JP7061089B2 - Classification device, classification method and classification program - Google Patents

Description

The present invention relates to a classification device, a classification method and a classification program.

Conventionally, there is known a transfer learning technique that applies a classification model for classifying events in the first domain to classification of events in a second domain different from the first domain. For example, a second domain sound classification model that classifies whether or not the sound is a predetermined environmental sound based on the time series data of the sound generated in the first household as the first domain. It is used to classify sounds generated in households.

In the transfer learning technique, the classification accuracy in the second domain can be improved by bringing the distribution of the feature amount in the first domain closer to the distribution of the feature amount in the second domain. For example, Non-Patent Document 1 proposes a transfer learning method that uses the Generative Adversarial Network (GAN) to bring the distribution of features between domains closer.

Eric Tzeng, et al. “Adversarial discriminative domain adaptation” The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 7167-7176

By the way, it is conceivable to apply the technique described in Non-Patent Document 1 to transfer learning between time series data corresponding to different domains. However, in the technique described in Non-Patent Document 1, when inputting time-series data corresponding to the second domain into the classification model of the first domain, it is considered to set the time scale of the time-series data to be input. Not. Therefore, with the technique described in Non-Patent Document 1, it is not possible to extract time-series data to be input on an appropriate time scale, and there is a possibility that transfer learning cannot be performed accurately.

Therefore, the present invention has been made in view of these points, and an object of the present invention is to provide a classification device, a classification method, and a classification program capable of accurately performing transfer learning between time-series data corresponding to different domains. And.

The classification device according to the first aspect of the present invention includes a first feature information acquisition unit that acquires first feature information, which is feature information indicating features of the first time series data, which is time series data of the first domain. A plurality of times from the time-series data acquisition unit that acquires the second time-series data that is the time-series data of the second domain different from the first domain and the second time-series data acquired by the time-series data acquisition unit. The second unit is based on the extraction unit that extracts a plurality of partial time-series data corresponding to each of the scales and the partial time-series data corresponding to any of the plurality of time scales. A plurality of generation units that generate second feature information, which is feature information indicating the characteristics of time-series data, and a plurality of identification units corresponding to each of the plurality of generation units, the first feature information or the correspondence. When the second feature information generated by the generation unit is input, a plurality of identification units for identifying whether the input feature information is the first feature information or the second feature information, and the above-mentioned identification unit. Learning to learn the plurality of identification units so that the identification accuracy in the identification unit is improved, and to learn the plurality of generation units so that the generated second feature information is similar to the first feature information. Each of the unit and the plurality of second feature information generated by the plurality of generation units is input to the classifier for classifying the first feature information, and the result obtained from the classifier and the learning are performed by the learning unit. It is provided with a classification result acquisition unit for acquiring the classification result corresponding to the second time-series data based on the identification result of each of the plurality of identification units after the data is generated.

The classification device is the contribution of each of the plurality of second feature information corresponding to each of the plurality of time scales based on the discrimination results of the plurality of identification units after learning by the learning unit. Further, the classification result acquisition unit is provided with a determination unit for determining the degree of contribution indicating the degree of influence on the classification result of the second time series data, and the classification result acquisition unit is a plurality of second feature information generated by the plurality of generation units. By synthesizing the results obtained from the classifier when each of them is input to the classifier based on the contribution degree determined by the determination unit, the classification result corresponding to the second time series data is acquired. May be good.

The classification device is the contribution of each of the plurality of second feature information corresponding to each of the plurality of time scales based on the discrimination results of the plurality of identification units after learning by the learning unit. Further, the classification result acquisition unit is provided with a determination unit for determining the degree of contribution indicating the degree of influence on the classification result of the second time-series data, and the classification result acquisition unit is a plurality of second feature information generated by the plurality of generation units. Among the classification results obtained from the classifier when each is input to the classifier, the second time series is based on the classification result corresponding to one or more second feature information having a relatively high contribution. The classification result corresponding to the data may be acquired.

The classification device is based on the discrimination results of the plurality of identification units until the learning by the learning unit is completed, and the contribution of each of the plurality of second feature information corresponding to each of the plurality of time scales. Further, a determination unit for determining the contribution degree is further provided by updating the contribution degree indicating the degree of influence on the classification result of the second time series data, and the classification result acquisition unit is the plurality of said. When each of the plurality of second feature information generated by the generation unit is input to the classifier, the result obtained from the classifier may be synthesized based on the contribution.

