JP7363910B2 - Display method, display program and information processing device - Google Patents

Description

The present invention relates to a display method, a display program, and an information processing device.

Conventionally, methods for detecting concept drift have been proposed. Here, concept drift refers to the change in the quality of data input to an AI system over time. AI systems use learning devices to perform various inferences on input data. Furthermore, the learning device is trained during development. Therefore, when concept drift occurs, the accuracy of the AI system may be relatively reduced.

Tegjyot Singh Sethi, Mehmed Kantardzic, "On the reliable detection of concept drift from streaming unlab ele d data", Expert Systems With Applications 82 (2017) 77-99

However, the above method has a problem in that the calculation cost may be high. For example, a method is conventionally known in which a plurality of learning devices are ensembled and a majority decision is made. In this method, since it is necessary to perform learning for each of a plurality of learning devices, the amount of calculation may increase.

One aspect aims to detect concept drift with a small amount of calculation.

In one embodiment, the computer obtains operational data that is input to the trained model. Based on the acquired operational data, the computer calculates an index indicating the magnitude of change in the output result of the model due to a temporal change in the trend of the operational data. Based on a comparison between the calculated index and a preset threshold, the computer determines whether a concept drift that requires a user response has occurred. If it is determined that a concept drift that requires a user's response has occurred, the computer causes the display screen to display information indicating that a concept drift that requires a user's response has occurred.

One aspect is that concept drift can be detected with a small amount of calculation.

FIG. 1 is a diagram for explaining concept drift. FIG. 2 is a diagram illustrating an example of the configuration of the detection device according to the first embodiment. FIG. 3 is a diagram for explaining the processing flow of the detection device. FIG. 4 is a diagram for explaining parameters of the neural network. FIG. 5 is a diagram for explaining the replica model. FIG. 6 is a diagram for explaining a method of acquiring the distribution of outputs for learning data. FIG. 7 is a diagram for explaining a method of acquiring the distribution of outputs for operational data. FIG. 8 is a diagram showing an example of a change in KLD. FIG. 9 is a diagram showing an example of a change in distribution. FIG. 10 is a diagram showing an example of a monthly display screen when concept drift does not occur. FIG. 11 is a diagram showing an example of a monthly display screen when incremental drift occurs. FIG. 12 is a diagram showing an example of a detailed display screen. FIG. 13 is a diagram showing an example of a monthly display screen when sudden drift occurs. FIG. 14 is a diagram showing an example of a monthly display screen when recurring drift occurs. FIG. 15 is a diagram showing an example of a daily display screen. FIG. 16 is a flowchart showing the flow of processing during learning of the detection device. FIG. 17 is a flowchart showing the flow of processing during operation of the detection device. FIG. 18 is a diagram for explaining parameters of a decision tree. FIG. 19 is a diagram illustrating an example of a hardware configuration.

Embodiments of a display method, a display program, and an information processing apparatus according to the present invention will be described in detail below with reference to the drawings. Note that the present invention is not limited to this example. Moreover, each embodiment can be combined as appropriate within a consistent range. Further, the detection device of the embodiment can be said to be an information processing device that executes the detection method and display method.

[About concept drift]
First, I will explain concept drift. Here, a trained model that has been trained using learning data with known labels is called an original model. Furthermore, the model in this embodiment is a classifier that estimates and classifies the class to which data belongs based on feature amounts. The classifier calculates a score for each class for each data and performs classification based on the score. Here, the score may be the probability that data belongs to each class. Furthermore, the classifier may be one that performs classification into two classes, such as whether it is normal or not, or may be one that performs classification into three or more classes.

FIG. 1 is a diagram for explaining concept drift. The left and right diagrams in Figure 1 depict data points for each class during learning and operation, and a curve representing the original model, on a plane whose horizontal and vertical axes represent different feature quantities, respectively. . Data points marked ○, Δ, and × in FIG. 1 correspond to data belonging to class A, class B, and class C, respectively.

As shown in the left diagram of FIG. 1, during learning, the original model was able to classify data in each class with high accuracy. On the other hand, during operation, the distribution of classes B and C in particular changes significantly, and the classification accuracy of the original model decreases. Concept drift is a phenomenon in which the quality of data changes during model operation compared to during training. In other words, changes in model output results due to time changes in trends in operational data become large. As shown in FIG. 1, when concept drift occurs, the accuracy of the model may be relatively reduced.

For example, consider a case in which the model classifies each stock into one of "buy," "sell," and "neutral" in stock trading. In this case, a model trained on 2018 data may show high accuracy in early 2019, but the data trend changes in mid-2019 and the accuracy decreases.

Additionally, if a model is trained to determine whether the indoor environment is hot or not using data from June, by August the model's judgment results may no longer match the user's feelings. Conceivable. This is because heat acclimation, which allows humans to experience heat, appears as concept drift. For example, a user may feel that the room temperature is hot in June but no longer feels hot in August.

Furthermore, the classification accuracy of a model that classifies whether a program is malware or not is likely to decline due to concept drift that occurs as malware evolves. Since malware that cannot be detected by existing methods is being developed day and night, malware with previously unknown characteristics may emerge over time.

For example, the model uses conditions such as the temperature and brightness of the manufacturing environment and predetermined measured values obtained in each process as features on the manufacturing line to determine whether or not the product is defective. be able to. In this case, changes in the environment, etc. may appear as concept drift, and the model's judgment accuracy may decrease.

The detection device of this embodiment can be applied to any model that is affected by concept drift as described above. The detection device of this embodiment detects concept drift using a plurality of replica models generated from an original model. Since no learning is required to generate the replica model, the detection device can generate the replica model with a small amount of calculation.

Note that in many cases, the true class to which data belongs during operation is unknown, but for the sake of explanation in Figure 1, the shapes of ○, △, and × corresponding to each class are assigned to the data points during operation. ing. The purpose of the detection device of this embodiment is to detect concept drift that the user needs to deal with even if the true class of data during operation is unknown.

Additionally, if a concept drift that requires a user response is detected, countermeasures may be taken, such as automatically or manually retraining the original model and excluding incorrect data that caused the concept drift. .

[Functional configuration]
The configuration of the detection device according to the embodiment will be explained using FIG. 2. FIG. 2 is a diagram illustrating an example of the configuration of the detection device according to the first embodiment. As shown in FIG. 2, the detection device 10 includes a communication section 11, an input section 12, an output section 13, a storage section 14, and a control section 15.

The communication unit 11 is an interface for communicating data with other devices. For example, the communication unit 11 may be a NIC (Network Interface Card) that communicates data via the Internet.

