WO2023218535A1 - 情報処理装置、情報処理方法及びプログラム - Google Patents
情報処理装置、情報処理方法及びプログラム Download PDFInfo
- Publication number
- WO2023218535A1 WO2023218535A1 PCT/JP2022/019840 JP2022019840W WO2023218535A1 WO 2023218535 A1 WO2023218535 A1 WO 2023218535A1 JP 2022019840 W JP2022019840 W JP 2022019840W WO 2023218535 A1 WO2023218535 A1 WO 2023218535A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- function
- unit
- parameters
- monotonically increasing
- survival
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
Definitions
- Embodiments relate to an information processing device, an information processing method, and a program.
- Some of these events occur only once (including cases that are not expected because the data changes significantly after they occur). Examples of such events include death, accident, marriage, recurrence of illness, and the like. Survival analysis is often used to predict such events.
- Prediction by survival analysis is usually performed using the following steps. 1. Model learning (or manual design) is performed using data in which an event has occurred and data in which an event has not occurred. 2. Using a model, a hazard function that represents the likelihood of an event occurring and/or a survival function that represents the probability that an event will occur over a certain period of time is determined for data that is actually desired to be predicted.
- the first problem is that there is not always enough data on the occurrence of the event that we want to predict.
- the second issue is that there are strong assumptions, such as being based on the COX proportional hazards model.
- the COX proportional hazards model the relative likelihood of occurrence is known, but the absolute time is not known.
- time cannot be estimated more accurately than the discretized granularity.
- the third problem is that if assumptions such as the COX proportional hazards model are not made, the likelihood includes an integral, making optimization difficult or requiring approximation.
- Non-Patent Document 1 and Non-Patent Document 2 have been proposed to address such issues.
- Non-Patent Document 1 discloses a method based on the COX proportional hazard model.
- the method in Non-Patent Document 1 solves the first problem by performing meta-learning using MAML (Model-Agnostic Meta-Learning), and solves the third problem by using the COX proportional hazard model. is avoided.
- MAML Model-Agnostic Meta-Learning
- the method of Non-Patent Document 1 cannot solve the second problem because it uses a COX proportional hazard model.
- Non-Patent Document 2 discloses a method of discretizing time. The method of Non-Patent Document 2 avoids the third problem by discretization. However, with the method of Non-Patent Document 2, the first problem is not solved, and the second problem is not solved because it is discretized.
- the present invention has been made in view of the above circumstances, and its purpose is to provide a means that enables calculation of at least one of a hazard function and a survival function without making any assumptions.
- An information processing device includes a latent expression calculation unit, a monotonically increasing neural network, and a function estimation unit.
- the latent expression calculation unit calculates a latent expression representing a feature amount from processing target data including a feature amount related to a prediction target event.
- the monotonically increasing neural network is modeled to output a scalar value according to a monotonically increasing function defined by the latent expression calculated by the latent expression calculation unit and time.
- the function estimator estimates at least one of a hazard function and a survival function based on the scalar value output from the monotonically increasing neural network.
- FIG. 1 is a block diagram showing an example of the hardware configuration of a survival analysis device as an information processing device according to the first embodiment.
- FIG. 2 is a block diagram showing an example of the configuration of a learning function of the survival analysis device as the information processing device according to the first embodiment.
- FIG. 3 is a block diagram illustrating an example of a configuration of a prediction function of a survival analysis device as an information processing device according to the first embodiment.
- FIG. 4A is a flowchart illustrating an example of a learning operation in the survival analysis device as the information processing device according to the first embodiment.
- FIG. 4B is a flowchart illustrating an example of a learning operation in the survival analysis device as the information processing device according to the first embodiment.
- FIG. 1 is a block diagram showing an example of the hardware configuration of a survival analysis device as an information processing device according to the first embodiment.
- FIG. 2 is a block diagram showing an example of the configuration of a learning function of the survival analysis device as the information processing device according to the first embodiment.
- FIG. 5 is a flowchart illustrating an example of a prediction operation in the survival analysis device as the information processing device according to the first embodiment.
- FIG. 6 is a block diagram illustrating an example of the configuration of a learning function of a survival analysis device as an information processing device according to the second embodiment.
- FIG. 7 is a block diagram illustrating an example of the configuration of a prediction function of a survival analysis device as an information processing device according to the second embodiment.
- FIG. 8A is a flowchart illustrating an example of a learning operation in the survival analysis device as an information processing device according to the second embodiment.
- FIG. 8B is a flowchart illustrating an example of a learning operation in the survival analysis device as an information processing device according to the second embodiment.
- FIG. 9 is a flowchart illustrating an example of a prediction operation in the survival analysis device as an information processing device according to the second embodiment.
- the survival analysis device has a learning function and a prediction function.
- the learning function is a function that meta-learns model parameters using data in which an event has occurred and data in which an event has not occurred.
- the prediction function is a function that calculates a hazard function, cumulative hazard function, and survival function for data that is actually desired to be predicted, based on the parameters of the model learned by the learning function.
- FIG. 1 is a block diagram showing an example of the hardware configuration of a survival analysis device 1 as an information processing device according to the first embodiment.
- the survival analysis device 1 includes a control circuit 10, a memory 11, a communication module 12, a user interface 13, and a drive 14.
- the control circuit 10 is a circuit that controls each component of the survival analysis device 1 as a whole.
- the control circuit 10 includes a CPU (Central Processing Unit), a RAM (Random Access Memory), a ROM (Read Only Memory), and the like. By using a multi-core and multi-threaded CPU, the CPU can execute multiple information processes at the same time. Further, the control circuit 10 may include a plurality of CPUs. Furthermore, the control circuit 10 may include an ASIC (Application Specific Integrated Circuit), a DSP (Digital Signal Processor), an FPGA (field-programmable gate array), a GPU (Graphics Processing Unit), etc. instead of or in addition to the CPU. Can include integrated circuits.
- ASIC Application Specific Integrated Circuit
- DSP Digital Signal Processor
- FPGA field-programmable gate array
- GPU Graphics Processing Unit
- the memory 11 is a storage device of the survival analysis device 1.
- the memory 11 includes, for example, an HDD (Hard Disk Drive), an SSD (Solid State Drive), a memory card, and the like.
- the memory 11 stores information used for learning operations and predictive operations in the survival analysis device 1.
- the memory 11 also stores a learning program for causing the control circuit 10 to perform a learning operation and a prediction program for causing the control circuit 10 to perform a predictive operation.
- the communication module 12 is a circuit used for transmitting and receiving data to and from the outside of the survival analysis device 1 via a network (not shown).
- the user interface 13 is a circuit for communicating information between the user and the control circuit 10.
- User interface 13 includes input equipment and output equipment.
- Input devices include, for example, a touch panel and operation buttons.
- the output device includes, for example, an LCD (Liquid Crystal Display) or an EL (Electroluminescence) display, and a printer.
- the user interface 13 outputs, for example, the execution results of various programs received from the control circuit 10 to the user.
- the drive 14 is a device for reading programs stored in the storage medium 15.
- the drive 14 includes, for example, a CD (Compact Disk) drive, a DVD (Digital Versatile Disk) drive, and the like.
- the storage medium 15 is a medium that stores information such as programs through electrical, magnetic, optical, mechanical, or chemical action.
