WO2016092218A1

WO2016092218A1 - Means for determining a level of relevance of a resource in an information-processing system

Info

Publication number: WO2016092218A1
Application number: PCT/FR2015/053418
Authority: WO
Inventors: Antoine Vincent BOISARD; Vincent Martin; Laurent PROVENAT; Emmanuel BRUNO; Elisabeth MURISASCO
Original assignee: Université De Toulon; Coexel
Priority date: 2014-12-10
Filing date: 2015-12-10
Publication date: 2016-06-16
Also published as: FR3030079A1

Abstract

The invention relates to means for personalizing, for a user, an access procedure to access a set of resources accessible via the intermediary of a information-processing system, comprising: collecting information relating to events occurring in connection with the set of resources; creating, as a function of the information collected relating to events occurring in connection with the set of resources, or obtaining a first profile associated with the user, the profile comprising: at least one context suitable for allowing representation of the set of conditions able to influence a choice of the user and/or a judgement of the relevance of the user relative to the resources of the set of resources accessible via the intermediary of the information-processing system; and, a first set comprising at least one resource initially suitable for allowing the determination, for each resource of the set of resources, of the level of relevance for the user according to at least one measurable property of said resource; and, a second set comprising at least one reference to a second profile associated with a reference user for the user; calculating a probability that the first resource, from the set of resources, is relevant for the user as a function of said first profile, and implementation of the access procedure as a function of said probability.

Description

Means for determining a level of relevance of a resource in an information processing system

The present invention relates to the field of systems for the personalized processing of information, and more particularly, but not exclusively, to that of information retrieval systems, search for reference users (for example experts), compiling a bibliography using collective knowledge, identifying sources of information or even researching data, for example medical data. It relates more particularly to user profile management means, as well as information processing means leveraging said user profiles.

The treatment of digital information in a specific field requires a dual competence: a first in the field and a second in the use of tools made available by the information processing system. In the example of a system for searching information in a database or documents, a user wishing to obtain a list of links to relevant resources for a given subject, must know and master the commands, the particular operators ( Boolean operators, specific filters, etc.), and their syntaxes, to make the best use of the capabilities of said information retrieval system. It is therefore not easy to obtain optimized results for a given user, the quality of the results obtained being a function of the user's ability to formulate optimally and adequately a relevant query.

In addition, we note that requests made by users are often short and therefore ambiguous. Thus, in the case of a Web search engine, it was observed that the average size of a query was 3.08 terms, as indicated in the document written by Mona Taghavi, Ahmed Patel, Nikita Schmidt, Christopher Wills, and Yiqi Tew, titled "An analysis of a web proxy with query distribution pattern approach for search engines", Comput. Stand. Interfaces, 34 (1): 162-170, January 2012.

In an attempt to answer this problem, it is known, for example from US Patent No. 8,538,955, to evaluate the relevance of the results obtained for a user by partially taking into account the online activity and returns. reactions from the latter on different websites, as well as from third-party users sharing the user's tastes / opinions / preferences common. However, such digital information processing systems are today insufficiently adapted to take into account the explicit and implicit information search needs of the users. This is especially true when a user has a very good knowledge of a part of his search domain.

This is why there is still a need for user profile management means, as well as information processing means leveraging said user profiles, in particular to exploit the skills of a community in an access process. to the information, and able to allow the joint consideration of a context and a priori specific to each user.

One of the objects of the invention is to make it possible, from data sources, to extract relevant information for a user. Another object of the invention is to allow the extraction of relevant information for a user, without previously requiring learning phases or specific configuration to said user. Another object of the invention is to enable the extraction of relevant information from various types of resources, for example from documents, information collected by other users, flows and / or sources of data. heterogeneous information, etc. Another object of the invention is to allow taking into account the context of the search of the user. Another object of the invention is to allow the integration of explicit preferences of users, difficult or not automatically extractable. Another object of the invention is to make it possible, from data sources such as databases or documents, to extract relevant information for a user, even when the latter has limited skills for formulating information. queries. Another object of the invention is to make it possible, from data sources, to extract relevant information for a user, by relying on reference users. Another object of the invention is to make it possible, from data sources, to extract relevant information for a user, taking into account, jointly, the skills of a community in a process of accessing the data. information, context and perceptions specific to each user.

One or more of these objects are filled by the method and the device according to the independent claims. The dependent claims further provide solutions to these objects and / or other advantages. More particularly, according to a first aspect, the invention relates to a method for customizing, for a user, a process for accessing a set of resources accessible via an information processing system.

The term "resource" denotes the information and data accessible to the user via the information processing system. Each resource can in particular be identified by a unique identifier, typically a unique resource identifier, more generally designated by the acronym "URI" for "Uniform Resource Identifier", defines in particular in the document RFC 3986. A resource can be an electronic document, for example a web page, an image or an office document. A resource can also be a system delivering a flow or a source of data or information or a user.

In particular, the term "customization, for a user, a process of accessing a set of resources accessible via an information processing system", the personalization for a user of an information search, search of reference users (for example experts), the constitution of a bibliography using collective knowledge, the identification of sources of information or the search for data, for example medical data, by a automated system.

The method comprises the following steps:

• gathering information on events occurring in relation to all resources;

• creation, based on the collected information relating to events occurring in relation to all the resources, or obtaining a first profile associated with the user, the profile comprising:

at least one context adapted to allow the representation of all the conditions likely to influence a user's choice and / or a judgment of relevance of the user relative to the resources of the set of accessible resources through the information processing system; and, o a first set comprising at least one a priori adapted to allow the determination, for each resource of the set of resources, of a level of relevance for the user according to at least one measurable property of said resource; and, a second set including at least one reference to a second profile associated with a reference user for the user;

Calculating a probability that a first resource, among the set of resources, is relevant to the user based on said first profile, and implementing the access process according to said probability.