The input unit 12 is an interface for receiving data input. For example, the input unit 12 may be an input device such as a keyboard or a mouse. Further, the output unit 13 is an interface for outputting data. The output unit 13 may be an output device such as a display or a speaker. Further, the input section 12 and the output section 13 may input/output data to/from an external storage device such as a USB memory.

The storage unit 14 is an example of a storage device that stores data, programs executed by the control unit 15, and the like, and is, for example, a hard disk or a memory. The storage unit 14 stores original model information 141, replica model information 142, and distribution information 143.

The original model information 141 is information such as parameters for constructing an original model. For example, if the original model is a neural network, the original model information 141 is the weight and bias of each layer.

The replica model information 142 is information such as parameters for constructing a replica model. In this embodiment, it is assumed that the replica model and the original model use the same algorithm. Therefore, for example, if the original model's algorithm is a neural network, the replica model's algorithm is also a neural network. In this case, the replica model information 142 is the weight and bias of each layer of the neural network. Further, there may be a plurality of replica models, in which case the replica model information 142 is a parameter corresponding to each of the plurality of replica models. A method for generating a replica model will be described later.

The distribution information 143 is information regarding the distribution of the output of the replica model during learning. The distribution information 143 is information that serves as a reference for evaluating data during operation and detecting concept drift that the user needs to respond to. The distribution information 143 may be sufficient information to calculate the evaluation value of the data during operation. For example, the distribution information 143 may be all or part of the output of the replica model, or may be a distribution parameter indicating the distribution of the output of the replica model. A method for obtaining the distribution of outputs of the replica model will be described later.

The control unit 15 executes a program stored in an internal storage device using RAM as a work area by, for example, a CPU (Central Processing Unit), MPU (Micro Processing Unit), GPU (Graphics Processing Unit), etc. This is realized by Further, the control unit 15 may be realized by, for example, an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array). The control unit 15 includes a learning unit 151, a generation unit 152, a calculation unit 153, a detection unit 154, and a display control unit 155.

The details of each part of the control unit 15 will be explained along the flow of processing shown in FIG. FIG. 3 is a diagram for explaining the processing flow of the detection device. The processing of the detection device 10 is divided into a learning phase and an operation phase. In the learning phase, the detection device 10 generates original model information 141, replica model information 142, and distribution information 143. On the other hand, in the operation phase, the detection device 10 uses the original model information 141, the replica model information 142, and the distribution information 143 to detect concept drift that the user needs to respond to.

(learning phase)
First, the learning phase processing will be explained. As shown in step S11 in FIG. 3, the learning unit 151 generates an original model based on the learning data. The learning unit 151 can perform learning using a known method. For example, if the original model is a neural network, the learning unit 151 performs learning using the error backpropagation method.

Next, the generation unit 152 generates a plurality of replica models by changing predetermined parameters of the learned original model. In the following description, the plurality of replica models generated by the generation unit 152 may be referred to as a replica model group. Further, the original model is an example of a first model. Further, the replica model is an example of the second model.

The generation unit 152 generates a plurality of neural networks as replica models by changing the weights and biases in predetermined layers of the neural network that is the trained original model.

FIG. 4 is a diagram for explaining parameters of the neural network. As shown in Figure 4, weights w_a1, w_a2, w_b1, w_b2, w_c1, w_c2, w_d1, w_d2 and biases b1 and b2 are set between the output layer and hidden layer of the neural network of the original model. . The generation unit 152 generates a replica model by multiplying the weights and biases by random numbers.

Let ε_i be a random number that follows a normal distribution with a mean of 1.0 and a standard deviation of 0.3. At this time, the generation unit 152 generates weights w´_a1, w´_b1, w´_c1, w´_d1, w´_a2, w´_b2 between the output layer and the hidden layer as the first replica model. , w´_c2, w´_d2, and biases b´1, b´2 are set to generate a neural network. However, each parameter is expressed as, for example, w´_a1 = w_a1 * 1.04, w´_a2 = w_a2 * 0.91, ..., w´_d2 = w_d2 * 1.13, b1´ = b1 * 0.83, , b2´ = b2 * 1.35. be done.

If the difference between the replica model and the original model is too large, false positives due to concept drift may occur. For this reason, the generation unit 152 generates a replica model such that the deviation from the original model is within a somewhat narrow range. For example, the generation unit 152 generates a replica model by multiplying a predetermined parameter by a random number that follows a normal distribution and has an average of 1.0 and a standard deviation of 0.3.

The calculation unit 153 calculates an output for data actually input to the model. As shown in step S12 in FIG. 3, the calculation unit 153 stores information regarding the distribution of the output obtained by inputting the learning data used for learning the original model into the replica model in the storage unit 14 as distribution information 143. . Further, the calculation unit 153 may use distribution information 143 as information regarding the distribution including not only the output of the replica model group but also the output of the original model.

Note that the learning unit 151 performs learning by optimizing the output of the model in conjunction with the calculation unit 153. On the other hand, the generation unit 152 can generate a replica model without the calculation unit 153 performing calculations.

Here, the replica model will be explained using FIG. 5. FIG. 5 is a diagram for explaining the replica model. When the replica model is drawn on a plane similar to that in FIG. 1, it is represented as shown in FIG. 5. That is, the replica model group can be said to be a group of curves that are distributed around the original model, which is a curve, with a predetermined variation, or a region formed by the group of curves.

As shown in the left diagram of FIG. 5, during learning, the replica model group shows the same classification results as the original model group. On the other hand, as shown in the right diagram of FIG. 5, during operation, each replica model may show different classification results for data points located near the curve of the original model. The detection device 10 of the present embodiment focuses on changes over time in the output of each replica model group to detect concept drift that the user needs to respond to.

FIG. 6 is a diagram for explaining a method of acquiring the distribution of outputs for learning data. In this embodiment, it is assumed that the original model classifies data into one of class A, class B, and class C. Further, it is assumed that the detection device 10 uses, for example, 70 pieces of learning data for each class for learning. Further, it is assumed that the detection device 10 generates, for example, 1,000 replica models.

At this time, as shown in FIG. 6, the calculation unit 153 converts each of the learning data s ₁ , s ₂ , ..., s ₇₀ that is known to belong to class C into the replica models r ₁ , r ₂ , r ₁₀₀₀ . Enter each to calculate the score for each class. For example, the score is the probability that data belongs to each class. For example, (A0.0, B0.0, C1.0) indicates that the probability that the data belongs to class A is 0.0, the probability that the data belongs to class B is 0.0, and the probability that the data belongs to class C is 1.0.

(operation phase)
Next, the operation phase processing will be explained. As shown in step S21 of FIG. 3, the calculation unit 153 inputs operational data to the original model and calculates the output. For example, the output is a score for each class to be classified. At this time, the output unit 13 can output the score for each class or information specifying the class with the highest score as the classification result.