- the storage medium 15 may store a learning program and a prediction program.
- FIG. 2 is a block diagram showing an example of the learning function configuration of the survival analysis device 1 as the information processing device according to the first embodiment.
- the CPU of the control circuit 10 loads the learning program stored in the memory 11 or the storage medium 15 into the RAM.
- the CPU of the control circuit 10 controls the memory 11, the communication module 12, the user interface 13, the drive 14, and the storage medium 15 by interpreting and executing the learning program developed in the RAM.
- the computer functions as a computer including sections 29 and 30.
- the memory 11 of the survival analysis device 1 functions as a learning data set storage section 20 and a learned parameter storage section 31 for storing information used for learning operations.
- the learning data set storage unit 20 stores a data set (hereinafter referred to as a learning data set) D k corresponding to an event to be predicted. Examples of events to be predicted include mechanical breakdowns, traffic accidents, and life events such as marriage.
- the learning data set D k is information including d survival time data X for each of k tasks, as described below.
- k is the task id
- d is the data id
- DS k is a data set of task k
- K is a task set.
- the survival time data X includes a feature amount x, an indicator variable ⁇ , and a time e.
- the feature amount x may be any information as long as it can be used for the event to be predicted. For example, it is sufficient that the feature quantity x can be handled by the same differentiable model for all tasks.
- Differentiable models include, for example, CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), Perceiver, and the like.
- Perceiver is disclosed in, for example, Andrew Jaegle, et al., “Perceiver: General Perception with Iterative Attention”, arXiv:2103.03206v2 [cs.CV] 23 Jun 2021.
- the events that we want to predict occur at most once for humans, such as life events, traffic accidents, equipment failures, etc. (including).
- the feature amount x may be static or time-series.
- the static feature quantity x is attribute information indicating the person's attributes, such as gender and age
- the time-series feature quantity x is, for example, money balance, location information history, SNS information, etc. This information includes posting history, etc.
- Task k in the life event learning data set D k is an event such as marriage, childbirth, moving, going to higher education, finding a job, etc. Taking an example of feature x and event for task k, if we write task k: (feature, event), for example, task 1: (money balance, marriage), task 2: (location information history + SNS history) Posting history, childbirth), Task 3: (expense history, moving), etc.
- d which is the data ID, is assigned to each person.
- the static feature quantity x is, for example, attribute information indicating attributes of the driver
- the time-series feature quantity x is, for example, the history of sensing data of various sensors, Information such as images from a drive recorder.
- Task k in the learning data set D k for traffic accidents is traffic accidents by country or region, by vehicle type (private car, truck, taxi, bus, etc.).
- the data ID d will be assigned for each drive.
- the event to be predicted, the feature amount x for each event, and the learning data set D k listed here are only examples.
- the event to be predicted is a device failure, and in that case, the feature quantity x is information such as the model, log data, temperature, humidity, etc., and is not limited to the above example.
- the data dividing unit 21 randomly selects the task k and sets the data set for the task k from the learning data set D k stored in the learning data set storage unit 20.
- the data dividing unit 21 randomly divides the extracted learning target data set to obtain a support set SS and a query set QS.
- the data division unit 21 transmits the support set SS to the latent expression calculation unit 23 and the query set QS to the latent expression calculation unit 24.
- the initialization unit 22 initializes the parameter set ⁇ based on an arbitrary rule R determined in advance.
- the parameter set ⁇ includes multiple parameters p1 and multiple parameters p2.
- the initialization unit 22 transmits the plurality of initialized parameters p1 to the latent expression calculation unit 23.
- the initialization unit 22 transmits the plurality of initialized parameters p2 to the function estimation unit 25. Further, the initialization unit 22 transmits the initialized parameter set ⁇ (a plurality of parameters p1 and p2) to the update unit 28.
- the plurality of parameters p1 and p2 will be described later.
- the latent expression calculation unit 23 calculates a latent expression z for the feature x of each piece of data X in the support set SS.
- the latent expression z is data representing the feature of the feature amount x in the data set.
- the latent expression calculation unit 23 transmits the calculated latent expression z to the function estimation unit 25.
- the latent expression calculation unit 23 includes a feature amount extraction unit 231 and a model 232.
- the feature extraction unit 231 extracts the feature x from the support set SS.
- the feature extraction unit 231 sends the feature x to the model 232.
- the model 232 is any differentiable model that can handle the feature quantity x. That is, the model 232 is a mathematical model modeled to input the feature quantity x and output the latent expression z.
- the model 232 can be, for example, CNN, RNN, Perceiver, or the like.
- the parameter ⁇ (multiple parameters p1) is applied to the model 232 as a weight and bias term.
- the model 232 to which the plurality of parameters p1 are applied receives the feature quantity x as input and outputs the latent expression z.
- the model 232 transmits the output latent expression z to the function estimator 25.
- the function estimation unit 25 calculates a hazard function h(t, z) based on the latent expression z and the predicted time t.
- the hazard function h(t,z) is a time function that represents the likelihood of occurrence of an event to be predicted with respect to data to be predicted.
- the function estimator 25 transmits the calculated hazard function h(t,z) to the updater 27.
- the function estimation section 25 includes a monotonically increasing neural network 251, a cumulative hazard function calculation section 252, and an automatic differentiation section 253.
- the monotonically increasing neural network 251 is a mathematical model modeled to calculate as an output a monotonically increasing function defined by the latent expression z and time t.
- Examples of the monotonically increasing neural network 251 include the one disclosed in Antoine Wehenkel, et al., “Unconstrained Monotonic Neural Networks”, arXiv:1908.05164v3 [cs.LG] 31 Mar 2021, where the activation function has a positive differential ( tanh, etc.) with non-negative constraints placed on the weights, etc. can be used.
- a plurality of weights and bias terms based on the parameter ⁇ (a plurality of parameters p2) are applied to the monotonically increasing neural network 251.
- the monotonically increasing neural network 251 to which the plurality of parameters p2 is applied calculates an output f(t, z) as a scalar value according to a monotonically increasing function defined by the latent expression z and the time t.
- the monotonically increasing neural network 251 transmits the output f(t, z) to the cumulative hazard function calculation unit 252.
- the cumulative hazard function calculation unit 252 calculates the cumulative hazard function H(t,z) based on the output f(t,z) according to the formula shown below.
- s is a scale parameter to compensate for the lack of expressive power of the monotonically increasing neural network.
- S(t) is a survival function and represents the probability that the survival time will be greater than or equal to t.
- the cumulative hazard function calculation unit 252 transmits the calculated cumulative hazard function H(t,z) to the automatic differentiation unit 253 and the update unit 27.
- the automatic differentiation section 253 calculates the hazard function h(t,z) by automatically differentiating the cumulative hazard function H(t,z).
- the automatic differentiator 253 transmits the calculated hazard function h(t,z) to the updater 27.
- the hazard function h(t,z) is expressed by the differential of the cumulative hazard function H(t,z) as follows.
- the updating unit 27 calculates an updated parameter set ⁇ (a plurality of parameters p1 and p2) based on the cumulative hazard function H(t, z) and the hazard function h(t, z).
- the updated parameter set is referred to as an updated parameter set ⁇ '(p1', p2').
- the updating section 27 transmits the updated parameter set ⁇ ' (a plurality of parameters p1' and p2') to the determining section 29.