The profile of the user thus created or obtained comprises three dimensions: at least one context, the first set, and the second set. Thus, by using the three dimensions of the profiles jointly, it is possible to exploit the profile of the user optimally, compared to a profile that would exploit only one of said three dimensions. Indeed, the three dimensions of the profile thus constructed by the method according to the invention interact with each other by completing and correcting themselves. In the same way that humans jointly exploit their environment (ie the context), their instinct and their acquired (that is to say the prion) and their peers (reference users) to its decision making, the present method exploits in the same way these three dimensions to allow the customization of resources provided to users for example following a request from the latter.

The method is particularly applicable to solving a technical problem which consists of obtaining optimized search results for a given user by a computerized system.

During the calculation step, in one embodiment, the probability that the first resource is relevant to the user can be obtained by:

• selecting, according to at least one criterion, in the second set, at least one profile associated with one of the reference users;

Calculating a second probability that the first resource is relevant, based on said at least one profile associated with one of the reference users;

Calculating the probability that the first resource is relevant to the user according to the second probability and said at least one context of the first profile and at least one prior of the first set.

For example, the criterion for selecting, in the second set, said at least one profile associated with one of the reference users may be the similarity between one of the contexts of the at least one profile associated with one of the reference users and the context of the first resource.

A priori is a measurable property of a resource: f ·. R → [o, i]. The calculation function of the a priori can be defined by the user or selected from a predefined list (size, date of creation, ...). The notion of a priori therefore makes it possible to characterize a resource according to a particular aspect and independently of any context. This notion allows a user to integrate their preferences into the system, preferences that are not necessarily represented or representable in the context. For example, a user may have a priori on the importance of a source of information, regardless of the resources it produces (ie: whether each of these resources is contextually close to the context of the user).

A priori, for a given user, can be a function returning a value, typically between 0 and 1, all the greater as the resource has an important level of relevance for the user.

Alternatively, a priori, for a given user, can be a function returning a value, typically between 0 and 1, involved in determining the level of relevance of a resource a user. Depending on the feedback from the user, it is the system itself that will determine, during the learning phase, the optimal value or combination of values so that the personalization process is as efficient as possible ( ie the closest to the actual expectations of the user).

Some users will prefer unpopular resources (value of a priori close to 0) while others will prefer very popular resources (value of the a priori close to 0). These different preferences can in particular be determined during the learning phase.

Said at least a priori can be a function adapted to allow the determination, for each resource of the set of resources, the level of relevance for the user according to at least one of the following properties: a size of said resource, a date of creating said resource, a number of links or references pointing to said resource, a signal-to-noise ratio of said resource, a type of address by which said resource is accessible; a level of popularity of said resource, a quality level of said resource.

Resource related event information may include one or more information from the list next: an event type, an identifier of the resource that produced the event; an ordered list of resulting resource identifiers provided by the information processing system; a set of metadata associated with the event.

During the collection stage, only information relating to events occurring in connection with the resources and meeting a set of user-specific criteria can be collected. For example, the events collected can be first categorized by their type on a scale with levels from 0 to 3. Level 0 indicates the obvious lack of interest of the user for the resources associated with the event, typically for a resource outside the user's domain. Level 1 indicates a moderate disinterest, for example for a resource in the domain of the user but irrelevant. Level 2 reflects the user's interest, for example user domain resources relevant to the user. Level 3 is of obvious interest, for example user domain resources that are strategic to the user. The chosen criterion can be the following one: only the events classified at a level 2 or 3, are selected.

The method may further comprise a step of updating said first profile associated with the user, following a modification of the context of the user.

The modification of the context of the user can be detected in particular using information collected during the collection step relating to events occurring in connection with interactions between the user and the treatment system. information. During the updating step, said first profile associated with the user can be updated according to the information, collected during the collection step, relating to events occurring in connection with the interactions between the user and the user. user and the information processing system. Thus, the profile of a user is thus dynamic, that is to say that it adapts automatically to the changing behavior of the user with the information processing system. For example, when the user interacts with the information processing system, he produces events that the system intercepts and then interprets to adapt the user's current context. Context being an element included in the profile, the latter is adapted accordingly and can evolve over time. An additional context can be added to the user's profile, the additional context being adapted to allow the modified representation of the user. In addition, the additional context is added to the user's profile only if the similarity between the additional context and the at least one context adapted to allow the representation of the user is less than a threshold. Thus, the additional context is added in case of break of context.

During the updating step, said first profile associated with the user is updated according to the profile (s) to which the second set refers.

According to a second aspect, the invention relates to a management device, adapted to be coupled, via a communication network, to at least one client device. The user profile management device is configured to access a set of resources. The management device is configured to customize for a user of said at least one client device, a process of accessing the set of resources by:

• collecting information about events occurring in relation to resources;

Creating, based on the collected information relating to events occurring in relation to all the resources, or obtaining a first profile associated with the user, the profile comprising:

at least one context adapted to allow the representation of all the conditions likely to influence a user's choice and / or a judgment of relevance of the user relative to the resources of the set of accessible resources through the information processing system; and, o a first set comprising at least one a priori adapted to allow the determination, for each resource of the set of resources, of a level of relevance for the user according to at least one measurable property of said resource; and,

a second set including at least one reference to a second profile associated with a reference user for the user;

Calculating a probability that a first resource, among the set of resources, is relevant to the user according to said first profile and implementing the access process according to said probability. According to a third aspect, the invention relates to a computer program comprising instructions for performing the steps of the method according to the first aspect, when said program is executed by a processor.

Each of these programs can use any programming language, and be in the form of source code, object code, or intermediate code between source code and object code, such as in a partially compiled form, or in any form what other form is desirable. In particular, it is possible to use scripting languages, such as notament tel, javascript, python, perl, which allow code generation "on demand" and do not require significant overhead for their generation or modification.

