Information processing device, information processing method, and information processing program

Abstract

In order to make it possible to improve classification accuracy even when there is a class imbalance, an information processing device (1) comprises: an acquisition means (11) that acquires series data; and a classification means (12) that performs a classification process which is for classifying the series data into any of a plurality of classes and which uses a score representing the certainty of classification, wherein the classification means has been trained using a loss function in accordance with a frequency distribution of the score that is obtained from a score series, which is a series of the score.

Description

[Technical Field]

[0001] This invention relates to an information processing device, an information processing method, and an information processing program. [Background technology]

[0002] Techniques are known for referring to sequential data and classifying that sequential data into one of several classes. For example, Patent Document 1 discloses an identification device that identifies which of several classes, such as pre-registered people or other objects, a set of time-series data, as an example of sequential data, belongs to. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2009-245314 [Overview of the project] [Problems that the invention aims to solve]

[0004] In general, classification processes that classify sequential data often suffer from class imbalance, where many sequential data are classified into certain classes, while others are hardly classified at all. In the classification device described in Patent Document 1, there is a problem in the device's learning process where, if class imbalance exists in the training data, learning to classify data into minority classes is difficult. As a result, there is a problem in that the accuracy of the classification process is difficult to improve.

[0005] One aspect of the present invention has been made in view of the above-mentioned problems, and one example of its objective is to provide a technology that can perform efficient learning and improve classification accuracy even when class imbalance exists in the training data. [Means for solving the problem]

[0006] An information processing device according to one aspect of the present invention comprises an acquisition means for acquiring sequential data, and a classification means for classifying the sequential data into one of a plurality of classes, wherein the classification means performs a classification process using a score that represents the certainty of the classification, and the classification means is learned using a loss function corresponding to the frequency distribution of the scores obtained from a score sequence which is a sequence of the scores.

[0007] An information processing device according to one aspect of the present invention comprises: an acquisition means for acquiring sequential data; a classification means for classifying the sequential data into one of a plurality of classes, wherein the classification process uses a score that represents the certainty of the classification; and a learning means for training the classification means using a loss function that corresponds to the frequency distribution of the scores obtained from a score sequence which is a sequence of the scores.

[0008] An information processing method relating to one aspect of the present invention involves at least one processor acquiring sequence data and performing a classification process using classification means to classify the sequence data into one of a plurality of classes, wherein the classification process uses a score representing the certainty of the classification. The classification means is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the score sequence, which is a sequence of scores. Information processing methods.

[0009] An information processing method according to one aspect of the present invention includes, at least one processor acquiring sequential data, and a classification process that classifies the sequential data into one of a plurality of classes, wherein the classification means performs a classification process using a score that represents the certainty of the classification, and the method is trained using a loss function that corresponds to the frequency distribution of the scores obtained from a score sequence which is a sequence of the scores.

[0010] An information processing program according to one aspect of the present invention causes a computer to execute an acquisition process of acquiring time-series data, and a classification process of classifying the time-series data into any one of a plurality of classes, the classification process using a score representing the certainty of classification, and the classification means for performing the classification process is learned using a loss function according to the frequency distribution of the scores obtained from the score series which is the series of the scores. Information processing program.

[0011] An information processing program according to one aspect of the present invention causes a computer to execute an acquisition process of acquiring time-series data, a classification process of classifying the time-series data into any one of a plurality of classes, the classification process using a score representing the certainty of classification, and a learning process of causing the classification means for performing the classification process to be learned using a loss function according to the frequency distribution of the scores obtained from the score series which is the series of the scores.

Advantages of the Invention

[0012] According to one embodiment of the present invention, even when class imbalance exists in learning data, efficient learning can be performed and classification accuracy can be improved.

Brief Description of the Drawings

[0013] [Figure 1] It is a block diagram showing the configuration of an information processing apparatus according to Exemplary Embodiment 1. [Figure 2] It is a flowchart showing the flow of an information processing method according to Exemplary Embodiment 1. [Figure 3] It is a block diagram showing the configuration of an information processing apparatus according to Exemplary Embodiment 1. [Figure 4] It is a flowchart showing the flow of an information processing method according to Exemplary Embodiment 1. [Figure 5] It is a block diagram showing the configuration of an information processing apparatus according to Exemplary Embodiment 2. [Figure 6] It is a diagram for explaining a score matrix and a score series according to Exemplary Embodiment 2. [Figure 7]This is a flowchart showing the flow of the classification process according to exemplary embodiment 2. [Figure 8] This is a diagram illustrating the classification process according to exemplary embodiment 2. [Figure 9] This is a flowchart showing the learning process flow according to exemplary embodiment 2. [Figure 10] This is a diagram illustrating the learning process according to exemplary embodiment 2. [Figure 11] This is a diagram illustrating the learning process according to exemplary embodiment 2. [Figure 12] This is a diagram illustrating the extraction process according to exemplary embodiment 2. [Figure 13] This is a block diagram showing the configuration of an information processing device according to exemplary embodiment 3. [Figure 14] This is a block diagram showing the configuration of a computer that functions as an information processing device according to each exemplary embodiment. [Modes for carrying out the invention]

[0014] [Exemplary Embodiment 1] A first exemplary embodiment of the present invention will be described in detail with reference to the drawings. This exemplary embodiment is the basic form of the exemplary embodiments described later.