- the update section 27 includes an evaluation function estimation section 271 and an optimization section 272.
- the evaluation function estimation unit 271 calculates the evaluation function L(SS) based on the cumulative hazard function H(t, z) and the hazard function h(t, z).
- the evaluation function L(SS) is, for example, the following negative log likelihood.
- the evaluation function estimation unit 271 transmits the calculated evaluation function L(SS) to the optimization unit 272.
- the optimization unit 272 optimizes the parameter set ⁇ , that is, the plurality of parameters p1 and p2, based on the evaluation function L(SS). For example, an error backpropagation method is used for the optimization.
- the optimization unit 272 transmits this optimized parameter set ⁇ (a plurality of parameters p1 and p2) to the determination unit 29 as an updated parameter set ⁇ ' (a plurality of parameters p1' and p2').
- the determination unit 29 determines whether the first condition is satisfied based on the update parameter set ⁇ ' (a plurality of parameters p1' and p2').
- the first condition may be, for example, that the number of times the update parameter set ⁇ ' is transmitted to the determination unit 29 (that is, the number of parameter update loops) is equal to or greater than a threshold value.
- the first condition may be, for example, that the amount of change in the value of the update parameter set ⁇ ' before and after the update is less than or equal to a threshold value.
- the determination unit 29 applies the updated parameter set ⁇ ' (a plurality of parameters p1' and p2') to the model 232 and the monotonically increasing neural network 251, and the latent expression calculation unit 23 and the function estimation
- the update section 25 and the update section 27 are caused to perform a parameter update operation based on this update parameter set ⁇ '.
- the determination unit 29 causes the latent expression calculation unit 23, function estimation unit 25, and update unit 27 to repeatedly execute a parameter update loop.
- the determination unit 29 terminates the parameter update loop, and transmits the last updated update parameter set ⁇ ′ (the plurality of parameters p1′ and p2′) to the latent expression calculation unit. 24 and the function estimator 26.
- the determination unit 29 initializes the parameters applied to the latent expression calculation unit 24 and the function estimation unit 26 to this updated parameter set ⁇ ' (a plurality of parameters p1' and p2').
- the latent expression calculation unit 24 calculates a latent expression z for the feature amount x of each data X of the query set QS.
- the latent expression calculation unit 24 transmits the calculated latent expression z to the function estimation unit 26.
- the latent expression calculation unit 24 has a configuration corresponding to the latent expression calculation unit 23. That is, the latent expression calculation unit 24 includes a feature amount extraction unit 241 and a model 242.
- the feature extraction unit 241 extracts the feature x from the query set QS.
- the feature extraction unit 241 sends the feature x to the model 242.
- the model 242 is any differentiable model that can handle the feature x.
- the updated parameters p1' are applied to the model 242 as weights and bias terms.
- the model 242 to which the plurality of parameters p1' is applied receives the feature amount x and outputs the latent expression z.
- the model 232 sends the output latent expression z to the function estimator 26.
- the function estimation unit 26 calculates the hazard function h(t, z) based on the latent expression z and the predicted time t.
- the function estimator 26 transmits the calculated hazard function h(t,z) to the updater 27.
- the function estimation unit 26 includes a monotonically increasing neural network 261, a cumulative hazard function calculation unit 262, and an automatic differentiation unit 263.
- the monotonically increasing neural network 261 is a mathematical model similar to the monotonically increasing neural network 251.
- a plurality of weights and bias terms based on the updated plurality of parameters p2' are applied to the monotonically increasing neural network 261.
- the monotonically increasing neural network 261 to which the plurality of parameters p2' is applied calculates an output f(t, z) as a scalar value according to a monotonically increasing function defined by the latent expression z and time t.
- the monotonically increasing neural network 261 transmits the output f(t,z) to the cumulative hazard function calculation unit 262.
- the cumulative hazard function calculation unit 262 is similar to the cumulative hazard function calculation unit 252, and calculates the cumulative hazard function H(t, z) based on the output f(t, z). The cumulative hazard function calculation unit 262 transmits the calculated cumulative hazard function H(t,z) to the automatic differentiation unit 263 and the update unit 28.
- the automatic differentiation section 263 is similar to the automatic differentiation section 253, and calculates the hazard function h(t,z) by automatically differentiating the cumulative hazard function H(t,z).
- the automatic differentiator 263 transmits the calculated hazard function h(t,z) to the updater 28.
- the update unit 28 updates the parameter set ⁇ (a plurality of parameters p1 and p2) from the initialization unit 22 based on the cumulative hazard function H (t, z) and the hazard function h (t, z), and Send to 30.
- the updating unit 28 includes an evaluation function estimating unit 281 and an optimizing unit 282, like the updating unit 27.
- the evaluation function estimation unit 281 calculates the evaluation function L(QS) based on the cumulative hazard function H(t, z) and the hazard function h(t, z).
- the evaluation function L(QS) is, for example, the following negative log likelihood.
- the evaluation function estimation unit 281 transmits the calculated evaluation function L(QS) to the optimization unit 282.
- the optimization unit 282 optimizes the parameter set ⁇ , that is, the plurality of parameters p1 and p2, based on the evaluation function L(QS). For example, an error backpropagation method is used for the optimization. More specifically, the optimization unit 282 calculates the second derivative of the evaluation function L2 (QS) with respect to the parameter set ⁇ (the multiple parameters p1 and p2) using the parameter set ⁇ (the multiple parameters p1 and p2). , ⁇ (multiple parameters p1 and p2). The optimization unit 282 transmits this optimized parameter set ⁇ (a plurality of parameters p1 and p2) to the determination unit 30 as an updated parameter set ⁇ (a plurality of parameters p1 and p2).
- QS evaluation function L2
- the determination unit 30 determines whether the second condition is satisfied based on the update parameter set ⁇ (a plurality of parameters p1 and p2).
- the second condition may be, for example, that the number of times the update parameter set ⁇ is transmitted to the determination unit 30 (that is, the number of parameter update loops) is equal to or greater than a threshold value.
- the second condition may be, for example, that the amount of change in the value of the update parameter set ⁇ before and after the update is less than or equal to a threshold value.
- the second condition will be explained using an example in which the number of times the update parameter set ⁇ is transmitted to the determination unit 30 is two or more times.
- the determination unit 30 determines that the update parameter set ⁇ (the plurality of parameters p1 and p2) is , is sent to the optimization unit 282 and applied to the model 232 and the monotonically increasing neural network 251. Thereby, the determining unit 30 causes the latent expression calculating units 23 and 24, the function estimating units 25 and 26, the updating units 27 and 28, and the determining unit 29 to perform parameter updating operations based on this updated parameter set ⁇ . In other words, if the second condition is not satisfied, the determining unit 30 causes the latent expression calculating units 23 and 24, the function estimating units 25 and 26, the updating units 27 and 28, and the determining unit 29 to execute the parameter update loop again. .
- the determination unit 30 transmits the update parameter set ⁇ (a plurality of parameters p1 and p2) are stored in the learned parameter storage section 31 of the memory 11 as a learned parameter set ⁇ * (a plurality of parameters p1 * and p2 * ).
- the survival analysis device 1 stores the learned parameter set ⁇ * (a plurality of It has a function of storing parameters p1 * and p2 * ).
