FR3005769A1 - MULTICRITERIC DECISION SUPPORT SYSTEM WITH AUTOMATIC GENERATION OF EXPLANATIONS AND CORRESPONDING METHOD - Google Patents

Abstract

The system comprises: - a knowledge base 1 forming a multi-criteria decision model, storing, in a hierarchical manner on several levels, criteria concerning a given decision-making, possible values associated with the criteria, criteria weights associated with the criteria, and interaction weights associated with the interactions between criteria, a given combination of possible values of the criteria forming an alternative, - an evaluation module 2, to determine a score for a given alternative, according to the criteria weights, the weights of interaction, and values. a module 4 for determining, for criteria belonging to several levels of the model, contributions of the criteria weights, interaction weights and values, in a score evaluated by the evaluation module for a given alternative; module 3 generation of explanations on the score of a given alternative. from contributions determined by the determination module.

Description

The present invention relates to a multicriterion decision support system, and a corresponding method. It finds an application in particular to the evaluation of alternatives relating to a decision to be made in a multicriteria-type context, for example when it comes to choosing the location of a plant, an investment strategy, evaluate projects, etc. Multi-criteria decision support systems and processes are based on a multi-criteria decision-making model that is drawn from the knowledge of experts in the field of application concerned. Methods and modeling tools exist that make it possible to build and organize the different elements constituting a multi-criteria decision model. These models group criteria and possibly aggregations of criteria. A combination of criteria determines an alternative, which can be quantitatively evaluated through utility functions or cost functions. Thus, multi-criteria decision support systems make it possible to evaluate the alternatives available to a decision-maker, and thus to make a decision quickly which can be complex, in particular because it involves many criteria, depending on the calculated scores for different alternatives. One of the problems with multicriteria decision support systems, with evaluation of alternatives, is that providing the user with raw evaluation data is not sufficient for informed decision-making. Thus, it can be very important for the user to know why an alternative is better or worse than another. Also, regardless of any comparison between alternatives, the user may wish to know the reasons for a given score for a given alternative. A decision support system that only evaluates alternatives is not enough to meet such a need, because it does not explain the origin of a score of an alternative. Document US 7,236,964 discloses a method for generating explanations for explaining why an alternative is good or bad, or why it is better than another. This process generates explanations based on the criteria or aggregations of a given level in the model, using a method using four anchors for generating explanations. These anchors are not related to the contribution of the different elements in the calculation of the difference between the scores of two compared alternatives, or in the calculation of the evaluation of an alternative. However, elements of different levels in the model can have a strong contribution to the outcome of the evaluation of an alternative. Such a method therefore does not make it possible to generate explanations based on the contribution of the various criteria of the model when this model is organized on several levels. One of the aims of the invention is therefore to solve the aforementioned problems, in particular to propose a system and a method for generating explanations on the score of an alternative, or on the comparison of an alternative with respect to another , based on the contributions of the different elements influencing this score, regardless of the level of these elements in the multi-criteria and multi-level decision-making model.

Thus, according to a first aspect, the subject of the invention is a multicriterion decision support system comprising a knowledge base, an alternative evaluation module, an explanations generating module, and a module determination of contributions. The knowledge base forms a multi-criteria decision model. In this database are stored criteria relating to a given decision-making, possible values associated with all or part of the criteria, criteria weights associated with all or part of the criteria, and interaction weights associated with the interactions between criteria, a given combination of possible values of at least a part of the criteria forming an alternative.

The alternative evaluation module is able to determine a score for a given alternative, depending on the criteria weights, and / or interaction weights, and / or values. The explanation generation module is able to construct explanations on the score of a given alternative in an absolute manner or in relation to the score of one or more other alternatives. Criteria are stored in the knowledge base in a multi-level hierarchical manner, with groups of terminal criteria grouped at the terminal level under higher level aggregated criteria, the aggregated criteria being themselves grouped at the top level under other criteria. aggregate criteria of even higher level.

The contribution determination module is able to determine, for criteria belonging to several levels of the multicriterion decision model, the contributions of the criteria weights, and / or the interaction weights and / or values, in a score evaluated by the alternative evaluation module for a given alternative.

