CN115062791A - Artificial intelligence interpretation method, device, equipment and storage medium - Google Patents

Abstract

An artificial intelligence interpretation method, device, equipment and storage medium. The invention discloses an artificial intelligence interpretation method, which comprises the following steps: acquiring an observation data set, analyzing the observation data set to generate a target rule, wherein the target rule comprises a group division rule, a behavior intervention rule and a law adaptation rule; acquiring an exploration target set corresponding to the observation data set, matching each target rule with the exploration target set, and generating a corresponding rule result pair according to a matching result; performing causal relationship exploration on each rule result pair to generate a corresponding rule causal relationship set; and (5) carrying out artificial intelligence explanation according to the rule cause and effect relationship. The artificial intelligence explanation method disclosed by the invention specifically divides the rule cause and effect processing method into three types, namely a group division rule, an action intervention rule and a law adaptation rule, takes the rule as the influence on the service performance due to the change of the analysis rule, better fits the practical application scene, leads the result to be more visual and reliable, avoids a large amount of manual analysis work through a universal rule cause and effect discovery method, and improves the working efficiency.

Description

Artificial intelligence interpretation method, device, equipment and storage medium

Technical Field

The present invention relates to the field of artificial intelligence technology, and in particular, to an artificial intelligence interpretation method, apparatus, device, and storage medium.

Background

With the development of artificial intelligence technology, artificial intelligence application shows explosive growth and is widely used in the fields of face recognition, automatic driving, dialogue systems, financial wind control and the like. However, unlike traditional IT applications, the external data of current artificial intelligence models (hereinafter referred to as "models") represented by machine learning and deep neural networks implicitly assumes a lot of assumptions, has a complex internal structure, and generally has a black box in ITs own operation mechanism, and thus, the influence relationship between the model input characteristics and the model results cannot be interpreted by human understandable language. XAI (interpretable artificial intelligence) can provide human-readable and understandable interpretations for model decisions, helping humans understand and trust the results and outputs produced by machine learning algorithms, establishing a direct understanding of the model's mechanism of operation.

The current main approaches to XAI are divided into the following: 1. providing a correlation interpretation based on statistical information, such as interpretation of the result by feature weights, which is intuitive and easy to understand, but plausible and not strict in interpretability; 2. providing procedural interpretations based on computational logic, such as computational processes that disclose models, such methods easily suggest trust, but computational logic is not equivalent to intelligible logic, interpretability is poor; 3. providing reasoning explanation based on rule reasoning, for example, the explanation aiming at a rule decision method has strong interpretability but large limitation, and is only suitable for a simple rule decision scene with definite meaning; 4. the method provides causal explanation based on causal relationship discovery, such as RCT test or AB test method, the result of the method is objective and stable, and can be reused and deduced, but the experimental method is more restricted by cost and ethical specifications, and the observational method mainly researches the relationship between factors and can not be directly applied to actual service scenes.

In summary, the existing XAI methods have advantages and disadvantages, and the interpretable output depends on the experience of an analyst to some extent, and the interpretation of the result often requires manual intervention, and the automation degree is not high.

Disclosure of Invention

The invention provides an artificial intelligence interpretation method, an artificial intelligence interpretation device, artificial intelligence equipment and a storage medium, which are used for providing human readable and understandable interpretation for artificial intelligence model decisions.

According to an aspect of the present invention, there is provided an artificial intelligence interpretation method, including:

acquiring an observation data set, and analyzing the observation data set to generate a target rule, wherein the target rule comprises a group division rule, a behavior intervention rule and a law adaptation rule;

acquiring an exploration target set corresponding to the observation data set, matching each target rule with the exploration target set, and generating a corresponding rule result pair according to a matching result;

performing causal relationship exploration on each rule result pair to generate a corresponding rule causal relationship set;

and carrying out artificial intelligence explanation according to the rule causal relationship.

Further, acquiring an observation data set, analyzing the observation data set to generate a target rule, including:

acquiring a feature set of the observation data set;

and analyzing the features in the feature set to generate the target rule.

Further, when the target rule is the group partition rule, analyzing the features in the feature set to generate the target rule, including:

if the characteristic is a sparse characteristic, generating the group division rule according to the group division mode of the characteristic;

and if the features are dense features, performing sparsification on the features, and generating the group division rule according to the group division mode of the features after the sparsification.

Further, when the target rule is the behavior intervention rule, analyzing the features in the feature set to generate the target rule, including:

if the characteristic is a sparse characteristic, generating the behavior intervention rule according to an intervention behavior executed on the characteristic;

and if the features are dense features, performing sparsification on the features, and generating the behavior intervention rule according to the intervention behavior executed on the sparsified features.

Further, when the target rule is the law adaptation rule, analyzing the features in the feature set to generate the target rule, including:

acquiring a law set, and matching the law set with the features;

and generating the law adaptation rule according to the matching result.

Further, acquiring an exploration target set corresponding to the observation data set, including:

expanding the observation data set to obtain an expanded data set;

and determining the exploration target set according to the observation data set and the expansion data set.

Further, expanding the observation data set includes:

the observation data set is expanded using a set-up tool comprising a knowledge graph, an expert knowledge base, and a database.

Further, the rule result pairs include a group division rule result pair, a behavior intervention rule result pair, and a law adaptation rule result pair, and a causal relationship is explored for each rule result pair to generate a corresponding rule causal relationship set, including:

grouping sampling inspection is carried out on the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair;

and carrying out causal relationship analysis according to the grouping sampling inspection result to generate a corresponding rule causal relationship set.

Further, performing a packet sampling check on the group partition rule result pair, including:

grouping the group division rule result pairs, and randomly sampling each group;

counting the random sampling results, and determining target distribution in different groups;

correspondingly, performing causal relationship analysis according to the packet sampling inspection result to generate a corresponding rule causal relationship set, including:

if the target distribution is different distribution, the group division rule result pair has causal effect, and if the target distribution is same distribution, the group division rule result pair does not have causal effect;

and if the group partitioning rule result pair has a causal effect, generating a group partitioning rule causal relationship set according to the causal effect.

Further, generating a group partitioning rule causal relationship set according to the causal effect, comprising:

acquiring at least one influence factor according to the causal effect;

determining the weight corresponding to each influence factor, and generating the group partition rule causal relationship set according to the at least one influence factor and the corresponding weight.

Further, performing a packet sampling check on the behavior intervention rule result pair, including:

dividing the behavior intervention rule result pair into a first behavior intervention rule result pair and a second behavior intervention rule result pair, wherein the first behavior intervention rule result pair is derived from observation data, and the second behavior intervention rule result pair is derived from model data or counterfactual data;

respectively carrying out grouping sampling inspection on the first behavior intervention rule result pair and the second behavior intervention rule result pair to obtain corresponding grouping sampling inspection results;

correspondingly, performing causal relationship analysis according to the packet sampling inspection result to generate a corresponding rule causal relationship set, including:

and if the grouping sampling test result is that the causal effect exists in the behavior intervention rule result pair, generating a behavior intervention rule causal relationship set according to the causal effect.

Further, generating a set of behavioral intervention rule causal relationships according to the causal effect, comprising:

acquiring at least one influence factor according to the causal effect;

determining the weight corresponding to each influence factor, and generating the behavior intervention rule causal relationship set according to the at least one influence factor and the corresponding weight.

Further, when the rule result pair is the law adaptation rule result pair, performing causal relationship exploration on each rule result pair to generate a corresponding rule causal relationship set, including:

dividing the law adaptation rules into a first law adaptation rule and a second law adaptation rule, wherein the first law adaptation rule comprises a grouping relation, and the second law adaptation rule comprises a causal relation;

taking the first law adaptation rule as a grouping condition, and carrying out grouping sampling inspection on the result of the law adaptation rule;

generating a first law adaptation rule causal relationship set according to a grouping sampling inspection result;

extracting the causal relationship contained in the second law adaptation rule and generating a set of causal relationships of the second law adaptation rule;

and determining the union of the first law adaptation rule causal relationship set and the second law adaptation rule causal relationship set as the law adaptation rule causal relationship set.

Further, performing a grouping sampling test on the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair, including:

stacking the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair;

and carrying out grouping sampling inspection on the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair after the stacking processing.

Further, the method further comprises:

and obtaining a custom rule, and generating at least one target rule according to the custom rule.

Further, generating at least one target rule according to the custom rule includes:

analyzing the self-defined rule to obtain at least one analysis rule;

and determining the analysis rule as the corresponding target rule according to the type of the analysis rule.

According to another aspect of the present invention, there is provided an artificial intelligence interpretation apparatus comprising:

and the target rule generating module is used for acquiring the observation data set, analyzing the observation data set and generating a target rule, wherein the target rule comprises a group division rule, a behavior intervention rule and a law adaptation rule.

Optionally, the target rule generating module is further configured to:

acquiring a feature set of an observation data set; and analyzing the features in the feature set to generate a target rule.