- FIG. 3 is a block diagram showing an example of the configuration of the prediction function of the survival analysis device 1 as the information processing device according to the first embodiment.
- the CPU of the control circuit 10 loads the prediction program stored in the memory 11 or the storage medium 15 into the RAM.
- the CPU of the control circuit 10 controls the memory 11, the communication module 12, the user interface 13, the drive 14, and the storage medium 15 by interpreting and executing the prediction program developed in the RAM.
- the survival analysis device 1 includes latent expression calculation units 32, 33, function estimation units 34, 35, updating unit 36, determination unit 37, conversion unit 38, and output unit 39. It also functions as a computer.
- the memory 11 of the survival analysis device 1 further functions as a prediction data set storage unit 40 and a prediction target data storage unit 41 for storing information used for prediction operations.
- FIG. 3 shows a case where a plurality of parameters p1 * and p2 * are applied to the model 322 and the monotonically increasing neural network 341 from the learned parameter storage unit 31, respectively.
- the prediction data set storage unit 40 stores a data set corresponding to the task to be predicted (hereinafter referred to as a prediction data set).
- k * is the ID of a task that is not included in the task set K in the learning data set D k . That is, the prediction data set D k * stored in the prediction data set storage unit 40 is a different data set from the learning data set D k .
- the prediction target data storage unit 41 stores prediction target data (hereinafter referred to as prediction target data).
- d k * is the id of data that is not included in the data set DS k * of task k * in the prediction data set D k *. That is, the prediction target data X * stored in the prediction target data storage unit 41 is data that is not included in the prediction data set D k * and the learning data set D k .
- the latent expression calculation unit 32 calculates a latent expression z for the feature quantity x of each data X of the prediction data set D k * based on the prediction data set D k * in the prediction data set storage unit 40. calculate.
- the latent expression calculation unit 32 transmits the calculated latent expression z to the function estimation unit 34.
- the latent expression calculation section 32 has a configuration corresponding to the latent expression calculation section 23. That is, the latent expression calculation unit 32 includes a feature amount extraction unit 321 and a model 322.
- the feature extraction unit 321 extracts the feature x * from the prediction data set D k *.
- the feature extraction unit 321 sends the feature x * to the model 322.
- the model 322 is any differentiable model that can handle the feature quantity x * .
- a plurality of parameters p1 * of the learned parameter set ⁇ * stored in the learned parameter storage unit 31 are applied to the model 322 as weights and bias terms.
- the model 322 to which the plurality of parameters p1 * is applied receives the feature quantity x * as input and outputs the latent expression z * .
- the model 322 sends the output latent expression z * to the function estimator 34.
- the function estimator 34 calculates the hazard function h * (t,z) based on the latent expression z * and the predicted time t.
- the function estimator 34 transmits the calculated hazard function h * (t,z) to the updater 36.
- the function estimator 34 includes a monotonically increasing neural network 341, a cumulative hazard function calculator 342, and an automatic differentiator 343.
- the monotonically increasing neural network 341 is a mathematical model similar to the monotonically increasing neural network 251.
- a plurality of weights and bias terms based on a plurality of parameters p2 * of the learned parameter ⁇ * stored in the learned parameter storage unit 31 are applied to the monotonically increasing neural network 341.
- the monotonically increasing neural network 341 to which the plurality of parameters p2 * is applied calculates an output f * (t,z) as a scalar value according to a monotonically increasing function defined by the latent expression z * and time t.
- the monotonically increasing neural network 341 transmits the output f * (t,z) to the cumulative hazard function calculation unit 342.
- the cumulative hazard function calculation unit 342 is similar to the cumulative hazard function calculation unit 252, and calculates the cumulative hazard function H * (t,z) based on the output f * (t,z).
- the cumulative hazard function calculation unit 342 transmits the calculated cumulative hazard function H * (t,z) to the automatic differentiation unit 343 and the update unit 36.
- the automatic differentiation section 343 is similar to the automatic differentiation section 253, and calculates the hazard function h * (t,z) by automatically differentiating the cumulative hazard function H * (t,z).
- the automatic differentiator 343 transmits the calculated hazard function h * (t,z) to the updater 36.
- the updating unit 36 is similar to the updating unit 27 , and updates the updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * ').
- the updating section 36 transmits this updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * ') to the determining section 37.
- the updating unit 36 includes an evaluation function estimating unit 361 and an optimizing unit 362, like the updating unit 27.
- the evaluation function estimation unit 361 calculates the evaluation function L*(D) based on the cumulative hazard function H * (t,z) and the hazard function h * (t,z).
- the evaluation function L * (D) is, for example, the following negative log likelihood.
- the evaluation function estimation unit 361 transmits the calculated evaluation function L * (D) to the optimization unit 362.
- the optimization unit 362 optimizes the parameter set ⁇ * , that is, the plurality of parameters p1 * and p2 * , based on the evaluation function L * (D). Similar to the optimization unit 272, the optimization uses, for example, the error backpropagation method.
- the optimization unit 362 uses this optimized parameter set ⁇ * (a plurality of parameters p1 * and p2 * ) as an updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * '), and passes it to the determination unit 37. Send to.
- the determining unit 37 determines whether the first condition is satisfied based on the updated parameter set ⁇ * ' (the plurality of parameters p1 * ' and p2 * '). If the first condition is not satisfied, the determination unit 37 applies the updated parameter set ⁇ * ′ (a plurality of parameters p1 * ′ and p2 * ′) to the model 322 and the monotonically increasing neural network 341, and the latent expression calculation unit 32 , the function estimation section 34 and the updating section 36 are caused to perform parameter updating operations based on this updated parameter set ⁇ * '.
- the determining unit 37 causes the latent expression calculating unit 32, the function estimating unit 34, and the updating unit 36 to repeatedly execute a parameter updating loop. Further, if the first condition is satisfied, the determination unit 37 ends the parameter update loop, and finally updates the updated parameter set ⁇ * ' (the plurality of parameters p1 * ' and p2 * ') to the latent expression calculation unit. 33 and the function estimator 35. In other words, the determining unit 37 initializes the parameters applied to the latent expression calculating unit 33 and the function estimating unit 35 to this updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * ').
- the latent expression calculation unit 33 calculates the latent expression z based on the prediction target data X k * input by the user through the user interface 13 and stored in the prediction target data storage unit 41, for example.
- the latent expression calculation unit 33 transmits the calculated latent expression z to the function estimation unit 34.
- the latent expression calculation unit 33 has a configuration corresponding to that of the latent expression calculation unit 23. That is, the latent expression calculation unit 33 includes a feature amount extraction unit 331 and a model 332.
- the feature amount extraction unit 331 extracts the feature amount x * from the prediction target data Xk *.
- the feature extraction unit 331 transmits the feature x * to the model 332.
- the model 332 is any differentiable model that can handle the feature quantity x * .
- the updated parameters p1 * ' are applied to the model 332 as weights and bias terms.
- the model 332 to which the plurality of parameters p1 * ' is applied receives the feature amount x * as input and outputs the latent expression z * .
- the model 332 sends the output latent expression z * to the function estimator 35.
- the function estimation unit 35 calculates the hazard function h * (t,z) based on the latent expression z * and the predicted time t.
- the function estimation unit 35 transmits the calculated hazard function h * (t,z) to the output unit 39.