In addition, the explanation generation module is capable of constructing explanations on a score of a given alternative, in an absolute manner or in relation to the score of one or more other alternatives, from contributions determined by the module of determination of contributions. According to other embodiments, the system further comprises one or more of the following characteristics, taken in isolation or in any technically possible combination: the alternative evaluation module uses an integral type cost function 2-additive Choquet or weighted sum type to determine a score for a given alternative; the contribution determination module is able to determine the contributions of the weights and / or interaction weights associated with the higher level aggregated criteria in the multicriteria decision model; the contribution determination module is able to determine the contributions of the weights and / or the interaction weights associated with the end criteria in the multicriteria decision model; the contribution determination module is able to determine the contributions of the values associated with the criteria; the contribution determination module is able to carry out a normalization of the determined contributions, preferably between the values 0 and 1; the contribution determination module is able to determine the positive contributions of the criteria weights, and / or interaction weights and / or values, in a score evaluated by the alternative evaluation module for a given alternative ; the contribution determination module is capable of determining the negative contributions, below a given threshold, of the criteria weights, and / or the interaction weights and / or values, in a score evaluated by the evaluation module alternatives for a given alternative; the module for determining the contributions is able to filter the contributions so as to keep only the N strongest contributions in absolute value, with N> 0, preferably N = 4; the explanation generation module is able to generate an explanation of the score of the alternative given with respect to the score of another alternative, the explanation presenting one or more contributions that contribute to the superiority of the score of the alternative; one of the two alternatives on the score of the other; the explanation generation module is capable of generating an explanation of the score of the alternative given with respect to the score of another alternative, the explanation presenting one or more contributions which contribute to the inferiority of the score of one of two alternatives on the score of the other; the possible scores of alternatives, the possible values of criteria, the possible values of weight and the possible values of interaction weight are grouped by interval, a level value being associated with each interval, and the generation module of explanation is able to generate an explanation presenting one or more contributions of a given criterion weight, or a given interaction weight, or a given criterion value, the explanation using the level value associated with the criterion interval to which the given weight of criteria belongs, or the given interaction weight, or the given criterion value; the system comprises a base of patterns in which at least one pattern containing one or more variables corresponding to an alternative name, and / or a value of criterion weight level and / or interaction weight between criterion, and / or an alternative value, and the explanation generation module is able to generate an explanation using the pattern; the explanation generation module is capable of generating an explanation by replacing the variables contained in the pattern by the values corresponding to the alternative name, and / or the value of the weight level and / or the interaction weight. , and / or the value of alternative, relating to a given alternative; the explanation generation module is suitable, when an explanation concerns the contribution of an interaction weight associated with two criteria presenting two values whose difference is smaller than a given threshold, to generate a complementary explanation on the discontinuity the explanation according to the variations of the values; the explanation generation module is capable of constructing explanations of the score of the alternative given with respect to a higher score corresponding to a fictitious alternative and / or compared to a lower score corresponding to a fictitious alternative.

The subject of the invention is also, according to a second aspect, a decision support method comprising a step of constructing a multicriterion decision model, a step of evaluating an alternative by means of a module of evaluating alternatives, a step of generating explanations by means of an explanation generating module, and a step of determining contributions by means of a contribution determining module. The multicriteria decision-making model includes criteria concerning a given decision-making, possible values associated with all or some of the criteria, criteria weights associated with all or some of the criteria, and interaction weights associated with the interactions between criteria, a combination data of possible values of at least part of the criteria forming an alternative, the multi-criteria decision model being stored in a knowledge base. The step of evaluating an alternative determines a score for a given alternative, based on criteria weights, and / or interaction weights, and / or values.

The explanation generation step constructs explanations of the score of the given alternative, either absolutely or in relation to the score of one or more other alternatives. The criteria are organized in the knowledge base in a hierarchical manner on a terminal level and at least a higher level, with groups of terminal criteria grouped at the terminal level under higher level aggregated criteria, the aggregated criteria being able to be themselves grouped at the top level under other aggregate criteria of even higher level. The contribution determination step determines, for criteria belonging to several levels of the multicriterion decision model, the contributions of the criteria weights, and / or the interaction weights and / or values, in the score evaluated during the evaluation. step of evaluating alternatives for the given alternative. In addition, the explanation generating step constructs the explanations of the score of the given alternative, in an absolute manner or with respect to the score of one or more other alternatives, from the contributions determined during the step of determination of contributions. According to other embodiments, the method further comprises one or more of the following characteristics, taken in isolation or in any technically possible combination: the evaluation step uses an integral type cost function of Choquet 2- additive or weighted sum type; the contribution determination step determines the contributions of the weights and / or interaction weights associated with the higher level aggregated criteria in the multicriteria decision model; the contribution determination step determines the contributions of the weights and / or interaction weights associated with the end criteria in the multicriteria decision model; - the contribution determination stage determines the contributions of the values associated with the criteria; the step of determining the contributions performs a normalization of the determined contributions, preferably between the values 0 and 1; the contribution determination step determines the positive contributions of the criteria weights, and / or the interaction weights and / or values, in the evaluated score during the alternatives evaluation step for the alternative given; the contribution determining step determines the negative contributions, below a given threshold, of the criteria weights, and / or interaction weights and / or values, in the score evaluated during the evaluation step alternatives for the given alternative; the contributions determined during the step of determining contributions contributions are filtered so as to keep only the N strongest contributions in absolute value, with N> 0, preferably N = 4; the explanation generation step generates an explanation of the score of the alternative given with respect to the score of another alternative, the explanation presenting one or more contributions that contribute to the superiority of the score of one of the two alternatives on the score of the other; - the explanation concerns the contribution of a weight that is not a minimum weight, or an interaction weight that is not a minimum interaction weight, or a value that is not a minimum value, associated with a criterion; the explanation generation step generates an explanation of the score of the alternative given with respect to the score of another alternative, the explanation presenting one or more contributions that contribute to the inferiority of the score of one two alternatives on the score of the other; the explanation presents the contribution of a weight that is a minimum weight, or an interaction weight that is a minimum interaction weight, or a value that is a minimum value, associated with a criterion; the possible scores of alternatives, the possible values of criteria, the possible values of weight and the possible values of interaction weight are grouped by interval, a level value is associated with each interval, and the generation step explanation generates an explanation presenting one or more contributions of a given criterion weight, or a given interaction weight, or a given criterion value, the explanation using the level value associated with the interval to which belongs the given weight of criteria, or the given interaction weight, or the given criterion value; the explanation generating step generates an explanation by using at least one pattern containing one or more variables corresponding to an alternative name, and / or a criterion weight level value and / or a weight value of interaction between criterion, and / or an alternative value, the patterns and variables being stored in a pattern database; the explanation generation step generates an explanation by replacing the variables contained in the pattern by the values corresponding to the alternative name, and / or the value of the weight level and / or the interaction weight, and / or the value of alternative, relating to a given alternative; when an explanation concerns the contribution of an interaction weight associated with two criteria having two values whose difference is smaller than a given threshold, the explanation generating step generates a complementary explanation on the discontinuity of the explanation according to the variations of the values; the explanation generation step constructs explanations on the score of the given alternative with respect to a higher score corresponding to a fictitious alternative and / or compared to a lower score corresponding to a fictitious alternative. Thus, by using the contribution of the elements of the model, without restricting itself to a given level, the system and the method of the invention make it possible to identify the strongest contributions and, secondly, to exploit them to generate explanations, the all without the need for language knowledge specific to one particular area or another. The features and advantages of the invention will appear on reading the description which follows, given solely by way of example, and not by way of limitation, with reference to the following appended figures: FIG. 1: schematic representation of a first example of multilevel multi-criteria decision model; FIG. 2: schematic representation of a second example of multi-criteria multi-criteria decision model; FIG. 3: schematic representation of a third example of multi-criteria multi-criteria decision model; FIG. 4: schematic representation of an exemplary system according to the invention.