Further, the learning data and the operational data may be time-series data. At this time, the learning data can be said to be data during learning of the original model among time-series data. On the other hand, operational data can be said to be data at a predetermined time after the learning time among time-series data.

Then, as shown in step S22 in FIG. 3, the detection unit 154 detects the difference between the distribution of outputs obtained by inputting the learning data into the replica model and the distribution of outputs obtained by inputting the operational data into the replica model. Detect. The operational data is an example of predetermined data different from the learning data.

The calculation unit 153 inputs operational data to each replica model and calculates the output. The detection unit 154 obtains the distribution of the output calculated by the calculation unit 153. The detection unit 154 also obtains the distribution of output obtained by inputting the learning data to the replica model from the distribution information 143.

FIG. 7 is a diagram for explaining a method of acquiring the distribution of outputs for operational data. Here, it is assumed that there is only one piece of operational data. Further, as described above, it is assumed that, for example, 1,000 replica models have been generated.

At this time, as shown in FIG. 7, the calculation unit 153 inputs the operational data u ₁ classified into class C by the original model to each of the replica models r ₁ , r ₂ , ..., r ₁₀₀₀ , and Calculate the score. Note that when there is a single piece of operational data, the calculation unit 153 can calculate the score by real-time processing, as shown in FIG. On the other hand, the calculation unit 153 may calculate the score by batch processing a plurality of pieces of operational data.

Returning to FIG. 3, the detection unit 154 calculates the KLD (Kullback-Leibler divergence) between the distribution of scores obtained by inputting learning data into the replica model and the distribution of scores obtained by inputting operational data into the replica model. Detect differences based on the KLD is an example of an index that indicates the magnitude of the difference between distributions. The detection unit 154 calculates the KLD between the probability distribution P(i) of the scores of the learning data and the probability distribution Q(i) of the scores of the operational data, as shown in equation (1). However, i is an identifier indicating the classification.

Further, the detection unit 154 calculates the sum of D _KL for class A, class B, and class C as shown in equation (2). Then, the detection unit 154 detects a concept drift that the user needs to deal with when the D _KL of each class or the calculated sum of D _KL is equal to or greater than a threshold value.

FIG. 8 is a diagram showing an example of a change in KLD. “Sample ID” is an ID for identifying individual data included in operational data. Furthermore, here, the "correct answer" is the true correct answer class of the operational data. Moreover, "prediction" is a classification result based on the original model.

As shown in FIG. 8, during learning or immediately after learning, the classification results and the correct answers match, and all KLDs take small values of 0.5 or less. On the other hand, during operation, there are many cases where the classification result does not match the correct answer, or where the KLD takes a large value of 30 or more. Note that KLD tends to increase when the classification result and the correct answer do not match, but even if the classification result and the correct answer match, KLD alone may increase.

Changes in distribution over time will be explained using FIG. 9. FIG. 9 is a diagram showing an example of a change in distribution. t represents a fixed time such as day, week, month, etc. Furthermore, each histogram in FIG. 9 is a plot of the scores output by each replica model with respect to data classified into class C by the original model. The histogram at t=0 is a plot of scores against the training data. Furthermore, the histogram after t=1 is a plot of scores against operational data.

As shown in Figure 9, in the histogram at t=0, all 1,000 replica models output scores of (A0.0, B0.0, C1.0) for all 70 pieces of training data. It shows. Furthermore, KLD at t=0 is 0. Note that depending on the learning method of the original model and the method of generating the replica model, some variation may occur in the histogram at t=0.

Then, as time passes, variations occur in the histograms of the scores of class B and the scores of class C. For example, at time t=1, many of the 1,000 replica models start outputting scores like (A0.0, B0.1, C0.9) for one piece of operational data. . Furthermore, at time t=2, the number of replica models that output large scores for class B increases. Furthermore, at time t=3, the number of replica models in which the scores of class C and class B are reversed increases. Furthermore, as the dispersion of the histogram increases, the KLD also increases.

The threshold value used by the detection unit 154 to detect a concept drift that the user needs to respond to varies depending on the type and content of the business in which data classification is performed, the required accuracy, and the like. For example, in the example of FIG. 9, if the KLD threshold is 10, the detection unit 154 detects a concept drift that the user needs to respond to at time t=2.

The display control unit 155 can display the distribution comparison results and concept drift detection results by the detection unit 154 on the screen via the output unit 13. FIG. 10 is a diagram showing an example of a monthly display screen when concept drift does not occur. As shown in FIG. 10, the display control unit 155 displays monthly changes in KLD in a line graph on the monthly display screen. Further, the display control unit 155 displays a screen transition menu when a data point is clicked. The screen transition menu includes items such as "Display details" and "Display by day." The transition destination screen will be described later.

(Display example of display device screen)
Here, display examples of the display screen of the display device will be explained using FIGS. 11 to 15. Information indicating that a concept drift that requires a user response has occurred is displayed on the screen of the display device. It is known that there are several types of concept drift: incremental drift, sudden drift, and recurring drift.

Incremental drift is a phenomenon in which the accuracy of a model gradually decreases due to concept drift. Sudden drift refers to a phenomenon in which the accuracy of a model sharply decreases due to concept drift. Recurring drift refers to a phenomenon in which the accuracy of a model decreases only at specific times on a periodic time axis due to concept drift.

Note that the accuracy of the model decreases relatively as the trends in the data change, but the accuracy of the model that a user can tolerate varies from user to user. For example, consider a case in which the model classifies each stock into one of "buy," "sell," and "neutral" in stock trading. In this case, it is necessary to maintain a high accuracy state of the model, so if the accuracy of the model decreases even slightly, it is necessary to immediately perform relearning using data with new trends. Then, it is necessary to update the currently operated model to the retrained model. On the other hand, consider a case where it is determined whether the indoor environment is hot or not. In this case, some users may tolerate some reduction in model accuracy due to their tolerance for heat, so when the change in the new trend data becomes large, the new trend data All you have to do is learn. Then, the currently operated model is updated to the retrained model.

In other words, the timing for relearning the model can be determined depending on the accuracy of the model that is acceptable to the user.

FIG. 11 is a diagram showing an example of a monthly display screen when incremental drift occurs. The generation unit 152 acquires operational data to be input to the trained model. Then, the generation unit 152 calculates, based on the acquired operational data, an index indicating the magnitude of change in the output result of the model due to a temporal change in the trend of the operational data.

Furthermore, the generation unit 152 determines whether or not a concept drift that the user needs to respond to has occurred, based on a comparison between the calculated index and a preset threshold. If it is determined that a concept drift that requires a user response has occurred, the generation unit 152 causes the display screen to display information indicating that a concept drift that requires a user response has occurred.