The classification device is based on the discrimination results of the plurality of identification units until the learning by the learning unit is completed, and the contribution of each of the plurality of second feature information corresponding to each of the plurality of time scales. Further, a determination unit for determining the contribution degree is further provided by updating the contribution degree indicating the degree of influence on the classification result of the second time series data, and the classification result acquisition unit is the plurality of said. Corresponds to one or more second feature information having a relatively high contribution among the classification results obtained from the classifier when each of the plurality of second feature information generated by the generation unit is input to the classifier. The classification result corresponding to the second time series data may be acquired based on the classification result.

The learning unit determines the number of times of learning between the generation unit that generates the second feature information having a relatively high contribution degree determined by the determination unit and the identification unit corresponding to the generation unit. The number of times of learning between the generation unit that generates relatively low second feature information and the identification unit corresponding to the generation unit may be increased.

The classification method according to the second aspect of the present invention includes a step of acquiring first feature information, which is feature information indicating features of the first time series data, which is time series data of the first domain, executed by a computer. From the step of acquiring the second time-series data which is the time-series data of the second domain different from the first domain and the acquired second time-series data, a plurality of partial times corresponding to each of the plurality of time scales. A step of extracting series data, a step of generating a plurality of second feature information which is feature information indicating the feature of the second time series data based on the partial time series data corresponding to each of the plurality of time scales, and a step of generating a plurality of second feature information. When the first feature information or the generated second feature information is input, the step of identifying whether the input feature information is the first feature information or the second feature information, and the step. A step of learning so as to improve the identification accuracy in the identification step, a step of learning so that the second feature information generated in the generation step is similar to the first feature information, and the plurality of steps. Each of the plurality of second feature information corresponding to each of the time scales is input to the classifier for classifying the first feature information, and the result obtained from the classifier and the plurality of the training after the learning is performed. A step of acquiring a classification result corresponding to the second time-series data based on the identification result of a plurality of second feature information corresponding to each of the time scales is provided.

The classification program according to the third aspect of the present invention is the first feature information acquisition in which the computer acquires the first feature information which is the feature information indicating the feature of the first time series data which is the time series data of the first domain. A plurality of units, a time-series data acquisition unit that acquires second time-series data that is time-series data of a second domain different from the first domain, and a plurality of second time-series data acquired by the time-series data acquisition unit. An extraction unit that extracts a plurality of partial time-series data corresponding to each of the time scales, the second one based on the partial time-series data corresponding to any of the plurality of time scales and corresponding to any one of the time scales. A plurality of generation units that generate second feature information, which is feature information indicating the characteristics of time-series data, and a plurality of identification units corresponding to each of the plurality of generation units, which are the first feature information or the corresponding ones. When the second feature information generated by the generation unit is input, a plurality of identification units for identifying whether the input feature information is the first feature information or the second feature information, the identification unit. A learning unit that learns the plurality of identification units so that the identification accuracy in the above-mentioned is improved, and learns the plurality of generation units so that the generated second feature information resembles the first feature information. Then, each of the plurality of second feature information generated by the plurality of generation units is input to the classifier for classifying the first feature information, and the result obtained from the classifier and the learning are performed by the learning unit. It functions as a classification result acquisition unit that acquires the classification result corresponding to the second time-series data based on the identification result of each of the plurality of identification units.

According to the present invention, there is an effect that transfer learning between time-series data corresponding to different domains can be performed with high accuracy.

It is a figure explaining the outline (the 1) of the classification apparatus which concerns on 1st Embodiment. It is a figure explaining the outline (the 2) of the classification apparatus which concerns on 1st Embodiment. It is a figure which shows the structure of the classification apparatus which concerns on 1st Embodiment. It is a flowchart which shows the process flow of the classification apparatus which concerns on 1st Embodiment. It is a figure which shows the relationship between the generation part and the identification part which concerns on 3rd Embodiment.

<First Embodiment>
[Overview of classification device]
1 and 2 are diagrams illustrating an outline of the classification device according to the first embodiment. The classification device is a computer that performs transfer learning to classify time-series data in a second domain different from the first domain by using a classification model of the first domain that classifies time-series data in the first domain.

The classification device acquires the first feature information which is the feature information indicating the feature of the first time series data which is the time series data of the first domain. The classification device acquires the second time-series data, which is the time-series data of the second domain. The time-series data is, for example, data in which a plurality of times are associated with a value indicating a predetermined event at the time, and is, for example, voice data in a house. The time-series data is not limited to voice data, but other data can be used as long as it is data in which a plurality of times and a value indicating a predetermined event at the time are associated with each other. There may be.