Figures 1 to 3 show three different examples of applications in which a multilevel multi-criteria decision model can be constructed. The schematic representation takes the form of a tree structure in which are organized the various criteria making it possible to model the making of a decision. It can be seen from the three figures 1 to 3 that the criteria A, C are grouped hierarchically. More precisely, criteria C are defined at the terminal level. Such criteria C have no children in the tree. Symbolically, we attach to these terminal criteria C universes U, which may or may not have the same name as the criteria C. These criteria C are grouped by aggregation under higher-level criteria, which are called aggregations A. Aggregations A can also be grouped by aggregation under higher-level A criteria, and so on. The highest-level aggregation A symbolically represents the application that is being modeled. For example, Figure 1 is for a site protection assessment model, Figure 2 is for a home purchase model, and Figure 3 is for a model for the assessment of a site protection. 'a remote meeting. The model is stored in a knowledge base 1. This knowledge base 1 is broken down into a base 1b in which the elements defining the model strictly speaking are stored: the criteria and aggregations A, C (with their hierarchical organization) , the values v of these criteria A, C, the weights ak of these criteria A, C as well as the weights bi, k relating to the interactions between criteria A, C. The other part la of the knowledge base 1 includes the description alternatives E, W that correspond to different possible decisions. A combination of value v of end criteria C determines a possible alternative.

Values v can be text values. It is emphasized that the representation, in FIG. 4, of the two parts 1a and 1b of the knowledge base 1 as two sub-parts of this knowledge base 1, is a purely functional, non-limiting representation of the invention. These two parts can be information storage areas that are physically separate or even distant from each other.

The system and method of the invention will now be described with reference to FIG. 4 and the application modeled in FIG. 3. More precisely, FIG. 3 looks at the case of meeting management tools, and more particularly to tools allowing remote access (via terminals: PC, SmartPhone, etc ...) and a training meeting recording. This model distinguishes at the first level (just below the main aggregation "Remote Meeting Evaluation") the functional aspects, which are related to management during the meeting and after the meeting, of non-functional aspects, which are related to the security and the type of terminals that can be used. Learning has made it possible to inform the different weights (including the interaction weights) of the model, which are not shown in FIG. 3. We will try to compare two alternatives E and W, defined in Table 1 below. Universes (U) associated with alternative E alternative W criteria (C) Indexation recording Semantic indexing Breakdown by chapters (criterion Search infos) Construction report CR construction possible no construction CR Shared screen part vision of the speaker and the screen support speaker and screen support Audio quality Very good, average, weak Very good, average, weak Exchange of information public or private exchanges public exchange in writing History History available No history available Safety Security simple Reinforced security Terminals usable All machines and all All machines and all operating systems Operating systems Table 1 Once the model is built, we will proceed to the evaluation of each alternative, by means of a module of evaluation 2, which allows to assign a score to each alternative.

Thus, the value of an alternative on a given aggregation A corresponds to the result given by a cost function, such as a Choquet integral or a weighted sum, on the criteria A or C aggregated under this aggregation A.

The overall satisfaction, or score, attached to an alternative corresponds to the result of the cost function applied to each aggregation. Preferably, for calculating the score of an alternative, a 2-additive Choquet integral is used, which makes it possible to take into account the possible interactions between criteria. In particular, we are interested in substitutability, or negative synergy. It is a negative interaction (only one criterion can be satisfied), whose weight is attributed to the maximum value of the alternative on the two interacting criteria. We are also interested in complementarity, or positive synergy. It is a positive interaction (both criteria must be satisfied), whose weight is attributed to the minimum value of the alternative on the two interacting criteria. For a given aggregation, the local score, ie on a single level, of an alternative E then corresponds to the computation of the following Choquet 2-additive integral: (1) H (xi, xn) = EkE M Vk X xk - k} emk X Ixj - xk1 / 2 with: - x1, ..., xn are the n criteria or sub-aggregations A, C aggregated under an aggregation A on which the alternative E is evaluated; vk is the weight of the criteria k; xk is the value of the alternative E on the criterion k; - Ij k represents the interaction between the criteria j and k; M represents all the criteria at the level considered. Another writing of this integral of 2-additive Choquet can be given as follows: (2) H (x1, xn) = EkE M ak X Xk E {l, k} EM bj, k X min (xj, Xk ) + k} EM cj, k X max (xj, Xk) with, in addition to the above definitions: ak corresponds to the minimum weight of criterion k; k corresponds to the complementarity between the criteria j and k; cj, k corresponds to the substitutability between the criteria j and k.

