Program, determination device, and method having explainable AI for determining the relationship between multiple events

Abstract

To provide a program, a determination program, and a method for determining the relationship among multiple events.SOLUTION: A determination device includes an explainable AI engine for calculating significance data Z including a significance (z) for each event (x) of input data X, a determination element determination section for determining a determination element ydec that becomes the highest score pydec for output data Y and an important event selection section for selecting a first event xi and a second event xj in which significance (z) becomes a predetermined selection condition or higher, and a pseudo data generation section for generating third pseudo input data Xrep i, j in which both of the first event xi and the second event xj are substituted in a pseudo manner for time-series data X, and determines that the events xi and xj of the input data X are in OR relationship when the score of third outputted pseudo output data Yrep i, j decreases much more than the score of the output data Y by a first predetermined threshold or larger by inputting third pseudo input data Xrep i, j to the explainable AI engine.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to the technology of explainable AI (Explainable Artificial Intelligence). [Background technology]

[0002] Advances in machine learning technologies such as deep learning have dramatically improved the performance of AI in classification, regression, and decision-making tasks. Classification tasks include behavior recognition, which collects sensor data from IoT devices installed in the home, such as motion sensors, door open / close sensors, temperature sensors, and humidity sensors, and estimates the resident's behavior (going out, sleeping, cooking, working, etc.), and image classification, which estimates the animals (cats, dogs, etc.) contained in an image. An example of a regression task is power consumption prediction, which predicts future trends based on past power consumption in a household. An example of a decision-making task is to understand the behavior of residents and the status of in-home devices and the environment from sensor data within the home, and then control the in-home devices.

[0003] However, the inside of an AI model is a complex black box, and humans cannot understand the basis for the AI's output results. This makes it difficult to use the output results for important, high-risk decision-making, such as controlling autonomous driving. Furthermore, even when controlling in-home devices does not pose a high risk, it can be difficult to persuade users to continue using the service because they cannot understand the basis for the AI's decisions. For example, users have a wide variety of preferences (energy saving, comfort, etc.) and lifestyle patterns, and the AI's device control may not match the user's preferences, especially at the start of the service. In such cases, if the AI ​​repeatedly controls devices in a way that the user cannot understand, the user may cancel the service. In such cases, if the AI ​​can present the basis for its device control decisions to the user, the user can change the service settings while understanding the AI's intentions. This will enable the AI ​​to adapt to the user's preferences, and it is expected that the service will continue to be used.

[0004] Against this background, "explainable AI," which calculates the basis for AI output results, has been attracting attention in recent years. A representative example is the Local Interpretable Model-agnostic Explanation (LIME) technology, which outputs the importance of input information relative to AI output results (see, for example, Non-Patent Document 1). With this technology, for example, in image classification, the importance of each pixel in an image to the classification result can be quantified, and pixels that the AI ​​considers important can be highlighted. Also, for example, in home appliance control, the importance of each sensor event contained in sensor data to the appliance control result can be quantified, and sensor events that the AI ​​considers important can be displayed. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Marco Tulio Ribeiro, Sameer Singh, Carlos Guestrin, “Why Should I Trust You? Explaining the Predictions of Any Classifier,” [online], [Retrieved March 25, 2023], Internet<URL:https: / / www.kdd.org / kdd2016 / papers / files / rfp0573-ribeiroA.pdf> Summary of the Invention [Problem to be solved by the invention]

[0006] Explainable AI, as described above, can present individual important events as the basis for the AI's decisions in time-series data. In response to this, the inventors of the present application thought that relationships such as "and" and "or" might be important between events in time-series data. Furthermore, according to the prior art, there is no suggestion at all that explainable AI can present relationships between events in time-series data.