[0015] <Overview of Information Processing Device 1> The information processing device 1 according to this exemplary embodiment has a function to classify sequential data input to the acquisition unit into one of several classes using a classification unit, and a function to train the classification unit. In other words, the information processing device 1 functions as both a classification device that classifies sequential data and a learning device that trains the classification unit that performs the classification process.

[0016] Here, the sequence data in this exemplary embodiment is ordered data containing one or more elemental data. In other words, the sequence data is composed of individual elemental data. The sequence data may, for example, be video data or audio data. The sequence data may also be data representing trends such as power consumption. Furthermore, the sequence data is not limited to time-series data composed of elemental data arranged in a time series, but is acceptable as long as the elemental data have some kind of sequential relationship with each other. Elemental data is data that constitutes the sequence data, and for example, is a frame in a video, but is not limited to this.

[0017] As an example, the information processing device 1 calculates a score for classifying each element data contained in a sequence of data. A series of scores corresponding to a given sequence of data is formed from multiple scores calculated for that sequence of data.

[0018] Then, as an example, the information processing device 1 classifies the sequence data corresponding to the score sequence into one of several predetermined classes by comparing the scores included in the score sequence with one or more predetermined thresholds.

[0019] These processes performed by the information processing device 1 can be considered a type of so-called SPRT (Sequential Probability Ratio Test), but are not limited to it. Furthermore, specific examples of classes are not limited to this exemplary embodiment, but examples include people, non-person objects, and the actions of people.

[0020] Information processing device 1 is, for example, a device that acquires video footage from a camera installed at the entrance of a building or event venue as sequential data, and classifies the subjects contained in the video footage into one of several pre-registered classes.

[0021] <Configuration of Information Processing Device 1> Next, the configuration of the information processing device 1 will be described with reference to Figure 1. Figure 1 is a block diagram showing the configuration of the information processing device 1. As shown in Figure 1, the information processing device 1 includes an acquisition unit 11, a classification unit 12, and a learning unit 13.

[0022] (Acquisition part 11) The acquisition unit 11 acquires series data. The acquisition unit 11 may acquire the series data all at once, or it may acquire the series data sequentially. An example of series data is as described above.

[0023] (Classification section 12) The classification unit 12 performs a classification process that classifies the sequence data acquired by the acquisition unit 11 into one of several classes, and performs a classification process using a score that represents the certainty of the classification. The details of the classification process performed by the classification unit 12 are not limited to this exemplary embodiment, but as an example, as described above, The acquisition unit 11 calculates a score from each element data included in the acquired series data, The system refers to the score series obtained from the calculated scores and compares the scores included in that score series with a predetermined threshold to classify the series data corresponding to that score series into one of several predetermined classes. One example of such a process is:

[0024] (Learning Section 13) The learning unit 13 trains the classification unit 12 using a loss function that corresponds to the frequency distribution of scores obtained from the score sequence, which is a series of scores.

[0025] Here, the score series referenced by the learning unit 13 may be a score series calculated by the classification unit 12 during the classification phase in which the information processing device 1 functions as a classification device, or it may be a score series calculated by the classification unit 12 during the learning phase in which the information processing device 1 functions as a learning device, by referring to the sequence data acquired as learning data.

[0026] Furthermore, the frequency distribution of the scores referenced by the learning unit is as follows: • Frequency distribution of scores obtained by ignoring the serial direction (or time direction in the case of time series data) in a given series of data (in other words, by making the serial direction zero-dimensional). You may use, • Frequency distribution of scores across a series of data at a specific point in the series direction (a specific time point in the case of time series data) in a series of data set containing multiple series data (in other words, a series of data within a minibatch). You may also use [this].

[0027] As described above, the information processing device 1 according to this exemplary embodiment employs a configuration in which a classification process is performed to acquire sequential data and classify the sequential data into one of a plurality of classes, the classification process is performed using a score that represents the certainty of the classification, and the classification means is trained using a loss function that corresponds to the frequency distribution of the scores obtained from the score sequence, which is a sequence of the scores.

[0028] Generally, in classification processes that classify sequential data, a problem of class imbalance often arises, where many sequential data are classified into certain classes, while others are hardly classified at all. In the information processing device 1 according to this exemplary embodiment, the classification unit 12 is trained using a loss function that corresponds to the frequency distribution of scores obtained from a score sequence, which is a sequence of scores. This configuration allows for efficient learning without being significantly affected by class imbalance. In other words, the information processing device 1 according to this exemplary embodiment can perform efficient learning and improve classification accuracy even when class imbalance exists in the training data.