Thus, the local coefficient in favor of an alternative corresponds to all the weights that highlight a force of the alternative. The local coefficient to the detriment of an alternative corresponds, for its part, to the set of "min", ie the set of minimum weights that highlight a weakness of the alternative rather than a force.

To compute the overall satisfaction or utility of an alternative, as part of a multilevel multicriterion model, the Choquet integral is applied at each level to obtain local satisfactions or scores on each aggregation. The computation of the global score of an alternative E, hence of its global utility, can then be written based on the local scores, as follows: (3) H (E) = N rik'eAk ak, (E) x ak (E) x Ek where Ak corresponds to the set of parent aggregations of criterion k, N is the set of criteria of the model and ak is the local coefficient of criterion k for alternative E (the expression 11 ke Ak ak '(E) x ak (E) corresponds to the overall coefficient of criterion k for alternative E).

In the case of the comparison between two alternatives E and W, we will be interested in the difference between H (E) and H (W). In the case of the evaluation of an alternative E, in an absolute way, that is to say without comparison with another alternative, this one is assigned to a level: "very bad", "bad", " medium "," good "," very good ".

To explain why an alternative is assigned to a given level, we explain why this alternative is no less good and why it is not better. For example, for a "good" alternative, explain why she is not "average" and why she is not "very good". For this, two comparisons are used: a comparison with a fictional alternative "average" (for example presenting a standardized score of 0.6 on all the criteria) and a comparison with a fictitious alternative "very good" (for example having a score of 0.8 on all criteria). By way of example, according to the own model used by the applicant in the context of the example of FIG. 3, the normalized score obtained for the alternative E, denoted U (E), is the following: U (E) = 77% Similarly, the normalized score obtained for the alternative W, denoted U (W), is the following: U (W) = 43%.

We thus distinguish the preferred alternative E, which has the highest score, from the other alternative W, which has the lowest score. The scores of the alternatives thus evaluated are stored in a storage area 7, and a first type of explanations EXP can be generated automatically by an explanation generation module 3, for example: "E" is preferred to "W" on "Remote Meeting Evaluation" A contribution determination module 4 also makes it possible to determine the contributions Cb of the criteria weights ak, and / or the interaction weights l0, k and / or values y, in an evaluated score. by the alternative evaluation module 2 for a given alternative E, W. The calculation of these contributions Cb is preferably performed by several different functions, which allows to multiply the points of view. The existence of interactions between criteria has a significant impact on the overall score of an alternative.

We are thus in the presence of several elements that may have a significant contribution to the score of an alternative: the minimum ak weights of the criteria and aggregations A, C, the interactions bj, k between criteria and aggregations A, C, the values v on criteria and aggregations. The weights and interactions (also called interaction weights) are on the same scale, while the values on the criteria and aggregations are comparable. Different calculations are therefore used to evaluate the contributions of criteria and aggregations (which group together minimum weights and interactions associated with values on criteria and aggregations), and the individual contributions of weights and interactions. In addition, weights and interactions are generally higher for higher-level A aggregations than for C-criteria, especially for models with more than three levels. Indeed, if we have a model containing three aggregations A at the higher level, they will be assigned an average weight of 0.33. If each of these aggregations A aggregates three criteria C, these will be assigned an average weight of 0.11. Also, different calculations are used to determine the contributions of the weights and interactions of the higher level and to determine the contributions of the weights and interactions related to the criteria C, in order not to "stifle" criteria that are nevertheless significant in the understanding of the overall assessment by the higher contributions of weights and interactions of the higher level. In addition, a standardization of the contribution calculations makes it possible to return to a common scale for the generation of explanations.

In an exemplary implementation, three contribution calculation functions are used, each function being associated with a perspective. The first calculation corresponds to a global perspective. It focuses only on the higher level, and allows to estimate the contribution of the weights and interactions of this level.

The second calculation corresponds to a "model" perspective and makes it possible to evaluate the contribution of each criterion and aggregation of the model. The third calculation corresponds to a detailed perspective. It focuses on the weights associated with the criteria: interactions, minimum weight and local coefficients. Thus, the contribution determination module 4 makes it possible to identify the elements having the greatest contribution, in the case of the comparison between two alternatives E, W, and in the case of the evaluation of an alternative. These elements with the highest contribution are then used for the generation of explanations EXP, by the module of generation of explanations 3. The calculation of the gross contribution Cb multi-level of a criterion (or an aggregation of criteria ) i in the preference for the alternative X with respect to an alternative Y, can be written: cbx> y (i) = II the E Ai ai ', pro (X) X ai, Pro (X) XX - pE Ai ap, p '(Y) x ai, pro (Y) X i Where A' corresponds to the set of parent aggregations of criterion i, ai ,, p, '(X) (respectively ar, p' (Y )) is the local coefficient of the criterion (or aggregation) i 'in a Choquet integral (or in a weighted sum) in favor of the alternative X (respectively Y), and X, (respectively g,) is the value of the alternative X (respectively Y) on the criterion (or the aggregation) i. When we want to determine the contribution of a first-level weight, we are interested in: the local coefficient in favor of X and Y (when ai, pro (X) = ai, pro (Y)), at the minimum weight , and negative synergy. More precisely, when ai, pro (X) = ai, pro (Y), we calculate the gross contribution of the local coefficient in favor of X and Y, noted ai, p, '(X, Y), concerning the criterion ( or the aggregation) of the first level i in the preference for the alternative X with respect to an alternative Y, in the following way: Cbx> y, ic M (a1, Pro (X, Y)) = ai, pro ( Moreover, when ai, pro (X) ai, pro (Y), we calculate the gross contribution of the minimum weight a, relating to the criterion (or the aggregation) of the first level i in the preference for the alternative X with respect to an alternative Y, in the following way: Cbx> y, ic M (ai) = a, X (X - i) Finally, always when ai, pro (X ) # ai, p '(Y), we calculate the gross contribution of the negative synergy c ,,, associated with the criteria (or aggregations) i and j in the preference for the alternative X with respect to an alternative Y, of the following: Cbx> y, ie M j) = Ci, j X (max (X, X j) - max (, i)) When one wants to determine the allocation of the terminal weights, we are interested in the weights directly related to the criteria: local coefficient has in favor of X and Y, minimum weight, and negative synergy.