According to a fourth aspect, the invention relates to a computer-readable recording medium on which is recorded a computer program comprising instructions for executing the steps of the method according to the first aspect.

The information carrier may be any entity or any device capable of storing the program. For example, the medium may comprise storage means, such as a ROM, for example a CD-ROM or a microelectronic circuit ROM, or a magnetic recording means, for example a diskette or a hard disk. On the other hand, the information medium may be a transmissible medium such as an electrical or optical signal, which may be conveyed by an electrical or optical cable, by radio or by other means. The program according to the invention can be downloaded in particular on an Internet or Intranet network.

Alternatively, the information carrier may be an integrated circuit in which the program is incorporated, the circuit being adapted to execute or to be used in the execution of the method in question.

Other features and advantages of the present invention will become apparent from the following description of embodiments with reference to the accompanying drawings, in which:

Figure 1 is an architecture diagram of a personalized information processing system according to an embodiment of the invention;

FIG. 2 is a diagram of a user profile management method in an information processing system, according to an embodiment of the invention;

Figure 3 is a diagram of a method of constructing a user graph; FIG. 4 is a diagram of a naive Bayesian network graphically representing the relationships between the characteristics of the user profile and the relevance of a resource (denoted I).

In the following description, the information and data accessible to a user through the information processing system are called resssources. More specifically, a resource is subsequently defined as an object adapted to be manipulated by the information system, regardless of any context of use. The set R of resources is denoted R = {r _1; ..., ¾} with M the number of resources. Each resource of the set R can in particular be identified by a unique identifier, typically a unique resource identifier, defined notably in the document RFC 3986. The unique resource identifier makes it possible in particular to uniquely and unambiguously identify a resource, physical or abstract, usually available through a communication network. A resource can be an electronic document, for example a web page, an image or an office document. A resource may also be a system delivering a stream or a source of data or information. A resource may also be a user of the information processing system. A user can be a full-fledged individual or a group of people - for example, an organization, a department of a company, etc. A user is a resource similar to other resources, to which a profile is associated and maintained by the information processing system.

In the following description, the construction of a user profile for a given user of the information processing system will be described. However, the method also applies to a plurality of user profiles corresponding to a plurality of users.

Figure 1 is an architecture diagram of a personalized information processing system according to an embodiment of the invention. The processing system includes a user profile management manager 30, coupled via a communication network 16, to a plurality of client devices 10a, 10b and a data network 20 from which resources are accessible. . Typically, the communication network 16 is adapted to implement data transfers between the user profile manager 30 and the plurality of client devices 10a, 10b. In a usual case, the user profile manager 30 and the client devices 10a, 10b are distributed over several sites. The communication network 16 uses for example the Internet protocol. The network 16 typically includes usual interconnect means for routing data, such as routers, bridges, firewalls, switches, etc. The network 20 is for example the Internet and / or a local network. In FIG. 1, are represented in particular:

A first client device 1 10a comprising two terminals 12 coupled to the communication network 16;

A second client device 10b comprising two terminals 12 coupled to the communication network 16.

Different users 14 may further access the terminals 12. The terminals 12 may be of one of the following types: computer, mobile phone, tablet, server, and more generally any device capable of exchanging data via the network 16. More particularly, the terminals 12 are configured so as to allow:

Collecting user information to enable the construction and / or updating of profiles;

• querying resources accessed through the data network 20;

• the transfer of the collected information and requests to the managers 30 of user profiles;

• the presentation of results obtained in response to requests.

The user profile manager 30 typically includes a protection module, for example a firewall module, to ensure security when the manager 30 is coupled to the networks 16 and 20. The user profile manager 30 further includes a storage device. data 36 relating to user profiles. The data storage device 36 is typically a database or document management system. The user profile manager 30 further comprises a data processing device 36 relating to user profiles, for example an application server. More particularly, the user profile manager 30 is configured to:

• receive and / or collect user information 14 to enable the construction and / or updating of profiles;

• create and update user profiles 14;

• receive and process requests for resources accessed through the data network 20;

• transfer the results obtained in response to requests from users 14; Accessing the resources of the data network 20 and obtaining information relating to said resources for processing the requests of the users 14.

With reference to FIG. 2, a method of managing user profiles in an information processing system, according to an embodiment of the invention, will now be described. The method for managing user profiles is particularly adapted to be implemented by the personalized information processing system according to one embodiment of the invention illustrated in FIG. The profiles thus managed can in particular be used, by a computerized system, to customize a process for accessing information, such as searching for information in databases or documents or making personalized recommendations. and optimized.

In the embodiment illustrated in FIG. 2, the method for managing user profiles comprises a step 1 10 for collecting information relating to all the resources R. In an embodiment of step 1, 10 between an initial moment and a moment t, a history H of events

-,> ,,, c? ^L] is such an ordered list of events s ^l _i fl ^i, _ul ^s: occurring between the original and the instant t moment.

An event e "can be represented by a quadruplet: e ^a = {type, source: r € i {, [rst .test, r € ii [meta: (key, value))} in which:

• t pe is the type of the event

for example a click, a bookmark, an impression, a request, etc;

• source is the unique identifier of the resource that produced the event fi ";

• result is the ordered list of the resulting resource identifiers provided by the information processing system;

• meta is a set of metadata associated with the event e ^i:.

An example of history H is given in the table below: source event source result metadata

e ^l click i \ - ΰ ^' - favorites - ϋ ^{' Α} query (loc, PARIS))

In an embodiment of step 1, only events responding to a set of criteria are added to the event history eS e ^ ^ e ¹ -. For example, the events collected can be first categorized by their type on a scale with levels from 0 to 3. Level 0 indicates the obvious lack of interest of the user for the resources associated with the event, typically for a resource outside the user's domain. Level 1 indicates a moderate disinterest, for example for a resource in the domain of the user but irrelevant. Level 2 reflects the user's interest, for example user domain resources relevant to the user. Level 3 is of obvious interest, for example user domain resources that are strategic to the user. The following table presents a possible classification of the events according to their type in relation to the previously presented scale (considering for example that a user expresses an interest for a resource when he visualizes it for more than 30 seconds):

In the example of an information retrieval system for information retrieval, the positive relevance returns (here the level 2 and 3 events) are much more important than the negative relevance returns. Also, in one embodiment, only events ranked at a level 2 or 3, are added to the history H. Alternatively, it is possible to add all the events to the history H and assign them a weighting coefficient relative to their level.