- the function estimator 35 includes a monotonically increasing neural network 351, a cumulative hazard function calculator 352, and an automatic differentiator 353.
- the monotonically increasing neural network 351 is a mathematical model similar to the monotonically increasing neural network 251.
- a plurality of weights and bias terms based on the updated plurality of parameters p2 * ' are applied to the monotonically increasing neural network 351.
- the monotonically increasing neural network 351 to which the plurality of parameters p2 * ' is applied calculates an output f * (t,z) as a scalar value according to a monotonically increasing function defined by the latent expression z * and time t.
- the monotonically increasing neural network 351 transmits the output f * (t,z) to the cumulative hazard function calculation unit 352.
- the cumulative hazard function calculation unit 352 is similar to the cumulative hazard function calculation unit 252, and calculates the cumulative hazard function H * (t,z) based on the output f * (t,z).
- the cumulative hazard function calculation unit 352 transmits the calculated cumulative hazard function H * (t,z) to the automatic differentiation unit 353, the conversion unit 38, and the output unit 39.
- the automatic differentiation section 353 is similar to the automatic differentiation section 253, and calculates the hazard function h * (t,z) by automatically differentiating the cumulative hazard function H * (t,z).
- the automatic differentiation section 353 transmits the calculated hazard function h * (t,z) to the output section 39.
- the conversion unit 38 converts the cumulative hazard function H * (t,z) sent from the cumulative hazard function calculation unit 352 into a survival function S * (t,z).
- the conversion unit 38 transmits the converted survival function S * (t,z) to the output unit 39.
- the output unit 39 sets the hazard function h * (t, z) transmitted from the automatic differentiation unit 353 as the hazard function h * (t
- the survival analysis device 1 can calculate the prediction target data X k stored in the prediction target data storage unit 41 based on the prediction data set D k * stored in the prediction data set storage unit 40 . It has a function to calculate the hazard function h * (t
- FIGS. 4A and 4B are a series of flowcharts showing an example of the learning operation in the survival analysis device 1 as the information processing device according to the first embodiment.
- the learning data set D k is stored in the learning data set storage unit 20 in the memory 11 in advance.
- the initialization unit 22 in response to a user's instruction to start a learning operation (start), the initialization unit 22 initializes a parameter set ⁇ (a plurality of parameters p1 and p2) based on an arbitrary rule R. (Step S10). For example, the initialization unit 22 initializes the plurality of parameters p1 and p2 based on an arbitrary rule R.
- the plurality of parameters p1 and p2 initialized by the process in step S10 are applied to the model 232 and the monotonically increasing neural network 251, respectively. Furthermore, this initialized parameter set ⁇ (a plurality of parameters p1 and p2) is sent to the optimization unit 282.
- the data dividing unit 21 randomly extracts a learning target data set for task k from the learning data set D k stored in the learning data set storage unit 20 . Subsequently, the data dividing unit 21 further extracts a support set SS and a query set QS from the extracted learning target data set (step S11).
- the feature extraction unit 231 extracts the feature x of each piece of data X in the support set SS from the support set SS extracted in the process of step S11 (step S12).
- the model 232 to which the plurality of parameters p1 initialized in the process of step S10 is applied uses as input the feature quantity x of the individual data X of the support set SS extracted in the process of step S12, and generates the latent expression z. Calculate (step S13).
- the monotonically increasing neural network 251 to which the plurality of parameters p2 initialized in the process of step S10 is applied outputs according to the monotonically increasing function defined by the latent expression z calculated in the process of step S13 and the time t. f(e,z) and f(0,z) are calculated (step S14).
- the cumulative hazard function calculating unit 252 calculates a cumulative hazard function H(e,z) based on the outputs f(e,z) and f(0,z) calculated in the process of step S14 (step S15). .
- the automatic differentiation unit 253 calculates a hazard function h(e,z) based on the cumulative hazard function H(e,z) calculated in the process of step S15 (step S16).
- the update unit 27 updates the update parameter set ⁇ ' (a plurality of parameters p1' and p2') are calculated (step S17). Specifically, the evaluation function estimation unit 271 calculates the evaluation function L(SS) based on the cumulative hazard function H(e, z) and the hazard function h(e, z). The optimization unit 272 uses the error backpropagation method to generate a plurality of optimized parameters p1' and p2' based on the evaluation function L(SS), that is, an update parameter set ⁇ ' (a plurality of parameters p1' and p2' ) is calculated.
- the determination unit 29 determines whether the first condition is satisfied based on the updated parameter set ⁇ ' (a plurality of parameters p1' and p2') (step S18).
- the determination unit 29 updates the parameters to be applied to the model 232 and the monotonically increasing neural network 251 from the parameter set ⁇ calculated in the process of step S17.
- the parameter set ⁇ ' (a plurality of parameters p1' and p2') is updated (step S19). Specifically, the determination unit 29 applies the plurality of optimized parameters p1' and p2' to the model 232 and the monotonically increasing neural network 251.
- the processes of steps S13 to S19 are executed.
- the updating process of the update parameter set ⁇ ' (the plurality of parameters p1' and p2') is repeated until it is determined in the process of step S18 that the first condition is satisfied.
- step S19 If the first condition is satisfied (step S19; YES), as shown in FIG. 4B, the determination unit 29 updates the parameters applied to the model 242 and the monotonically increasing neural network 261 in the process of step S18.
- the updated parameter set ⁇ ' (a plurality of parameters p1' and p2') is initialized (step S20).
- the feature amount extraction unit 241 extracts the feature amount x of each data X of the query set QS from the query set QS extracted in the process of step S11 above (step S21).
- the model 242 to which the plurality of parameters p1' initialized in the process of step S20 is applied uses as input the feature quantity x of the individual data X of the query set QS extracted in the process of step S21, and generates a latent expression z. is calculated (step S22).
- the monotonically increasing neural network 261 to which the plurality of parameters p2' initialized in the process of step S20 is applied follows the monotonically increasing function defined by the latent expression z calculated in the process of step S22 and the time t.
- Outputs f(e, z) and f(0, z) are calculated (step S23).
- the cumulative hazard function calculating unit 262 calculates the cumulative hazard function H(e,z) based on the outputs f(e,z) and f(0,z) calculated in the process of step S23 (step S24). .
- the automatic differentiation unit 263 calculates the hazard function h(e,z) based on the cumulative hazard function H(e,z) calculated in the process of step S24 (step S25).
- the update unit 28 updates the parameter set ⁇ (a plurality of parameters p1 and p2) initialized in the process of step S10, the cumulative hazard function H(e, z) calculated in the process of step S24, and the cumulative hazard function H(e,z) calculated in step S25.
- An updated parameter set ⁇ (a plurality of parameters p1 and p2) is calculated based on the calculated hazard function h(e, z) (step S26).
- the evaluation function estimation unit 281 calculates the evaluation function L(QS) based on the cumulative hazard function H(e, z) and the hazard function h(e, z).
- the optimization unit 282 uses the error backpropagation method to calculate an optimized update parameter set ⁇ (a plurality of parameters p1 and p2) based on the evaluation function L(QS).
- the determination unit 30 determines whether the second condition is satisfied based on the updated parameter set ⁇ (a plurality of parameters p1 and p2) (step S27).