More precisely, when ai, pro (X) = ai, p, - ,, (Y), we calculate the gross contribution of the local coefficient in favor of X and Y, denoted ai, pro (X, Y), modulo the product of weight of parent aggregations, denoted n β P-, relating to the criterion i in the preference for the alternative X with respect to an alternative Y, as follows: Cbx> y, iG N (ai, Pro ( X, Y) ° A Pi ') = 11 i'E A Pi' X ai, pro (X, Y) x - I) where pr is the weight of a parent aggregation i 'of criterion i. When ai ,, pro (X) = ap, pro (Y), then pi '= ai ,, p' (X, Y). Otherwise pc = ai, or pr = Moreover, when ai, pro (X) # ai, pro (Y), one calculates the multilevel gross contribution of the minimum weight a,, modulo the product of weight of parent aggregations, denoted H PG Ai p, -, relating to the criterion i in the preference for the alternative X with respect to an alternative Y, in the following way: Cbx> y, the N (ai ° re Ai Pr) = the G Ai Pi Finally, always, when ai, pro (X) # ai, p '(Y), we compute the multilevel gross contribution of the negative synergy c ,, i, modulo the weight product of parent aggregations. , denoted by the E A pc, associated with the criteria i and j in the preference for the alternative X with respect to an alternative Y, as follows: Cbx> y, the N (ci, 1 ° 11 i'E Ai Pi) = 11 re Ai pi, x ci, x (max (X i, X j) - max (, As the final weight contribution calculations presented above show, the contribution of the weight of a criterion, or terminal weight, is therefore a function of the weights on the parent aggregations To illustrate these contribution calculations of a terminal weight, we can use the following simplified multilevel multi-criteria model: Level 1: Aggregation i "(associated with a minimum weight ar) Level 2: Aggregation i '(associated with a local coefficient ar, pro (X, Y)) Level 3 or terminal level: criterion i and criterion j It is also assumed that Xi> Xj and that, <j. And we posit that a, is the minimum weight on criterion i, ci, j is a substitutability (or negative syngery) between criteria i and j The proportion of weights associated with criterion i in the difference between the global score of X and the overall score of Y is highlighted by the following calculation: H (X) - H (Y) - (nk,. Ak pk '(X) X ak (X) x X k - 1 1 kE Ak MY ) x ak (Y) xk) + ai .. x ar, pro (X, Y) x ai x -;) + ai ,. For this example, we obtain two raw contributions associated with criterion i: one relating to the minimum weight of criterion i, the other relating to the interaction with the criterion j: Cbx> y, ie NIE N, the E, i "E Al '(ai ° ai, pro (X, Y)) ar) x ar, pro (X, Y) x ai x (X i - i) Cbx> y, JE N, jE N, re Ai, 1 "E Ai '(Ci, j ° ai', Pro (X, Y) ° ai") = ai ,,, X ai ,, pro ( X, Y) X ci, x X The distinction of the weights ai and ci, j can thus make it possible to emphasize the specific role of the interaction between the criteria i and j in a preference between two alternatives X and Y, for example when x, = y, = 0.9, and when the combination of weights a, - x ai, p '(X, Y) xc ,, 1 is large (high value).

For each contribution calculation Cb, the result is normalized between 0 and 1 to obtain a normalized contribution Cn. The normalization can constitute to carry out the following operation: Cnx> y (i) = Cb »Y (i) I Eke N ICbX > Y (k) 1 Here, i corresponds to a criterion or a combination of elements, N corresponds to the set of criteria or combinations of elements. The standardized contribution thus obtained is expressed as a percentage, which is simpler to understand, and which retains the sign of the corresponding gross contribution.

Thus, even if contribution calculations of different types are used, covering different elements (criteria or aggregations, element of the first level or terminal elements, ...), the standardized contributions are comparable on the same scale. In the example corresponding to FIG. 3, a first calculation, focused on the elements (weights and interactions) of the first level, makes it possible to obtain the following normalized contributions: Contribution (Weight (Functional Aspects)) = 0.56 Contribution (Complementarity ( Functional aspects, Non-functional aspects)) = 0.27 A second calculation, focused on the end elements (weight and interactions associated with the criteria), makes it possible to obtain the following standardized contributions: Contribution (Overall weight (CR preparation)) = 0.188 Contribution (Overall weight (Search infos)) = 0.176 A third calculation, focused on the criteria and aggregations, allows to obtain the following standardized contributions: - Contribution (Functional Aspects) = 0.418 Contribution (Management After) = 0.128 Contribution (Management During) = 0.07 Preferably, once all positive contributions (and possibly strong negative contributions) have been grouped together, keep only a limited number of explanations for the explanations, for example four, which have the largest contribution, in ascending order. ^. ^ Similarly, we group together certain elements (in the example of Figure 3: "Functional Aspects" is present via its weight in the first calculation, and via the third calculation), in order to avoid repetitions of names of criteria or aggregations. The filtering and this grouping are performed by a filtering module 6.