Specifically, the generation unit 152 generates a plurality of replica models by changing a predetermined parameter of an original model, which is a model that calculates a score for classifying data into one of a plurality of classes. Then, the detection unit 154 detects that the index indicating the size of the difference between the distribution of scores obtained by inputting the learning data to the replica model and the distribution of outputs obtained by inputting the operational data to the replica model is equal to or greater than the threshold value. If so, it is detected that a concept drift has occurred that the user needs to respond to.

In other words, the detection unit 154 calculates the distribution of scores obtained by inputting the learning data during learning of the original model into the replica model among the time-series data, and the replica model's data at a plurality of times after the learning time. Obtain the distribution of scores obtained by inputting them into the model. Then, when the index indicating the magnitude of the difference between the respective distributions is equal to or greater than the threshold value, the detection unit 154 detects that a concept drift that needs to be dealt with by the user has occurred. Further, the display control unit 155 displays information regarding each index at a plurality of times on the screen. Then, the display control unit 155 further displays information regarding the index of the time when it was detected that a concept drift that the user needs to respond to has occurred. Display information about time indicators that are not available in a different manner.

At this time, the detection unit 154 sets a threshold value for the index indicating the magnitude of the difference between each distribution. The threshold value can be set depending on the accuracy of the model that is acceptable to the user. First, the detection unit 154 receives a value related to acceptable model accuracy from the user. Then, the detection unit 154 sets a value corresponding to the received value as a threshold value. Note that the threshold value can be set using information regarding the accuracy of the model required by the industry. The detection unit 154 identifies the accuracy of the model required for the industry from the user's industry information. Then, the detection unit 154 determines a threshold value based on the accuracy of the identified model. For example, the accuracy of a model that detects defective products from images of products must not overlook even one defective product, so it is necessary to set a threshold according to the high accuracy.

In the example of FIG. 11, it is assumed that the original model was trained using training data before July 2018. Here, months such as July 2018 and August 2018 are examples of multiple times. As shown in FIG. 11, since KLD reaches the threshold of 10 or more in November 2018, the detection unit 154 detects incremental drift in November 2018.

As shown in FIG. 11, the display control unit 155 displays each index at a plurality of times in a line graph. In the line graph, the display control unit 155 sets a point corresponding to a time when a concept drift that needs to be dealt with by the user is detected, and a point corresponding to a time when a concept drift that needs to be dealt with by the user is not detected; The display is made to differ in at least one of color, shape, and size.

In other words, the display control unit 155 displays information indicating the concept drift that has occurred on the display screen, and displays concept drifts that the user needs to respond to and concept drifts that the user does not need to respond to in different ways. to be displayed.

Specifically, the generation unit 152 causes the display screen to display each index at a plurality of times in the form of a line graph. The generation unit 152 determines, in the line graph, a point corresponding to a time when a concept drift that requires a user response is detected to have occurred, and a point that corresponds to a point where a concept drift that requires a user response is detected to have occurred. A point corresponding to a time when no time is present is displayed so as to be different in at least one of color, shape, and size.

In the example of FIG. 11, the display control unit 155 displays data points before October 2018, for which no concept drift that the user needs to respond to, is displayed as a white circle. On the other hand, the display control unit 155 displays the data points for November 2018 where a concept drift that the user needs to respond to is detected as a black circle and a large-sized circle surrounding the circle. Further, the display control unit 155 displays a message indicating that an incremental drift has occurred near the data point in November 2018 where a concept drift that needs to be addressed by the user is detected. Further, the display control unit 155 may display a message using a tool tip.

The display control unit 155 displays a detailed display screen when "Display details" in the transition menu is selected. FIG. 12 is a diagram showing an example of a detailed display screen. The display control unit 155 displays a histogram of the distribution of outputs obtained by inputting learning data into the replica model and a histogram showing the distribution of outputs obtained by inputting data at a specified time into the replica model on a detailed display screen. to be displayed. Each histogram may be similar to that described using FIG. 9. Furthermore, the display control unit 155 may display the value of each variable of the input data corresponding to the selected data point.

When the learning unit 151 detects that a concept drift that requires a user's response has occurred, the learning unit 151 causes the display screen of the display device to display information indicating that a concept drift that requires a user's response has occurred. At this time, the learning unit 151 receives from the user whether or not to perform relearning. When receiving a request from the user to perform relearning, the learning unit 151 extracts the latest operational data at the time when a concept drift that needs to be addressed by the user occurs from among the time-series data. . Specifically, the learning unit 151 extracts operational data for a preset period before the point in time when a concept drift that needs to be dealt with by the user occurs. The learning unit 151 then performs relearning using the operational data and generates a new model. For example, the learning unit 151 performs relearning using a neural network. At this time, the learning unit 151 can perform relearning using the input data displayed on the detailed display screen as learning data. Note that the labels of the learning data in relearning may be assigned manually, or may be automatically assigned based on the outputs of the original model and the replica model.

The learning unit 151 inputs the extracted operational data from the input layer and forward propagates it to the neural network. Then, the classification result obtained from the output layer is compared with the correct answer (positive example/negative example) of the user's operation to find an error. Furthermore, by using the error backpropagation method, the error between the classification result and the correct answer is propagated to the neural network in the opposite direction to that during classification, and the parameters of each layer of the neural network are changed to approach the optimal solution. Although an example using a neural network has been described as a learning model, the present invention is not limited to this, and other machine learning methods such as a logistic regression model and a support vector machine may also be employed. The learning unit 151 changes the current original model to a newly generated model.

FIG. 13 is a diagram showing an example of a monthly display screen when sudden drift occurs. The generation unit 152 acquires operational data to be input to the trained model. Then, the generation unit 152 calculates an index indicating the magnitude of change in the model output result due to a temporal change in the trend of the operational data based on the acquired operational data, and uses the calculated index to calculate a time-series index. Identify changes in metrics. Furthermore, the generation unit 152 determines whether a concept drift that requires a user response has occurred based on a comparison between the change in the specified index and a preset threshold. If it is determined that a concept drift that requires a user response has occurred, the generation unit 152 causes the display screen to display information indicating that a concept drift that requires a user response has occurred.

At this time, the display control unit 155 displays information indicating the concept drift that has occurred on the display screen, and displays concept drifts that the user needs to respond to and concept drifts that the user does not need to respond to in different ways. Display it on the screen.

Specifically, the generation unit 152 causes the display screen to display each index at a plurality of times in the form of a line graph. The generation unit 152 determines, in the line graph, a point corresponding to a time when a concept drift that requires a user response is detected to have occurred, and a point that corresponds to a point where a concept drift that requires a user response is detected to have occurred. A point corresponding to a time when no time is present is displayed so as to be different in at least one of color, shape, and size.