[0007] Therefore, an object of the present invention is to provide a program, a determination device, and a method for determining the relationship between multiple events. [Means for solving the problem]

[0008] According to the present invention, there is provided a program for causing a computer to function as an explainable AI engine that receives input data X (x1 to xN) including a plurality of time-series events x and outputs output data Y (p1 to pM) representing scores p for each judgment element, the program comprising: The explainable AI engine further calculates importance data Z (z1 to zN) including the importance z for each event x in the input data X, A decision element determination means for determining a decision element ydec that has the highest score pydec for the output data Y (p1 to pM); an important event selection means for selecting a first event xi and a second event xj whose importance z is equal to or greater than a predetermined selection condition; pseudo data generating means for generating third pseudo input data Xrep_i,j by pseudo-replacing both the first event xi and the second event xj for the time series data X; and make it work, The explainable AI engine receives third pseudo input data Xrep_i,j and outputs third pseudo output data Yrep_i,j; a first condition determination means for determining whether a score pydec_rep_i,j of a determination element ydec of the third pseudo output data Yrep_i,j has decreased by a first predetermined threshold value or more compared to the score pydec of the determination element ydec of the output data Y; a relationship determining means for determining that the events xi and xj of the input data X are in an OR relationship when the first condition determining means determines that the result is true; The present invention is characterized by the fact that it makes a computer function by using the above-mentioned method.

[0009] According to another embodiment of the program of the present invention, the pseudo data generating means further generates first pseudo input data Xrep_i by pseudo-substituting only the first event xi and second pseudo input data Xrep_j by pseudo-substituting only the second event xj; The explainable AI engine receives first pseudo input data Xrep_i and outputs first pseudo output data Yrep_i, receives second pseudo input data Xrep_j and outputs second pseudo output data Yrep_j, The first condition determination means determines, as the first predetermined threshold value, a 21st score (pydec-pydec_rep_i) that is lowered from the score pydec of the decision element ydec of the output data Y to the score pydec_rep_i of the decision element ydec of the first pseudo output data Yrep_i; The 22nd score (pydec-pydec_rep_j) is lowered from the score pydec of the decision element ydec of the output data Y to the score pydec_rep_j of the decision element ydec of the second pseudo output data Yrep_j. Set a second score, which is the sum of the above, and judge as follows: (pydec-pydec_rep_i,j)> {(pydec-pydec_rep_i)+(pydec-pydec_rep_j)} It is also preferable to make the computer function in this way.

[0010] According to another embodiment of the program of the present invention, If the first condition determining means determines that the result is false, then, as in the following formula: The 21st score (pydec-pydec_rep_i) is greater than a second predetermined threshold, and The 22nd score (pydec-pydec_rep_j) is greater than the second predetermined threshold. (pydec-pydec_rep_i)>second predetermined threshold, and (pydec-pydec_rep_j)>second predetermined threshold a second condition determination means for determining whether and further enhance its functionality. The relationship determining means determines that the events xi and xj of the input data X are in an and relationship when the second condition determining means determines that the condition is true. It is also preferable to make the computer function in this way.

[0011] According to another embodiment of the program of the present invention, If the second condition determining means determines that the condition is false, the relationship determining means determines that the events xi and xj of the input data X have no relationship. It is also preferable to make the computer function in this way.

[0012] According to another embodiment of the program of the present invention, The important event selection means sets the predetermined selection conditions as follows: Select the first event xi and the second event xj in descending order of importance z. It is also preferable to make the computer function in this way.

[0013] According to another embodiment of the program of the present invention, The important event selection means sets the predetermined selection conditions as follows: A predetermined number of events x or events x having a predetermined degree or more are selected in descending order of importance z, and a plurality of combinations of first events xi and second events xj are selected. It is also preferable to make the computer function in this way.

[0014] According to another embodiment of the program of the present invention, The relationship determination means determines a relationship for each combination of a plurality of first events xi and second events xj; a decision basis output means for outputting decision basis combining the relationships between combinations of multiple events; It is also preferable to make the computer function as follows.

[0015] According to another embodiment of the program of the present invention, The time-series event x included in the input data X is the time-series sensor data output from an arbitrary sensor, The score p of the decision element included in the output data Y is the score for each sensor data of the sensor. It is also preferable to make the computer function in this way.

[0016] According to the present invention, there is provided a relationship determination device having an explainable AI engine that receives input data X (x1 to xN) including a plurality of time-series events x and outputs output data Y (p1 to pM) representing a score p for each determination element, The explainable AI engine further calculates importance data Z (z1 to zN) including the importance z for each event x in the input data X, A decision element determination means for determining a decision element ydec that has the highest score pydec for the output data Y (p1 to pM); an important event selection means for selecting a first event xi and a second event xj whose importance z is equal to or greater than a predetermined selection condition; pseudo data generating means for generating third pseudo input data Xrep_i,j by pseudo-replacing both the first event xi and the second event xj for the time series data X; and The explainable AI engine receives third pseudo input data Xrep_i,j and outputs third pseudo output data Yrep_i,j; a first condition determination means for determining whether a score pydec_rep_i,j of a determination element ydec of the third pseudo output data Yrep_i,j has decreased by a first predetermined threshold value or more compared to the score pydec of the determination element ydec of the output data Y; a relationship determining means for determining that the events xi and xj of the input data X are in an OR relationship when the first condition determining means determines that the result is true; The present invention is characterized by having the following.