The generation of explanations EXP, through the module of generation of explanations 3, then makes it possible to transform into text of each element, relying on knowledge relating to the importance of the weight, which depends on each model , and the value of the element. This knowledge of the importance of the weights and the value of the elements is grouped in a storage area 8.

For the weights, we define, according to the weights of the model, value slices associated with "very important", "important", "little important" and "very little important". For the values y, we define thresholds at 0, 0.2, 0.4, 0.6, 0.8 and 1 to which we associate the terms "null" (v = 0), "very bad" (0 <v <0.2), "bad" , "Medium", "good", "very good" (0.8 <v <1), "maximal" (v = 1).

For the element with the highest contribution (weight of the aggregation "Functional aspects"): weight is considered very important. The value of "E" here is very good while the value of "W" is bad here. Hole sentences or patterns, stored as sentences and variables in a pattern database 5 are used.

From the following hole phrase "on the aggregationWeight aqreqationName, preferredAlternativeName is value0fPreferredAltOnAq while otherAlternativeName is valueOfOtherAltOnAq. We obtain the following sentence "on the very important Functional aspects, E is very qood while W is bad. ". For the element with the second highest contribution (complementarity between the "functional aspects" and "non-functional aspects" aggregations), weight is considered very important. A complementarity highlighting a weakness, we only look at the value for "W" which is the one on "functional aspects", evaluated as bad. The following sentence is used: "The complWeight complementarity between aggregation1Name and agregation2Name penalizes otherAlternativeName which is valueOfOtherAltOnComplementarity on aggregationNameVal0f0therAltOnCompl. ". We obtain the following sentence: "The very important complementarity between Functional aspects and Not functional aspects penalizes W which is bad on Functional aspects. ".

If the complementarity was not managed separately, we would look at the contribution of the overall weight of "non-functional aspects" for "E" (light red and turquoise surfaces) compared to the overall weight for "W" (light red surface), the contribution would be the most important and the generated explanation would look like "" not functional aspects "reinforces" E "which is good while" W "is maximal. ". For the element with the third highest contribution (criterion "CR preparation"): the weight is considered important, and the value is maximum for "E". The following sentence is used: "On the CriterionWeight" criterionName "," preferredAlternative "is reinforced by its valueOfPrefAltOnCriterion score. ". We obtain the following sentence: "On the important" Preparation of the report "," E "is reinforced by its maximal score. Link words and enhancements (bold) are finally used to obtain the following explanatory text: "E" is preferred to "W" on "Evaluation of the remote training meetings" since - on the very important "Functional aspects", "E" is very good while "W" is bad. - The very important complementarity between "Functional aspects" and "Not functional aspects" penalizes "W" which is bad on "Functional aspects". - On the important "Preparation of the report", "E" is reinforced by its maximum score. Moreover, in the case of an interaction between two criteria (or aggregations) having close values (the difference between these values being less than a determined threshold), to allow a user having slightly modified the value of an alternative of to understand why new explanations are generated that no longer relate to the same elements, a complementary explanation can be added to the explanation associated with these criteria or aggregations. This is an explanation that indicates to the user that the contribution of the element in question may change if the values close to the alternative on one criterion and on another criterion are reversed because of their interaction.

The present description is given by way of non-limiting example of the invention. In particular, the representation of the architecture of the system as given in the example of FIG. 4 is a functional representation, which does not necessarily prejudge the exact organization of the different functional elements (separate, shared, remote, etc.). ..) included in the system.

Likewise, the calculations for determining the contributions Cb given in the present description are examples of nonlimiting calculations of the invention, which notably make it possible to take into account the multi-criteria and multi-level nature of a decision model. Other types of calculations can be used, which allow to take into account this character.

Claims (34)