[0029] <Flow of information processing method S1> Next, the flow of the information processing method S1 according to this exemplary embodiment 1 will be explained with reference to Figure 2. Figure 2 is a flowchart showing the flow of the information processing method S1. As shown in Figure 2, in step S11, the acquisition unit 11 acquires sequential data. In step S12, the learning unit 13 trains a classification unit 12, which performs a classification process that classifies the sequential data into one of a plurality of classes, using a loss function corresponding to the frequency distribution of the scores obtained from the score sequence, which is a series of the scores.

[0030] As described above, the information processing method S1 according to this exemplary embodiment employs a configuration in which a classification means that acquires sequential data and classifies the sequential data into one of a plurality of classes is trained using a loss function that corresponds to the frequency distribution of the scores obtained from a score sequence, which is a sequence of the scores. Therefore, according to the information processing method S1 according to this exemplary embodiment, even if there is class imbalance in the training data, efficient learning can be performed and the classification accuracy can be improved.

[0031] <Configuration of Information Processing Device 2> Next, the configuration of the information processing device 2 according to this exemplary embodiment will be described with reference to Figure 3. Figure 3 is a block diagram showing the configuration of the information processing device 2. As shown in Figure 3, the information processing device 2 includes an acquisition unit 21 and a classification unit 22. On the other hand, the information processing device 2 does not have the learning unit that the information processing device 1 had. However, the classification unit 22 of the information processing device 2 is pre-trained with a configuration equivalent to that of the learning unit 13 of the information processing device 1. The information processing device 2 according to this exemplary embodiment functions as a classification device that classifies sequential data.

[0032] (Acquisition part 21) The acquisition unit 21 acquires sequential data, similar to the acquisition unit 11 of the information processing device 1 described earlier. The specific processing performed by the acquisition unit 21 is the same as that of the acquisition unit 11 of the information processing device 1 described earlier, so it will not be explained here.

[0033] (Classification section 22) The classification unit 22 performs a classification process that classifies the sequence data into one of several classes, and performs a classification process using a score that represents the certainty of the classification. Here, the classification unit 22 is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the score sequence, which is a sequence of scores. The specific processing performed by the classification unit 22 is the same as the processing performed by the classification unit 12 of the information processing device 1 that has already been described, so the explanation is omitted here.

[0034] As described above, the information processing device 2 according to this exemplary embodiment acquires sequential data and performs a classification process to classify the sequential data into one of a plurality of classes, wherein the classification process uses a score that represents the certainty of the classification. Here, the classification unit that performs the classification process is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the score sequence, which is a sequence of scores.

[0035] Therefore, according to the information processing device 2 of this exemplary embodiment, even when class imbalance exists in the training data, it is possible to improve classification accuracy using an efficiently trained classification unit.

[0036] <Flow of information processing method S2> Next, the flow of the information processing method S2 according to this exemplary embodiment 1 will be explained with reference to Figure 4. Figure 4 is a flowchart showing the flow of the information processing method S2. As shown in Figure 4, in step S21, the acquisition unit 21 acquires sequential data. In step S22, the classification unit 22 performs a classification process to classify the sequential data into one of a plurality of classes, using a score that represents the certainty of the classification. Here, the classification unit 22 is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the score sequence, which is a series of scores.

[0037] As described above, the information processing method S2 according to this exemplary embodiment is a classification process that acquires sequential data and classifies the sequential data into one of a plurality of classes, and performs a classification process using a score that represents the certainty of the classification. Furthermore, in the information processing method S2 according to this exemplary embodiment, the classification unit 22 is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the score sequence, which is a series of scores.

[0038] Therefore, according to the information processing method S2 of this exemplary embodiment, even if class imbalance exists in the training data, it is possible to improve classification accuracy by using an efficiently trained classification unit.

[0039] [Exemplary Embodiment 2] A second exemplary embodiment of the present invention will be described in detail with reference to the drawings. Components having the same function as those described in Exemplary Embodiment 1 will be denoted by the same reference numerals, and their descriptions will not be repeated.

[0040] <Configuration of Information Processing Device 1A> Figure 5 is a block diagram showing the configuration of the information processing device 1A according to this exemplary embodiment. The information processing device 1A comprises a control unit 10A, a storage unit 20A, a communication unit 30A, and an input / output unit 40A.

[0041] (Communications Unit 30A) The communication unit 30A communicates with external devices of the information processing device 1A via a communication line. The specific configuration of the communication line is not limited to this exemplary embodiment, but examples of communication lines include wireless LAN (Local Area Network), wired LAN, WAN (Wide Area Network), public telephone network, mobile data communication network, or a combination thereof. The communication unit 30A transmits data supplied from the control unit 10A to other devices and supplies data received from other devices to the control unit 10A.