[0017] According to the present invention, there is provided a relationship determination method for a device having an explainable AI engine that receives input data X (x1 to xN) including multiple time-series events x and outputs output data Y (p1 to pM) representing scores p for each determination element, comprising: The explainable AI engine further calculates importance data Z (z1 to zN) including the importance z for each event x in the input data X, a first step of determining a decision element ydec that gives the highest score pydec for the output data Y (p1 to pM); a second step of selecting first events xi and second events xj whose importance z is equal to or greater than a predetermined selection condition; a third step of generating third pseudo input data Xrep_i,j by pseudo-replacing both the first event xi and the second event xj for the time series data X; a fourth step of inputting third pseudo input data Xrep_i,j and outputting third pseudo output data Yrep_i,j using an explainable AI engine; a fifth step of determining whether the score pydec_rep_i,j of the decision element ydec of the third pseudo output data Yrep_i,j has decreased by a first predetermined threshold or more compared to the score pydec of the decision element ydec of the output data Y; If the fifth step determines that the result is true, the sixth step determines that the events xi and xj of the input data X are in an or relationship. The present invention is characterized by carrying out the following. [Effects of the Invention]

[0018] According to the program, determination device, and method of the present invention, it is possible to determine the relationship between a plurality of events. [Brief explanation of the drawings]

[0019] [Figure 1] This is an example of a user's home where multiple sensors are installed. [Figure 2] FIG. 2 is a functional configuration diagram of a determination device according to the present invention. [Figure 3] FIG. 1 is an explanatory diagram showing examples of input data, output data, and importance data in an explainable AI engine. [Figure 4] FIG. 2 is an explanatory diagram illustrating an important event selection unit and a pseudo data generation unit in the present invention. [Figure 5] FIG. 1 is an explanatory diagram showing an example of pseudo input data and pseudo output data in an explainable AI engine. [Figure 6] FIG. 4 is an explanatory diagram illustrating the processing of a first condition determination unit and a second condition determination unit. DETAILED DESCRIPTION OF THE INVENTION

[0020] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0021] FIG. 1 shows an example of a user's home where multiple sensors are installed.

[0022] According to Figure 1, a wide variety of sensors such as those listed below are placed in a user's home as IoT devices. Bedroom lighting (ON / OFF), bedroom human sensor (ON / OFF), bedroom temperature sensor (temperature) Living room lighting (ON / OFF), living room motion sensor (ON / OFF), Living temperature sensor (temperature), living humidity sensor (temperature), Living room air conditioner sensor (ON / OFF), living room TV (ON / OFF), Kitchen lighting (ON / OFF) Sensors detect the status and changes of the indoor environment and indoor devices as events. For example, a motion sensor is installed in a room such as the living room or kitchen, and outputs an ON signal when it detects human movement and an OFF signal when it no longer detects human movement. For example, a temperature sensor is installed in a room and measures the temperature periodically. Lighting devices such as smart plugs and televisions output their status when the switch is turned ON / OFF.

[0023] Each sensor transmits the displaced sensor data as an event to the determination device 1. For example, this time-series sensor data can be said to be the result of the actions (sleeping, working, cooking, watching, etc.) of the person staying in the room. By analyzing the patterns of this sensor data, it becomes possible to recognize human actions. This allows the determination device 1 to control home appliances as a decision-making task according to the actions of the user in the room.