The method for managing user profiles comprises a step 120 for initializing, building and / or updating a profile for the user.

By profile On is meant an associated profile, at a time t, with a resource; user type. The profile ¾ ₍ allows in particular to define formally the expectations of said resource a; in information processing by the information processing system. The profile · ¾ makes it possible to customize, for the resource i, the processing of the information by the system, in particular to customize the information that is provided to the resource tj The profile ¾ is for example used to estimate the expected probability that a resource r is relevant to resource t.

In one embodiment of the invention, at time t, a profile ¾ comprises a sequence of contexts J ^L , a set A _Ut of a priori, and a set R _Ui of reference users is obtained, generated or updated. This embodiment has the particular advantage of allowing effective monitoring of frequent changes in the user's context. But it has been found that users tend to be multi-tasking - that is, needing to manipulate diverse and varied contexts - and to change contexts relatively frequently. In an alternative embodiment, the context sequence can be reduced to a single context. This alternative embodiment may be particularly interesting for applications where the user is certain not to change context.

"Context" refers to information relating to a formal interpretation of past events relating to a resource and characterizing said resource. A context therefore comprises the information relating to a set of conditions that can influence the choices and judgments of relevance, taking into account the feedback obtained by the information processing system.

From the history H, it is possible to deduce a restricted history H _Tl to a resource n ER corresponding to the set of events for which the resource is identified as the source. Thus, Hr _t = [e ^v E

of a resource n at time t in the sequence of contexts s, also referred to as the current context as opposed to the local context, is a representation of the resource i function of the restricted history _Ti to the resource n to l moment t. In one embodiment, the context C ¹ of a resource t at time t in the sequence of contexts s is represented by a distribution of probabilities on themes. Alternatively, the context can be represented by a vector of weight in the vector space of the terms, or by a vector of weight on the concepts of an ontology etc. A context sequence is an ordered list of at least one context in time. The index s in the notation f represents a number in the ordered list in time of the contexts. The notion of sequence of contexts makes it possible to represent the changes of context that users may have during their experience through the information system.

By local context C i, at an instant t, an event whose source is the resource t is a representation of said event independent of past events. This representation is determined from the resources resulting from said event and its type. By local context C i _{r |} a resource r, - a representation of this resource.

For example, calculating the local context of an event where the user places the favorite resource η is equivalent to calculating the local context Ci,., Of the resource knowing the type of the event.

In one embodiment, the local context ϋί ^ of an event or the local context ^_((a resource is represented by a probability distribution on thematic. Alternatively, the local context can be represented by a weight vector in the vector space of the terms, or by a weight vector on the concepts of an ontology etc.

When new events occur, the context of a resource can change in two ways:

• the local context of this event is sufficiently close to a context of C, specifying or supplementing it;

• The local context of this event is too different, and a break in context potentially occurs.

When a context break occurs, a new context can be created and added in a sequence of contexts of the resource \. When the break of context concerns a user, it can be proposed to the latter to validate or invalidate said break. When it concerns another type of resource n, this validation can be done automatically or, depending on the type of resource, manually by a system administrator.

Also, an event is considered broken when the similarity between the local context of the event and the current context of the resource at the origin of the event is less than a threshold e. In one embodiment of the invention, the similarity is calculated by the Jensern-Shannon divergence.

During a step 122, the context ^ of a resource at time t is obtained by applying the following formula: with:

• t ^{L -1} the context at time t-1 in the sequence s of resource n, ie the current context until the context ^ ^L is obtained;

· CLI _t the local context of the resource at time t

• f a function adapted to produce a new context by combining two contexts, according to their relative importance controlled by the parameter a; in this case, the function combines the local context Ct ^ and the current context ^{~ l} , as a function of the parameter a.

In one embodiment, the function is a function adapted to produce a new context by combining two contexts, and giving importance to the events according to their novelty. Indeed, it can be considered that the more recent the events are, the more they are representative of the current need of a user. For example, the function can be an exponential moving average function, described in particular in the Harry V Roberts document, entitled "Stock-market patterns" and financial analysis: Methodologica suggestions. The Journal of Finance, 14 (1): 1-10, 1959. Thus, in step 122, the context c ^ ¹ of a resource at time t can be obtained by applying the following formula :

with ϋ <a <1 the smoothing constant determines the importance of the current context in relation to the local context of the event.

Obtaining the context ^ ^L of a resource r, at time t, during step 122, from the history H constructed during step 1 10, is typically a recursive process . An example of such a recursive process of constructing the context will now be briefly discussed.

At the initial moment, the history H has no event.

At time t = 1, the resource t produced an event e, ^1, collected in step 1 10, and added to the history H. In step 122, the local context. ^ Of the event is calculated. Since the context C ^ ¹ of the resource \ does not exist at time t = 1, the context L is equal to the local context Ct ^ of the resource r, -.

At a time t = 2, the resource n produces an event o ^' , collected during the step 1 10, and added to the history H. During the step 122, the new current context of the The event is computed according to the current context, the local context C. ^ of the event ¾, and the parameter se. For example, in the case of a current context represented by a probability distribution on thematics, the function f used to compute the new current context t: / of the event er, is for example a linear combination of thematic distributions. of the context C, ^ ¹ and the local context i of the event e, -,.