Specifically, the generation unit 152 generates a plurality of replica models by changing a predetermined parameter of an original model, which is a model that calculates a score for classifying data into one of a plurality of classes. Of the time-series data, the detection unit 154 calculates the distribution of scores obtained by inputting the learning data during learning of the original model into the replica model, and the replica model data at a plurality of times after the learning time. Obtain the distribution of scores obtained by inputting them into the model. Further, the detection unit 154 detects that a concept drift, which the user needs to respond to, has occurred when the time-series change in the index indicating the magnitude of the difference in each distribution satisfies a predetermined condition.

In other words, when the amount of change in the index over time is equal to or greater than the threshold, the detection unit 154 detects that a concept drift that the user needs to respond to has occurred. As shown in FIG. 13, since KLD increases rapidly in November 2018, the detection unit 154 detects that sudden drift occurred in November 2018. In this case, KLD is less than the incremental drift threshold of 10, but the amount of change between October 2018 and November 2018 is greater than or equal to the sudden drift threshold.

The display control unit 155 displays a message indicating that a sudden drift has occurred near the data point in November 2018 where a concept drift that needs to be addressed by the user is detected. On the other hand, the display control unit 155 displays data points before October 2018, in which no concept drift that the user needs to deal with is detected, as white circles.

Furthermore, when the learning unit 151 detects that a concept drift that needs to be dealt with by a trained user has occurred, the learning unit 151 displays information indicating that a concept drift that needs to be dealt with by the user has occurred on the display screen of the display device. to be displayed. At this time, the learning unit 151 receives from the user whether or not to perform relearning. When receiving a request from the user to perform relearning, the learning unit 151 extracts the latest operational data at the time when a concept drift that needs to be addressed by the user occurs from among the time-series data. . Specifically, the learning unit 151 extracts operational data for a preset period before the point in time when a concept drift that needs to be dealt with by the user occurs. The learning unit 151 performs relearning using operational data from a preset period ago.

The KLD on the monthly display screen described above may be, for example, the total value of KLD on a daily basis. In this case, the generation unit 152 generates a plurality of replica models by changing predetermined parameters of the original model, which is a model that calculates a score for classifying data into one of a plurality of classes. Of the time-series data, the detection unit 154 calculates the distribution of scores obtained by inputting the learning data during learning of the original model into the replica model, and the replica model data at a plurality of times after the learning time. Obtain the distribution of scores obtained by inputting them into the model. Here, the display control unit 155 displays on the screen the total value for each fixed period of the index indicating the magnitude of the difference in each distribution acquired by the detection unit 154.

If the index or total value is equal to or greater than the threshold, the detection unit 154 detects that a concept drift that requires a user to respond has occurred. In other words, the detection unit 154 detects that a concept drift that the user needs to respond to has occurred based on the daily KLD and/or the monthly KLD. In this case, for example, even if the total KLD for a certain month does not exceed the first threshold, if the KLD for one day of the month is greater than or equal to the second threshold, the detection unit 154 , it is possible to detect that a concept drift has occurred that the user needs to respond to.

When a concept drift that requires a user response is detected, the display control unit 155 causes information regarding the concept drift to be displayed on the screen. For example, even if the display control unit 155 detects that a concept drift that requires a user response has occurred from the daily KLD, the display control unit 155 displays a message to the effect that a concept drift has occurred on the monthly display screen. You may.

Further, the display control unit 155 switches between a screen that displays information regarding a plurality of time-based indicators and a screen that displays information regarding total values for each fixed period. For example, when "display by day" is selected in the monthly display transition menu, the display control unit 155 displays a daily display screen.

The daily display screen and detection of recurring drift will be explained using FIGS. 14 and 15. FIG. 14 is a diagram showing an example of a monthly display screen when recurring drift occurs. Further, FIG. 15 is a diagram showing an example of a daily display screen.

In the example of FIG. 14, it is assumed that the total value of daily KLD is displayed as monthly KLD on the monthly display screen. It is assumed that the detection unit 154 detects recurring drift for daily KLD included in November 2018.

The generation unit 152 acquires operational data to be input to the trained model. Then, the generation unit 152 calculates, based on the acquired operational data, a first index indicating the magnitude of change in the output result of the model due to a temporal change in the trend of the operational data. Furthermore, the generation unit 152 calculates a second indicator by dividing the calculated first indicator into periodic unit times of the operational data, and compares the calculated second indicator with a preset threshold. Based on the comparison, it is determined whether a concept drift has occurred that needs to be addressed by the user. If it is determined that a concept drift that requires a user response has occurred, the generation unit 152 causes the display screen to display information indicating that a concept drift that requires a user response has occurred.

At this time, the display control unit 155 displays information indicating the concept drift that has occurred on the display screen, and displays concept drifts that the user needs to respond to and concept drifts that the user does not need to respond to in different ways. Display it on the screen.

Specifically, the generation unit 152 causes the display screen to display each index at a plurality of times in the form of a line graph. The generation unit 152 determines, in the line graph, a point corresponding to a time when a concept drift that requires a user response is detected to have occurred, and a point that corresponds to a point where a concept drift that requires a user response is detected to have occurred. A point corresponding to a time when no time is present is displayed so as to be different in at least one of color, shape, and size.

The display control unit 155 displays a message indicating that a recurring drift has occurred near the data point in November 2018 where a concept drift that needs to be addressed by the user is detected. On the other hand, the display control unit 155 displays data points before October 2018, in which no concept drift that the user needs to deal with is detected, as white circles.

The generation unit 152 generates a plurality of replica models by changing predetermined parameters of the original model, which is a model that calculates a score for classifying data into one of a plurality of classes. Of the time-series data, the detection unit 154 calculates the distribution of scores obtained by inputting the learning data during learning of the original model into the replica model, and the replica model data at a plurality of times after the learning time. Obtain the distribution of scores obtained by inputting them into the model. Further, the detection unit 154 detects that a concept drift, which the user needs to respond to, has occurred when the index indicating the magnitude of the difference in each distribution satisfies a predetermined condition at a constant cycle. For example, the detection unit 154 detects that a concept drift, which the user needs to deal with, has occurred when the index becomes equal to or greater than a threshold value at regular intervals.

As shown in FIG. 15, the display control unit 155 displays KLD along with the date and day of the week on the daily display screen. In the example of FIG. 15, the KLD on November 6th (Wednesday), November 13th (Wednesday), November 20th (Wednesday), and November 27th (Wednesday) is greater than or equal to the threshold value of 0.4. From this, the detection unit 154 detects that a concept drift that the user needs to respond to occurs periodically on Wednesdays every other week.