[0024] 1 shows time-series input data X input to the determination device 1. The time-series data representation is configured by arranging n events in chronological order. The input data X includes, for example, 10 events {x1, x2, . . . , x10} in chronological order. Note that x1 represents the oldest event, and x10 represents the newest event. x1: Living room air conditioner OFF x2: Bedroom motion sensor ON x3: Bedroom motion sensor OFF x4: Bedroom motion sensor ON x5: Bedroom motion sensor OFF x6: Living room humidity sensor 30% x7: Living room temperature sensor 15 degrees x8: Bedroom temperature sensor 14 degrees x9: Living room motion sensor ON x10: Living room TV on Here, the air conditioner in the living room is off, the temperature in the living room is low at 15 degrees, and the human sensor has reacted, so it can be inferred that the user is in the living room. An example of an AI decision that can be expected based on such sensor data is "turn on the living room air conditioner."

[0025] FIG. 2 is a functional configuration diagram of the determination device according to the present invention.

[0026] The determination device 1 is based on classification tasks, regression tasks, and decision-making tasks that target time-series data. 2, the determination device 1 includes an explainable AI engine 10, a determination element determination unit 11, an important event selection unit 12, a pseudo-data generation unit 13, a first condition determination unit 141, a second condition determination unit 142, a relationship determination unit 15, and a determination basis output unit 16. These functional components are realized by executing a program that causes a computer installed in the device to function. The processing flow of these functional components can also be understood as a determination method of the device.

[0027] [Explanation of AI Engine 10: First Estimation Stage] Description The AI ​​engine 10 receives input data X (x1 to xN) including multiple events x in a time series, and outputs output data Y (p1 to pM) representing a score p for each decision element. Here, N represents the number of events in the input data X. M represents the number of decision elements (for example, the number of classes in a classification task, or the number of decision elements in a decision-making task).

[0028] FIG. 3 is an explanatory diagram showing an example of input data, output data, and importance data in an explainable AI engine.

[0029] According to FIG. 3 (similar to FIG. 1 described above), the event x in the input data X is, for example, as follows: (event) x1: Living room air conditioner OFF x2: Bedroom motion sensor ON x3: Bedroom motion sensor OFF x4: Bedroom motion sensor ON x5: Bedroom motion sensor OFF x6: Living room humidity sensor 30% x7: Living room temperature sensor 15 degrees x8: Bedroom temperature sensor 14 degrees x9: Living room motion sensor ON x10: Living room TV on

[0030] Moreover, the determination elements in the output data Y are the control items of the home appliances, and the score p is an expected value. According to FIG. (Judgment element) (Score) y1: Living room air conditioner ON 0.8 y2: Living room air conditioner OFF 0 y3: Living room lighting ON 0.03 y4: Living room lighting OFF 0.01 y5: Bedroom light ON 0.01 y6: Bedroom lighting OFF 0.01 y7: Kitchen light ON 0.02 y8: Kitchen lights OFF 0.02 y9: Do nothing 0.1 Here, the control item is a binary value of ON / OFF, but of course it can also be a numerical value. Also, for example, for regression tasks, the score is a predicted value.

[0031] Furthermore, the explainable AI engine 10 further calculates importance data Z (z1 to zN) including the importance z for each event x of the input data X. N corresponds to the event of the input data X. According to FIG. 3, for example, it is as follows. (Event) (Importance) z1: Living room air conditioner OFF 0.3 z2: Bedroom motion sensor ON 0.01 z3: Bedroom motion sensor OFF 0.01 z4: Bedroom motion sensor ON 0.05 z5: Bedroom human sensor OFF 0.01 z6: Living room humidity sensor 30% 0.01 z7: Living room temperature sensor 15 degrees 0.15 z8: Bedroom temperature sensor 14 degrees 0.01 z9: Living room motion sensor ON 0.15 z10: Living room TV ON 0.2 The importance z represents the basis for the judgment when a score p is predicted as output data Y for input data X.

[0032] The explainable AI engine 10 may be, for example, LIME (Local Interpretable Model-agnostic Explanations), SHAP (SHapley Additive Explanations), or Anchor. This also estimates the permutation importance (sometimes called confidence), which is an index for evaluating "how much an event (feature) contributes to improving the accuracy (score) of a prediction (expected value)."

[0033] The explainable AI engine 10 is pre-trained for classification, regression and decision-making tasks. Importance can be estimated by perturbing some of the input data and observing the change in the prediction. This looks at a counterfactual situation with respect to the perturbed part. That is, if the score drops significantly due to the perturbation of some events, then we can say that some events are important for the prediction.