CLAIMS1.- A multicriterion decision support system comprising: - a knowledge base (1) forming a multi-criteria decision model, in which are stored criteria (A, C) concerning a given decision-making (D), values possible (v) associated with some or all of the criteria (A, C), criteria weights (ak) associated with some or all of the criteria (A, C), and interaction weights (bi, k, q, k) associated with the interactions between criteria (A, C), a given combination of possible values (v) of at least part of the criteria (A, C) forming an alternative (E, W), - an evaluation module alternatives (2), able to determine a score for a given alternative (E, W), according to the criterion weights (ak), and / or interaction weights (bJ, k, ci, k), and / or values (y), - an automatic explanations generation module (3) capable of automatically constructing explanations (EXP) on the score of a given alternative (E ) in absolute terms or with respect to the score of one or more other alternatives (W), characterized in that - the criteria (A, C) are stored in the knowledge base (1) in a hierarchical manner on several levels, with groups of terminal criteria (C) grouped at the terminal level under higher level aggregated criteria (A), the aggregated criteria (A) which can themselves be grouped at the higher level under other aggregated criteria (A) of level still superior, - the system also comprises a contribution determination module (4), able to determine, for criteria (A, C) belonging to several levels of the multicriterion decision model, the contributions (Cb) of the criteria weights (ak ), and / or interaction weights (bi, k, ci, k) and / or values (y), in a score evaluated by the alternative evaluation module (2) for a given alternative (E , W), - the explanation generation module (3) is apt e to automatically construct explanations (EXP) on a score of a given alternative (E), absolutely or in relation to the score of one or more other alternatives (W), from contributions (Cb) determined by the contribution determination module (4). 2. System according to claim 1, characterized in that the alternative evaluation module (2) uses a cost function (H) integral type Choquet 2-additive or weighted sum type to determine a score for a given alternative (E, W). 3. System according to any one of claims 1 and 2, characterized in that the contribution determination module (4) is able to determine the contributions (Cb) weights (ak) and / or interaction weights (bj, k, ci, k) associated with higher level aggregated criteria (A) in the multicriteria decision model. 4. System according to any one of claims 1 to 3, characterized in that the contribution determination module (4) is able to determine the contributions (Cb) weights (ak) and / or interaction weights (bj, k, ci, k) associated with the terminal criteria (C) in the multicriteria decision model. 5. System according to any one of claims 1 to 4, characterized in that the contribution determination module (4) is able to determine the contributions (Cb) of the values (v) associated with the criteria (A, C). . 6. System according to any one of claims 1 to 5, characterized in that the contribution determination module (4) is adapted to perform a normalization of the contributions (Cb) determined, preferably between the values 0 and 1. 7. System according to any one of claims 1 to 6, characterized in that the contribution determination module (4) is able to determine the positive contributions (Cb) of the criteria weights (ak), and / or interaction weight (bj, k, cpk) and / or values (y), in a score evaluated by the alternative evaluation module (2) for a given alternative (E, W). 8. System according to any one of claims 1 to 7, characterized in that the contribution determination module (4) is able to determine the contributions (Cb) negative, below a given threshold, of the weights of criteria ( ak), and / or interaction weights (bj, k, ci, k) and / or values (v), in a score evaluated by the alternative evaluation module (2) for a given alternative ( E, W). 9. System according to any one of claims 1 to 8, characterized in that the contribution determination module (4) is able to filter the contributions (Cb) so as to keep only the N contributions (Cb) most strong in absolute value, with N> 0, preferably N = 4. 10. System according to any one of claims 1 to 9, characterized in that the explanation generating module (3) is able to generate an explanation (EXP) on the score of the given alternative (E) by compared to the score of another alternative (W), the explanation (EXP) presenting one or more contributions (Cb) which contribute to the superiority of the score of one (E) of the two alternatives (E, W) on the score of the other (W). 11. System according to any one of claims 1 to 10, characterized in that the explanation generating module (3) is able to generate an explanation (EXP) on the score of the given alternative (W) by compared to the score of another alternative (E), the explanation (EXP) presenting one or more contributions (Cb) which contribute to the inferiority of the score of one (W) of the two alternatives (E, W) on the score of the other (E). 12. System according to any one of claims 1 to 11, the possible scores of alternatives (E, W), the possible values (y) of criteria (A, C), the possible values of weight (ak) and the possible values of interaction weight (1) ;, k, ci, k) being grouped by interval, a level value being associated with each interval, characterized in that the explanation generating module (3) is suitable generating an explanation (EXP) having one or more contributions (Cb) of a given criterion weight (ak), or a given interaction weight (bbk, cpk), or a given value (y) criterion (A, C), the explanation (EXP) using the level value associated with the interval to which said given criterion weight (ak) belongs, or said interaction weight (bi, k, ci, k) given, or said given value (v) of criterion (A, C). 13.- System according to any one of claims 1 and 12, characterized in that it comprises a base of patterns (5) in which is stored at least one pattern containing one or more variables corresponding to an alternative name ( E, W), and / or a criterion weight level value (ak) (A, C) and / or interaction weight (bi, k, ci, k) between criterion (A, C), and / or at an alternative value (E, W), and in that the explanation generation module (3) is able to generate an explanation (EXP) using said pattern. 14.- System according to claim 13, characterized in that the explanation generation module (3) is able to generate an explanation (EXP) by replacing the variables contained in the pattern by the values corresponding to the alternative name ( E, W), and / or the weight level value (ak) and / or interaction weight (bj, k, ci, k), and / or the alternative value (E, W) , relating to a given alternative. 15.- System according to any one of claims 1 to 14, characterized in that the explanations generating module (3) is fit, when an explanation (EXP) relates to the contribution (Cb) of a weight d interaction (bj, k, cj, k) associated with two criteria presenting two values (v) whose difference is smaller than a given threshold, to generate a complementary explanation on the discontinuity of the explanation (EXP) according to the variations said values (v). 16. System according to any one of claims 1 to 15, characterized in that the explanations generating module (2) is able to construct explanations (EXP) on the score of the given alternative (E). compared to a higher score corresponding to a fictitious alternative and / or compared to a lower score corresponding to a fictitious alternative. 