- the second condition is, for example, that the number of times the update parameter set ⁇ has been transmitted to the determination unit 30 is two or more times.
- the determination unit 30 determines that the second condition is not satisfied. If the second condition is not satisfied (step S27; NO), the determination unit 30 selects the parameters to be applied to the model 232 and the monotonically increasing neural network 251 from the parameter set ⁇ ' (the plurality of parameters p1' and p2') to the updated parameter set ⁇ (a plurality of parameters p1 and p2) calculated in the process of step S26 (step S28). Specifically, the determination unit 30 applies the plurality of optimized parameters p1 and p2 to the model 232 and the monotonically increasing neural network 251.
- the above steps S11 to S26 are executed based on the updated parameter set ⁇ (a plurality of parameters p1 and p2) updated in the process of step S28.
- the updated parameter set ⁇ (the plurality of parameters p1 and p2) is calculated again.
- step S27 the determination unit 30 determines that the second condition is satisfied. If the second condition is satisfied in this way (step S27; YES), the determination unit 30 converts the updated parameter set ⁇ (the plurality of parameters p1 and p2) calculated in step S26 into the learned parameter set ⁇ * ( A plurality of parameters p1 * and p2 * ) are stored in the learned parameter storage unit 31 (step S29).
- step S29 When the process of step S29 is finished, the learning operation in the survival analysis device 1 ends (end).
- FIG. 5 is a flowchart showing an example of prediction operation in the survival analysis device 1 as the information processing device according to the first embodiment.
- the prediction data set D k * is stored in the prediction data set storage unit 40 in the memory 11 due to a learning operation performed in advance.
- prediction target data X k * is stored in the prediction target data storage section 41 in the memory 11.
- the parameters to be applied to the model 322 and the monotonically increasing neural network 341 are the learned parameters stored in the learned parameter storage unit 31.
- the set ⁇ * (a plurality of parameters p1 * and p2 * ) is initialized (step S30).
- the feature amount extraction unit 321 extracts the feature amount x * of each data X of the prediction data set D k * from the prediction data set D k * stored in the prediction data set storage unit 40 (step S31 ).
- the model 322 to which the plurality of parameters p1 initialized in the process of step S30 is applied calculates a latent expression z * by inputting the feature quantity x * extracted in the process of step S31 (step S32).
- the monotonically increasing neural network 341 to which the plurality of parameters p2 initialized in the process of step S30 is applied follows the monotonically increasing function defined by the latent expression z * calculated in the process of step S32 and the time t.
- Outputs f * (e,z) and f * (0,z) are calculated (step S33).
- the cumulative hazard function calculating unit 342 calculates a cumulative hazard function H*(e,z) based on the outputs f * (e,z) and f * (0,z) calculated in the process of step S33 ( Step S34).
- the automatic differentiation unit 343 calculates a hazard function h * (e,z) based on the cumulative hazard function H * (e,z) calculated in the process of step S34 (step S35).
- the updating unit 36 updates the updated parameter set ⁇ * ( A plurality of parameters p1 * and p2 * ) are calculated (step S36). Specifically, the evaluation function estimation unit 361 calculates the evaluation function L * (D) based on the cumulative hazard function H * (e,z) and the hazard function h * (e,z). The optimization unit 362 uses the error backpropagation method to optimize the plurality of parameters p1 *' and p2 *' based on the evaluation function L(D), that is, the updated parameter set ⁇ * ' (the plurality of parameters p1 * ' and p2 * ').
- the determining unit 37 determines whether the first condition is satisfied based on the updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * ') (step S37).
- step S37 If the first condition is not satisfied (step S37; NO), the determination unit 37 selects the parameters to be applied to the model 322 and the monotonically increasing neural network 341 from the parameter set ⁇ * calculated in the process of step S36.
- the updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * ') is updated (step S38). Specifically, the determination unit 37 applies the plurality of optimized parameters p1 * ′ and p2 * ′ to the model 322 and the monotonically increasing neural network 341.
- step S37 If the first condition is satisfied (step S37; YES), the determination unit 37 sets the parameters to be applied to the model 332 and the monotonically increasing neural network 351 to the updated parameter set ⁇ * that was last updated in the process of step S36 above. ' (a plurality of parameters p1 *' and p2 *' ) (step S39).
- the feature amount extraction unit 331 extracts the feature amount x * from the prediction target data Xk * stored in the prediction target data storage unit 41 (step S40).
- the model 332 to which the plurality of parameters p1 *' initialized in the process of step S39 is applied calculates a latent expression z * by inputting the feature quantity x * extracted in the process of step S40 (step S41 ).
- the monotonically increasing neural network 351 to which the plurality of parameters p2 *' initialized in the process of step S39 is applied has a monotonically increasing function defined by the latent expression z * calculated in the process of step S41 and the time t. Accordingly, outputs f * (e,z) and f * (0,z) are calculated (step S42).
- the cumulative hazard function calculating unit 352 calculates a cumulative hazard function H*(e,z) based on the outputs f * (e,z) and f * (0,z) calculated in the process of step S42 ( Step S43).
- the automatic differentiation unit 353 calculates a hazard function h * (e,z) based on the cumulative hazard function H * (e,z) calculated in the process of step S42 (step S44).
- the conversion unit 38 calculates the survival function S * (e,z) based on the cumulative hazard function H * (e,z) calculated in the process of step S43 (step S45).
- the output unit 39 sets the hazard function h * (t,z) calculated in the process of step S44 above as the hazard function h* ( t
- step S46 ends, the prediction operation in the survival analysis device 1 ends (end).
- the monotonically increasing neural network 351 performs latent expression calculation that calculates a latent expression representing a feature amount from processing target data including a feature amount related to a prediction target event. It is configured to output a scalar value according to a monotonically increasing function defined by the latent expression calculated by the unit 33 and the time.
- the cumulative hazard function calculation unit 352 and automatic differentiation unit 353 of the function estimation unit 35 estimate a hazard function based on the scalar value output from the monotonically increasing neural network 351. In this way, by modeling using the monotonically increasing neural network 351, integral calculations by approximation can be avoided. Therefore, it is possible to calculate the hazard function for the prediction target data without making any assumptions.
- the learning function configuration (for learning The data set storage unit 20 to determination unit 29) are further provided. Therefore, even if there is not enough prediction data in which an event to be predicted has occurred, it is possible to calculate a hazard function for the prediction target data.
- parameters that update the parameters learned by the learning function configuration based on a plurality of prediction data including feature amounts related to the prediction target event stored in the prediction data set storage unit 40 It includes a latent expression calculating section 32, a function estimating section 34, and an updating section 36, which function as an updating section. Therefore, by updating the parameters learned through meta-learning using MAML to parameters according to the prediction target data, a more accurate hazard function can be calculated.
- the cumulative hazard function calculation unit 352 of the function estimation unit 35 calculates the cumulative hazard function based on the scalar value output from the monotonically increasing neural network 351, and the automatic differentiation unit 353 of the function estimation unit 35 calculates the cumulative hazard function based on the scalar value output from the monotonically increasing neural network 351.
- the hazard function is calculated by automatically differentiating the cumulative hazard function calculated by the function calculation unit 352. In this way, a hazard function can be calculated based on a monotonically increasing function.