The classification device extracts a plurality of partial time-series data corresponding to each of the plurality of time scales from the acquired second time-series data. In the example shown in FIG. 1, the classification device extracts m partial time series data D1 to Dm. The classification device includes a plurality of generators that generate second feature information, which is feature information indicating the features of the second time series data, based on the partial time series data. The classification device generates the second feature information corresponding to each of the plurality of time scales by the plurality of generation units. In the example shown in FIG. 1, the classification device generates m second feature information FI1 to FIm corresponding to each of m partial time series data D1 to Dm.

The classification device includes a plurality of identification units corresponding to each of the plurality of generation units. When the first feature information or the second feature information is input, each of the plurality of identification units outputs an identification result that identifies whether the input feature information is the first feature information or the second feature information. do.

The classification device learns a plurality of identification units so that the identification accuracy is improved, and learns a plurality of generation units so that the generated second feature information resembles the first feature information. The classification device has a plurality of second feature information FI1 to corresponding to each of the plurality of time scales based on the discrimination results of the plurality of identification units after learning the plurality of identification units and the plurality of generation units. Each contribution degree P1 to Pm of FIm is calculated. The degree of contribution is information indicating the degree to which the second feature information of each of the plurality of time scales contains information that affects the classification result of the second time series data.

Moving on to FIG. 2, the classifier inputs a plurality of second feature information FI1 to FIm generated from a plurality of time scales into a classifier that classifies the first feature information in the first domain. The classification device synthesizes the classification results y1 to ym corresponding to each of the plurality of second feature information FI1 to FIm output from the classifier based on the contribution P1 to Pm, thereby performing the second time series data. The classification result y corresponding to is acquired. In the classification device, the classification result of the second feature information corresponding to the relatively high contribution corresponds to the second time series data as compared with the second feature information corresponding to the relatively low contribution. The classification results corresponding to each of the plurality of second feature information FI1 to FIm are synthesized so as to affect the classification results. By doing so, the classification device utilizes the knowledge of the first domain with the label by performing transfer learning even for the second time series data without the label, and the time series data. Classification of can be realized. Further, the classification device can strongly reflect the classification result of the second feature information corresponding to the appropriate time scale to the classification result corresponding to the second time series data. Therefore, the classification device can accurately perform transfer learning between time series data corresponding to different domains.
Hereinafter, the configuration of the classification device will be described.

[Configuration example of classification device 1]
FIG. 3 is a diagram showing the configuration of the classification device 1 according to the first embodiment. The classification device 1 includes a storage unit 11 and a control unit 12.

The storage unit 11 is, for example, a ROM (Read Only Memory), a RAM (Random Access Memory), or the like. The storage unit 11 stores various programs for operating the classification device 1. For example, the storage unit 11 uses the control unit 12 of the classification device 1 as a first feature information acquisition unit 121, a time series data acquisition unit 122, an extraction unit 123, a plurality of generation units 124, a plurality of identification units 125, and a learning unit 126. , The determination unit 127, and the classification program to function as the classification result acquisition unit 128.

The control unit 12 is, for example, a CPU (Central Processing Unit). The control unit 12 controls the function related to the classification device 1 by executing various programs stored in the storage unit 11. By executing the program stored in the storage unit 11, the control unit 12 executes the first feature information acquisition unit 121, the time series data acquisition unit 122, the extraction unit 123, the generation unit 124, the identification unit 125, and the learning unit 126. , The determination unit 127, and the classification result acquisition unit 128.

The first feature information acquisition unit 121 acquires the first feature information indicating the features of the first time series data. For example, a plurality of first feature information is stored in advance in the storage unit 11 or an external device. When the first feature information acquisition unit 121 receives the acquisition operation of the first feature information via the operation unit (not shown) that receives the operation from the user, the first feature information acquisition unit 121 is stored in the storage unit 11 or the external device. Get multiple pieces of information.

Here, the first feature information is a data set including a plurality of partial time-series data extracted from the first time-series data corresponding to each of the plurality of time scales. Further, the first feature information includes a plurality of partial time series data so that the classifier contains a lot of information indicating the features of the first time series data when the classifier classifies the first time series data with high accuracy. Is preset. Further, since the time scales of the plurality of partial time series data are different, it is assumed that the data is adjusted to the reference sampling scale, for example. Here, it is assumed that a predetermined label is attached to the first time series data in advance. For example, when the first time-series data is time-series data indicating environmental sounds in a predetermined home, it is assumed that the first time-series data is labeled with cleaning sounds, cooking sounds, and the like.