Optionally, when the target rule is a group partition rule, the target rule generating module is further configured to:

if the characteristic is a sparse characteristic, generating a group division rule according to a group division mode of the characteristic; and if the features are dense features, performing sparsification on the features, and generating a group division rule according to the group division mode of the features after the sparsification.

Optionally, when the target rule is the behavior intervention rule, the target rule generating module is further configured to:

if the characteristic is a sparse characteristic, generating a behavior intervention rule according to an intervention behavior executed on the characteristic; if the features are dense features, performing sparsification on the features, and generating behavior intervention rules according to intervention behaviors executed on the sparsified features.

Optionally, when the target rule is the law adaptation rule, the target rule generating module is further configured to:

acquiring a law set, and matching the law set with the characteristics; and adapting the rule according to the matching result.

And the rule result pair generation module is used for acquiring the exploration target set corresponding to the observation data set, matching each target rule with the exploration target and generating a corresponding rule result pair according to the matching result.

Optionally, the rule result pair generating module is further configured to:

expanding the observation data set to obtain an expanded data set; and determining an exploration target set according to the observation data set and the expansion data set.

Optionally, the rule result pair generating module is further configured to:

and expanding the observation data set by using a setting tool, wherein the setting tool comprises a knowledge map, an expert knowledge base and a database.

And the rule causal relationship set generation module is used for carrying out causal relationship exploration on each rule result pair to generate a corresponding rule causal relationship set.

Optionally, the rule result pair includes a group division rule result pair, an action intervention rule result pair, and a law adaptation rule result pair, and the rule causal relationship set generating module is further configured to:

grouping sampling inspection is carried out on the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair; and carrying out causal relationship analysis according to the grouping sampling inspection result to generate a corresponding rule causal relationship set.

Optionally, the rule causal relationship set generating module is further configured to:

grouping the result pairs of the group division rules, and randomly sampling each group; and counting the random sampling results, and determining the target distribution in different groups.

Correspondingly, if the target distribution is different distribution, the causal effect exists in the group division rule result pair, and if the target distribution is the same distribution, the causal effect does not exist in the group division rule result pair; and if the group partitioning rule result pair has a causal effect, generating a group partitioning rule causal relationship set according to the causal effect.

Optionally, the rule causal relationship set generating module is further configured to:

acquiring at least one influence factor according to the causal effect; determining the weight corresponding to each influence factor, and generating a group partitioning rule causal relationship set according to at least one influence factor and the corresponding weight.

Optionally, the rule causal relationship set generating module is further configured to:

dividing the behavior intervention rule result pair into a first behavior intervention rule result pair and a second behavior intervention rule result pair, wherein the first behavior intervention rule result pair is derived from observation data, and the second behavior intervention rule result pair is derived from model data or counterfactual data; and respectively carrying out grouping sampling inspection on the first behavior intervention rule result pair and the second behavior intervention rule result pair to obtain corresponding grouping sampling inspection results.

Correspondingly, if the grouped sampling inspection result is that the causal effect exists in the action intervention rule result pair, generating an action intervention rule causal relationship set according to the causal effect.

Optionally, the rule causal relationship set generating module is further configured to:

acquiring at least one influence factor according to the causal effect; determining the weight corresponding to each influence factor, and generating a behavior intervention rule causal relationship set according to at least one influence factor and the corresponding weight.

Optionally, when the rule result pair is a law adaptation rule result pair, the rule causal relationship set generating module is further configured to:

dividing the law adaptation rules into a first law adaptation rule and a second law adaptation rule, wherein the first law adaptation rule comprises a grouping relation, and the second law adaptation rule comprises a causal relation; taking the first law adaptation rule as a grouping condition, and carrying out grouping sampling inspection on the result of the law adaptation rule; generating a first law adaptation rule causal relationship set according to a grouping sampling inspection result; extracting the causal relationship contained in the second law adaptation rule and generating a second law adaptation rule causal relationship set; and determining the union of the first law adaptation rule causal relationship set and the second law adaptation rule causal relationship set as a law adaptation rule causal relationship set.

Optionally, the rule causal relationship set generating module is further configured to:

stacking the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair; and carrying out grouping sampling inspection on the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair after the stacking processing.

And the artificial intelligence interpretation module is used for carrying out artificial intelligence interpretation according to the rule cause and effect relationship.

Optionally, the artificial intelligence interpretation apparatus further includes a second target rule generation module, configured to obtain a user-defined rule, and generate at least one target rule according to the user-defined rule.

Optionally, the second target rule generating module is further configured to:

analyzing the self-defined rule to obtain at least one analysis rule; and determining the analysis rule as a corresponding target rule according to the type of the analysis rule.

According to another aspect of the present invention, there is provided an electronic apparatus including:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein the content of the first and second substances,

the memory stores a computer program executable by the at least one processor, the computer program being executable by the at least one processor to enable the at least one processor to perform the artificial intelligence interpretation method of any of the embodiments of the invention.

According to another aspect of the present invention, there is provided a computer-readable storage medium storing computer instructions for causing a processor to implement the artificial intelligence interpretation method according to any one of the embodiments of the present invention when the computer instructions are executed.

The artificial intelligence interpretation method provided by the embodiment of the invention comprises the steps of firstly obtaining an observation data set, analyzing the observation data set to generate a target rule, wherein the target rule comprises a group division rule, a behavior intervention rule and a law adaptation rule; then acquiring an exploration target set corresponding to the observation data set, matching each target rule with the exploration target set, and generating a corresponding rule result pair according to a matching result; performing causal relationship exploration on each rule result pair to generate a corresponding rule causal relationship set; and finally, carrying out artificial intelligence explanation according to the rule causal relationship. The invention discloses an artificial intelligence explanation method, which specifically divides a rule causal processing method into three types, namely a group division rule, a behavior intervention rule and a law adaptation rule, analyzes the influence of rule change on service performance by taking the rule as a factor, is more suitable for practical application scenes, avoids the complex steps of firstly carrying out factor causal analysis and then analyzing the influence of service process change based on a factor analysis result in the traditional mode, further optimizes the service performance through the rule change in one step, and has more visual and reliable results and higher implementation efficiency; in addition, the invention provides a universal rule cause and effect automatic discovery method, which avoids a large amount of manual analysis work and further improves the engineering application efficiency of XAI machine learning.

It should be understood that the statements in this section do not necessarily identify key or critical features of the embodiments of the present invention, nor do they necessarily limit the scope of the invention. Other features of the present invention will become apparent from the following description.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

FIG. 1 is a flowchart of an artificial intelligence interpretation method according to an embodiment of the present invention;

FIG. 2 is a process diagram of an artificial intelligence interpretation method according to an embodiment of the present invention;

FIG. 3 is a flowchart of an artificial intelligence interpretation method according to a second embodiment of the present invention;

FIG. 4 is a schematic diagram of characteristic items in AI planting according to a second embodiment of the invention;

FIG. 5 is a schematic structural diagram of an artificial intelligence interpretation apparatus according to a third embodiment of the present invention;

fig. 6 is a schematic structural diagram of an electronic device implementing an artificial intelligence interpretation method according to a fourth embodiment of the present invention.

Detailed Description

In order to make the technical solutions of the present invention better understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the invention described herein are capable of operation in sequences other than those illustrated or described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

Example one

Fig. 1 is a flowchart of an artificial intelligence interpretation method according to an embodiment of the present invention, which is applicable to a case where human-readable and understandable interpretation is provided for decisions of an artificial intelligence model, and the method can be performed by an artificial intelligence interpretation apparatus, which can be implemented in hardware and/or software, and the artificial intelligence interpretation apparatus can be configured in an electronic device. As shown in fig. 1, the method includes:

s110, acquiring an observation data set, and analyzing the observation data set to generate a target rule.

The observation data set is data collected by investigation or observation, and is a sample data set used for rule search and generation.

In this embodiment, the target rules generated after the observation data set is used for rule exploration and analysis may be divided into three types, which are a group division rule, a behavior intervention rule and a law adaptation rule, where the group division rule is a rule related to the intrinsic attributes of the group, the behavior intervention rule is a rule related to one or some specific operations, and the law adaptation rule is a rule or axiom known to people. According to the generated target rules, causal effects existing between the rules and the results can be explored, namely rule causal relationships, the rule causal relationships can be further applied to provide attributive explanations for the results, human readable and understandable explanations are provided for model decisions, the results and outputs generated by a machine learning algorithm are assisted and understood by humans, and direct understanding of the working mechanism of the model is established.

Optionally, the observation data set may be obtained by selecting a sample object, investigating or observing the sample object, collecting the required data, and sorting the data into the observation data set. Furthermore, the experimental exploration data result is an important supplement of the observation data set, and the rule exploration can be expanded from an observability method to an experimental discovery method, namely the observation data set is established in an experimental mode. For example, a related experiment can be designed according to a sample object, a perception controller is arranged, experimental environment and result data are collected through the perception controller, the experimental environment and the result data are processed to generate an observation data set, and the perception controller can be connected with a third-party system to control the experiment in the experimental process.