17. A method for decision support comprising: a step of constructing a multicriterion decision model comprising criteria (A, C) concerning a given decision-making (D), possible values (v) associated with any or part of the criteria (A, C), criteria weights (ak) associated with all or some of the criteria (A, C), and interaction weights (bj, k, ci, k) associated with the interactions between criteria ( A, C), a given combination (E, W) of possible values (y) of at least part of the criteria (A, C) forming an alternative (E, W), said multi-criteria decision model being stored in a base of knowledge (1), - a step of evaluating an alternative (E) by means of an alternative evaluation module (2), which determines a score for a given alternative (E, W), in according to criteria weights (ak), and / or interaction weights (bj, k, ci, k), and / or values (v), an automatic explanations generation step (EXP), a by means of an automatic explanations module (3), which automatically constructs explanations (EXP) on the score of the given alternative (E), absolutely or in relation to the score of one or more others alternatives (W), characterized in that: - the construction step of the multicriterion decision model organizes, in the knowledge base (1), the criteria (A, C) in a hierarchical manner on a terminal level and at least one level higher, with groups of terminal criteria (C) grouped at the terminal level under higher level aggregated criteria (A), the aggregated criteria (A) which can themselves be grouped at the higher level under other aggregated criteria (A). ), the method also comprises a step of determining contributions (Cb), by means of a contribution determination module (4), which determines, for criteria (A, C) belonging to several levels of a multicriteria decision model, the contributions (Cb) of criteria weights (ak), and / or interaction weights (bpk, cpk) and / or values (y), in the score evaluated during step d evaluation of alternatives (2) for the given alternative (E, W),. the automatic explanations generation step (EXP) automatically constructs the explanations (EXP) on the score of the given alternative (E), absolutely or in relation to the score of one or more other alternatives (W) from the contributions (Cb) determined during the contribution determination step. 18. A method according to claim 17, characterized in that the evaluation step uses a cost function (H) integral type Choquet 2-additive or weighted sum type. 19. A process according to any one of claims 17 and 18, characterized in that the step of determining the contributions (Cb) determines the contributions (Cb) of the weights (ak) and / or the interaction weights ( bbk, ci, k) associated with higher level aggregated criteria (A) in the multicriteria decision model. 20. A process according to any one of claims 17 to 19, characterized in that the step of determining the contributions (Cb) determines the contributions (Cb) of the weights (ak) and / or the interaction weights (bi , k, ci, k) associated with the end criteria (C) in the multicriteria decision model. 21.- Method according to any one of claims 17 to 20, characterized in that the step of determining the contributions (Cb) determines the contributions (Cb) of the values (y) associated with the criteria (A, C). 22. The method according to claim 17, wherein the step of determining the contributions (Cb) performs a normalization of the contributions (Cb) determined, preferably between the values 0 and 1. 23.- Process according to any one of claims 17 to 22, characterized in that the step of determining the contributions determines the positive contributions (Cb) of the criteria weights (ak), and / or the interaction weights ( bi, k, ci, k) and / or values (v), in the score evaluated in the alternative evaluation step for the given alternative (E, W). 24.- Method according to any one of claims 17 to 23, characterized in that the step of determining the contributions determines the contributions (Cb) negative, below a given threshold, weight criteria (ak), and / or interaction weights (bi, k, ci, k) and / or values (v), in the score evaluated in the alternative evaluation step for the given alternative (E, W). 25. Method according to any one of claims 17 to 24, characterized in that the contributions (Cb) determined during the step of determining the contributions (Cb) the contributions (Cb) are filtered so as to keep only the N contributions (Cb) highest in absolute value, with N> 0, preferably Nz --- 4. 26.- Method according to any one of claims 17 to 25, characterized in that the explanation generating step (EXP) generates an explanation (EXP) on the score of the given alternative (E) with respect to the score of another alternative (W), the explanation (EXP) presenting one or more contributions (Cb) that contribute to the superiority of the score of one (E) of the two alternatives (E, W) on the score of the other (W). 27. The method as claimed in claim 26, characterized in that the explanation (EXP) concerns the contribution of a weight (ak) that is not a minimum weight, or an interaction weight (bj, k). , cj, k) which is not a minimum interaction weight, or a value (v) which is not a minimum value (v), associated with a criterion (A, C). 28.- Method according to any one of claims 17 to 27, characterized in that the explanation generation step (EXP) generates an explanation (EXP) on the scoree the given alternative (W) with respect to the score of another alternative (E), the explanation (EXP) presenting one or more contributions (Cb) which contribute to the inferiority of the score of one (W) of the two alternatives (E, W) on the score of the other (E). 29. The method as claimed in claim 28, characterized in that the explanation (EXP) presents the contribution of a weight (ak) which is a minimum weight, or of an interaction weight (bj, k, cj, k) which is a minimum interaction weight, or a value (v) which is a minimum value (v), associated with a criterion (A, C). 30.- Process according to any one of claims 17 to 29, characterized in that the possible scores of alternatives (E, W), the possible values (v) of criteria (A, C), the possible values of weight (ak) and the possible values of interaction weight (bj, k, cj, k) are grouped by interval, a level value is associated with each interval, and in that the generation step of explanation (EXP) generates an explanation (EXP) having one or more contributions (Cb) of a given criterion weight (ak), or a given interaction weight (bj, k, cj, k), or of a given value (v) of criterion (A, C), the explanation (EXP) using the level value associated with the interval to which said given criterion weight (ak) belongs, or said interaction weight ( bj, k, cj, k), or said given value (y) of criterion (A, C). 31.- Method according to any one of claims 17 to 30, characterized in that the explanation generation step (EXP) generates an explanation (EXP) using at least one pattern containing one or more variables corresponding to a alternative name (E, W), and / or a criterion weight level value (ak) (A, C) and / or interaction weight (bj, k, cj, k) between criterion ( A, C), and / or an alternative value (E, W), the patterns and variables being stored in a pattern database (5). 32.- Method according to claim 31, characterized in that the explanation generation step (EXP) generates an explanation (EXP) by replacing the variables contained in the pattern by the values corresponding to the alternative name (E, W), and / or the value of weight level (ak) and / or interaction weight (4, k, cj, k), and / or the value of alternative (E, W), relative to a given alternative. 33.- Method according to any one of claims 17 to 32, characterized in that, when an explanation (EXP) relates to the contribution (Cb) of an interaction weight (bj, k, ci, k) associated with two criteria having two values (v) whose difference is smaller than a given threshold, the explanation generating step (EXP) generates a complementary explanation on the discontinuity of the explanation (EXP) as a function of the variations of said values (y). 34.- Method according to any one of claims 17 to 33, characterized in that, the explanation generation step (EXP) builds explanations (EXP) on the score of the given alternative (E) relative to at a higher score corresponding to a fictitious alternative and / or compared to a lower score corresponding to a fictitious alternative.