The time-series data acquisition unit 122 acquires one or more second time-series data which is time-series data of the second domain different from the first domain. For example, when the first time-series data is time-series data indicating an environmental sound in a predetermined home, the time-series data acquisition unit 122 obtains time-series data indicating an environmental sound in a home different from the predetermined home at the second time. Acquire as series data. Here, it is assumed that the second time-series data is not labeled with a predetermined label like the first time-series data.

The extraction unit 123 extracts a plurality of partial time-series data corresponding to each of the plurality of time scales from the second time-series data acquired by the time-series data acquisition unit 122. The extraction unit 123 may extract the partial time-series data so that a part of the partial time-series data overlaps with the other partial time-series data. Further, the extraction unit 123 may extract the partial time-series data so that a part of the partial time-series data does not overlap with the other partial time-series data.

Here, when the extraction unit 123 extracts the partial time-series data corresponding to each of the m time scales, the partial time-series data corresponding to the m-th time scale TSm is referred to as the partial time-series data Dm. ..

The plurality of generation units 124 correspond to any of the plurality of time scales, and generate second feature information indicating the features of the second time series data based on the partial time series data corresponding to the one of the time scales. .. Here, when the extraction unit 123 extracts partial time-series data corresponding to each of the m time scales, the mth generation unit 124 is referred to as the generation unit 124-m. Further, the second feature information generated by the third generation unit 124 is referred to as a second feature information FIm.

The plurality of identification units 125 correspond to each of the plurality of generation units 124. When the extraction unit 123 extracts partial time-series data corresponding to each of the m time scales, the m-th identification unit 125 is referred to as the identification unit 125-m. When the first feature information or the second feature information generated by the generation unit 124 corresponding to itself is input to each of the plurality of identification units 125, whether the input feature information is the first feature information or the first feature information. 2 Identify whether it is feature information and output information indicating the identification result.

The learning unit 126 learns a plurality of identification units 125 so that the identification accuracy in the identification unit 125 is improved, and at the same time, a plurality of generation units 124 so that the generated second feature information resembles the first feature information. To learn. For example, the learning unit 126 alternately performs learning of the plurality of identification units 125 and learning of the plurality of generation units 124. The learning unit 126 learns the plurality of identification units 125 and the plurality of generation units 124 until the amount of change in the binary cross entropy based on the identification results output by each of the plurality of identification units 125 converges within a predetermined amount. Repeat learning.

The determination unit 127 determines the contribution of each of the plurality of second feature information corresponding to each of the plurality of time scales based on the identification results of the plurality of identification units 125 after the learning by the learning unit 126. Specifically, the determination unit 127 has the first feature information and the second identification information generated by the generation unit 124 corresponding to the identification unit 125 for each of the plurality of identification units 125 after learning by the learning unit 126. By inputting multiple times, multiple identification results are acquired. The determination unit 127 calculates the binary cross entropy for each of the plurality of identification units 125 based on the plurality of identification results.

By learning the plurality of generation units 124, each of the plurality of generation units 124 generates the second feature information similar to the first feature information, but the first is according to the information contained in the partial time series data. 1 The degree of similarity to the feature information is different. As a result, the correct answer rates of the identification results in the plurality of identification units 125 are also different. The larger the binary cross entropy, the more similar the second feature information generated by the generation unit 124 to the first feature information.

The determination unit 127 determines the contribution of each of the plurality of second feature information so that the larger the binary cross entropy calculated based on the identification result of each of the plurality of identification units 125, the higher the contribution. ..

Assuming that the contribution corresponding to the second feature information FIi (however, 1 ≦ i ≦ m) is Pi and the binary cross entropy corresponding to the identification unit 125-i is Li, the determination unit 127 is the second feature information FIi. The contribution degree Pi corresponding to is calculated by the following equation (1).
Pi = Li / (L1 + L2 +… + Lm) ・ ・ ・ (1)

The determination unit 127 calculated the degree of contribution based on the above equation (1), but the present invention is not limited to this. For example, the determination unit 127 may calculate the degree of contribution by the following equation (2). Note that exp (x) is the x-th power of the exponential function e.
Pi = exp (Li) / (exp (L1) + exp (L2) + ・ ・ ・ + exp (Lm)) ・ ・ ・ (2)

The classification result acquisition unit 128 causes each of the plurality of generation units 124 to generate the second feature information after the learning is performed by the learning unit 126. The classification result acquisition unit 128 inputs each of the second feature information generated by the plurality of generation units 124 into the classifier of the first domain that classifies the first feature information, and classifies the result from the classifier of the first domain. To get.