[0042] (Input / output section 40A) Input / output devices such as keyboards, mice, displays, printers, and touch panels are connected to the input / output unit 40A. The input / output unit 40A receives various types of information from the connected input devices to the information processing device 1A. The input / output unit 40A also outputs various types of information to the connected output devices under the control of the control unit 10A. An interface such as USB (Universal Serial Bus) can be used for the input / output unit 40A.

[0043] (Acquisition part 11) The acquisition unit 11 acquires sequential data SD. For example, the acquisition unit 11 acquires sequential data SD from another device via the communication unit 30A. Alternatively, the acquisition unit 11 may acquire sequential data SD input via the input / output unit 40A. Furthermore, the acquisition unit 11 may acquire sequential data SD by reading it from the storage unit 20A or an externally connected storage device.

[0044] In this exemplary embodiment, the series data SD acquired by the acquisition unit 11 is, for example, time series data having a temporal seriality, but this is not limited to this exemplary embodiment.

[0045] Furthermore, in this exemplary embodiment, the sequence data SD acquired by the acquisition unit 11 is, for example, a video image data containing multiple frames, but this is not limited to this exemplary embodiment.

[0046] (Classification unit 12) The classification unit 12 performs a classification process that classifies the series data acquired by the acquisition unit 11 into one of a plurality of classes, and performs a classification process using a score representing the accuracy of classification. Here, as an example, the score is calculated by the classification unit 12. A specific example of the score calculated by the classification unit 12 does not limit this exemplary embodiment, but as an example, a log-likelihood ratio can be mentioned.

[0047] The upper part of FIG. 6 shows the log-likelihood ratio as a specific example of the score calculated by the classification unit 12. In the example shown in the upper part of FIG. 6, classification into three classes is assumed. As shown in FIG. 6, a score matrix λ (1,t) , (t) is defined. Here, k is an index specifying the row of the score matrix, l is an index specifying the column of the score matrix, and t is an index indicating time.

[0048] The score matrix λ kl (t) is defined for each time of one series X (1,t) ={x (1) ,x (2) ,···,x (t)} Also, in the expression of the score matrix, y indicates a class label. <​​​​​​​​​​​​​​​​​​​​​​​​​ p(X (1,t) |y=0) It is defined as the logarithmic ratio of the score matrix λ. As is clear from these definitions, kl (t) is an antisymmetric matrix.

[0050] The specific configuration used by the classification unit 12 to calculate the above probability (likelihood) is not limited to this exemplary embodiment, but as an example, a deep learning algorithm using a neural network can be used.

[0051] Furthermore, the classification unit 12 uses the score matrix λ defined above. kl From each component of (t), as shown in the lower part of Figure 6, the values ​​at each time t are

number

[0052] (Flow of classification processing by classification unit 12) The classification unit 12 classifies the sequence data into a class if the score corresponding to a certain data point in the sequence data acquired by the acquisition unit 11 reaches a threshold of any of the multiple classes. If the score corresponding to a certain data point does not reach any of the thresholds of the multiple classes, the classification unit 12 updates the score using the data following the data point in the sequence data, in other words, it updates the score sequence. By performing the above processing, the classification unit 12 can efficiently classify the sequence data.

[0053] The specific flow of the classification process of sequential data by the classification unit 12 will be explained below with reference to Figure 7. Figure 7 is a flowchart showing the flow of the classification process of sequential data by the classification unit 12.

[0054] (S101) As shown in Figure 7, in step S101, the classification unit 12 calculates a score from the series data. More specifically, the classification unit 12 calculates the score series as defined in the above description.

[0055] (S102) Next, in step S102, the classification unit 12 determines whether or not there are any scores in the score series that have reached a predetermined threshold Th. If there are any scores in the score series that have reached a predetermined threshold Th, the unit proceeds to step S103; otherwise, it proceeds to step S104.

[0056] Here, the threshold Th may be defined for each class. For example, a configuration may be set in which a threshold Th0 is referenced for classification into class 0 (y=0), a threshold Th1 is referenced for classification into class 1 (y=1), and a threshold Th2 is referenced for classification into class 2 (y=2).

[0057] (S103) If any of the scores in the score series reach a predetermined threshold Th, in step S103, the classification unit 12 classifies the series data into a class corresponding to the score series containing the score that reached the threshold Th, and terminates the classification process. The classification unit 12 stores the classification result in the storage unit 20A, as an example.

[0058] (S104) On the other hand, if none of the scores in the score series reach a predetermined threshold Th, in step S104, the classification unit 12 updates the scores and updates the score series using those scores. After these update processes, the process in step S102 is performed.

[0059] Figure 8 is a diagram illustrating the results of the classification process performed by the classification unit 12 as described above. In the example in Figure 8, the horizontal axis represents time, and the vertical axis represents the score value. In the example shown in Figure 8, the score series s1 indicates classification to class 1. (1,t) The score sequence s1 approaches the threshold Th over time, and reaches the threshold Th at time τ. This is because, in step S102 described above, the score sequence s1 at time τ (1,t) The threshold was reached, corresponding to the classification to class 1 in step S103. On the other hand, as shown in Figure 8, the score sequence s0 indicates classification to class 0. (1,t) , and the score sequence s2 which indicates classification to Class 2 (0,t) The value does not approach the threshold Th even as time passes, and does not reach the threshold Th at time τ. Therefore, in this example, the sequence data to be processed is not classified into either class 0 or class 2.