[0034] [Determination element determination unit 11] The decision factor determining unit 11 determines the decision factor ydec that will give the highest score for the output data Y (p1 to pM). According to FIG. 3, the following decision factors are determined: (Decision element ydec) (Score pydec) y1: Living room air conditioner ON 0.8 (max) As a result, the decision element ydec=y1 and the score pydec=0.8 are determined.

[0035] FIG. 4 is an explanatory diagram showing the important event selection unit and the pseudo data generation unit in the present invention.

[0036] [Important Event Selection Section 12] The important event selection unit 12 selects the first event xi and the second event xj whose importance z is equal to or greater than the "predetermined selection condition."

[0037] Here, the predetermined selection conditions include, for example, the following three conditions. (First selection condition) The first event xi and the second event xj are selected in descending order of importance z. According to Figure 4, the following selections are made: Event x1 = Importance 0.3 Event x 10 = Importance 0.2

[0038] (Second selection condition) A "predetermined number k" of events x are selected in descending order of importance z, and multiple combinations of first events xi and second events xj are selected. For example, the predetermined number k=4. According to Figure 4, the following selections are made: Event x1 = Importance 0.3 Event x 7 = Importance 0.15 Event x 9 = Importance 0.15 Event x 10 = Importance 0.2

[0039] (Third selection condition) Events x with a "predetermined level or higher" are selected in descending order of importance z, and multiple combinations of first event xi and second event xj are selected. For example, the predetermined level is set to 0.1. According to Figure 4, the following selections are made: Event x1 = Importance 0.3 Event x 7 = Importance 0.15 Event x 9 = Importance 0.15 Event x 10 = Importance 0.2

[0040] For example, when Xtop_4={x1, x7, x9, x10} is selected according to the second selection condition, there are six combinations of the first event xi and the second event xj as follows: Pattern 1: (x1, x7) Pattern 2: (x1, x9) Pattern 3: (x1, x10) Pattern 4: (x7, x9) Pattern 5: (x7, x10) Pattern 6: (x9, x10)

[0041] [Pseudo Data Generation Unit 13] The pseudo data generating unit 13 generates at least the following three pseudo input data for the time series data X. First pseudo input data Xrep_i: Only the first event xi is pseudo-replaced Second pseudo input data Xrep_j: Only the second event xj is pseudo-replaced Third pseudo input data Xrep_i,j: the first event xi and the second event xj Pseudo-replace both

[0042] According to FIG. 4, for example, in the case of the above-mentioned pattern 1: (x1, x7), three pseudo input data are generated as follows. First pseudo input data Xrep_1: Only event x1 is pseudo-replaced with null Second pseudo input data Xrep_7: Only event x7 is pseudo-replaced with null Third pseudo input data Xrep_1,7: Both events x1 and x7 Pseudo-replace with null If the pseudo value is set to null, it is ignored by the explainable AI engine 10. Furthermore, the pseudo value is not limited to null, and may be any value randomly selected from all events.

[0043] [Explanation AI Engine 10 - Second estimation stage] The explainable AI engine 10 again estimates as follows: Input the first pseudo input data Xrep_i and output the first pseudo output data Yrep_i. Input the second pseudo input data Xrep_j and output the second pseudo output data Yrep_j. Input the third pseudo input data Xrep_i,j and output the third pseudo output data Yrep_i,j.

[0044] For example, in the case of the above-mentioned pattern 1: (x1, x7), three pseudo output data are output as follows. Input the first pseudo input data Xrep_1 and output the first pseudo output data Yrep_1. Input the second pseudo input data Xrep_7 and output the second pseudo output data Yrep_7. Input the third pseudo input data Xrep_1,7 and output the third pseudo output data Yrep_1,7.

[0045] FIG. 5 is an explanatory diagram showing an example of pseudo input data and pseudo output data in an explainable AI engine. According to FIG. 5, the score p1 of the decision element ydec=1 in the pseudo input data Xrep_1, Xrep_7, and Xrep_1,7 in pattern 1 is estimated, for example, as follows: p1_rep_1=0.2 p1_rep_7=0.15 p1_rep_1,7=0.1

[0046] FIG. 6 is an explanatory diagram illustrating the processing of the first condition determination unit and the second condition determination unit.