Then, the classification result acquisition unit 128 is based on the classification result acquired from the classifier of the first domain and the identification result of each of the plurality of identification units 125 after learning is performed by the learning unit 126. Acquire the classification result corresponding to the time series data. Specifically, the classification result acquisition unit 128 inputs each of the plurality of second feature information generated by the plurality of generation units 124 into the classifier of the first domain as one data set. Then, the classification result acquisition unit 128 determines the classification results of the plurality of second feature information included in the one data set obtained from the classifier of the first domain based on the contribution degree determined by the determination unit 127. By synthesizing, the classification result corresponding to the second time series data is acquired.

For example, when the classification result obtained from the classifier by inputting the second feature information FIi into the classifier of the first domain is set to yi, the classification result acquisition unit 128 uses the following formula (3) to obtain the second. The classification result y corresponding to the time series data is calculated.
y = y1 * P1 + y2 * P2 +… + ym * Pm ・ ・ ・ (3)

The classification result acquisition unit 128 calculates the classification result y based on the above equation (3), but the classification result y is not limited to this. The classification result acquisition unit 128 has a relative contribution among the classification results obtained from the classifier when each of the plurality of second feature information generated by the plurality of generation units 124 is input to the classifier of the first domain. The classification result corresponding to the second time series data may be acquired based on the classification result corresponding to one or more second feature information which is relatively high.

For example, the classification result acquisition unit 128 secondly obtains the classification result corresponding to the second feature information having the highest contribution among the classification results of each of the plurality of second feature information obtained from the classifier of the first domain. It may be a classification result corresponding to time series data. Further, the classification result acquisition unit 128 inputs only the second feature information having the highest contribution to the classifier of the first domain, and the classification result obtained from the classification result is used as the classification result corresponding to the second time series data. May be good. By doing so, it is possible to acquire the classification result corresponding to the second time series data while reducing the processing load of the classification device 1.

[Processing flow in classification device 1]
Subsequently, an example of the processing flow in the classification device 1 will be described. FIG. 4 is a flowchart showing the flow of processing according to the classification device 1 according to the present embodiment.

First, the first feature information acquisition unit 121 acquires the first feature information indicating the features of the first time-series data (S1).
Subsequently, the time-series data acquisition unit 122 acquires the second time-series data, which is the time-series data of the second domain (S2).

Subsequently, the extraction unit 123 extracts a plurality of partial time-series data corresponding to each of the plurality of time scales from the second time-series data acquired by the time-series data acquisition unit 122 in S2 (S3).

Subsequently, the plurality of generation units 124 generate second feature information indicating the features of the second time series data based on the partial time series data extracted on the time scale corresponding to the plurality of generation units 124 (S4).
Further, the learning unit 126 inputs the second feature information or the first feature information generated by the generation unit 124 to the identification unit 125, and whether the feature information input to the identification unit 125 is the first feature information or not. 2 It is made to identify whether it is characteristic information (S5).
Subsequently, the learning unit 126 learns the plurality of identification units 125 and also learns the plurality of generation units 124 (S6). The processes S4 to S6 are repeatedly executed, for example, until the amount of change in the binary cross entropy corresponding to each of the plurality of identification units 125 converges within a predetermined amount.

Subsequently, the determination unit 127 repeats the processes of S4 to S6, learns the plurality of generation units 124 and the plurality of identification units 125, and then, based on the identification results of each of the plurality of identification units 125, the determination unit 127. The contribution of each of the plurality of second feature information corresponding to each of the plurality of time scales is determined (S7).

Subsequently, the classification result acquisition unit 128 causes the plurality of generation units 124 to generate the second feature information after the learning of the plurality of generation units 124 and the plurality of identification units 125 is performed (S8).
Subsequently, the classification result acquisition unit 128 acquires the classification results of the plurality of second feature information by inputting the plurality of second feature information generated by the plurality of generation units 124 into the classifier of the first domain. (S9).
Subsequently, the classification result acquisition unit 128 acquires the classification result corresponding to the second time-series data by synthesizing the classification results of the acquired plurality of second feature information based on the contribution degree (S10).