Further, after the observation data set is obtained, the group division rule, the behavior intervention rule and the law adaptation rule can be explored and generated respectively. Optionally, different rules may be explored in sequence according to the observation data set, or may be processed in parallel to improve rule exploration efficiency.

And S120, acquiring an exploration target set corresponding to the observation data set, matching each target rule with the exploration target set, and generating a corresponding rule result pair according to a matching result.

Wherein, the exploration target set is from the needs of real service scene, for example, in the AI planting scene, the sensing controller of the greenhouse can sense the data of indoor and outdoor illumination, temperature, humidity, etc., and can also control the CO2 generator, windowing controller, curtain controller, irrigation controller, etc. in the greenhouse, the characteristics included in the observation data set can be ID (serial number), Time (Time), specifices (variety), Subclass (Subclass), OutLI (outdoor illumination intensity), OutTMP (outdoor temperature), Hum (indoor humidity), TMP (indoor temperature), CO2 (indoor carbon dioxide solubility), LI (indoor illumination intensity), LiCtrl (illumination time control), VentCtrl (ventilation duration control), fertlerctrrl (fertilization irrigation control), and the exploration target set may be result indexes RHeight (fruit length/height), RWidth (fruit stem width), RWeight (fruit weight), and the like.

Optionally, the Group partitioning Rule, the behavior intervention Rule, and the Law adaptation Rule are respectively expressed by Group _ Rule, Do _ Rule, and Law _ Rule, the exploration Target set is expressed by Target, and the Rule result pair is generated according to a { Group _ Rule | Do _ Rule | Law _ Rule } - { Target } set matching mode: group _ Rule-Target pairs (Group partitioning Rule result pairs), Do _ Rule-Target pairs (behavior intervention Rule result pairs), and Law _ Rule-Target pairs (Law adaptation Rule result pairs). The Group _ Rule-Target pairs and the Do _ Rule-Target pairs can be generated by combining in a Cartesian product mode; for the Law _ Rule-Target pair, the value of Target must be the output item of Law _ Rule, and on the basis of the constraint, Cartesian product mode combination is performed. And recording the overall Rule result pair as Rule-Target { Group _ Rule-Target | Do _ Rule-Target | Law _ Rule-Target }.

Preferably, in the generation process of the target rule, the target rule can be generated in a reverse order according to the exploration target set, so that the result is guided, unnecessary calculation is reduced, and the processing efficiency can be further improved.

S130, performing causal relationship exploration on each rule result pair to generate a corresponding rule causal relationship set.

Wherein the set of rule causal relationships is generated according to causal effects between each rule and an object or effect in the set of exploration objects. For different rules, respective explorers may be created, for example, for Group partitioning rules, behavioral intervention rules, and Law adaptation rules, which may be a Group explorer, a Do explorer, and a Law explorer, respectively, explore three different rule effect pairs, respectively, and generate corresponding rule causal relationships.

Preferably, the three rule cause-and-effect relationship sets may be respectively represented as cause _ G, cause _ D and cause _ L, which respectively represent the group partitioning rule cause-and-effect relationship set, the behavior intervention rule cause-and-effect relationship set and the law adaptation rule cause-and-effect relationship set, and the overall rule cause-and-effect relationship set may be represented as: cause (Rule- > Target) { cause _ G, cause _ D, cause _ L }.

Furthermore, based on the automated rule causal relationship exploration framework, the processing links such as rule type support, rule support, result type support, analysis method support and the like can be further subjected to plug-in, the flexibility of the method is further improved through the mode of the plug-in, and the compatible processing capability of the method for unknown conditions is expanded. Preferably, to further improve efficiency, the processing logic can be migrated to a coprocessor (GPU, FPGA, etc.) and a multi-machine cluster for computation without violating the core processing logic of the present invention.

And S140, carrying out artificial intelligence explanation according to the rule causal relationship.

In this embodiment, based on the rule causal relationship discovery, the existing result may be explained in an attributive way, and the specific explanatory steps may be: the result can be attributed to the intrinsic property influence of a specific group according to cause _ G, i.e. the Target is a specific result of a certain class of specific group; the result can be attributed to the influence of a specific factor according to Causal _ D, that is, the Target is the comprehensive influence result of operation of a certain or some specific factors; target results can be attributed to the well-known regularity-affecting results according to cause _ L.

Further, due to the fact that the causal relationship has certainty and stability, causal reasoning can be further conducted on the basis of the causal relationship, and therefore the future influence result generated by the rule can be inferred and intervened.

Further, the method further comprises: and obtaining a custom rule, and generating at least one target rule according to the custom rule.

The self-defined rule is an existing rule, and the rule types, namely the group division rule, the behavior intervention rule and the law adaptation rule, can be obtained by analyzing the self-defined rule.

Optionally, the manner of generating at least one target rule according to the custom rule may be: analyzing the self-defined rule to obtain at least one analysis rule; and determining the analysis rule as a corresponding target rule according to the type of the analysis rule.

Specifically, the customized rule can be analyzed in a syntax analysis mode, the customized rule is decomposed into one or more sub-rules according with the rule types, then the analyzed sub-rules are subjected to causal analysis, and finally the causal analysis result of the customized rule is combined and output in a causal reasoning mode.

Fig. 2 is a schematic processing procedure diagram of an artificial intelligence interpretation method provided in this embodiment, as shown in the figure, for an obtained observation data set, firstly, the observation data set is analyzed to generate a set of target rules, where the target rules include a group division rule, a behavior intervention rule, and a law adaptation rule, and correspondingly, different rule explorers may be used for analysis; then acquiring an exploration target set corresponding to the observation data set, matching each target rule with the exploration target set, and generating a corresponding rule result pair according to a matching result; performing causal relationship exploration on each rule result pair to generate a corresponding rule causal relationship set, wherein the rule causal relationship can be generated by adopting corresponding explorers for different rule result pairs; and finally, carrying out artificial intelligence explanation according to the causal relationship of the rules to obtain the attributive XAI.

The invention discloses an artificial intelligence explanation method, which specifically divides a rule causal processing method into a group division rule, a behavior intervention rule and a law adaptation rule, analyzes the influence of rule change on service performance by taking the rule as a cause, is more suitable for practical application scenes, avoids the complicated steps of firstly performing factor causal analysis and then analyzing the influence of service flow change based on the factor analysis result in the traditional mode, thereby optimizing the service performance by rule change in one step, and having more visual and reliable results and higher efficiency in implementation; in addition, the invention provides a universal rule cause and effect automatic discovery method, which avoids a large amount of manual analysis work and further improves the engineering application efficiency of XAI machine learning.

Example two

Fig. 3 is a flowchart of an artificial intelligence interpretation method according to a second embodiment of the present invention, which is a refinement of the above-mentioned embodiments. As shown in fig. 3, the method includes:

s210, acquiring a feature set of the observation data set.

In this embodiment, the observation data set may be composed of data corresponding to one or more features of the sample object. For example, if the sample object is a certain population, the characteristic fields of the observation dataset, i.e. the statistical data of the sample object for each characteristic, may include gender, age, education level, etc. according to the needs of the research objective.

Optionally, the feature set of the observation data set is set to { F1, F2 … Fn }, where F1, F2 … Fn are feature items, and each feature item may be dense feature (dense feature) or sparse feature (sparse feature).

And S220, analyzing the features in the feature set to generate a target rule.

In the present embodiment, for different types of target rules (group partitioning rule, behavior intervention rule, and law adaptation rule), different analysis methods may be adopted to generate corresponding rules.

Optionally, when the target rule is a group partition rule, the manner of analyzing the features in the feature set to generate the target rule may be: if the characteristic is a sparse characteristic, generating a group division rule according to a group division mode of the characteristic; and if the features are dense features, performing sparsification on the features, and generating a group division rule according to the group division mode of the features after the sparsification.

Specifically, when the target rule is a group partition rule, the feature set { F1, F2 … Fn } of the observation data set may be traversed. If Fi is sparse feature, Fi assignment set may be set as strategies set Fi _ sT of group partition, that is, Fi _ sT ═ Fi _ V1, Fi _ V2 …, where Fi _ V1 and Fi _ V2 … are respective assignments of Fi, and the corresponding group partition rule is denoted as Fi { Vj }. If Fi is dense feature dense, the Fi value may be thinned by a quantile or other thinning method, the thinning method is denoted as sF, the strategies set for group division is Fi _ dT, Fi _ dT ═ Fi _ sF1, Fi _ sF2 …, where Fi _ sF1 and Fi _ sF2 … are the results of the thinning of Fi assignments, and the corresponding group division rule is denoted as Fi { sFk }. Finally, the output overall group partition rule set can be expressed as: group _ Rule ═ { Fi { Vj | sFk } }.

Optionally, when the target rule is a behavior intervention rule, the manner of analyzing the features in the feature set to generate the target rule may be: if the characteristic is a sparse characteristic, generating a behavior intervention rule according to an intervention behavior executed on the characteristic; and if the features are dense features, performing sparsification on the features, and generating behavior intervention rules according to intervention behaviors executed on the sparsified features.

