electronic address ADD1, the metadata of the second enriched data set include contact data relating to a telephone number. telephone ADD2 while the metadata of the third enriched data set comprises contact data relating to a postal address ADD3.

With regard to these addresses, it should also be noted that an enriched data set may be, for verification purposes, obviously transmitted to the individual or entity being the subject of these data but may also be sent to the source of the dataset. For example, by considering again the field of application of mobile banking, the generation of a set of data and then the transmission to the SYS system may have been triggered by a user's terminal, for example during a payment. . More precisely, this data is generated from a user account of the user on the payment application. These data do not relate to the user in question but to the trade, business or society. During step S9, the enriched data set can therefore be transmitted for verification of course to the business, the company or the company via contact data included in the metadata, but can also, still for verification, also be sent to the user account at the origin of the generation of the data set as received by the SYS system and more particularly by the communication module COM.

The enriched data sets are then transmitted to these addresses provided by the contact data, for example via the communication module COM, for verifications of the label, and possibly of the fundamental / processed data and of the data. 'enriched data set transmitted. In particular, the processing unit UNT is for example provided with technologies to automatically send an email or use a call bot to automatically call the retrieved phone number.

Of course, if at least part of the data in a data set is found to be erroneous after verification, this erroneous data can be corrected and then sent back to the SYS system. For example, still in the case where a set of data has been generated following the payment made by a user via a payment application executed on a terminal, this application also allows him to receive the enriched set of data at the output of the system and to access, at least in part, certain data in the enriched data set for verification purposes. If a data, for example the fundamental data or the processed data or a metadata, is erroneous, the user has the possibility to correct this data then to send this correction to the SYS system.

Upon receipt of a corrected enriched data set, the SYS system can then re-implement certain steps of the method described above. For example, the SYS system can re-group or cluster on multiple corrected rich data sets or re-search one or more databases.

Claims

Claims 1. A method of enriching data implemented by computer means and comprising: a) receiving (SI) several sets of data (DAT1, DAT2, DAT3), a set of data comprising fundamental data (Di, D ₂ , D ₃ ) and one or more metadata relating to said fundamental data (MDi ¹ , MDi ^m , MD ₂ \ MD ₂ ", MD3 ¹ , MD ₃ ^P ), b) grouping (S3) the data sets according to the fundamental data respectively associated with said data sets according to a similarity function (F), c) enrich (S4) each data set with an additional data (label (Ci), label (C ₂ ), label (C ₃ )) called label characterizing the group to which said data set belongs, d) search (S5) for each enriched data set, in at least one database (DB 1, DB2) storing data sets each comprising metadata and a label, a combination of at least part of the metadata and the label of said data set enriched, and e) if said combination of at least part of the metadata and the label of the enriched dataset is absent from the at least one database, removing (S6) the label from the dataset enriched. 2. The method of claim 1, wherein the fundamental data is a sequence of characters, or a sound signal or a digital image. 3. Method according to claim 1 or 2, further comprising, following the reception of the data sets: generating (S2) for each data set, by applying a processing for reducing a noise level to the data. fundamental, a processed data item (Di ', D ₂ ', D ₃ ') associated with said data set, the grouping of the data sets being implemented as a function of the processed data respectively associated with said data sets. 4. Method according to claim 3, in which the fundamental data is a series of characters and the processed data item is generated by deleting said series of characters from one or more characters from a list of predetermined characters. 5. Method according to one of the preceding claims, in which the grouping of the data sets uses an unsupervised learning algorithm. 6. Method according to one of the preceding claims, wherein each set of data stored in the at least one database further comprises fundamental data and, if the combination of at least part of the metadata and the label d 'an enriched data set is present in the at least one database in a corresponding data set, the fundamental data of the enriched data set is replaced if necessary by the fundamental data of the corresponding data set. 7. Method according to one of the preceding claims, in which each set of data stored in the at least one database further comprises fundamental data, in which the search is carried out on a plurality of databases, each base. data being characterized by a reliability coefficient, and, if the combination of at least part of the metadata and the label of the enriched data set is present in corresponding data sets respectively stored in separate databases from the plurality of databases, the label of the enriched data set is removed if the fundamental data of said enriched data set is distinct from the fundamental data of the corresponding data set stored in the database characterized by the most high reliability coefficient. 8. Method according to one of claims 1 to 6, in which each set of data stored in the at least one database further comprises fundamental data, in which the search is carried out on a plurality of databases, each database being characterized by a reliability coefficient, and, if the combination of at least part of the metadata and of the label of the enriched dataset is present in corresponding datasets respectively stored in databases distinct from the plurality of databases, each fundamental datum present in at least one of the corresponding datasets is associated with a likelihood factor determined as a function of the reliability coefficient of each database storing a corresponding dataset comprising said datum fundamental, and the label of the enriched data set is removed if the fundamental data of said enriched data set is is distinct from the fundamental data associated with the highest likelihood factor. 9. Method according to one of the preceding claims, each metadata of an enriched data set being associated with a weight, in which the combination of at least part of the metadata and of the label is present in a database if and only if a value a presence function (G), calculated as a function of the respective weights of the metadata of said combination present in said database, is greater than or equal to a predetermined threshold. 10. Method according to one of the preceding claims, in which if, after the search in the at least one database, an enriched data set retains its label, said enriched data set is again enriched by representative data. of the similarity function and / or of at least one database within which the combination of at least part of the metadata and of the label of said enriched data set has been found. 11. Method according to one of the preceding claims, in which steps b) to e) are repeated for the data sets from which the label has been removed with a new similarity function, so that a data set cannot be enriched by a label already aggregated then withdrawn previously. 12. The method of claim 10, wherein the repetition of steps b) to e) is limited to a predetermined maximum number of iterations. 13. Method according to one of the preceding claims, in which the fundamental data relates to an individual or of an entity, and the metadata comprises at least contact data of said individual or of said entity, and in which the whole of enriched data is transmitted (S9), using contact data, to said individual or to said entity for verification of the aggregated label. 14. The method of claim 13, wherein the contact data is a postal address (ADD3), a telephone number (ADD2), an electronic address (ADD1) and / or an address of a user account of an application. . 15. Computer program comprising instructions for implementing the method according to one of the preceding claims, when said instructions are executed by at least one processor (PROC). 16. Data enrichment system (SYS) comprising: a communication module (COM) designed to receive several sets of data, a set of data comprising a fundamental datum and one or more metadata relating to said fundamental datum, - a processing unit (UNT) designed to: group the data sets according to the fundamental data respectively associated with said data sets according to a similarity function, enrich each data set with an additional data called label characterizing the group to which belongs said set of data, and - at least one database (DB1, DB2) configured to store data sets each comprising metadata and a label, the processing unit being further arranged to: search for each enriched data set, in the au minus a database, a combination of at least part of the metadata and the label of said enriched data set, and if the combination of at least part of the metadata and the label of the enriched data set is absent from at least one database, remove the label from the enriched dataset.