[Effect in the first embodiment]
As described above, the classification device 1 according to the first embodiment is based on a plurality of partial time-series data extracted from the second time-series data of the second domain corresponding to each of the plurality of time scales. A plurality of generation units 124 that generate second feature information indicating the features of the time series data, first feature information showing the features of the first time series data of the first domain, or the generated second feature information are input. And a plurality of identification units 125 for identifying whether the input feature information is the first feature information or the second feature information. Further, the classification device 1 inputs each of the learning unit 126 for learning between the plurality of generation units 124 and the plurality of identification units 125 and the plurality of second feature information into the classifier for classifying the first feature information. The classification result acquisition unit 128 that acquires the classification result corresponding to the second time-series data based on the result obtained from the classifier and the identification result of each of the plurality of identification units 125 after learning is provided. ..

By doing so, the classification device 1 can strongly reflect the classification result of the second feature information corresponding to the appropriate time scale to the classification result corresponding to the second time series data. As a result, the classification device 1 can accurately perform transfer learning between time-series data corresponding to different domains.

<Second Embodiment>
[Determine the degree of contribution while learning the generation unit 124 and the identification unit 125]
Subsequently, the second embodiment will be described. The classification device 1 according to the second embodiment is different from the first embodiment in that the contribution degree is determined while learning the generation unit 124 and the identification unit 125. Hereinafter, the classification device 1 according to the second embodiment will be described. The same parts as those of the first embodiment will be omitted as appropriate.

In the second embodiment, the determination unit 127 is an identification output by each of the plurality of identification units 125 from the start to the end of learning of the plurality of generation units 124 and the plurality of identification units 125 by the learning unit 126. Based on the result, the contribution is determined by updating the contribution of each of the plurality of second feature information corresponding to each of the plurality of time scales.

Specifically, the determination unit 127 is binary based on the relationship between the identification result of the plurality of identification units 125 and the correct answer during the learning of the plurality of generation units 124 and the plurality of identification units 125 by the learning unit 126. Calculate the cross entropy. The determination unit 127 calculates the contribution degree based on the binary cross entropy of the plurality of identification units 125 during learning. The determination unit 127 calculates the degree of contribution based on, for example, the formula (1) or the formula (2) shown in the first embodiment.

The learning unit 126 determines the ratio of inputting the plurality of second feature information generated by the plurality of generation units 124 into the identification unit 125 based on the calculated contribution degree, thereby determining the plurality of generation units 124 and the plurality of identification units. The number of learnings of each part 125 is made different. For example, the learning unit 126 determines the number of times of learning between the generation unit 124 that generates the second feature information with a relatively high contribution determined by the determination unit 127 and the identification unit 125 corresponding to the generation unit 124. The number of times of learning between the generation unit 124 that generates the second feature information having a relatively low value and the identification unit 125 corresponding to the generation unit 124 is increased. By doing so, it is possible to enhance the learning of the generation unit 124 and the identification unit 125 corresponding to the second feature information of the time scale, which has a high degree of contribution and easily affects the classification of the second time series data.

Similar to the first embodiment, the classification result acquisition unit 128 uses each of the plurality of second feature information generated by the plurality of generation units 124 as a classifier of the first domain after learning is performed by the learning unit 126. The results obtained from the classifier when input are combined based on the degree of contribution. For example, the classification result acquisition unit 128 calculates the classification result corresponding to the second time series data by the formula (3) shown in the first embodiment.

The classification result acquisition unit 128 calculates the classification result based on the formula (3) shown in the first embodiment, but the classification result is not limited to this. Similar to the first embodiment, the classification result acquisition unit 128 is a classification obtained from the classifier when each of the plurality of second feature information generated by the plurality of generation units 124 is input to the classifier of the first domain. Among the results, the classification result corresponding to the second time series data may be acquired based on the classification result corresponding to one or more second feature information having a relatively high contribution.

[Effect in the second embodiment]
As described above, the classification device 1 according to the second embodiment corresponds to each of the plurality of time scales based on the discrimination results of the plurality of identification units 125 until the learning by the learning unit 126 is completed. The contribution is determined by updating the contribution of each of the plurality of second feature information. Then, the classification device 1 synthesizes the results obtained from the classifier when each of the plurality of second feature information is input to the classifier based on the degree of contribution. By doing so, the classification device 1 according to the second embodiment can accurately perform transfer learning between time-series data corresponding to different domains, as in the first embodiment.