Specifically, when the target rule is a behavior intervention rule, the feature set { F1, F2 … Fn } of the observation data set may be traversed. If Fi is sparse feature, Set intervention action may be performed on the observation data Set (for example, a specific value of Fi may be Set to value b from value a), and the corresponding action intervention rule is Fi _ sD ═ Fi _ sdo (Vj)) }, where Vj belongs to an assigned value Set of Fi, which may be abbreviated as Fi { do (Vj)) } ═ Fi { Vj- > Vj'. If Fi is dense feature dense, Fi may be thinned by a quantile or other thinning methods, and then, based on a result set after the thinning, Tune intervention operation is performed, that is, specific values of Fi are increased or decreased, so as to change a thinning result, where a corresponding behavioral intervention rule is Fi _ tD ═ Fi _ tDo (sFn _ Vk)) }, where sFn _ Vk belongs to a result set of Fi assignment thinning operation sf (Vk), and may be abbreviated as Fi { do (sfk)) } ═ Fi { sFk- > sFk' }. Finally, the overall intervention behavior Rule set of the output is Do _ Rule ═ { Fi { Do (Vj | sFk) }.

Optionally, when the target rule is a law adaptation rule, the manner of analyzing the features in the feature set to generate the target rule may be: acquiring a law set, and matching the law set with the characteristics; and generating a law adaptation rule according to the matching result.

Specifically, when the target rule is a law adaptation rule, the semantic analysis module may traverse the feature set { F1, F2 … Fn } of the observation data set, perform semantic matching with the law set, and output the law adaptation rule corresponding to the law. The semantic analysis and semantic matching may be an existing Natural Language Processing (NLP) technology, and the content of the law set may be a theorem, a law, an experience, and the like, and is stored in a rule form. Further, the law convergence law may correspond to a form of reasoning: { … FC (Input [ ] - > Output [ ]) … }, where FC is a causal function or inference rule, there may be multiple groups of FCs, and each group of FCs may have multiple inputs and outputs. And matching the input field and the output field with the feature item, if the input field and the output field are matched with the feature item, performing inference operation on the feature item, and if the input field and the output field are not matched with the feature item, adding the new output to the new feature item. The final set of Law adaptation rules is Law _ Rule ═ { FC (I- > O) }.

And S230, expanding the observation data set to obtain an expanded data set.

In this embodiment, the observation data set includes one or more feature items, and the observation data set may be expanded in a manner of expanding the feature items. The feature item expansion mode can be that the expansion feature item with similar semantics or reasoning logic with the original feature item is obtained according to the original feature item, and the method for obtaining the expansion feature item can be a semantic similar technology or a reasoning method based on a knowledge graph, and the like. The original feature items and the extended feature items may be denoted as Fi and EXT (Fi), respectively, and the corresponding observed data set and extended data set are { F1, F2 … Fn } and { EXT (F1), EXT (F2) … EXT (Fn) }.

Further, the manner of expanding the observation data set may be: and expanding the observation data set by using a setting tool, wherein the setting tool comprises a knowledge map, an expert knowledge base and a database.

Optionally, for the characteristics of "one effect and multiple causes", "multiple causes and one effect", "causes behind the back" and the like existing in the cause and effect relationship, in order to further improve the accuracy of the rule cause and effect, the feature items may be expanded by means of tools such as a knowledge graph, an expert knowledge base, a relational database and the like, new features associated with the existing feature items are introduced, and features and feature relationships of the data set are expanded.

And S240, determining an exploration target set according to the observation data set and the expansion data set.

The exploration target set is from the needs of a real service scene, and the setting of the exploration target can be directly the feature item or the expansion of the feature item.

Optionally, the exploration Target set may be denoted as Target ═ { Fi | ext (Fi) }, where Fi is an existing feature item in the original observation data set, ext (Fi) is an extended feature item, and there may be multiple feature items in the exploration Target set.

And S250, matching each target rule with the exploration target set, and generating a corresponding rule result pair according to a matching result.

And the rule result pair is a combination of the target rule and the target with matching relation in the exploration target set.

Optionally, Group _ Rule, Do _ Rule, and Law _ Rule respectively represent a Group partitioning Rule, a behavior intervention Rule, and a Law adaptation Rule, and { Target } represents an exploration Target set, then a Rule result pair may be generated according to a { Group _ Rule | Do _ Rule | Law _ Rule } set matching manner, where the Rule result pair includes a Group partitioning Rule result pair, a behavior intervention Rule result pair, and a Law adaptation Rule result pair, and may be represented by Group _ Rule-Target, Do _ Rule-Target, and Law _ Rule-Target, respectively.

Further, for the Group _ Rule-Target pair and the Do _ Rule-Target pair, the generation can be combined in a Cartesian product mode; for the Law _ Rule-Target pair, the value of Target must be the output item of Law _ Rule, and on the basis of the constraint, Cartesian product mode combination is performed. The final Rule result pair may be denoted as Rule-Target, Rule-Target { Group _ Rule-Target | Do _ Rule-Target | Law _ Rule-Target }.

Taking an AI planting scene as an example, a sensing controller of the greenhouse can sense indoor and outdoor illumination, temperature, humidity and other data, and can also control a CO2 generator, a windowing controller, a curtain controller, an irrigation controller and other devices in the greenhouse. Fig. 4 is a schematic diagram of characteristic items in AI planting according to an embodiment of the present invention, as shown in the figure, characteristics included in the observation data set may be ID (serial number), Time (Time), specifices (varieties), subcategories (subclasses), OutLI (outdoor illumination intensity), OutTMP (outdoor temperature), Hum (indoor humidity), TMP (indoor temperature), CO2 (indoor carbon dioxide solubility), LI (indoor illumination intensity), LiCtrl (illumination Time control), VentCtrl (ventilation Time control), fertlerctrrl (fertigation control), and the exploration target set may be result indexes RHeight (fruit length/height), rwidtth (fruit stem width), RWeight (fruit weight), and the like. After receiving the group partitioning rule generating task group result (RWeight) with RWeight as an effect, the rule explorer performs causal feature engineering processing on a feature item set of the observation data set with RWeight as an effect and feature items other than the result feature items rheeight (effect length/height), RWidth (fruit stem width), and RWeight (effect weight) as a causal exploration set, wherein the causal feature engineering processing includes constraint-based causal feature engineering methods GSMB, IAMB, iambpc, MMMB, BAMB, EEMB, KIAMB, and the like, and score-based causal feature engineering methods GSBN, MMHC, and pcbybb.

In the generation of the group division rule result pair, a processing object generated directly by being used as a rule set is selected, and the selected characteristic items are as follows: subclass, LI (indoor light Intensity), Hum (indoor humidity), TMP (indoor temperature), CO2 (indoor carbon dioxide solubility), fert irrctrl (fertigation irrigation control), Causal Intensity (CI) is ordered as: CI (Subclas- > RWEIGHT) > CI (FertIrrctrl- > RWEIGHT) > CI (LI- > RWEIGHT) > CI (TMP- > RWEIGHT) > CI (Hum- > RWEIGHT) > CI (CO2- > RWEIGHT), a causal relationship generation group division rule of TOP2 strength can be obtained according to a twenty-eight principle, wherein Subclas is sparse and FertIrrctrl and LI are dense. The Subclass value set is { sweet peer, color peer, Large free match, cherry match, Fruit cut, and Dense dyes cut }, the corresponding code is Subclass ID ═ 0,1,2,3,4,5,6}, and the group division rule corresponding to Subclass is Subclass {0,1,2,3,4,5,6 }; and converting the dense value into a sparse value by FerrtIrrctrl and LI according to a 3-quantile method, wherein the corresponding group division rule is FerrtIrrctrl { LE, BE and GT }. Finally, the generated group partition rule and the exploration target index RWEight are combined to output a group partition rule causal pair as follows: group _ Rule-Target { Subclas {0,1,2,3,4,5,6} - > RWEIGHT, FertIrrCtrl { LE, BE, GT } - > RWEIGHT }.

In the generation of the action intervention rule result pair, the cause-effect feature engineering results in the above steps can BE directly multiplexed, but all the cause-effect rule generation processing objects are taken, the sparse features are processed according to coding combination, the dense features are processed according to a 3-quantile method uniformly, the processed feature item set is Subclas {0,1,2,3,4,5,6}, FerrtIrCtrCtrl { LE, BE, GT }, LE { BE, GT }, TMP { LE, BE, GT }, Hum { LE, BE, GT }, CO2{ LE, BE, GT }, and the Do rule cause-effect pair of final combined output is: do _ Rule-target [ { Do (Subclas {0,1,2,3,4,5,6}) - > RWEight, Do (FerrtIrCtrl { LE, BE, GT }) - > RWEight, Do (LI { LE, BE, GT }) - > RWEight, Do (TMP { LE, BE, GT }) - > RWEight, Do (Hum { LE, BE, GT }) - > RWEight), Do (CO2{ LE, BE, GT }) - > RWEight }.