[0047] [First condition determination unit 141] The first condition determination unit 141 may use any of the following determination methods, as shown in FIGS. (First judgment method) The first condition determination unit 141 determines whether the score pydec_rep_i,j of the determination element ydec of the third pseudo output data Yrep_i,j has decreased below the score pydec of the determination element ydec of the output data Y by a first predetermined threshold or more. (pydec-pydec_rep_i,j)>first predetermined threshold

[0048] According to the example of FIG. 5 described above, the first condition determination unit 141 performs the calculation as follows. (p1-p1_rep_1,7)>first predetermined threshold (0.8-0.1) > First predetermined threshold Here, the first predetermined threshold may be set to, for example, 0.72, which is 90% of p1. In this case, according to the example of FIG. 5, the first condition decision unit 141 makes a decision of "false."

[0049] (Second judgment method) For the first judgment method, the first condition judgment unit 141 sets the first predetermined threshold to a second score {(pydec-pydec_rep_i)+(pydec-pydec_rep_j)}, which is the sum of the 21st score (pydec-pydec_rep_i) and the 22nd score (pydec-pydec_rep_j). 21st score (pydec-pydec_rep_i): From the score pydec of the decision element ydec of the output data Y, The score pydec_rep_i of the decision element ydec of the first pseudo output data Yrep_i is Decreased score 22nd score (pydec-pydec_rep_j): From the score pydec of the decision element ydec of the output data Y, The score pydec_rep_j of the decision element ydec of the second pseudo output data Yrep_j is Decreased score As a result, the first condition determination unit 141 determines whether the first score (pydec-pydec_rep_i,j) is greater than the second score as shown in the following formula: (pydec-pydec_rep_i,j)> {(pydec-pydec_rep_i)+(pydec-pydec_rep_j)}

[0050] According to the example of FIG. 5 described above, the first condition determination unit 141 performs the calculation as follows. (p1-p1_rep_1,7)>{(p1-p1_rep_1)+(p1-p1_rep_7)} (0.8-0.1) > {(0.8-0.2)+(0.8-0.15)}

[0051] When the first condition determination unit 141 determines "true" by the first determination method or the second determination method, it outputs to the relationship determination unit 15 that the first condition determination is "true." According to the example of FIG. 5 described above, the first condition decision unit 141 decides "false."

[0052] [Second condition determination unit 142] If the first condition decision unit 141 decides "false", the second condition decision unit 142 makes a decision according to the following formula. The 21st score (pydec-pydec_rep_i) is greater than a second predetermined threshold, and Is the 22nd score (pydec-pydec_rep_j) greater than a second predetermined threshold? (pydec-pydec_rep_i)>second predetermined threshold, and (pydec-pydec_rep_j)>second predetermined threshold

[0053] According to the example of FIG. 5 described above, the second condition determination unit 142 performs the calculation as follows. (p1-p1_rep_1)>second predetermined threshold (0.8-0.2) > Second predetermined threshold (p1-p1_rep_7)>Second predetermined threshold (0.8-0.15) > Second predetermined threshold Here, the second predetermined threshold may be set to, for example, 0.4, which is 50% of p1. In this case, according to the example of FIG. 5, the second condition decision unit 142 decides that the result is "true."

[0054] [Relationship Determination Unit 15] If the first condition determination unit 141 determines that the result is "true," the relationship determination unit 15 determines that the events xi and xj of the input data X are in an "or relationship." This means that if either the events xi or xj are included in the data, the determination result ydec does not change significantly, but if both are missing, the result changes significantly. Furthermore, if the second condition determination unit 142 determines that the condition is "true," the relationship determination unit 15 determines that the events xi and xj of the input data X are in an "and relationship." This means that if either the events xi or xj is missing, the determination result ydec will change significantly. If the second condition determining unit 142 determines that the result is "false," the relationship determining unit 15 determines that the events xi and xj of the input data X have "no relationship." In this way, the relationship determination unit 15 determines the relationship for each combination of a first event xi and a second event xj. In this case, according to the example of FIG. 5, the relationship determination unit 15 determines that the relationship is an "and relationship."

[0055] Such relationships between events can be obtained by analyzing how the score pydec of the judgment result ydec of the explainable AI engine 10 for actual input data X changes in the case of pseudo input data.