[0060] In this manner, the classification unit 12 sequentially calculates score sequences equal to the number of classes from the acquired sequence data, and classifies the sequence data into one of several classes by sequentially comparing the calculated score sequences with a threshold. It should be noted that the classification process described above can be considered a type of SPRT (Sequential Probability Ratio Test) algorithm, but this statement is not limited to this exemplary embodiment.

[0061] (Learning Section 13) The learning unit 13 trains the classification unit 12 using a loss function that corresponds to the frequency distribution of the scores obtained from the score sequence, which is a series of scores calculated by the classification unit 12. Here, the loss function referenced by the learning unit 13 is, as an example, a loss function expressed using the partial AUC (partial Area Under the Curve) of the ROC curve (Receiver Operating Characteristic Curve) obtained from the frequency distribution of the scores. Specific examples of the ROC curve and partial AUC will be described later.

[0062] (Flow of learning process by learning unit 13) The following describes the learning process of the classification unit 12 by the learning unit 13, with reference to Figure 9. Figure 9 is a flowchart showing the learning process of the classification unit 12 by the learning unit 13.

[0063] (S201) As shown in Figure 9, the learning unit 13 acquires a score sequence in step S201. Here, the score sequence acquired by the learning unit 13 is, as an example, the score sequence calculated by the classification unit 12 described above. An example of a score sequence acquired by the learning unit 13 is shown in the upper part of Figure 10. The upper part of Figure 10 shows the score sequence s0, which indicates classification to class 0, as an example of a score sequence. (0,t) , the score sequence s1 which indicates classification to Class 1 (0,t) , and the score sequence s2 which indicates classification to Class 2 (0,t) This is shown.

[0064] (S202) Next, in step S202, the learning unit 13 calculates the cumulative frequency distribution ADD from the score series obtained in step S201. In this step, the learning unit 13 first calculates a frequency distribution (histogram) showing the relationship between each score and its frequency by ignoring (or making 0-dimensional) the sequence direction (time direction in this example) of the score series, as shown in the lower left of Figure 10. Here, the learning unit 13 generates the frequency distribution including a histogram corresponding to the true class and a histogram corresponding to the false classes. In this example, the histogram corresponding to the true class is the score series s1 which indicates classification to class 1. (0,t) This is a histogram generated from the data, and the histogram corresponding to the false class is the score sequence s0, which indicates classification to class 0. (0,t) and the score sequence s1 which indicates classification into class 2 (0,t) It is calculated from the score series obtained by summing up and

[0065] Next, in this step, the learning unit 13 calculates histograms for multiple other sequential data included in the mini-batch (sequence data stored in the memory unit 20A) using the same method, and calculates the cumulative frequency distribution by summing the calculated histograms within the mini-batch. This summing is performed for each class. An example of the calculated cumulative frequency distribution is shown in the lower right of Figure 10.

[0066] (S203) Next, in step S203, the learning unit 13 calculates ROC curves for each class using the cumulative frequency distribution calculated in step S202. Subsequently, in this step, the learning unit 13 calculates pAUC (partial Area Under the Curve) for each class from the ROC curves described above. The learning unit 13 stores the calculated pAUC (partial area under the curve) in the memory unit 20A, for example. Figure 11 is a schematic diagram showing the ROC curves and pAUC calculated by the learning unit 13 in this step. Here, the pAUC calculated in this step is also referred to as pAUC-ROC.

[0067] Note that the specific example of pAUC calculation by the learning unit 13 is not limited to this exemplary embodiment, but as an example, • pAUC obtained by limiting the FPR (False Positive Rate) • pAUC obtained by limiting the TPR (True Positive Rate) • pAUC obtained by restricting both FPR and TPR • pAUC obtained by restricting to the Region of Interest These are some examples. The pAUC shown in Figure 11 is an example of a pAUC obtained by restricting both FPR and TPR.

[0068] Next, in this step, the learning unit 13 transforms the class-specific pAUCs described above so that they become differentiable functions. Here, this transformation process may include replacing the step function contained in the pAUC before the transformation with a smooth function.

[0069] (S204) Next, in step S204, the learning unit 13 calculates the value of the loss function defined using the converted pAUC described above. The learning unit 13 stores the calculated loss function value VLF in the storage unit 20A, for example.