In the generation of the law fitting rule result pair, a knowledge base storage structure and information, for example, Large free tomato, are shown in table 1.

TABLE 1

As shown in table 1, the key fields of the law adaptation knowledge base have time, day temperature, night temperature, illumination intensity, soil water content, humidity and soil pH value, and the keywords of the knowledge base are semantically matched with the names of the feature items, and algorithms such as DSSM, MV-DSSM and ARC-I, ARC-I can be used, and the semantic matching result is: Time-Time, Time + TMP-day temperature/night temperature, LI-illumination intensity, Hum-humidity, soil water content in a knowledge base and soil pH value do not find matched characteristic items, law adaptation rules are extracted based on matching results, the reasoning form of the law is centralized as { FC (Time- > TMP), FC (Time- > LI) and FC (Time- > Hum) } and the optimal experience setting of the Large front tomato on temperature, illumination and humidity corresponding to different growth periods (germination period, seedling period, growth period, flowering period and fruit period) can be carried out on the basis, the system can use historical experience accordingly, for example, the Large Time front tomato of the current greenhouse is judged to be in the flowering period according to a field, and the optimal value of the greenhouse humidity is in a range of 60-65. The final combined output law adaptation rule causal pair is: raw _ Rule-Target { FC (Time- > TMP) - > RWeight, FC (Time- > LI) - > RWeight, FC (Time- > Hum) - > RWeight }.

And S260, carrying out grouping sampling inspection on the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair.

The rule result pair comprises a group division rule result pair, a behavior intervention rule result pair and a law adaptation rule result pair.

In this embodiment, the way of performing packet sampling check on the result pair of the group partitioning rule may be: grouping the result pairs of the group division rules, and randomly sampling each group; and counting the random sampling results, and determining the target distribution in different groups.

Optionally, the rule causal relationship may be automatically discovered based on an explorer, where the explorer includes a Group explorer, a Do explorer, and a Law explorer, and explores three different rule causal pairs respectively.

Specifically, for the Group explorer, first, the values of the rules Fi { Vj | sFk } in the Group partitioning Rule Group _ Rule may be respectively sent to the independent sample distributors, and the independent sample distributors perform Group sampling, where the sampling is performed in a Group random sampling manner, and sampling results of the sampling should satisfy causal analysis assumptions. Wherein the causal analysis hypothesis is usually a negligible hypothesis or an individual process stability hypothesis, and the hypothesis can be appropriately relaxed according to the business requirements in actual operation. The packet sampling inspection method may be: the P-value (confidence) of each feature variable between different groupings is greater than a threshold (typically 0.05). Then, the Target distribution in different groups can be counted, and if the Target distribution among the groups is different, the causal effect is shown; if the Target distribution among the groups is the same, no causal effect exists. The mode of testing the distribution among the components can be T test, KS test, KL divergence test and the like.

In this embodiment, the way of performing the packet sampling check on the behavior intervention rule result pair may be: dividing the behavior intervention rule result pair into a first behavior intervention rule result pair and a second behavior intervention rule result pair, wherein the first behavior intervention rule result pair is derived from observation data, and the second behavior intervention rule result pair is derived from model data or counterfactual data; and respectively carrying out grouping sampling inspection on the first behavior intervention rule result pair and the second behavior intervention rule result pair to obtain corresponding grouping sampling inspection results.

Specifically, for the Do explorer, it may be assumed that the data has 2 types of sources, a first type derived from real-world observation data and a second type derived from model space data or counterfactual experimental data. The causal relationship discovery method based on the first type of source data comprises the following steps: firstly, each Rule { Vj | sFk- > Vj ' | sFk ' } in the behavior intervention Rule Do _ Rule is sent to an independent sample distributor, the independent sample distributor carries out grouping sampling according to Vj | sFk and Vj ' | sFk ', sampling results are respectively marked as G and G ', and the sampling results should meet the assumption of causal analysis. The G and G' packet sampling inspection method may be: the P-value (confidence) of each feature variable between different groupings is greater than a threshold (typically 0.05). The causal relationship discovery method based on the second type of source data comprises the following steps: firstly, model Learning fitting { Vj | sFk } and { Targeti } is performed based on an observation data set, and the used model can be a Machine Learning (ML) model such as Logistic Regression (LR), XGboost and the like, and can also be a Deep Learning (DL) model such as a neural network model and the like. The learned fitting model is denoted as M, and the output value of the fitting model M can be used as the value of the Do counterfactual hypothesis space, i.e., Target 'is M (Vj | sFk), or the result of the counterfactual hypothesis is obtained experimentally, and is also denoted as Target'. Further, aiming at the counter-fact result, the result of the pre-training model can be further introduced as the counter-fact result, and the existing knowledge and experience are introduced into the processing process of the rule causal discovery in the mode of the pre-training model.

Further, the grouping and sampling inspection for the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair may be performed in the following manner: stacking the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair; and carrying out grouping sampling inspection on the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair after the stacking processing.

In this embodiment, to further improve efficiency, multiple sets of rule cause and effect discovery processes may be stacked. The stacking method can be divided into physical stacking and logic stacking, and can be implemented simultaneously. Physical stacking is based on different sampling groups, and different rule causal discovery tasks are executed in parallel; the logic stacking is to combine the rule cause and effect discovery tasks which do not influence each other, and execute the tasks in one sampling group. In short, the parallelism of the rule cause and effect discovery can be improved through the physical stacking, and the workload of the processing task is reduced through the logic stacking, so that the efficiency of the rule cause and effect discovery is further improved.

And S270, performing causal relationship analysis according to the grouping sampling inspection result to generate a corresponding rule causal relationship set.

In this embodiment, the packet sampling tests corresponding to the different rule result pairs in the previous step may analyze the results of the packet tests, respectively, and generate corresponding rule causal relationships according to the causal effects therein.

Optionally, for the Group explorer, performing causal relationship analysis according to the grouped sampling inspection result, and generating the corresponding rule causal relationship set may be: if the target distribution is different distribution, the causal effect exists in the group division rule result pair, and if the target distribution is the same distribution, the causal effect does not exist in the group division rule result pair; and if the group partitioning rule result pair has a causal effect, generating a group partitioning rule causal relationship set according to the causal effect.

Specifically, Target distributions in different groups can be counted, and if the Target distributions among the groups are different, a causal effect is shown; if the Target distribution among the groups is the same, no causal effect exists. The mode of testing the distribution among the components can be T test, KS test, KL divergence test and the like. Finally, extracting a Causal relationship set according to the Causal effect, and representing the Causal relationship set of the Group partitioning rule by Group _ cause 1, wherein the calculation mode is

Further, the manner of generating the group partitioning rule causal relationship set according to the causal effect may be: acquiring at least one influence factor according to the causal effect; determining the weight corresponding to each influence factor, and generating a group division rule causal relationship set according to at least one influence factor and the corresponding weight.

Specifically, for the case of one or multiple causes, the Causal relationship set may be obtained by model fitting (such as linear fitting), where the weight coefficient is the influence weight of different factors, and the Causal relationship set of the Group partitioning rule in this case is represented by Group _ cause 2, and is calculated by

Where δ i is a causal weight.

Finally, the overall output of the Group explorer is the Group partitioning rule Causal relationship set, i.e., cause _ G ═ Group _ cause 1, Group _ cause 2.

Optionally, for the Do explorer, performing causal relationship analysis according to the packet sampling inspection result, and generating the corresponding rule causal relationship set may be: and if the grouped sampling inspection result is that the causal effect exists in the behavior intervention rule result pair, generating a behavior intervention rule causal relationship set according to the causal effect.

Specifically, Target 'in the G' packet may be taken as the Target value of the Do (Vj | sFk) operation. Extracting a causal relationship set according to the causal effect of the real space and the assumed space, wherein the calculation mode is

Extracting a causal relationship set according to the causal effect of an observation space and an anti-fact space in the following calculation mode

Further, the manner of generating the behavior intervention rule causal relationship set according to the causal effect may be: acquiring at least one influence factor according to the causal effect; determining the weight corresponding to each influence factor, and generating a behavior intervention rule causal relationship set according to at least one influence factor and the corresponding weight.

Specifically, for the case of one or more factors, the causal influence relationship set may be obtained by model fitting (such as linear fitting, tendency score regression, etc.), the weight coefficients are influence weights of different factors, and the calculation method is

Where δ i is a causal weight.

Finally, the overall output of the Do explorer is the behavioral intervention rule Causal relationship set, namely cause _ D ═ { Do _ cause 1| Do _ cause 2| Do _ cause 3 }.

Further, when the rule result pair is a law adaptation rule result pair, a causal relationship is explored for each rule result pair, and a manner of generating a corresponding rule causal relationship set may be: dividing the law adaptation rules into a first law adaptation rule and a second law adaptation rule, wherein the first law adaptation rule comprises a grouping relation, and the second law adaptation rule comprises a causal relation; taking the first law adaptation rule as a grouping condition, and carrying out grouping sampling inspection on the result of the law adaptation rule; generating a first law adaptation rule causal relationship set according to a grouping sampling inspection result; extracting the causal relationship contained in the second law adaptation rule and generating a second law adaptation rule causal relationship set; and determining the union of the first law adaptation rule causal relationship set and the second law adaptation rule causal relationship set as a law adaptation rule causal relationship set.