The display control unit 155 displays the data points of 11/6 (Wednesday), 11/13 (Wednesday), 11/20 (Wednesday), and 11/27 (Wednesday) where concept drifts that need to be addressed by the user are detected. Display a message nearby.

Furthermore, when the detection unit 154 detects that a concept drift that requires a user to respond has occurred, the learning unit 151 creates a new model using data at a time that satisfies a predetermined condition among the time-series data. Learn about. For example, in the example shown in FIG. 15, the learning unit 151 may learn a new model for Wednesday using only data from Wednesday every week. In other words, the learning unit 151 learns the operational data for Wednesday in November 2018 that the user needs to respond to.

Specifically, when the learning unit 151 detects that a concept drift that requires a user's response has occurred, the learning unit 151 displays information indicating that a concept drift that requires a user's response has occurred on the display screen of the display device. to be displayed. At this time, the learning unit 151 receives from the user whether or not to perform relearning. When receiving a request from the user to perform relearning, the learning unit 151 extracts the latest operational data at the time when a concept drift that needs to be addressed by the user occurs from among the time-series data. . Specifically, the learning unit 151 extracts operational data for the same time unit a preset period before, starting from the point in time when a concept drift that needs to be addressed by the user occurs. The learning unit 151 performs relearning using operational data from a preset period ago.

[Processing flow]
The flow of processing during learning of the detection device 10 will be explained using FIG. 16. FIG. 16 is a flowchart showing the flow of processing during learning of the detection device. As shown in FIG. 16, first, the detection device 10 generates an original model by learning learning data (step S101). Next, the detection device 10 generates a group of replica models in which predetermined parameters of the original model are changed (step S102). Then, the detection device 10 stores information regarding the distribution of outputs when learning data is input to the replica model group (step S103).

The flow of processing during learning of the detection device 10 will be explained using FIG. 17. FIG. 17 is a flowchart showing the flow of processing during learning of the detection device. As shown in FIG. 17, the detection device 10 first calculates the output when operational data is input to the original model (step S201). The detection device 10 can obtain an inference result based on the output calculated in step S201.

Next, the detection device 10 obtains the distribution of outputs when operational data is input to the replica model group (step S202). Then, the detection device 10 calculates an index for evaluating the difference between the distribution acquired during operation and the distribution acquired during learning (step S203). The distribution acquired during operation is the distribution acquired by the detection device 10 in step S202. On the other hand, the distribution acquired during learning is acquired by the detection device 10 in step S103 of FIG. 16. Also, for example, the indicator is KLD.

Here, the detection device 10 determines whether the index satisfies the condition (step S204). If the index satisfies the condition (step S204, Yes), the detection device 10 detects concept drift and outputs an alert (step S205). The alert may be a message as shown in FIG. 11, for example. Furthermore, when the index does not satisfy the condition (step S204, No), the detection device 10 outputs an inference result based on the output of the original model (step S206). Note that the conditions here are the conditions for detecting each type of concept drift described above.

[effect]
As described above, the generation unit 152 generates a plurality of replica models by changing predetermined parameters of the learned original model. The detection unit 154 determines the distribution of output obtained by inputting learning data used for learning the original model into the replica model, and the distribution of output obtained by inputting predetermined data different from the learning data into the replica model. Detect differences. In this way, if the original model has been generated, the detection device 10 can generate a replica model and detect concept drift without performing learning. As a result, the detection device 10 can detect concept drift with a small amount of calculation.

The detection unit 154 acquires a distribution of outputs obtained by inputting learning data, which is data during learning of the original model and used for learning the original model, out of the time-series data to the replica model. The detection unit 154 inputs data of a predetermined time after the learning time out of the time-series data to the replica model and acquires the distribution of outputs obtained. The detection unit 154 detects differences in each distribution. In this way, the detection device 10 can evaluate the difference in output distribution between time-series data during learning and during operation. As a result, the detection device 10 can detect concept drift in time series data with a small amount of calculation.

The generation unit 152 generates a plurality of replica models by changing predetermined parameters of the original model, which is a model that calculates a score for classifying data into one of a plurality of classes. The detection unit 154 calculates KLD (Kullback-Leibler divergence) between the distribution of scores obtained by inputting learning data into the replica model and the distribution of outputs obtained by inputting predetermined data different from the learning data into the replica model. Detect differences based on the In this way, the detection device 10 can easily evaluate differences between distributions by using KLD.

The generation unit 152 generates a replica model by multiplying a predetermined parameter by a random number that follows a normal distribution and has an average of 1.0 and a standard deviation of 0.3. In this way, the detection device 10 can prevent the difference between the original model and the replica model from becoming too large. As a result, the detection device 10 can reduce erroneous detection of concept drift.

The generation unit 152 generates a plurality of neural networks as replica models by changing the weights and biases in predetermined layers of the neural network that is the trained original model. As a result, the detection device 10 can generate a replica model when the original model is a neural network.

The generation unit 152 generates a plurality of SVMs as replica models by changing the weight and bias for each feature in the discriminant function of the SVM, which is the trained original model. As a result, the detection device 10 can generate a replica model when the original model is an SVM.

The generation unit 152 generates a plurality of logistic regression models as replica models by changing the weight and bias for each feature in the decision function of the trained original model, the logistic regression model. As a result, the detection device 10 can generate a replica model when the original model is a logistic regression model.

The generation unit 152 generates a plurality of decision trees as replica models by changing the information gain threshold of the decision tree that is the trained original model. As a result, the detection device 10 can generate a replica model when the original model is a decision tree.

When the detection unit 154 detects a difference, the learning unit 151 re-learns the original model using predetermined data. As a result, the detection device 10 can adapt the original model to concept drift.

(Effects related to detection of incremental drift)
The generation unit 152 generates a plurality of replica models by changing predetermined parameters of the original model, which is a model that calculates a score for classifying data into one of a plurality of classes. The detection unit 154 detects when an index indicating the size of the difference between the distribution of scores obtained by inputting learning data to the replica model and the distribution of outputs obtained by inputting operational data to the replica model is equal to or greater than a threshold value. , detects that a concept drift has occurred that the user needs to respond to. As a result, the detection device 10 can detect incremental drift with a small amount of calculation.

Among the time-series data, the detection unit 154 calculates the distribution of scores obtained by inputting learning data during learning of the original model into the replica model, and inputs each data at a plurality of times after the learning into the replica model. Obtain the distribution of scores obtained by inputting. The detection unit 154 detects that a concept drift, which the user needs to deal with, has occurred when the index indicating the magnitude of the difference between each distribution is equal to or greater than a threshold value. The display control unit 155 displays information regarding each index at a plurality of times on the screen. The display control unit 155 further displays information regarding an index of time when a concept drift to which the user needs to respond has been detected, and information regarding an index of time when a concept drift to which the user must respond has not been detected. Display information in a different manner. As a result, the detection device 10 can detect incremental drift in time-series data and present the detection results in an intuitive manner.