[0070] Note that the specific example of the loss function LF calculated by the learning unit 13 is not limited to this exemplary embodiment, but as an example, LF = (1 - Z) 2 ...(Formula 1) (Here, Z is the average of the transformed pAUCs described above across multiple classes.) This can be used. Here, as the averaging process, a simple average may be used, or a weighted average may be used by setting a weight coefficient for each class. For example, as Z=w0*pAUC0+ w1*pAUC1+ w2*pAUC2 You may also use the following. Here, w0, w1, and w2 represent the weight coefficients set for classes 0, 1, and 2, respectively, and pAUC0, pAUC1, and pAUC2 represent the transformed pAUC for classes 0, 1, and 2, respectively, and * is the symbol for product.

[0071] (S205) Next, in step S205, the learning unit 13 trains the classification unit 12 using the loss function calculated as described above. More specifically, the learning unit 13 trains the classification unit 12 by updating the parameters of the classification unit 12 (or, if a neural network is used as the classification unit 12, the parameters of the neural network) so that the value of the loss function calculated as described above becomes larger or smaller. In other words, the learning unit 13 trains the classification unit 12 by updating the parameters of the classification unit 12 so that the loss function calculated as described above becomes larger or smaller. Note that while stochastic gradient descent and error propagation methods can be used as parameter optimization methods using the loss function, the system is not limited to these.

[0072] (Effects of using pAUC) As will be explained in more detail later, pAUC tends to be less affected by class imbalance. Therefore, by performing training using the loss function LF as described above, the learning unit 13 can train the classification unit 12 with almost no impact from class imbalance. Thus, with the above configuration, even if class imbalance exists in the training data, efficient training can be performed, and the classification accuracy can be improved.

[0073] (Other examples of loss functions) Furthermore, the learning unit 13 may adopt a linear sum of a loss function that depends on pAUC and a loss function that does not depend on pAUC as the loss function LF, as described above.

[0074] For example, the learning unit 13 uses the following loss function LF: LF = a * (1 - Z) 2 + b*L LLLR ...(Formula 2) You may also use this. Here, a and b represent weight coefficients, and L LLLR This represents a loss function that does not depend on pAUC, and as an example,

number

[0075] Therefore, by using the loss function LF defined by Equation 2 above, the classification unit 12 can be trained to effectively perform hard sample mining while being resilient to class imbalances.

[0076] (Other examples of cumulative frequency distributions) In the above explanation, the cumulative frequency distribution calculated in step S202 is as follows: By ignoring (or reducing to 0-dimensionality) the sequence direction of the score series (in this example, the time direction), we can calculate a frequency distribution (histogram) that shows the relationship between each score and its frequency. The same method is used to calculate histograms for other series data included in the mini-batch, and the cumulative frequency distribution is calculated by summing the calculated histograms within the mini-batch. While this is one example, the exemplary embodiments described here are not limited to this.

[0077] The learning unit 13, as a process for calculating the cumulative frequency distribution, instead of the above example, • A frequency distribution of multiple data series contained in a mini-batch, where the frequency distribution of scores at a specific point in the series direction (a specific time point in the case of time series data) is the frequency distribution of the score at that point. The frequency distribution may be calculated for each class, and the calculated frequency distribution may be used as the cumulative frequency distribution.

[0078] (Explanation of not being affected by class imbalance) Below, we will elaborate on why using the pAUC-dependent loss function described above allows for learning that is less susceptible to class imbalance.

[0079] Traditionally, the loss function has been:

number

[0080] On the other hand, as explained in this exemplary embodiment, AUC (pAUC) is, by definition,

number

[0081] (Extraction part 14) Next, the extraction unit 14 provided in the control unit 10A according to this exemplary embodiment will be described. The extraction unit 14 extracts sequence data from the sequence data acquired by the acquisition unit 11 that corresponds to the region in the frequency distribution of scores calculated by the classification unit 12 in which the frequency distribution obtained from one score series overlaps with the frequency distribution obtained from another score series.

[0082] Figure 12 is a diagram illustrating the extraction process performed by the extraction unit 14. As shown in Figure 12, the region defined by pAUC on the ROC graph corresponds to the region in the (cumulative) frequency distribution where the histograms of the true class and the false class overlap. The extraction unit 14 extracts sequential data corresponding to such overlapping regions from the data acquired by the acquisition unit 11. Then, the learning unit 13 uses the sequential data extracted by the extraction unit 14 to train the classification unit 12.

[0083] In this exemplary embodiment, the information processing device 1A, by including the extraction unit 14 as described above, can train the classification unit 12 to effectively classify difficult-to-discriminate sequence data without being affected by class imbalance, in other words, to perform hard sample mining.

[0084] [Exemplary Embodiment 3] A third exemplary embodiment of the present invention will be described in detail with reference to the drawings. Components having the same function as those described in exemplary embodiments 1 and 2 will be denoted by the same reference numerals, and their descriptions will not be repeated.

[0085] <Configuration of Information Processing Device 2A> Figure 13 is a block diagram showing the configuration of the information processing device 2A according to this exemplary embodiment. As shown in Figure 13, the information processing device 2A comprises a control unit 10A, a storage unit 20A, a communication unit 30A, and an input / output unit 40A. The information processing device 2A functions as a classification device for classifying sequential data.