Specifically, for the Law explorer, the Law adaptation Rule default includes a grouping relationship or a causal relationship, so Law _ Rule can be used as a grouping condition to perform comparison among different groups, thereby obtaining the causal relationship, outputting the causal effect, and recording the causal effect as the causal relationship

For the Law adaptation Rule containing the causal relationship, the result factor can also be extracted from Law _ Rule containing the causal relationship, and is used as the causal effect output and is marked as I->And O. So that the output of the Law explorer is

The causal effect based on the Law explorer output can be used directly as an outcome or as an input for new causal inferences.

Furthermore, aiming at the law adaptation rule processing, the method can be expanded based on form and semantics, and can further establish a mapping relation between the feature items and nodes in the existing knowledge base and knowledge graph based on semantic analysis and semantic matching technologies, so that more historical knowledge and experience can be utilized and expanded, the reasoning capability of a third-party knowledge system can be utilized, and the expansibility can be further improved.

Through the above discovery of automated rule causality, all rule causality in the observation dataset can be expressed as: cause (Rule- > Target) { cause _ G, cause _ D, cause _ L }.

For example, in an AI planting scenario, for a Group partition Rule causal relationship, the Group explorer receives a Group partition Rule result pair Group _ Rule-target { Subclass {0,1,2,3,4,5,6} - > RWeight, ferrtctrl { LE, BE, GT } - > RWeight }, parses the Rule causal pair and decomposes the Rule causal pair into 2 random sample set generation tasks Dataset _ Subclass {0,1,2,3,4,5,6} and Dataset _ ferrtctrl { LE, BE, GT }, and sends the sample set generation task to the sample independent distributor. And after receiving the random sample set generation task, the sample independent distributor carries out random sampling and detection, and then sends the random sample set to the Group explorer. The sample set generation process takes Dataset _ Subclass {0,1,2,3,4,5,6} as an example, the sample independent distributor performs grouped random sampling in the measurement data set by Subclass, in this embodiment, the number of samples in each group is 1000, and after the sampling is finished, 7 groups of random sample sets are output respectively and are respectively marked as Dataset _ Subclass _0 and Dataset _ Subclass _1 … … Dataset _ Subclass _ 6. And then, performing P value test on the characteristic items of the random sample set except for the subclass and the result item, wherein if all P values are greater than 0.05, the sample set passes the test, and otherwise, performing sampling again. After receiving the random sample set, the Group explorer judges the causal relationship of the subclass and FertIrrctrl characteristic items on the yield RWEIGht through the average causal effect. Also describing this process by way of example of Subclas, RWEight mean data for 7 sets of random samples are shown in Table 2,

TABLE 2

The group with the minimum average yield is taken as a control group, namely the group with Subclas ID ═ 2 is taken as a control group, so that

Is represented by subbclass _ i instead of subclass _2 was the subject,

the average causal effect ATE (subclass i) replacing subclass _2 with subclass i is RWeight _ i-RWeight _ 2. Therefore, aiming at the Subclases feature item grouping, the final output causal relationship set is as

Group _ cause _ FertIrrCtrl can be similarly derived. The final output result of the Group explorer is cause _ G ═ { Group _ cause _ subclass, Group _ cause _ fertlrctrl }.

For the behavior intervention Rule causal relationship, the Do explorer receives Do _ Rule-target ═ Do (Subclas {0,1,2,3,4,5,6}) ->RWeight,Do(FertIrrCtrl{LE,BE,GT})->RWeight,Do(LI{LE,BE,GT})->RWeight,Do(TMP{LE,BE,GT})->RWeig ht,Do(Hum{LE,BE,GT})->RWeight,Do(CO2{LE,BE,GT})->RWEight }, the Do explorer analyzes the cause and effect of the rule, then combines 6 Do exploration subtasks into 1 Do exploration task, and analyzes the influence relationship of a plurality of factors on a result at one time. The merged Do exploration task is { RWEight | Do (Subclss, FertIrrCtrl, LI, TMP, Hum, CO2) }. The Do exploration is counterfactual exploration, and counterfactual results can be derived from observation data or experimental data. For case 1 "counterfactual results are derived from observed data": the Do explorer generates a random sample set generation task Dataseset _ { Subclas, FertIrrCtrl, LI, TMP, Hum, CO2}, and sends it to the sample independent distributor. And after receiving the random sample set generation task, the sample independent distributor carries out random sampling and detection, and then sends the random sample set to the Do explorer. The counter fact model M adopts a linear model, a Do explorer performs linear regression learning based on random sample set data to obtain the counter fact model M, the counter fact model M is marked as M ═ alpha 1Subclas + alpha 2FertIrrCtrl + alpha 3LI + alpha 4TMP + alpha 5Hum + alpha 6CO2, and the counter fact result is the reasoning result of the counter fact model. RWEi of random sample setThe mean value of light is 15kg/m2, for example to calculate the causal effect of LI,

that is, if the settings of Subclas, FertIrrCtrl, etc. are not changed, only changing the illumination intensity to LE degree will result in an average decrease of 3kg/m2 in yield. Similarly, all other Do exploration results can be obtained, and the final results are summarized as

For the case 2 "counterfactual results are derived from experimental data", the Do explorer generates a random sample set G generation task Dataset _ { Subclass, fertirctrl, LI, TMP, Hum, CO2} and a test sample set G' collection task, and sends to the sample independent assigner. After receiving the random sample set generation task, the sample independent distributor carries out random sampling and detection, and then sends the random sample set G to the Do explorer, wherein the RWEight mean value of the random sample set is 15kg/m 2. The experimental data generation follows the basic requirements of a control experiment, a single parameter adjustment experiment mode can be adopted, the FerrtIrrctrl level is set as GT for example, a Do explorer sends a control instruction to a greenhouse sensing controller to change the FerrtIrrctrl level of the greenhouse, after the FerrtIrrctrl of the greenhouse is changed, the sensing controller collects various data regularly and sends the data to a regular causal attributive XAI system to form an experimental sample data set G ', the RWEIGHT mean value of the G ' data set is calculated to be 17kg/m2, and then the RWEIGHT mean value of the G ' data set is calculated to be 17kg/m2

I.e. the level of fertigation is set to GT, which results in a 2kg/m2 product lift. Similarly, all other Do exploration results can be obtained, and the final results are summarized as

For the Law adaptation Rule causal relationship, the Law explorer receives Raw _ Rule-Target ═ FC (Time->TMP)->RWeight,FC(Time->LI)->RWeight,FC(Time->Hum)->RWEight }, the causal pair of the parsing rules is decomposed into 3 random sample sets to generate a task Dataset _ FC{Time->TMP}、Dataset_FC{Time->LI and Dataset _ FC { Time->Hum, and sends the sample set generation task to the sample independent distributor. The mode of direct generation from the observation set can be adopted, the sample independent distributor receives the random sample set generation task, and performs grouping sampling according to different LAW rules to form a plurality of grouping sampling sample sets. Using Dataset _ FC (Time->Hum), for example, the specimen is Large fresh tomato, the Time can be divided into germination stage, seedling stage, growth stage, flowering stage and fruit stage, the independent sample distributor takes 2 groups of sample sets according to different periods, the sample sets are respectively sample set G meeting the law and sample set G' not meeting the law, for example, the Large fresh tomato subcategory is generated, firstly, the samples are screened under the condition that the Time is 3-5, and then the Hum is 80-90 and the Hum is | in the screening result! (80-90) are divided into 2 groups, random sampling is carried out in the 2 groups, the sampling result needs to meet the sufficient random detection condition, the samples conforming to the germination period humidity law are grouped into G, the non-conforming groups are G ', the average values of the groups RWEight of G and G' are respectively calculated to be 8 and 6, and the causal effect of the germination period humidity law is that for the Large fresh tomato products

That is, meeting the germination law would result in an increase in yield of 2kg/m 2. Similarly, all other LAW exploration results can be obtained, and the final results are summarized as

Through the automatic discovery of the three types of Rule Causal relationships, all Rule Causal relationship sets are summarized as cause (Rule- > Target) { cause _ G, cause _ D, and cause _ L }, and then the Rule Causal relationship automatic discovery module sends the cause (Rule- > Target) to the XAI interpretation module for artificial intelligence interpretation.

And S280, carrying out artificial intelligence explanation according to the rule causal relationship.

In this embodiment, based on the rule causal relationship discovery, the existing result may be explained in an attributive way, and the specific explanatory steps may be: the result can be attributed to the intrinsic property influence of a specific group according to cause _ G, i.e. the Target is a specific result of a certain class of specific group; the result can be attributed to the influence of a specific factor according to Causal _ D, that is, the Target is the comprehensive influence result of operation of a certain or some specific factors; target results can be attributed to the well-known regularity-affecting results according to cause _ L.