[0056] [Decision basis output unit 16] The decision basis output unit 16 outputs a decision basis that combines the relationships between combinations of multiple events selected by the important event selection unit 12. The relationships between all combinations of events xi, xj (∈Xtop_k) are combined with "and."

[0057] 5 and 6, the relationships are combined as follows, for example. Pattern 1: (x1, x7) and relation Pattern 2: (x1, x9) No relationship Pattern 3: (x1, x10) No relationship Pattern 4: (x7, x9) No relationship Pattern 5: (x7, x10) No relationship Pattern 6: (x9, x10) or relation By combining these relationships with "and," we can obtain the following basis for judgment: x1 and x7 and (x9 or x10)

[0058] According to the present invention, it can be determined that the determination element "living room air conditioner ON" has the following event relationship. <Determination element: "living room air conditioner ON"> x1 (living room air conditioner OFF) and x7 (living room temperature 15 degrees) and (x9 (living room sofa motion sensor ON) or x10 (living room TV motion sensor ON)) (x9 or x10) means that it is not important whether the user is near the living room (sofa) motion sensor or the living room TV, but it is important that the user is in the living room. Also, since x1 and x7 are in an AND relationship with respect to (x9 or x10), the "living room air conditioner ON" is controlled based on the conditions that the user is in the living room, the living room air conditioner is OFF, and the temperature in the living room is low.

[0059] According to Non-Patent Document 1, the basis for the judgment is not disclosed, and only the following basis for the judgment is obtained. <Determination element: "living room air conditioner ON"> x1 (living room air conditioner OFF) x7 (living room temperature 15 degrees) x9 (living room sofa motion sensor ON) x10 (living room TV motion sensor ON)

[0060] As described above in detail, the program, determination device, and method of the present invention make it possible to determine the relationship between a plurality of events.

[0061] Furthermore, this will enable, for example, "determining the relationships between multiple events," which will contribute to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs), which is to "build resilient infrastructure, promote sustainable industrialization and foster innovation."

[0062] With respect to the various embodiments of the present invention described above, various changes, modifications, and omissions that fall within the scope of the technical spirit and aspects of the present invention may be easily made by those skilled in the art. The above description is merely illustrative and is not intended to be limiting in any way. The present invention is limited only by the claims and their equivalents. [Explanation of symbols]

[0063] 1 Judgment device 10. Explained AI Engine 11. Judgment element determination unit 12 Important Event Selection Section 13 Pseudo Data Generation Unit 141 First condition determination unit 142 Second condition determination section 15 Relationship determination section 16. Decision Basis Output Unit

Claims

1. A program that causes a computer to function as an explainable AI engine that receives input data X (x1 to xN) including multiple time-series events x and outputs output data Y (p1 to pM) representing scores p for each judgment element, The explainable AI engine further calculates importance data Z (z1 to zN) including importance z for each event x of the input data X, a decision element determining means for determining a decision element ydec that will give the highest score pydec for the output data Y (p1 to pM); an important event selection means for selecting a first event xi and a second event xj whose importance z is equal to or greater than a predetermined selection condition; pseudo data generating means for generating third pseudo input data Xrep_i,j by pseudo-replacing both the first event xi and the second event xj for the time series data X; and make it work, The explainable AI engine receives third pseudo input data Xrep_i,j and outputs third pseudo output data Yrep_i,j; a first condition determination means for determining whether a score pydec_rep_i,j of a determination element ydec of the third pseudo output data Yrep_i,j has decreased by a first predetermined threshold value or more compared to the score pydec of the determination element ydec of the output data Y; relationship determining means for determining that the events xi and xj of the input data X are in an OR relationship when the first condition determining means determines that the result is true; A program that causes a computer to function.

2. the pseudo data generating means further generates first pseudo input data Xrep_i by pseudo-replacing only the first event xi, and second pseudo input data Xrep_j by pseudo-replacing only the second event xj; The explainable AI engine receives first pseudo input data Xrep_i and outputs first pseudo output data Yrep_i, receives second pseudo input data Xrep_j and outputs second pseudo output data Yrep_j, The first condition determination means determines, as the first predetermined threshold value, a 21st score (pydec-pydec_rep_i) that is lowered from the score pydec of the decision element ydec of the output data Y to the score pydec_rep_i of the decision element ydec of the first pseudo output data Yrep_i; A 22nd score (pydec-pydec_rep_j) that is lowered from the score pydec of the decision element ydec of the output data Y to the score pydec_rep_j of the decision element ydec of the second pseudo output data Yrep_j; A second score is set as the sum of the above and is judged as follows: (pydec−pydec_rep_i,j)> {(pydec−pydec_rep_i)+(pydec−pydec_rep_j)} 2. The program according to claim 1, which causes a computer to function as follows:

3. If the first condition determining means determines that the result is false, then, as shown in the following formula: The twenty-first score (pydec-pydec_rep_i) is greater than a second predetermined threshold, and The twenty-second score (pydec-pydec_rep_j) is greater than a second predetermined threshold. (pydec-pydec_rep_i)>second predetermined threshold, and (pydec-pydec_rep_j)>second predetermined threshold a second condition determination means for determining whether and further enhance its functionality. The relationship determining means determines that the events xi and xj of the input data X are in an and relationship when the second condition determining means determines that the condition is true.

3. The program according to claim 1, wherein the program causes a computer to function as described above.

4. If the second condition determining means determines that the condition is false, the relationship determining means determines that the events xi and xj of the input data X have no relationship.

4. The program according to claim 3, which causes a computer to function as follows.

5. The important event selection means sets the predetermined selection conditions as follows: Select the first event xi and the second event xj in descending order of importance z.

2. The program according to claim 1, which causes a computer to function as follows:

6. The important event selection means sets the predetermined selection conditions as follows: A predetermined number of events x or events x having a predetermined degree or more are selected in descending order of importance z, and a plurality of combinations of first events xi and second events xj are selected.

2. The program according to claim 1, which causes a computer to function as follows:

7. The relationship determination means determines a relationship for each combination of a plurality of first events xi and second events xj; a decision basis output means for outputting decision basis combining the relationships between combinations of multiple events; 7. The program according to claim 6, which causes a computer to function as described above.

8. The time-series event x included in the input data X is time-series sensor data output from an arbitrary sensor, The score p of the judgment element included in the output data Y is the score for each sensor data of the sensor.

2. The program according to claim 1, which causes a computer to function as follows:

9. A relationship determination device having an explainable AI engine that receives input data X (x1 to xN) including a plurality of time-series events x and outputs output data Y (p1 to pM) representing scores p for each determination element, The explainable AI engine further calculates importance data Z (z1 to zN) including importance z for each event x of the input data X, a decision element determining means for determining a decision element ydec that will give the highest score pydec for the output data Y (p1 to pM); an important event selection means for selecting a first event xi and a second event xj whose importance z is equal to or greater than a predetermined selection condition; pseudo data generating means for generating third pseudo input data Xrep_i,j by pseudo-replacing both the first event xi and the second event xj for the time series data X; and The explainable AI engine receives third pseudo input data Xrep_i,j and outputs third pseudo output data Yrep_i,j; a first condition determination means for determining whether a score pydec_rep_i,j of a determination element ydec of the third pseudo output data Yrep_i,j has decreased by a first predetermined threshold value or more compared to the score pydec of the determination element ydec of the output data Y; relationship determining means for determining that the events xi and xj of the input data X are in an OR relationship when the first condition determining means determines that the result is true; A relationship determination device comprising:

10. A relationship determination method for a device having an explainable AI engine that receives input data X (x1 to xN) including a plurality of time-series events x and outputs output data Y (p1 to pM) representing scores p for each determination element, The explainable AI engine further calculates importance data Z (z1 to zN) including importance z for each event x of the input data X, a first step of determining a decision element ydec that results in the highest score pydec for the output data Y (p1 to pM); a second step of selecting a first event x and a second event x whose importance z is equal to or greater than a predetermined selection condition; a third step of generating third pseudo input data Xrep_i,j by pseudo-replacing both the first event x and the second event x for the time series data X; a fourth step of inputting third pseudo input data Xrep_i,j and outputting third pseudo output data Yrep_i,j using an explainable AI engine; a fifth step of determining whether the score p ydec_rep_i,j of the decision element ydec of the third pseudo output data Yrep_i,j has decreased by a first predetermined threshold or more compared to the score p ydec of the decision element ydec of the output data Y; If the fifth step determines that the result is true, a sixth step determines that the events xi and xj of the input data X are in an or relationship. A method for determining a relationship between devices, comprising:

Images