<Third Embodiment>
[The second feature information corresponding to each of the plurality of time scales is identified by one identification unit 125]
Subsequently, the third embodiment will be described. In the first embodiment and the second embodiment, a plurality of generation units 124 and identification units 125 are provided corresponding to each of the plurality of time scales, and learning of these generation units 124 and identification unit 125 is performed. However, since the number and variety of the second feature information corresponding to one time scale is smaller than the number and variety of the second feature information corresponding to each of the plurality of time scales, when learning is performed. In addition, the identification accuracy of the identification unit 125 may be higher than the generation accuracy of the second feature information in the generation unit 124. On the other hand, the classification device 1 according to the third embodiment includes only one identification unit 125, and the second feature information corresponding to each of the plurality of time scales is identified by one identification unit 125. It is different from the first embodiment. Hereinafter, the classification device 1 according to the third embodiment will be described. The same parts as those of the first embodiment will be omitted as appropriate.

FIG. 5 is a diagram showing the relationship between the generation unit 124 and the identification unit 125 according to the third embodiment. As shown in FIG. 5, the classification device 1 according to the third embodiment includes a plurality of generation units 124 and one identification unit 125.

When the identification unit 125 receives the input of the first feature information or the second feature information corresponding to each of the plurality of time scales generated by each of the plurality of generation units 124, the input feature information becomes the first feature. Identify whether it is information or second feature information.

The learning unit 126 learns the identification unit 125 so that the identification accuracy in the identification unit 125 is improved, and learns a plurality of generation units 124 so that the generated second feature information resembles the first feature information. I do. Since the learning of the identification unit 125 is performed based on the characteristic information of all the time scales, it is influenced by the characteristic information of all the time scales. Therefore, as compared with the case where the identification unit 125 is learned based on the feature information of one time scale, the improvement in accuracy due to the learning of the identification unit 125 is suppressed, and the state of overfitting is achieved as compared with the generation unit 124. Can be suppressed.

Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments, and various modifications and changes can be made within the scope of the gist. be. For example, all or part of the device can be functionally or physically distributed / integrated in any unit. Also included in the embodiments of the present invention are new embodiments resulting from any combination of the plurality of embodiments. The effect of the new embodiment produced by the combination has the effect of the original embodiment together.

1 ... Classification device, 11 ... Storage unit, 12 ... Control unit, 121 ... First feature information acquisition unit, 122 ... Time series data acquisition unit, 123 ... Extraction unit, 124 ... Generation unit, 125 ... Identification unit, 126 ... Learning unit, 127 ... Decision unit, 128 ... Classification result acquisition unit

Claims (8)