- the converter further includes a converter 38 that converts the cumulative hazard function calculated by the cumulative hazard function calculator 352 of the function estimator 35 into a survival function. Therefore, it is possible to calculate the survival function as well. In this way, according to the first embodiment, it is possible to calculate at least one of the hazard function and the survival function for the prediction target data without any assumptions.
- ⁇ is a monotonically increasing and second-order differentiable function whose range is defined by [0, 1], such as a sigmoid function.
- the hazard function h(t) is calculated by automatically differentiating the survival function S(t). Therefore, the hazard function and survival function can be calculated without calculating the cumulative hazard function as in the first embodiment.
- a survival analysis device will be described as an example of the information processing device according to the second embodiment, similar to the first embodiment.
- the configuration and operation that are different from the first embodiment will be mainly explained. Descriptions of configurations and operations equivalent to those of the first embodiment will be omitted as appropriate.
- FIG. 6 is a block diagram showing an example of the learning function configuration of the survival analysis device 1 as an information processing device according to the second embodiment.
- FIG. 6 corresponds to FIG. 2 in the first embodiment.
- the survival analysis device 1 includes a data division section 51, an initialization section 52, latent expression calculation sections 53 and 54, function estimation sections 55 and 56, update sections 57 and 58, and a determination section 59, 60 functions as a computer. Furthermore, the memory 11 of the survival analysis device 1 functions as a learning data set storage section 50 and a learned parameter storage section 61 for storing information used for learning operations.
- the configurations of the learning dataset storage unit 50 and the data division unit 51 are the same as the configurations of the learning dataset storage unit 20 and the data division unit 21 in FIG. 2 of the first embodiment. That is, the data division unit 51 extracts the support set SS and the query set QS from the learning data set storage unit 50.
- the configuration of the initialization unit 52 is equivalent to the configuration of the initialization unit 22 in FIG. 2 of the first embodiment. That is, the initialization unit 52 initializes the parameter set ⁇ (a plurality of parameters p1 and p2) based on a predetermined arbitrary rule R. The initialization unit 52 transmits the plurality of initialized parameters p1 to the latent expression calculation unit 53, and transmits the plurality of initialized parameters p2 to the function estimation unit 55. Furthermore, the initialization unit 52 transmits the initialized parameter set ⁇ (a plurality of parameters p1 and p2) to the update unit 58.
- the configuration of the latent expression calculation unit 53 is equivalent to the configuration of the latent expression calculation unit 23 in FIG. 2 of the first embodiment, and includes a feature amount extraction unit 531 and a model 532. That is, the latent expression calculation unit 53 calculates a latent expression z for the feature amount x of each piece of data X in the support set SS, based on the support set SS. The latent expression calculation unit 53 transmits the calculated latent expression z to the function estimation unit 55.
- the function estimation unit 55 calculates the survival function S(t, z) and the hazard function h(t, z) based on the latent expression z and time t.
- the function estimating unit 55 transmits the calculated survival function S(t, z) and hazard function h(t, z) to the updating unit 57.
- the function estimation section 55 includes a monotonically increasing neural network 551, a survival function calculation section 552, and an automatic differentiation section 553.
- the configurations of the monotonically increasing neural network 551 and the automatic differentiator 553 are the same as the configurations of the monotonically increasing neural network 251 and the automatic differentiator 253 in FIG. 2 of the first embodiment.
- the monotonically increasing neural network 551 to which the plurality of parameters p2 are applied calculates the output f(t, z) according to a monotonically increasing function defined by the latent expression z and time t.
- the monotonically increasing neural network 551 transmits the calculated output f(t,z) to the survival function calculation unit 552.
- the survival function calculation unit 552 calculates the survival function S(t,z) based on the output f(t,z) from the monotonically increasing neural network 551.
- the survival function calculation unit 552 transmits the calculated survival function S(t,z) to the automatic differentiation unit 553. Furthermore, the survival function calculating unit 552 transmits the calculated survival function S(t,z) to the updating unit 57.
- the automatic differentiation section 553 calculates the hazard function h(t,z) by automatically differentiating the survival function S(t,z).
- the automatic differentiator 553 transmits the calculated hazard function h(t,z) to the updater 57.
- the updating unit 57 calculates an updated parameter set ⁇ ' (a plurality of parameters p1' and p2') based on the survival function S(t, z) and the hazard function h(t, z).
- the updating section 57 transmits the updated parameter set ⁇ ' (a plurality of parameters p1' and p2') to the determining section 59.
- the update section 57 includes an evaluation function estimation section 571 and an optimization section 572.
- the configuration of the evaluation function estimator 571 is the same as the evaluation function estimator 271 in FIG. 2 of the first embodiment, except that the survival function S(t, z) is used instead of the cumulative hazard function H(t, z). This is equivalent to the configuration of The evaluation function estimation unit 571 calculates the evaluation function L(SS) based on the survival function S(t, z) and the hazard function h(t, z). The evaluation function estimation unit 571 transmits the calculated evaluation function L(SS) to the optimization unit 572.
- the optimization unit 572 optimizes the parameter set ⁇ , that is, the plurality of parameters p1 and p2, based on the evaluation function L(SS). For example, an error backpropagation method is used for the optimization.
- the optimization unit 572 transmits this optimized parameter set ⁇ (a plurality of parameters p1 and p2) to the determination unit 59 as an updated parameter set ⁇ ' (a plurality of parameters p1' and p2').
- the determining unit 59 is equivalent to the determining unit 29 in FIG. 2 of the first embodiment. That is, the determination unit 59 determines whether the first condition is satisfied based on the update parameter set ⁇ ' (a plurality of parameters p1' and p2'). If the first condition is not satisfied, the determination unit 59 causes the latent expression calculation unit 53, function estimation unit 55, and update unit 57 to repeatedly execute a parameter update loop. If the first condition is satisfied, the determination unit 59 terminates the parameter update loop, and updates the last updated parameter set ⁇ ′ (the plurality of parameters p1′ and p2′) to the latent expression calculation unit 54 and It is transmitted to the function estimator 56. In other words, the determining unit 59 initializes the parameters applied to the latent expression calculating unit 54 and the function estimating unit 56 to this updated parameter set ⁇ ' (a plurality of parameters p1' and p2').
- the configuration of the latent expression calculation unit 54 is equivalent to the configuration of the latent expression calculation unit 24 in FIG. 2 of the first embodiment, and includes a feature amount extraction unit 541 and a model 542. That is, the latent expression calculation unit 54 calculates a latent expression z for the feature amount x of each data X of the query set QS based on the query set QS. The latent expression calculation unit 54 transmits the calculated latent expression z to the function estimation unit 56.
- the function estimation unit 56 calculates the survival function S(t, z) and the hazard function h(t, z) based on the latent expression z and the predicted time t.
- the function estimating unit 56 transmits the calculated survival function S(t, z) and hazard function h(t, z) to the updating unit 58.
- the function estimator 56 includes a monotonically increasing neural network 561, a survival function calculator 562, and an automatic differentiator 563.
- the configurations of the monotonically increasing neural network 551 and the automatic differentiator 553 are equivalent to the configurations of the monotonically increasing neural network 261 and the automatic differentiator 263 in FIG. 2 of the first embodiment.
- the monotonically increasing neural network 561 to which the plurality of parameters p2' is applied calculates the output f(t, z) according to a monotonically increasing function defined by the latent expression z and time t.