Taking AI planting as an example, the XAI interpretation module receives the Rule Causal relationship set cause (Rule- > Target), searches for relevant Rule Causal relationships in the cause (Rule- > Target) according to the setting requirements of the interpretability report and the feature item segment, and gives an interpretation for an exploration Target according to the Rule Causal relationships. Wherein, the interpretable report display content is defined as the following 3 parts: the method comprises the steps of selecting quality, greenhouse control and normative regularity, wherein the quality of a selected product is matched with characteristic item fields of specials (variety) and Subclas (Subclass), the greenhouse control is matched with characteristic item fields of Time (Time), Hum (indoor humidity), TMP (indoor temperature), CO2 (indoor carbon dioxide solubility), LI (indoor illumination intensity), LiCrlHours (illumination Time control), VentCtrl (ventilation duration control) and FertIrrCtrL (fertilization and irrigation control), the normative regularity is matched with the content of a knowledge base of a system, the setting of the matching range of the part mainly comes from the scene of upper-layer service application, and if the matching range is not set, all characteristic items or the content of the knowledge base can be defaulted.

The method comprises the steps of firstly obtaining a feature set of an observation data set, then analyzing features in the feature set to generate target rules, then expanding the observation data set to obtain an expanded data set, then determining a search target set according to the observation data set and the expanded data set, then matching each target rule with the search target set, generating corresponding rule result pairs according to matching results, then performing grouping sampling inspection on the group division rule result pairs, the behavior intervention rule result pairs and the law adaptation rule result pairs, then performing causal relationship analysis according to grouping sampling inspection results to generate corresponding rule causal relationship sets, and finally performing artificial intelligence explanation according to rule causal relationships. The invention discloses an artificial intelligence explanation method, which specifically divides a rule causal processing method into a group division rule, a behavior intervention rule and a law adaptation rule, analyzes the influence of rule change on service performance by taking the rule as a cause, is more suitable for practical application scenes, avoids the complicated steps of firstly performing factor causal analysis and then analyzing the influence of service flow change based on the factor analysis result in the traditional mode, thereby optimizing the service performance by rule change in one step, and having more visual and reliable results and higher efficiency in implementation; in addition, the invention provides a universal rule cause and effect automatic discovery method, which avoids a large amount of manual analysis work and further improves the engineering application efficiency of XAI machine learning.

EXAMPLE III

Fig. 5 is a schematic structural diagram of an artificial intelligence interpretation apparatus according to a third embodiment of the present invention. As shown in fig. 5, the apparatus includes: a target rule generation module 310, a rule effect pair generation module 320, a rule causal set generation module 330, and an artificial intelligence interpretation module 340.

And the target rule generating module 310 is configured to obtain the observation data set, analyze the observation data set, and generate a target rule, where the target rule includes a group division rule, a behavior intervention rule, and a law adaptation rule.

Optionally, the target rule generating module 310 is further configured to:

acquiring a feature set of an observation data set; and analyzing the features in the feature set to generate a target rule.

Optionally, when the target rule is a group partition rule, the target rule generating module 310 is further configured to:

if the characteristic is a sparse characteristic, generating a group division rule according to a group division mode of the characteristic; and if the features are dense features, performing sparsification on the features, and generating a group division rule according to the group division mode of the features after the sparsification.

Optionally, when the target rule is the behavior intervention rule, the target rule generating module 310 is further configured to:

if the characteristic is a sparse characteristic, generating a behavior intervention rule according to an intervention behavior executed on the characteristic; and if the features are dense features, performing sparsification on the features, and generating behavior intervention rules according to intervention behaviors executed on the sparsified features.

Optionally, when the target rule is the law adaptation rule, the target rule generating module 310 is further configured to:

acquiring a law set, and matching the law set with the characteristics; and adapting the rule according to the matching result.

And the rule result pair generating module 320 is configured to obtain an exploration target set corresponding to the observation data set, match each target rule with an exploration target, and generate a corresponding rule result pair according to a matching result.

Optionally, the rule result pair generating module 320 is further configured to:

expanding the observation data set to obtain an expanded data set; and determining an exploration target set according to the observation data set and the expansion data set.

Optionally, the rule result pair generating module 320 is further configured to:

and expanding the observation data set by using a setting tool, wherein the setting tool comprises a knowledge map, an expert knowledge base and a database.

And a rule causal relationship set generating module 330, configured to perform causal relationship exploration on each rule result pair, and generate a corresponding rule causal relationship set.

Optionally, the rule result pair includes a group division rule result pair, an action intervention rule result pair, and a law adaptation rule result pair, and the rule causal relationship set generating module 330 is further configured to:

grouping sampling inspection is carried out on the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair; and carrying out causal relationship analysis according to the grouping sampling inspection result to generate a corresponding rule causal relationship set.

Optionally, the rule causal relationship set generating module 330 is further configured to:

grouping the result pairs of the group division rules, and randomly sampling each group; and counting the random sampling results, and determining the target distribution in different groups.

Correspondingly, if the target distribution is different distribution, the causal effect exists in the group division rule result pair, and if the target distribution is the same distribution, the causal effect does not exist in the group division rule result pair; and if the group partitioning rule result pair has a causal effect, generating a group partitioning rule causal relationship set according to the causal effect.

Optionally, the rule causal relationship set generating module 330 is further configured to:

acquiring at least one influence factor according to the causal effect; determining the weight corresponding to each influence factor, and generating a group partitioning rule causal relationship set according to at least one influence factor and the corresponding weight.

Optionally, the rule causal relationship set generating module 330 is further configured to:

dividing the behavior intervention rule result pair into a first behavior intervention rule result pair and a second behavior intervention rule result pair, wherein the first behavior intervention rule result pair is derived from observation data, and the second behavior intervention rule result pair is derived from model data or counterfactual data; and respectively carrying out grouping sampling inspection on the first behavior intervention rule result pair and the second behavior intervention rule result pair to obtain corresponding grouping sampling inspection results.

Correspondingly, if the grouped sampling inspection result is that the causal effect exists in the action intervention rule result pair, generating an action intervention rule causal relationship set according to the causal effect.

Optionally, the rule causal relationship set generating module 330 is further configured to:

acquiring at least one influence factor according to the causal effect; determining the weight corresponding to each influence factor, and generating a behavior intervention rule causal relationship set according to at least one influence factor and the corresponding weight.

Optionally, when the rule result pair is a law adaptation rule result pair, the rule causal relationship set generating module 330 is further configured to:

dividing the law adaptation rules into a first law adaptation rule and a second law adaptation rule, wherein the first law adaptation rule comprises a grouping relationship, and the second law adaptation rule comprises a causal relationship; taking the first law adaptation rule as a grouping condition, and carrying out grouping sampling inspection on the result of the law adaptation rule; generating a first law adaptation rule causal relationship set according to a grouping sampling inspection result; extracting the causal relationship contained in the second law adaptation rule and generating a second law adaptation rule causal relationship set; and determining the union of the first law adaptation rule causal relationship set and the second law adaptation rule causal relationship set as a law adaptation rule causal relationship set.

Optionally, the rule causal relationship set generating module 330 is further configured to:

stacking the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair; and carrying out grouping sampling inspection on the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair after the stacking processing.

And the artificial intelligence interpretation module 340 is used for carrying out artificial intelligence interpretation according to the rule cause-effect relationship.

Optionally, the artificial intelligence interpreting apparatus further includes a second target rule generating module 350, configured to obtain a user-defined rule, and generate at least one target rule according to the user-defined rule.

Optionally, the second target rule generating module 350 is further configured to:

analyzing the self-defined rule to obtain at least one analysis rule; and determining the analysis rule as a corresponding target rule according to the type of the analysis rule.

The artificial intelligence interpretation device provided by the embodiment of the invention can execute the artificial intelligence interpretation method provided by any embodiment of the invention, and has corresponding functional modules and beneficial effects of the execution method.

Example four

FIG. 6 illustrates a schematic structural diagram of an electronic device 10 that may be used to implement an embodiment of the present invention. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be exemplary only, and are not meant to limit implementations of the inventions described and/or claimed herein.

As shown in fig. 6, the electronic device 10 includes at least one processor 11, and a memory communicatively connected to the at least one processor 11, such as a Read Only Memory (ROM)12, a Random Access Memory (RAM)13, and the like, wherein the memory stores a computer program executable by the at least one processor, and the processor 11 may perform various suitable actions and processes according to the computer program stored in the Read Only Memory (ROM)12 or the computer program loaded from the storage unit 18 into the Random Access Memory (RAM) 13. In the RAM 13, various programs and data necessary for the operation of the electronic apparatus 10 may also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other via a bus 14. An input/output (I/O) interface 15 is also connected to the bus 14.

A number of components in the electronic device 10 are connected to the I/O interface 15, including: an input unit 16 such as a keyboard, a mouse, or the like; an output unit 17 such as various types of displays, speakers, and the like; a storage unit 18 such as a magnetic disk, an optical disk, or the like; and a communication unit 19 such as a network card, modem, wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information/data with other devices via a computer network such as the internet and/or various telecommunication networks.