At a time t = j, the resource ï produces an event e, ¹ ,, collected during the step 1 10, and added to the history H. During step 122, the new current context C ^ of the event α 'is computed according to the current context ^ "' ^{^} , of the local context i ^ of the event e. '^' ,, and parameter a.

Thus, the collection of the events, the update of the history H, as well as the current context, continues thus during the stages 1 10 and 122, with each new event.

The set A _u , a priori included in the profile ¾ is determined and / or updated during a step 124. A priori is a measurable property of a resource. It is defined by a function f R → il.

A priori makes it possible to characterize a resource according to a particular aspect and independently of any context. Also, the use of a priori in the profile ¾ makes it possible to take into account the preferences a priori of the user, preferences that could not necessarily be represented in the context. For example, a user may have a priori on the importance of an information source, regardless of the resources it produces, that is to say independently of the proximity of each of said resources with the context of the user. The priori can be shared among several resources or be specific to a given resource. An illustrative and non exhaustive list is given below:

• size: a resource containing many words in its representation is more relevant;

• creation date: the more recent a resource is, the more relevant the resource is;

• number of incoming links: the more a resource has inbound links (e.g., many pages including references to said resources), plus said resource is relevant;

• noise information report: a resource with low noise - for example, little redundancy - is more relevant; • type of address: a resource of type input page whose address, typically a uniform resource locator more generally designated by the acronym "URL" for "Uniform Resource Locator", is root type is more relevant than another type of page; the type of the address (root, sub-root, path, file) is an indicator of the relevance of a resource;

• overall popularity of a resource: a popular resource is more relevant; this measure represents the amount of events (clicks, favorite ...) associated with a resource; this measure can be calculated as follows:

• quality of the source: a resource from a quality source is more relevant; the quality of an information source can be manually specified by a user. For example: pcnincnceiso rce} = 0.8.

A priori is a measurable property of a resource:

f - R <| 0. u

The calculation function of the a priori is defined by the user or selected from a predefined list (size, date of creation, ...). The notion of a priori therefore makes it possible to characterize a resource according to a particular aspect and independently of any context. This notion allows a user to integrate their preferences into the system, preferences that are not necessarily represented or representable in the context. For example, a user may have a priori on the importance of a source of information, regardless of the resources it produces (ie: whether each of these resources is contextually close to the context of the user). A priori function can be interpreted according to two alternatives:

1. A priori, for a given user, is a function returning a value, typically between 0 and 1, all the greater as the resource has an important level of relevance for the user.

2. A priori, for a given user, is a function returning a value, typically between 0 and 1, involved in determining the level of relevance of a resource a user. Depending on the feedback from the user, it is the system itself that will determine, during the learning phase, the optimal value or combination of values so that the process of customization is the most efficient possible (that is to say the closest to the real expectations of the user).

Some users will prefer unpopular resources (value of a priori close to 0) while others will prefer very popular resources (value of the a priori close to 0). These different preferences will be determined during the learning phase.

The user profile management method may comprise an optional step 126 of identifying at least one profile associated with a reference user for the resource u; .

The set of reference users for the resource is the set of users whose profiles can be used to improve the relevance of the information transmitted by the information processing system to the resource. Collaborative filtering recommendation in which the relevance of a resource is a function of the assessments of other users on said resource, the use of the set of reference users on a profile makes it possible to determine the relevance of a resource to the resource. help profiles of a limited number of users, not necessarily close to the resource.

In one embodiment, the set of user-type resources Uj is represented by an oriented graph, each node of the graph representing a user. In the oriented graph, an arc (^ /) between the resource and the resource indicates that the user w; is a reference user for the user A user u is referred to as u _; s \ - is a predator of .i- in the user graph.

The set ¾, reference users for the user Ui, represented by the oriented graph, can be obtained or constructed manually by the user Ui or created automatically or semi-automatically by a search method and selection of reference users. An example of creating the user reference set R _Ui for the user will now be described.

Initially, the set is empty and the user t-ij has therefore not designated reference users. Subsequently, the user u: chooses, manually or according to recommendations formulated by the system, one or more reference users. Then, the user u; can add another reference user, or remove a present one from the set R _u,. Intuitively, a reference user is a user on whom another user can rely in order to improve the relevance of the information returned to him by the system. Unlike a recommendation system - collaborative filtering - where the objective is to determine if a resource (video, music, book, etc.) will please a user who knows what others have liked, The approach implemented by the invention focuses on a restricted set of users who are not necessarily close to the initial user and whose contribution would therefore not have been important in a conventional recommendation system. Assuming, for example, that a user, an expert in Near Field Communication (NFC) technology - is recruited by a company offering medical traceability solutions, it is assumed that this user has very little knowledge in the medical field. He will therefore find it difficult to express his need on the medical side due to his weaker knowledge in this field of application. The fact that this user chooses expert referrers in the medical field will guide the system to provide the user with relevant information from the point of view of referrals that may not be his because of his lack of knowledge. in the medical field. Since the context is dynamic, the influence of a reference user on another user also changes over time.

Formally, the set of users ^UI, J is represented in the form of a directed graph. Each node of the graph represents a user and an arc! ¾. means that user ^ is a reference user for user ^. A user is called a reference for another user ^ if is a predecessor of ^ r in the user graph. The set of reference users of the user ^υ ί is noted

The user graph can be constructed manually by the users or semi-automatically by a method of searching and selecting reference users. In both cases, the method of constructing the user graph is illustrated by the example of FIG. 3, in which 4 users are considered to be ^u -.

1. At initialization, users do not have a reference user. The user graph is then represented by a tree where the node R is the root of the tree and where the other nodes, son of R, are the users. The graph on the left of Figure 3 illustrates its state at initialization.