[0086] As shown in Figure 13, the control unit 10A of the information processing device 2A includes an acquisition unit 11 and a classification unit 12, but does not include a learning unit 13 or an extraction unit 14. The processing performed by the acquisition unit 11 and the classification unit 12 of the control unit 10A of the information processing device 2A is the same as the processing performed by the acquisition unit 11 and the classification unit 12 of the information processing device 1A. The classification unit 12 of the control unit 10A of the information processing device 2A is learned by the learning unit 13 described in Exemplary Embodiment 2. Alternatively, the classification unit 12 may be learned by the learning unit 13 using sequential data extracted by the extraction unit 14 described in Exemplary Embodiment 2.

[0087] With the information processing device 2A configured as described above, since it uses a classification unit 12 that has been trained in the same way as the information processing device 1A to classify sequential data, it can perform efficient classification processing without being affected by class imbalance.

[0088] [Examples of implementation using software] The control blocks of the information processing devices 1, 2, 1A, and 2A (particularly the classification units 12, 22, the learning unit 13, and the extraction unit 14) may be implemented by logic circuits (hardware) formed on an integrated circuit (IC chip) or by software.

[0089] In the latter case, the information processing devices 1, 2, 1A, and 2A are equipped with a computer that executes instructions for a program, which is software that realizes each function. This computer is equipped with, for example, at least one processor (control device) and at least one computer-readable recording medium that stores the program. The object of the present invention is achieved when the processor reads the program from the recording medium and executes it in the computer. For example, a CPU (Central Processing Unit) can be used as the processor. As the recording medium, a "tangible medium that is not temporary," such as ROM (Read Only Memory), can be used, as well as tape, disk, card, semiconductor memory, programmable logic circuit, etc. It may also be further equipped with RAM (Random Access Memory) for deploying the program. Furthermore, the program may be supplied to the computer via any transmission medium capable of transmitting the program (such as a communication network or broadcast wave). In one aspect of the present invention, the program can also be realized in the form of a data signal embedded in a carrier wave, which is embodied by electronic transmission.

[0090] [Additional Note 1] The present invention is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining the technical means disclosed in the embodiments described above are also included in the technical scope of the present invention.

[0091] [Additional Note 2] Some or all of the embodiments described above may also be described as follows. However, the present invention is not limited to the embodiments described below.

[0092] (Note 1) A means of acquiring sequential data, A classification process for classifying the aforementioned sequence data into one of several classes, comprising a classification means that performs a classification process using a score representing the certainty of the classification, Equipped with, The classification means is This is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the aforementioned score sequence. Information processing device.

[0093] (Note 2) The loss function is expressed using the partial AUC of the ROC curve obtained from the frequency distribution of the score. The information processing device described in Appendix 1.

[0094] (Note 3) The classification means is Series data extracted from the series data acquired by the acquisition means, which corresponds to the region in the score frequency distribution where the frequency distribution obtained from one score series overlaps with the frequency distribution obtained from another score series. This was learned using The information processing device described in Appendix 1 or 2.

[0095] (Note 4) The classification means is When the score corresponding to a certain data point in the aforementioned series data reaches a threshold in any of several classes, the series data is classified into that class. If the score corresponding to the aforementioned data does not reach any of the thresholds for the multiple classes, the score is updated using the data following the aforementioned data from the sequence data. An information processing device as described in any one of the items 1 to 3 in the appendix.

[0096] (Note 5) A means of acquiring sequential data, A classification process for classifying the aforementioned sequence data into one of several classes, comprising a classification means that performs a classification process using a score representing the certainty of the classification, A learning means that trains the classification means using a loss function corresponding to the frequency distribution of the scores obtained from the score sequence, which is the score sequence. It is equipped with Information processing device.

[0097] (Note 6) The loss function is expressed using the partial AUC of the ROC curve obtained from the frequency distribution of the score. The information processing device described in Appendix 5.

[0098] (Note 7) From the series data acquired by the acquisition means, In the frequency distribution of the aforementioned scores, the sequence data corresponding to the region where the frequency distribution obtained from one score series overlaps with the frequency distribution obtained from another score series. It is equipped with an extraction means for extracting [something]. The information processing device described in Appendix 5 or 6.

[0099] (Note 8) The classification means is trained using the sequential data extracted by the extraction means. The information processing device described in Appendix 7.

[0100] (Note 9) The classification means is When the score corresponding to a certain data point in the aforementioned series data reaches a threshold in any of several classes, the series data is classified into that class. If the score corresponding to the aforementioned data does not reach any of the thresholds for the multiple classes, the score is updated using the data following the aforementioned data from the sequence data. An information processing device as described in any one of the items 5 to 8 of the appendix.

[0101] (Note 10) At least one processor, To obtain series data, A classification process that uses a classification means to classify the aforementioned sequence data into one of several classes, wherein the classification process uses a score that represents the certainty of the classification. Includes, The classification means is This is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the aforementioned score sequence. Information processing methods.