The processor 11 may be a variety of general and/or special purpose processing components having processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various specialized Artificial Intelligence (AI) computing chips, various processors running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, or the like. The processor 11 performs the various methods and processes described above, such as the interpretation of artificial intelligence.

In some embodiments, the interpretation method of artificial intelligence may be implemented as a computer program tangibly embodied on a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and/or installed onto the electronic device 10 via the ROM 12 and/or the communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the interpretation of artificial intelligence described above may be performed. Alternatively, in other embodiments, the processor 11 may be configured by any other suitable means (e.g., by means of firmware) to perform artificial intelligence interpretation methods.

Various implementations of the systems and techniques described here above may be implemented in digital electronic circuitry, integrated circuitry, Field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (ASSPs), system on a chip (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.

A computer program for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the computer programs, when executed by the processor, cause the functions/acts specified in the flowchart and/or block diagram block or blocks to be performed. A computer program can execute entirely on a machine, partly on a machine, as a stand-alone software package partly on a machine and partly on a remote machine or entirely on a remote machine or server.

In the context of the present invention, a computer-readable storage medium may be a tangible medium that can contain, or store a computer program for use by or in connection with an instruction execution system, apparatus, or device. A computer readable storage medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. Alternatively, the computer readable storage medium may be a machine readable signal medium. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

To provide for interaction with a user, the systems and techniques described here can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the electronic device. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, speech, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), Wide Area Networks (WANs), blockchain networks, and the internet.

The computing system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server can be a cloud server, also called a cloud computing server or a cloud host, and is a host product in a cloud computing service system, so that the defects of high management difficulty and weak service expansibility in the traditional physical host and VPS service are overcome.

It should be understood that various forms of the flows shown above, reordering, adding or deleting steps, may be used. For example, the steps described in the present invention may be executed in parallel, sequentially, or in different orders, and are not limited herein as long as the desired results of the technical solution of the present invention can be achieved.

The above-described embodiments should not be construed as limiting the scope of the invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made in accordance with design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (19)

1. An artificial intelligence interpretation method, comprising:

acquiring an observation data set, and analyzing the observation data set to generate a target rule, wherein the target rule comprises a group division rule, a behavior intervention rule and a law adaptation rule;

acquiring an exploration target set corresponding to the observation data set, matching each target rule with the exploration target set, and generating a corresponding rule result pair according to a matching result;

performing causal relationship exploration on each rule result pair to generate a corresponding rule causal relationship set;

and carrying out artificial intelligence explanation according to the rule causal relationship.

2. The method of claim 1, wherein obtaining an observation data set, analyzing the observation data set to generate a target rule comprises:

acquiring a feature set of the observation data set;

and analyzing the features in the feature set to generate the target rule.

3. The method of claim 2, wherein when the target rule is the group partition rule, analyzing the features in the feature set to generate the target rule comprises:

if the characteristic is a sparse characteristic, generating the group division rule according to the group division mode of the characteristic;

and if the features are dense features, performing sparsification on the features, and generating the group division rule according to the group division mode of the features after the sparsification.

4. The method of claim 2, wherein when the target rule is the behavioral intervention rule, analyzing the features in the feature set to generate the target rule comprises:

if the characteristic is a sparse characteristic, generating the behavior intervention rule according to an intervention behavior executed on the characteristic;

and if the features are dense features, performing sparsification on the features, and generating the behavior intervention rule according to the intervention behavior executed on the sparsified features.

5. The method of claim 2, wherein when the target rule is the law adaptation rule, analyzing the features in the feature set to generate the target rule comprises:

acquiring a law set, and matching the law set with the features;

and generating the law adaptation rule according to the matching result.

6. The method of claim 1, wherein obtaining an exploration target set corresponding to the observation data set comprises:

expanding the observation data set to obtain an expanded data set;

and determining the exploration target set according to the observation data set and the expansion data set.

7. The method of claim 6, wherein expanding the observation data set comprises:

the observation data set is expanded using a set-up tool comprising a knowledge graph, an expert knowledge base, and a database.

8. The method of claim 1, wherein the rule effect pairs comprise a group partition rule effect pair, a behavioral intervention rule effect pair, and a law adaptation rule effect pair, and wherein performing a causal relationship exploration on each of the rule effect pairs to generate a corresponding set of rule causal relationships comprises:

grouping sampling inspection is carried out on the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair;

and carrying out causal relationship analysis according to the grouping sampling inspection result to generate a corresponding rule causal relationship set.

9. The method of claim 8, wherein performing a packet sample check on the pair of group partition rule results comprises:

grouping the group division rule result pairs, and randomly sampling each group;

counting the random sampling results, and determining target distribution in different groups;

correspondingly, performing causal relationship analysis according to the packet sampling inspection result to generate a corresponding rule causal relationship set, including:

if the target distribution is different distribution, the group division rule result pair has causal effect, and if the target distribution is same distribution, the group division rule result pair does not have causal effect;

and if the group partitioning rule result pair has a causal effect, generating a group partitioning rule causal relationship set according to the causal effect.

10. The method of claim 9, wherein generating a set of group partitioning rule causal relationships from the causal effect comprises:

acquiring at least one influence factor according to the causal effect;

determining the weight corresponding to each influence factor, and generating the group partition rule causal relationship set according to the at least one influence factor and the corresponding weight.

11. The method of claim 8, wherein performing a packet sample check on the behavioral intervention rule result pair comprises:

dividing the behavior intervention rule result pair into a first behavior intervention rule result pair and a second behavior intervention rule result pair, wherein the first behavior intervention rule result pair is derived from observation data, and the second behavior intervention rule result pair is derived from model data or counterfactual data;

respectively carrying out grouping sampling inspection on the first behavior intervention rule result pair and the second behavior intervention rule result pair to obtain corresponding grouping sampling inspection results;

correspondingly, performing causal relationship analysis according to the packet sampling inspection result to generate a corresponding rule causal relationship set, including:

and if the grouping sampling test result is that the causal effect exists in the behavior intervention rule result pair, generating a behavior intervention rule causal relationship set according to the causal effect.

12. The method of claim 11, wherein generating a set of behavioral intervention rule causal relationships from the causal effects comprises:

acquiring at least one influence factor according to the causal effect;

determining the weight corresponding to each influence factor, and generating the behavior intervention rule causal relationship set according to the at least one influence factor and the corresponding weight.

13. The method of claim 8, wherein, when the rule effect pairs are the law adaptation rule effect pairs, performing a causal relationship exploration on each of the rule effect pairs to generate a corresponding rule causal relationship set, comprises:

dividing the law adaptation rules into a first law adaptation rule and a second law adaptation rule, wherein the first law adaptation rule comprises a grouping relation, and the second law adaptation rule comprises a causal relation;

taking the first law adaptation rule as a grouping condition, and carrying out grouping sampling inspection on the result of the law adaptation rule;

generating a first law adaptation rule causal relationship set according to a grouping sampling inspection result;

extracting the causal relationship contained in the second law adaptation rule and generating a set of causal relationships of the second law adaptation rule;

and determining the union of the first law adaptation rule causal relationship set and the second law adaptation rule causal relationship set as the law adaptation rule causal relationship set.

14. The method of claim 8, wherein performing a packet sample test on the group partition rule result pair, the behavior intervention rule result pair, and the law adaptation rule result pair comprises:

stacking the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair;

and carrying out grouping sampling inspection on the group division rule result pair, the behavior intervention rule result pair and the law adaptation rule result pair after the stacking processing.

15. The method of claim 1, further comprising:

and obtaining a custom rule, and generating at least one target rule according to the custom rule.

16. The method of claim 15, wherein generating at least one of the target rules according to the custom rule comprises:

analyzing the self-defined rule to obtain at least one analysis rule;

and determining the analysis rule as the corresponding target rule according to the type of the analysis rule.

17. An artificial intelligence interpretation apparatus, comprising:

the target rule generating module is used for acquiring an observation data set, analyzing the observation data set and generating a target rule, wherein the target rule comprises a group division rule, a behavior intervention rule and a law adaptation rule;

the rule result pair generation module is used for acquiring an exploration target set corresponding to the observation data set, matching each target rule with the exploration target and generating a corresponding rule result pair according to a matching result;

the rule causal relationship set generation module is used for carrying out causal relationship exploration on each rule result pair to generate a corresponding rule causal relationship set;

and the artificial intelligence interpretation module is used for carrying out artificial intelligence interpretation according to the rule causal relationship.

18. An electronic device, characterized in that the electronic device comprises:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein the content of the first and second substances,

the memory stores a computer program executable by the at least one processor to enable the at least one processor to perform the artificial intelligence interpretation method of any one of claims 1-16.

19. A computer-readable storage medium storing computer instructions for causing a processor to perform the artificial intelligence interpretation method of any one of claims 1-16 when executed.

Images