2. Then, the user chooses (manually or as recommended by the system) his or her reference user (s). The choice of a reference user ^u ; by the user ^v t results in the creation of an arc in the user graph. The graph on the right of Figure 3 shows an example of this construction. Node R has been voluntarily removed for readability. The graph can then evolve: a reference user can be added (creation of an arc in the graph) or deleted (deletion of an arc) by a user. In all cases, the system ensures that any action performed on the graph makes it move from a coherent state to another coherent state (ie: no violation of the constraints of the graph.) At the end of step 124 or the optional step 126, the profile of the user is constructed and combines three dimensions: the context sequence c ^L , the set A _Ui of priori, and the set R ^ of reference users.

Thus, by using the three dimensions of the profile jointly, it is possible to exploit the profile of the user optimally, compared to a profile that would exploit only one of said three dimensions. Indeed, the three dimensions of the profile thus constructed by the method according to the invention interact with each other by completing and correcting themselves. For example, two use cases in which the synergistic effect of using the three dimensions of the profile is highlighted:

· If a resource does not correspond to the contexts of the context sequence of the user, but it corresponds to one or more contexts associated with the reference users of the set, the relevance of this resource for the user can be sufficient. high so that it is interested: the sequence of contexts of the user will evolve so to adapt to that of its users of reference. In fact, for the following resources, the context will have evolved and will make it possible to refine the scheduling of the relevant resources for the reference users; similarly, when the contexts of the sequence of contexts do not correspond to the local context of a resource although the user has a positive assumption on the resource, the sequence of contexts J _; ^The user's ^L will therefore evolve to adapt accordingly;

• A user may have a negative prior to a resource but by definition, a priori, is not a truth; taking into account the profiles of reference users of the set _Ui , the estimation of the probability that the resource satisfies the user can however be composed accordingly, despite the negative priori of the user.

During a step 130, the probability that a resource r satisfies the user is determined For this, the probability

at time t r a resource is relevant to resource i is calculated according to a local CI _r context, the context of the set A _Ut of a priori, and all users reference or ^i> (r | _£ IFF) = i '(r | n _r.' ¾.) -

In an embodiment of step 130, the probability that a resource r satisfies the user u is determined; in :

In a step 132, selecting a group R _U1 (r) in the set of reference users whose context is close to the resource r;

In a step 134, calculating the probability that the resource satisfies the group of reference users selected during step 132, according to a heuristic;

In a step 136, calculating the probability that the resource satisfies the user tij, according to the local context of the resource r, the context of the user of the user's priori ¾ and the probability that the resource r satisfies the reference user group calculated in step 134.

In step 132, a reference user may be added to the selected reference user group based on its relevance for a specific domain. In particular, it is possible to filter the set of reference for users by generating resource r given a list of pairs li _^ i) with C; ¹ the context of the reference users closest to the context of the resource r, then retaining only the pairs ΐΐί, €; ) whose similarity between the context c and the context of the resource r is greater than a threshold e fixed for example by the user; .

During step 134, for example, the probability P is calculated.

Cr for the group ¾ (Γ) of reference users selected in step 132, determining, for each reference user of the set, the probability J "(r | ^ y) that the resource r is appropriate reference user audit in the context closest to the resource r The probabilities of each reference user of the group R _Ui (r) are then combined according to To achieve the probability that the resource will satisfy the group In one embodiment of the invention, the heuristic used is the average probability that corresponds to the average needs of the users. Alternatively, the possible heuristics are of the minimum probability type and the average probability.

During the step 136, the probability that the resource satisfies the user can be calculated by using for example a learning model in which the local context of the resource r, the context of the user; the user a priori t.j and the probability that resource r satisfies the reference user group are the input characteristics.

The estimation of the probability of a resource satisfying a user consists in solving the equation 1 that we recall below:

To do this, we proceed in several steps:

1. Filtering the reference users to keep only those having one of their contexts close to that of the resource (the result of this filtering is called reference user group);

2. Estimation of the probability that the resource satisfies the reference user group according to a heuristic;

3. Estimation of the probability that the resource will satisfy the user: taking into account the local context of the resource, the context of the user, its assumptions and the probability that this resource will satisfy the reference user group .

These steps are detailed below.

Filtering reference users: A reference user can be chosen for his qualities in a specific area. However, since a user can have multiple contexts, a reference user may not be relevant for certain contexts. For example, if a reference user is selected for his medical skills, it is this context of use that must be exploited and not another. We therefore propose to filter the reference users of a user by generating for a given resource r a list of tu-couples.

with ^f ^£ the context of the reference user closest to the context of the resource. Then we only keep couples whose similarity between their

end context and that of the resource is greater than a threshold ^£ set by the user. We formalize this process by algorithm 2.

Probability that a resource satisfies the reference user group: The user group filtered for the resource noted ^ Λ ⁷ is used to calculate the probability »A ^? }} a resource ^r satisfies it. To calculate this probability, we estimate for each reference user of the probability ^p

that this resource suits him in his context closest to the resource. The probabilities of the reference users are then combined heuristically to obtain the probability that the resource will satisfy the group. We list below three heuristics:

· Using the maximum probability

• Use of the minimum probability

ίψ ί ΐ ti ί ΐ r r r r r r

• Using the average of the probabilities '° ί _ϋ . · Ι)) AVERAGE {Pir \ ref * ref]

where Qu 'is a user of ¾. ('^' ).

System Resolution: The calculation of the probability H ^ U can be seen as a learning problem where the local context of the resource, the context of the user, its assumptions and the probability that the resource satisfies the group of Reference users are the input characteristics of a learning model. We represent these characteristics in the form of a vector:

\ ci {V, iJ (fc) .i (L), c (frj. s [1) a (n), P (| H _u r)} \

OR

, ⁱ i (¾. .. d {} _S have the ^k dimensions of the local context of resource r;

. that _S have the ^k context dimensions of the user;

• * ί ^ι > ^ 'ι are the ¹³ prior values of the user;

•

} is the probability that the resource satisfies the reference user group.