The display control unit 155 displays each index at a plurality of times in a line graph. The display control unit 155 places a point in the line graph corresponding to a time when a concept drift that requires a user's response is detected at a time when a concept drift that requires a user's response is not detected. The display is made to be different from the corresponding point in at least one of color, shape, and size. As a result, the detection device 10 can detect incremental drift in time-series data and present the detection results in an intuitive manner.

The display control unit 155 displays on the screen a histogram of the distribution of outputs obtained by inputting learning data into the replica model, and a histogram showing the distribution of outputs obtained by inputting data at a specified time into the replica model. let As a result, the detection device 10 can present the occurrence of concept drift from various aspects and facilitate the user's understanding.

When the detection unit 154 detects that a concept drift to which the user needs to respond has occurred, the learning unit 151 selects the most recent time-series data at the time when the concept drift to which the user needs to respond is detected. Retrain the original model using the data. As a result, the detection device 10 can adapt the original model to incremental drift.

(Effects related to detection of sudden drift)
The generation unit 152 generates a plurality of replica models by changing predetermined parameters of the original model, which is a model that calculates a score for classifying data into one of a plurality of classes. Among the time-series data, the detection unit 154 calculates the distribution of scores obtained by inputting learning data during learning of the original model into the replica model, and inputs each data at a plurality of times after the learning into the replica model. Obtain the distribution of scores obtained by inputting. The detection unit 154 detects that a concept drift, which the user needs to respond to, has occurred when a time-series change in the index indicating the magnitude of the difference in each distribution satisfies a predetermined condition. As a result, the detection device 10 can detect sudden drift in time series data with a small amount of calculation.

The detection unit 154 detects that a concept drift that requires a user to respond has occurred when the amount of change in the index over time is equal to or greater than a threshold value. As a result, the detection device 10 can detect sudden drift in time series data.

When the detection unit 154 detects that a concept drift to which the user needs to respond has occurred, the learning unit 151 selects the most recent time-series data at the time when the concept drift to which the user needs to respond is detected. Retrain the original model using the data. As a result, the detection device 10 can adapt the original model to sudden drift.

(Effects related to total value display)
The generation unit 152 generates a plurality of replica models by changing predetermined parameters of the original model, which is a model that calculates a score for classifying data into one of a plurality of classes. Among the time-series data, the detection unit 154 calculates the distribution of scores obtained by inputting learning data during learning of the original model into the replica model, and inputs each data at a plurality of times after the learning into the replica model. Obtain the distribution of scores obtained by inputting. The display control unit 155 displays on the screen the total value of the index indicating the magnitude of the difference in each distribution acquired by the detection unit 154 for each fixed period. As a result, the detection device 10 can detect the occurrence of a concept drift that requires a user to respond within a predetermined period with a small amount of calculation.

If the index or total value is equal to or greater than the threshold, the detection unit 154 detects that a concept drift that requires a user to respond has occurred. When a concept drift that requires a user response is detected, the display control unit 155 causes information regarding the concept drift to be displayed on the screen. As a result, the detection device 10 can detect the occurrence of concept drift that requires the user to respond in both a narrow time range and a wide time range.

The display control unit 155 switches between a screen that displays information regarding a plurality of time-based indicators and a screen that displays information regarding total values for each fixed period. As a result, the detection device 10 is able to present information regarding concept drift in both narrow and wide time ranges.

When the detection unit 154 detects that a concept drift to which the user needs to respond has occurred, the learning unit 151 selects the most recent time-series data at the time when the concept drift to which the user needs to respond is detected. Retrain the original model using the data. As a result, the detection device 10 can adapt the original model to concept drift.

(Effects related to detection of recurring drift)
The generation unit 152 generates a plurality of replica models by changing predetermined parameters of the original model, which is a model that calculates a score for classifying data into one of a plurality of classes. Among the time-series data, the detection unit 154 calculates the distribution of scores obtained by inputting learning data during learning of the original model into the replica model, and inputs each data at a plurality of times after the learning into the replica model. Obtain the distribution of scores obtained by inputting. The detection unit 154 detects that a concept drift, which the user needs to deal with, has occurred when the index indicating the magnitude of the difference in each distribution satisfies a predetermined condition at a constant period. As a result, the detection device 10 can detect recurring drift in time series data with a small amount of calculation.

The detection unit 154 detects that a concept drift, which the user needs to deal with, has occurred when the index becomes equal to or greater than the threshold value at regular intervals. As a result, the detection device 10 can detect recurring drift in time-series data.

When the detection unit 154 detects that a concept drift that requires a user response has occurred, the learning unit 151 learns a new model using data at a time that satisfies a predetermined condition among the time series data. conduct. As a result, the detection device 10 can adapt the new model to the recurring drift.

[Other Examples]
In the above embodiment, an example was explained in which the algorithms for the original model and the replica model were neural networks. On the other hand, the detection device 10 can detect concept drift even when the algorithm of each model is an algorithm other than a neural network.

For example, the algorithm may be SVM (Support Vector Machine). In this case, the generation unit 152 generates a plurality of SVMs as replica models by changing the weight and bias for each feature in the discriminant function of the SVM, which is the trained original model. Equation (3) is the discriminant function of SVM. The generation unit 152 generates a replica model by multiplying the weight a _i and bias b of the discriminant function by a random number. Similar to the neural network example described above, the generation unit 152 can multiply each parameter by a random number that follows a normal distribution with an average of 1.0 and a standard deviation of 0.3.

Similarly, the generation unit 152 can generate a plurality of logistic regression models as replica models by changing the weight and bias for each feature in the decision function of the trained original model, the logistic regression model. . Equation (4) is a determination function of logistic regression. The generation unit 152 generates a replica model by multiplying the weight w _i and bias b of the decision function by a random number. Similar to the neural network example described above, the generation unit 152 can multiply each parameter by a random number that follows a normal distribution with an average of 1.0 and a standard deviation of 0.3.

Further, the generation unit 152 can generate a plurality of decision trees as replica models by changing the information gain threshold of the decision tree, which is the trained original model. The generation unit 152 generates a replica model by multiplying the information gain threshold of each question by a random number. Similar to the neural network example described above, the generation unit 152 can multiply each parameter by a random number that follows a normal distribution with an average of 1.0 and a standard deviation of 0.3.

FIG. 18 is a diagram for explaining parameters of a decision tree. For example, the information gain threshold for the question "Is the temperature over 25 degrees Celsius?" is 25. The generation unit 152 uses values such as 25.1 and 24.5 obtained by multiplying 25 by random numbers to generate a decision tree as a replica model with questions such as "Is the temperature 25.1°C or higher?" and "Is the temperature 24.5°C or higher?" Can be done.