The first feature information acquisition unit that acquires the first feature information, which is the feature information indicating the features of the first time series data, which is the time series data of the first domain, and
A time-series data acquisition unit that acquires second time-series data, which is time-series data of a second domain different from the first domain.
An extraction unit that extracts a plurality of partial time-series data corresponding to each of a plurality of time scales from the second time-series data acquired by the time-series data acquisition unit.
A feature of a plurality of generation units corresponding to each of the plurality of time scales, showing the characteristics of the second time series data based on the partial time series data extracted by the extraction unit with respect to the corresponding time scale. A plurality of generators that generate second feature information, which is information,
When the first feature information or the second feature information generated by the corresponding generation unit is input, the input feature information is the plurality of identification units corresponding to each of the plurality of generation units. A plurality of identification units for discriminating between the first feature information and the second feature information,
The plurality of identification units are learned so that the identification accuracy in the identification unit is improved, and the plurality of generation units are learned so that the generated second feature information is similar to the first feature information. With the learning department
Each of the plurality of second feature information generated by the plurality of generation units is input to a classifier that classifies the first feature information, and the result obtained from the classifier and the learning performed by the learning unit are performed. A classification result acquisition unit that acquires a classification result corresponding to the second time-series data based on the identification result of each of the plurality of identification units.
A classification device equipped with. Based on the discrimination results of the plurality of identification units after learning by the learning unit, the contribution of each of the plurality of second feature information corresponding to each of the plurality of time scales is the second time. It also has a decision-making part that determines the degree of contribution that indicates the degree of influence on the classification result of the series data.
The classification result acquisition unit transfers the result obtained from the classifier when each of the plurality of second feature information generated by the plurality of generation units is input to the classifier to the contribution degree determined by the determination unit. By synthesizing based on the above, the classification result corresponding to the second time series data is acquired.
The classification device according to claim 1. Based on the discrimination results of the plurality of identification units after learning by the learning unit, the contribution of each of the plurality of second feature information corresponding to each of the plurality of time scales is the second time. It also has a decision-making part that determines the degree of contribution that indicates the degree of influence on the classification result of the series data.
The classification result acquisition unit has a relative degree of contribution among the classification results obtained from the classifier when each of the plurality of second feature information generated by the plurality of generation units is input to the classifier. Based on the classification result corresponding to one or more high second feature information, the classification result corresponding to the second time series data is acquired.
The classification device according to claim 1. Based on the discrimination results of the plurality of identification units until the learning by the learning unit is completed, the contribution of each of the plurality of second feature information corresponding to each of the plurality of time scales is described above. By updating the contribution degree indicating the degree of influence on the classification result of the second time series data, a determination unit for determining the contribution degree is further provided.
The classification result acquisition unit synthesizes the results obtained from the classifier when each of the plurality of second feature information generated by the plurality of generation units is input to the classifier based on the contribution degree.
The classification device according to claim 1. Based on the discrimination results of the plurality of identification units until the learning by the learning unit is completed, the contribution of each of the plurality of second feature information corresponding to each of the plurality of time scales is described above. By updating the contribution degree indicating the degree of influence on the classification result of the second time series data, a determination unit for determining the contribution degree is further provided.
The classification result acquisition unit has a relative degree of contribution among the classification results obtained from the classifier when each of the plurality of second feature information generated by the plurality of generation units is input to the classifier. Based on the classification result corresponding to one or more high second feature information, the classification result corresponding to the second time series data is acquired.
The classification device according to claim 1. The learning unit determines the number of times of learning between the generation unit that generates the second feature information having a relatively high contribution degree determined by the determination unit and the identification unit corresponding to the generation unit. The number of times of learning between the generation unit that generates relatively low second feature information and the identification unit corresponding to the generation unit is increased.
The classification device according to claim 4. Computer runs,
The step of acquiring the first feature information, which is the feature information indicating the features of the first time series data, which is the time series data of the first domain, and
The step of acquiring the second time-series data, which is the time-series data of the second domain different from the first domain,
A step of extracting a plurality of partial time series data corresponding to each of a plurality of time scales from the acquired second time series data, and
A step of generating a plurality of second feature information, which is feature information indicating the features of the second time series data, based on the partial time series data corresponding to each of the plurality of time scales.
When the first feature information or the generated second feature information is input, a step of identifying whether the input feature information is the first feature information or the second feature information,
A step of learning so as to improve the identification accuracy in the identification step,
A step of learning so that the second feature information generated in the generation step is similar to the first feature information,
Each of the plurality of second feature information corresponding to each of the plurality of time scales is input to the classifier that classifies the first feature information, and the result obtained from the classifier and the learning are performed. A step of acquiring a classification result corresponding to the second time series data based on the identification result of a plurality of second feature information corresponding to each of the plurality of time scales, and a step of acquiring the classification result.
Classification method with. Computer,
The first feature information acquisition unit, which acquires the first feature information which is the feature information indicating the feature of the first time series data which is the time series data of the first domain.
A time-series data acquisition unit that acquires a second time-series data that is time-series data of a second domain different from the first domain.
An extraction unit that extracts a plurality of partial time-series data corresponding to each of a plurality of time scales from the second time-series data acquired by the time-series data acquisition unit.
A plurality of generation units corresponding to each of the plurality of time scales, and the characteristics of the second time series data are shown based on the partial time series data extracted by the extraction unit with respect to the corresponding time scales. Multiple generators that generate the second feature information, which is the feature information,
When the first feature information or the second feature information generated by the corresponding generation unit is input, the input feature information is the plurality of identification units corresponding to each of the plurality of generation units. A plurality of identification units for discriminating between the first feature information and the second feature information,
The plurality of identification units are learned so that the identification accuracy in the identification unit is improved, and the plurality of generation units are learned so that the generated second feature information is similar to the first feature information. Learning department and
Each of the plurality of second feature information generated by the plurality of generation units is input to a classifier that classifies the first feature information, and the result obtained from the classifier and the learning performed by the learning unit are performed. A classification result acquisition unit that acquires a classification result corresponding to the second time-series data based on the identification result of each of the plurality of identification units.
A classification program that functions as.

Images