Each pair (essmree, uiUsateur) corresponds to a truth in the field, reflecting the relevance of the resource for the user. This ground truth can be either manually specified on a restricted set of resources, or determined automatically based on the type of events that this resource is a result of. In the second case, the truth is represented by one of the 4 levels of relevance. Automatic ground truth determination is the method we use because it allows the system to adapt to the needs of the user without the user needing to explicitly update their expectations. To simplify the explanations, we can reduce the relevance to a binary level: events whose type is classified by levels 0 and 1 are irrelevant (0) and events whose type is classified by levels 2 and 3 are relevant. (1). The learning problem consists of determining a function * ^{1 y} where ^x represents the observations (the values of the characteristics, see above) and where ^Y represents a judgment of relevance (real number or discrete value as in the present example ). The function f is determined via a learning algorithm such as the Naïve Bayes algorithm, logistic regression, support vector machines, neural networks, and so on. Note that the inventive aspect does not lie in the algorithm for determining this function f but in the construction of the input characteristics of this algorithm. The set of resources for which ground truth is known forms the learning base of the user, as shown in Table 1. We note ¾. the user's learning base ^. i relevance

Resource ^■ \ - tlljj .... iiix) - il) · *:.: '

: 0.7 i .. ϋΑ \ ΰ. \ ... \ na 1. where a.o 1.

| VS'. l 03! I2 ... II4 to 0.3 a. -! at

7 ..>>> 0.7 0.7 na na 1 1 1.

Table 1: Basic learning example

We describe below an example of implementation of the resolution process via a Bayesian network. A Bayesian network is a graphical model representing cause-and-effect relationships between random variables. It makes it possible to calculate conditional probabilities and, in our invention, it is used to calculate the probability ^p 1 ¾). The structure of the Bayesian network is defined by the graphical model in Figure 4. In this representation, I is a binary random variable representing the relevance of a resource for a user. This relevance generates the profile of the user (generative model). This structure is more commonly known as the Naive Bayesian Model.

In this model, estimate the probability ^p L ^r l¾i is to estimate the probability:

P σ lk <(l) ci (Jf). ). aOn "(he). P (r | " _¾ (r)))

The learning phase consists of estimating for each user his maximum likelihood estimator, ie:

advice ¹ | J 1 1 I and

P r \--

for each value ^r1, of a random variable of

^■ -ifi) c ■■■ ^, ). AT . a to. PM RM) and where

f ^fj3.fi tCr ^y | /: ij _es t | _e number of times a resource is relevant (' ¹ ) for a value of one of the random variables (denoted ^rv ) in the learning base ^ε' .

™ .I f ⁱ (fj _{rs: ii:} if _e st number of relevant resources ^(J ^L) in the learning base • ^a is a smoothing factor.

The method, described above, is based on a dynamic model, adapting automatically to the changing needs of users through the actions they perform on the system but also by the influence of their reference users. Thus, when a user interacts with the system, his profile is updated and then exploited so that the information provided to him corresponds to his current need.

This dynamic can be illustrated by the example described below. In a system comprising three users u1, u2, u3, in an initial state, no influence link exists between the users. The profile of each user then includes a context as well as a priori. Each user can then indicate which reference users are suitable for him. This results in creating links between users. Thus, in the example, the user u1 has the user u2 as the reference user. The user u2 has the user u3 as the reference user. User u3 does not have a reference user. A resource, denoted r, is initially relevant for the user u3, but not very relevant for the users u1 and u2. If the user u3 performs an action (click, bookmark, etc.) on r that is of interest to the user u3 for that resource, a new event will be listed in the system history, and generate the update user profile u3. Since the user u3 is a reference user of u2, the relevance of r for the user u2 will be influenced by the updated profile of the user u3 and the relevance of r for the user u2 will increase. At this time, the relevance of the resource r for the user u1 remains unchanged because its reference user, the user u2, has not yet performed action on the resource r. To confirm its interest for the resource r, the user u2 performs an action on the resource r. The resource r will then see its relevance for the user u1 to increase because the profile of the user u2 will also have been updated, the relevance of the resource r for the user u1 remaining unchanged. If the user u2 does not perform any action on r, the resource r is not of interest to the user u2. In this example, the relevance of the resource r for the user u1 and u2 will increase as long as these two users perform actions on the resource r, because the latter both mark their interest for this resource and because this interest is reinforced. by that of their reference user. However, when a resource no longer interests the user u1 and / or u2, no action will perform the resource and the influence as a reference user will not change. A detailed use case will now be presented. The following example refers to four users: u1, u2, u3, u4. The user u1 is considered as an expert user for the user u2, in the maritime domain. The user u3 is considered as an expert user for the user u4, in the textile field.

In an initial state, the system accesses the data recorded in the user's web browser u1, u2, u3, u4, to initialize the context specific to each user. Thus, the system makes it possible to obtain, in this initial state, personalized responses for each user. For example, following the receipt of a user's authorization, the user's browsing history can be scanned, as well as the pages marked as favorites in the browser. The user's event history can be initialized. The events produced by the user enrich in real time the history and context relative to said user. The favorites and user browsing history u1, u2, u3 and u4 refer to web pages related to the maritime domain and the textile field. The resulting contexts for each of the users after the initialization phase are shown in the following table (the number in parenthesis next to each term represents the weight of the term in the user's context):

The expert users u1 and u3 respectively define their a priori, to specify their preferences, for example those concerning the sources of information. It is furthermore assumed that the users u2 and u3 have not defined an a priori. The following table presents these a priori:

user source relevance u1 planetsolar.org 0.8 u1 leparisien.fr 0.5 u3 ifm-paris.com 0.9 Thus, the user u1 specifies that the source planetsolar.org and the resources that said resource produces have a priori probability of being relevant of 80% against 50% for the source leparisien.fr and the resources produced by leparisien.fr. The user u2 specifies that the ifm-pahs.conn source and the resources it produces have a priori probability of being 90% relevant. As the relevance of the other sources is not specified, these other sources are therefore neither favored nor disadvantaged when calculating the relevance of a resource.