[0102] (Note 11) At least one processor, To obtain series data, A classification process that classifies the aforementioned sequence data into one of several classes, wherein a classification means that performs classification using a score representing the certainty of the classification is trained using a loss function that corresponds to the frequency distribution of the scores obtained from the score sequence, which is a sequence of the scores. Information processing methods including

[0103] (Note 12) On the computer, The acquisition process for obtaining series data, A classification process for classifying the aforementioned series data into one of several classes, comprising a classification process using a score representing the certainty of the classification, Make it run, The classification means that performs the classification process is: This is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the aforementioned score sequence. Information processing program.

[0104] (Note 13) On the computer, The acquisition process for obtaining series data, A classification process for classifying the aforementioned series data into one of several classes, comprising a classification process using a score representing the certainty of the classification, A learning process that trains the classification means performing the classification process using a loss function corresponding to the frequency distribution of the scores obtained from the score sequence, which is a series of scores. An information processing program that executes [something].

[0105] [Additional Note 3] Some or all of the embodiments described above can also be expressed as follows:

[0106] (Note 14) It comprises at least one processor, the processor being, The acquisition process for obtaining series data, A classification process for classifying the aforementioned series data into one of several classes, comprising a classification process using a score representing the certainty of the classification, Execute, The classification means that performs the classification process is: This is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the aforementioned score sequence. Information processing device.

[0107] (Note 15) It comprises at least one processor, the processor being, The acquisition process for obtaining series data, A classification process for classifying the aforementioned series data into one of several classes, comprising a classification process using a score representing the certainty of the classification, A learning process that trains the classification means performing the classification process using a loss function corresponding to the frequency distribution of the scores obtained from the score sequence, which is a series of scores. An information processing device that performs the following actions.

[0108] These information processing devices may also be equipped with memory, which may store a program that causes the processor to perform the acquisition process, the classification process, and the learning process. This program may also be recorded on a computer-readable, non-temporary, tangible recording medium. [Explanation of symbols]

[0109] 1,2,1A,2A ···Information Processing Device 11,21 ···Acquisition part 12,22...Classification section 13 ···Learning Department 14. Extraction Section 20A ···Memory Department 30A···Communication Department 40A ···Input / Output Force Section

Claims

1. A means of acquiring sequential data, A classification process for classifying the aforementioned sequence data into one of several classes, comprising a classification means that performs a classification process using a score representing the certainty of the classification, Equipped with, The classification means is This is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the aforementioned score sequence. Information processing device.

2. The loss function is a loss function expressed using the partial AUC of the ROC curve obtained from the frequency distribution of the score. The information processing apparatus according to claim 1.

3. The classification means is Series data extracted from the series data acquired by the acquisition means, which corresponds to the region in the score frequency distribution where the frequency distribution obtained from one score series overlaps with the frequency distribution obtained from another score series. This was learned using The information processing apparatus according to claim 1 or 2.

4. The classification means is When the score corresponding to a certain data point in the aforementioned series data reaches a threshold in any of several classes, the series data is classified into that class. If the score corresponding to the aforementioned data does not reach any of the thresholds for the multiple classes, the score is updated using the data following the aforementioned data from the sequence data. The information processing apparatus according to claim 1 or 2.

5. A means of acquiring sequential data, A classification process for classifying the aforementioned sequence data into one of several classes, comprising a classification means that performs a classification process using a score representing the certainty of the classification, A learning means that trains the classification means using a loss function corresponding to the frequency distribution of the scores obtained from the score sequence, which is the score sequence. It is equipped with Information processing device.

6. The loss function is a loss function expressed using the partial AUC of the ROC curve obtained from the frequency distribution of the score. The information processing apparatus according to claim 5.

7. From the series data acquired by the acquisition means, In the frequency distribution of the aforementioned scores, the sequence data corresponding to the region where the frequency distribution obtained from one score series overlaps with the frequency distribution obtained from another score series. It is equipped with an extraction means for extracting [something]. The information processing apparatus according to claim 5 or 6.

8. The classification means is When the score corresponding to a certain data point in the aforementioned series data reaches a threshold in any of several classes, the series data is classified into that class. If the score corresponding to the aforementioned data does not reach any of the thresholds for the multiple classes, the score is updated using the data following the aforementioned data from the sequence data. The information processing apparatus according to claim 5 or 6.

9. At least one processor, To obtain series data, A classification process that uses a classification means to classify the aforementioned sequence data into one of several classes, wherein the classification process uses a score that represents the certainty of the classification. Includes, The classification means is This is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the aforementioned score sequence. Information processing methods.

10. On the computer, The acquisition process for obtaining series data, A classification process for classifying the aforementioned series data into one of several classes, comprising a classification process using a score representing the certainty of the classification, Make it run, The classification means that performs the classification process is: This is learned using a loss function that corresponds to the frequency distribution of the scores obtained from the aforementioned score sequence. Information processing program.

Images