[system]
Information including processing procedures, control procedures, specific names, and various data and parameters shown in the above documents and drawings can be changed arbitrarily unless otherwise specified. Furthermore, the specific examples, distributions, numerical values, etc. described in the examples are merely examples, and can be changed arbitrarily.

Furthermore, each component of each device shown in the drawings is functionally conceptual, and does not necessarily need to be physically configured as shown in the drawings. That is, the specific form of distributing and integrating each device is not limited to what is shown in the drawings. In other words, all or part of them can be functionally or physically distributed and integrated into arbitrary units depending on various loads, usage conditions, and the like. Furthermore, all or any part of each processing function performed by each device can be realized by a CPU and a program that is analyzed and executed by the CPU, or can be realized as hardware using wired logic.

[hardware]
FIG. 19 is a diagram illustrating an example of a hardware configuration. As shown in FIG. 19, the detection device 10 includes a communication interface 10a, an HDD (Hard Disk Drive) 10b, a memory 10c, and a processor 10d. Furthermore, the parts shown in FIG. 19 are interconnected via a bus or the like.

The communication interface 10a is a network interface card or the like, and communicates with other servers. The HDD 10b stores programs and DB that operate the functions shown in FIG.

The processor 10d reads a program that executes the same processing as each processing unit shown in FIG. It is a hardware circuit. That is, this process executes the same functions as each processing unit included in the detection device 10. Specifically, the processor 10d reads a program having the same functions as the learning section 151, the generation section 152, the calculation section 153, the detection section 154, and the display control section 155 from the HDD 10b or the like. The processor 10d then executes a process similar to that of the learning unit 151, the generation unit 152, the calculation unit 153, the detection unit 154, the display control unit 155, and the like.

In this way, the detection device 10 operates as an information processing device that executes a detection method or a display method by reading and executing a program. Further, the detection device 10 can also realize the same functions as in the above-described embodiments by reading the program from a recording medium using a medium reading device and executing the read program. Note that the programs in other embodiments are not limited to being executed by the detection device 10. For example, the present invention can be similarly applied to cases where another computer or server executes a program, or where these computers or servers cooperate to execute a program.

This program can be distributed via a network such as the Internet. Additionally, this program is recorded on a computer-readable recording medium such as a hard disk, flexible disk (FD), CD-ROM, MO (Magneto-Optical disk), or DVD (Digital Versatile Disc), and is read from the recording medium by the computer. It can be executed by being read.

10 detection device 11 communication unit 12 input unit 13 output unit 14 storage unit 15 control unit 151 learning unit 152 generation unit 153 calculation unit 154 detection unit 155 display control unit

Claims (9)

Obtain operational data to be input to the trained model,
generating a plurality of second models in which predetermined parameters of the first model, which is the model that calculates a score for classifying data into one of a plurality of classes, are changed;
The distribution of scores obtained by inputting the learning data used for learning the first model into the second model, and the operational data, which is predetermined data different from the learning data, are input into the second model. Calculate an index that shows the size of the difference between the input and output distribution ,
Based on a comparison between the calculated index and a preset threshold, determine whether a concept drift that requires a user to respond has occurred;
A display method characterized in that a computer executes a process of displaying on a display screen information indicating that a concept drift that requires a user to respond has occurred when the index is equal to or greater than the threshold value . Displaying information indicating a concept drift that has occurred on the display screen, and displaying concept drift that the user needs to respond to and concept drift that the user does not need to respond to in different forms on the display screen. The display method according to claim 1, further comprising performing the following steps. Receive instructions from the user to perform relearning,
Based on instructions from the user, extract operational data for a preset period before the point in time when a concept drift that the user needs to respond to occurs,
Create a model based on relearning using the extracted operational data before the preset period,
The display method according to claim 1, further comprising: changing the model to a model based on the relearning. accept values related to acceptable model accuracy from the user;
The display method according to claim 1, further comprising: setting the predetermined threshold value based on the accepted value related to the accuracy of the acceptable model. The index is data during learning of the first model among time-series data, and is obtained by inputting the learning data used for learning the first model into the second model. and the distribution of output obtained by inputting the operational data, which is data for each of a plurality of times after the learning time, into the second model among the time-series data. an index indicating the magnitude of the index , if the index is equal to or greater than the threshold, detecting that a concept drift that the user needs to respond to has occurred;
Information regarding the index for each of the plurality of times is displayed on the screen, and information regarding the index for the time when concept drift has been detected is displayed for the time when concept drift has not been detected. The display method according to claim 1 , further comprising: displaying the information in a manner different from the information regarding the index. The indicators for each of a plurality of times are displayed in a line graph, and further, the user corresponds to a point on the line graph that corresponds to a time when a concept drift that needs to be addressed by the user is detected to have occurred. The display according to claim 1, characterized in that the display is made to differ in at least one of color, shape, and size from a point corresponding to a time when it is not detected that the necessary concept drift has occurred. Method. A histogram showing the distribution of the output obtained by inputting the learning data into the second model, and a histogram showing the distribution of the output obtained by inputting the data at a specified time into the second model are displayed on the screen. 2. The display method according to claim 1 , further comprising: displaying the image on the screen. Obtain operational data to be input to the trained model,
generating a plurality of second models in which predetermined parameters of the first model, which is the model that calculates a score for classifying data into one of a plurality of classes, are changed;
The distribution of scores obtained by inputting the learning data used for learning the first model into the second model, and the operational data, which is predetermined data different from the learning data, are input into the second model. Calculate an index that shows the size of the difference between the input and output distribution ,
Based on a comparison between the calculated index and a preset threshold, determine whether a concept drift that requires a user to respond has occurred;
A display program that causes a computer to execute a process of displaying on a display screen information indicating that a concept drift that requires a user response has occurred when the index is equal to or greater than the threshold value . an acquisition unit that acquires operational data to be input to the trained model;
a generation unit that generates a plurality of second models in which predetermined parameters of the first model, which is the model that calculates a score for classifying data into one of a plurality of classes, are changed;
The distribution of scores obtained by inputting the learning data used for learning the first model into the second model, and the operational data, which is predetermined data different from the learning data, are input into the second model. an index calculation unit that calculates an index indicating the magnitude of the difference between the input and the output distribution ;
a determination unit that determines whether a concept drift that requires a user to respond has occurred based on a comparison between the calculated index and a preset threshold;
a display control unit that displays on a display screen information indicating that a concept drift that requires a user to respond has occurred when the index is greater than or equal to the threshold ;
An information processing device comprising:

Images