To select the users considered as reference users, the users search, through the system, the users who are similar to them. For example, the user u2 is offered the user u1. The user u4 is offered the user u3. Users u2 and u4 then validate manually or by thresholding the users they wish to have as reference users. The so-called u1 and u3 users do not use this feature.

User profiles u1, u2, u3 and u4 are then used to customize the responses to queries. For example, for a "small boat" request submitted by user u2 (non maritime expert). The use of the u2 user profile makes it possible to target the request as well as possible. After the initialization phase, the contexts of the users are empty. In fact, when the user u2 submits the request "small boat", he will obtain mainly documents related to the field of textile (i.e. related to the mark "small boat") that will not correspond to his need. Subsequently, when the reference users have carried out research, their context will target their respective domains (maritime for u1, textile for u3) because they will have, for example, defined a priori sufficiently. Thus, when the user u2 (re) submits the request "small boat", its reference user (u1) will act implicitly on the relevance of the resources returned and the actions of u2 on these resources will enrich his own context. In subsequent searches, the context of u2 has evolved and the relevance of the resources that will be delivered to it will be greater. By exploiting the assumptions, it is still possible to overweight the results coming from a priori reliable sources.

Claims

1. A method for customizing, for a user, a process of accessing a set of resources accessible through an information processing system, characterized in that it comprises the following steps:

• collection (1 10) of information relating to events occurring in relation to all resources;

• creation, according to the collected information relating to events occurring in relation to all the resources, or obtaining (120, 122, 124, 126) of a first profile associated with the user, the profile comprising:

Calculating (130, 132, 134, 136) a probability that a first resource, among the set of resources, is relevant to the user according to said first profile, and implementing the access process according to said probability.

The method of claim 1, wherein, during the computing step (130, 132, 134, 136), the probability that the first resource is relevant to the user is obtained by:

• selecting (132), according to at least one criterion, in the second set, at least one profile associated with one of the reference users; Calculating (134) a second probability that the first resource is relevant, based on said at least one profile associated with one of the reference users;

Calculating (136) the probability that the first resource is relevant to the user based on the second probability and said at least one context of the first profile and at least one prior of the first set.

The method according to any of the preceding claims, wherein said at least one priori is a function adapted to allow the determination, for each resource of the set of resources, of the level of relevance for the user according to at least one one of the following properties: a size of said resource, a date of creation of said resource, a number of links or references pointing to said resource, a signal-to-noise ratio of said resource, a type of address by which said resource is accessible; a level of popularity of said resource, a quality level of said resource.

A method as claimed in any one of the preceding claims, wherein the information relating to events occurring in connection with the resources includes one or more of the following: an event type, an identifier of the resource having produced the event; an ordered list of resulting resource identifiers provided by the information processing system; a set of metadata associated with the event.

5. Method according to any one of the preceding claims, wherein, during the collection step (1 10), only the information relating to events occurring in connection with the resources and meeting a set of criteria specific to the user are collected

The method of any one of the preceding claims, further comprising a step of updating (120, 122, 124, 126) said first profile associated with the user, following a modification of the context of the user. 'user.

The method according to claim 6, wherein the modification of the context of the user is detected using information collected during the collection step (1 10) relating to events occurring in connection with interactions between the user and the information processing system.

The method of claim 7, wherein during the updating step (120, 122, 124, 126), said first profile associated with the user is updated according to the information collected during the collection step (1 10), relating to events occurring in connection with the interactions between the user and the information processing system.

The method according to any one of claims 6 to 8, wherein during the updating step (120, 122, 124, 126) of said first profile associated with the user, an additional context is added to the user profile, the additional context being adapted to allow the modified representation of the user

The method of claim 9, wherein during the updating step (120, 122, 124, 126), the additional context is added to the user's profile only if the similarity between the additional context and the said at least one context adapted to allow the representation of the user is less than a threshold.

1 1. A method as claimed in any one of claims 6 to 10, wherein during the updating step (120, 122, 124, 126), said first user associated profile is updated according to the or profiles to which the second set refers.

12. Management device (30), adapted to be coupled, via a communication network (16), to at least one client device (10a, 10b), and configured to access a set of resources, characterized in that the management device is configured to customize for a user of said at least one client device, a process of accessing the set of resources by:

• collecting (1 10) information relating to events occurring in connection with the resources;

Creating, based on the collected information relating to the events occurring in relation to all the resources, or obtaining (120, 122, 124, 126) a first profile associated with the user, the profile comprising:

at least one context adapted to allow the representation of all the conditions likely to influence a user's choice and / or a judgment of relevance of the user relative to the resources of the set of accessible resources through the information processing system; and, o a first set including at least one a priori adapted to allow the determination, for each resource of the set of resources, a level of relevance for the user according to at least one measurable property of said resource; and,

Calculating (130, 132, 134, 136) a probability that a first resource, among the set of resources, is relevant to the user based on said first profile and implementing the access process as a function of said probability.

13. Management device according to claim 12, configured so that the probability that the first resource is relevant to the user is obtained by:

Calculating (134) the probability that the first resource is relevant to the reference users of the second set;

Calculating (136) the probability that the first resource is relevant to the user Ui based on the first profile and the probability that the first resource is relevant to the reference users of the second set.

The management apparatus of claim 12, configured so that said first profile associated with the user is updated (120, 122, 124, 126) as a result of a modification of the user's context.

15. Computer program comprising instructions for performing the steps of the method according to any one of claims 1 to 1 1, when said program is executed by a processor.

16. A computer-readable recording medium on which is recorded a computer program comprising instructions for carrying out the steps of the method according to any one of claims 1 to 11.