- the monotonically increasing neural network 561 transmits the output f(t,z) to the survival function calculation unit 562.
- the survival function calculation unit 562 is similar to the survival function calculation unit 552, and calculates the survival function S(t,z) based on the output f(t,z) from the monotonically increasing neural network 561.
- the survival function calculation unit 562 transmits the calculated survival function S(t,z) to the automatic differentiation unit 563 and the update unit 58.
- the automatic differentiation section 563 is similar to the automatic differentiation section 553, and calculates the hazard function h(t, z) by automatically differentiating the survival function S(t, z).
- the automatic differentiator 563 transmits the calculated hazard function h(t,z) to the updater 58.
- the update unit 58 updates the parameter set ⁇ (a plurality of parameters p1 and p2) from the initialization unit 52 based on the survival function S (t, z) and the hazard function h (t, z), and Send to.
- the update unit 58 includes an evaluation function estimation unit 581 and an optimization unit 582.
- the evaluation function estimator 581 has the same configuration as the evaluation function estimator 281 in FIG. 2 of the first embodiment, except that the survival function S(t, z) is used instead of the cumulative hazard function H(t, z). is equivalent to
- the evaluation function estimation unit 581 calculates the evaluation function L (QS) based on the survival function S (t, z) and the hazard function h (t, z).
- the evaluation function estimation unit 581 transmits the calculated evaluation function L(QS) to the optimization unit 582.
- the optimization unit 582 optimizes the parameter set ⁇ , that is, the plurality of parameters p1 and p2, based on the evaluation function L(QS). For example, an error backpropagation method is used for the optimization. More specifically, the optimization unit 582 calculates the second derivative of the evaluation function L2 (QS) with respect to the parameter set ⁇ (the multiple parameters p1 and p2) using the parameter set ⁇ (the multiple parameters p1 and p2). , optimize the parameter set ⁇ (a plurality of parameters p1 and p2). The optimization unit 582 transmits this optimized parameter set ⁇ (a plurality of parameters p1 and p2) to the determination unit 60 as an updated parameter set ⁇ (a plurality of parameters p1 and p2).
- QS evaluation function L2
- the determining unit 60 is equivalent to the determining unit 30 in FIG. 2 of the first embodiment. That is, the determination unit 60 determines whether the second condition is satisfied based on the updated parameter set ⁇ (a plurality of parameters p1 and p2).
- the determination unit 60 transmits the updated parameter set ⁇ (a plurality of parameters p1 and p2) to the optimization unit 582 and applies it to the model 532 and the monotonically increasing neural network 551. Thereby, the determining unit 60 causes the latent expression calculating units 53 and 54, the function estimating units 55 and 56, the updating units 57 and 58, and the determining unit 59 to perform parameter updating operations based on this updated parameter set ⁇ . In other words, if the second condition is not satisfied, the determination unit 60 causes the latent expression calculation units 53, 54, function estimation units 55, 56, update units 57, 58, and determination unit 59 to execute the parameter update loop again. .
- the determination unit 60 sets the updated parameter set ⁇ (the plurality of parameters p1 and p2) as the learned parameter set ⁇ * (the plurality of parameters p1 * and p2 * )
- the parameters are stored in the parameter storage unit 61.
- the survival analysis device 1 stores the learned parameter set ⁇ * (a plurality of It has a function of storing parameters p1 * and p2 * ).
- FIG. 7 is a block diagram illustrating an example of a prediction function configuration of the survival analysis device 1 as an information processing device according to the second embodiment.
- FIG. 7 corresponds to FIG. 3 in the first embodiment.
- the survival analysis device 1 further functions as a computer including latent expression calculation units 62 and 63, function estimation units 64 and 65, an update unit 66, a determination unit 67, and an output unit 68. Furthermore, the memory 11 of the survival analysis device 1 further functions as a prediction data set storage unit 69 and a prediction target data storage unit 70 for storing information used for prediction operations. Note that FIG. 7 shows a case where a plurality of parameters p1 * and p2 * are applied to the model 622 and the monotonically increasing neural network 641 from the learned parameter storage unit 61, respectively.
- the configurations of the prediction data set storage unit 69 and the prediction target data storage unit 70 are equivalent to the configurations of the prediction data set storage unit 40 and the prediction target data storage unit 41 in FIG. 3 of the first embodiment.
- the configuration of the latent expression calculation unit 62 is the same as the configuration of the latent expression calculation unit 32 in FIG. 3 of the first embodiment, and includes a feature amount extraction unit 621 and a model 622. That is, the latent expression calculation unit 62 calculates a latent expression for the feature x of each data X of the prediction data set D k * based on the prediction data set D k * in the prediction data set storage unit 69. Calculate z. The latent expression calculating unit 62 transmits the calculated latent expression z to the monotonically increasing neural network 641 in the function estimating unit 64.
- the function estimation unit 64 calculates the survival function S * (t, z) and the hazard function h * (t, z) based on the latent expression z * and the predicted time t.
- the function estimator 64 transmits the calculated survival function S * (t,z) and hazard function h * (t,z) to the updater 66.
- the function estimation section 64 includes a monotonically increasing neural network 641, a survival function calculation section 642, and an automatic differentiation section 643, similar to the function estimation section 55.
- the configurations of the monotonically increasing neural network 641 and the automatic differentiator 643 are equivalent to the configurations of the monotonically increasing neural network 341 and the automatic differentiator 343 in FIG. 3 of the first embodiment.
- the monotonically increasing neural network 641 to which the plurality of parameters p2 * are applied calculates the output f * (z,t) according to a monotonically increasing function defined by the latent representation z * and the time t.
- the monotonically increasing neural network 641 transmits the calculated output f * (z,t) to the survival function calculation unit 642.
- the survival function calculation unit 642 is similar to the survival function calculation unit 552, and calculates the survival function S * (t,z) based on the output f * (t,z) from the monotonically increasing neural network 641.
- the survival function calculation unit 642 transmits the calculated survival function S * (t,z) to the automatic differentiation unit 643. Furthermore, the survival function calculating unit 642 transmits the calculated survival function S * (t,z) to the updating unit 66.
- the automatic differentiation section 643 is similar to the automatic differentiation section 553, and calculates the hazard function h * (t,z) by automatically differentiating the survival function S * (t,z).
- the automatic differentiator 643 transmits the calculated hazard function h * (t,z) to the updater 66.
- the update unit 66 calculates an updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * ') based on the survival function S * (t,z) and the hazard function h * (t,z).
- the updating section 66 transmits this updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * ') to the determining section 67.
- the updating section 66 is similar to the updating section 57 and includes an evaluation function estimation section 661 and an optimization section 662.
- the evaluation function estimation unit 661 is similar to the evaluation function estimation unit 571, and calculates the evaluation function L * (D) based on the survival function S * (t, z) and the hazard function h * (t, z). calculate.
- the evaluation function estimation unit 661 transmits the calculated evaluation function L * (D) to the optimization unit 662.
- the optimization unit 662 is similar to the optimization unit 572, and optimizes the parameter set ⁇ * , that is, the plurality of parameters p1 * and p2 * , based on the evaluation function L * (D). For example, an error backpropagation method is used for the optimization.
- the optimization unit 662 uses this optimized parameter set ⁇ * (a plurality of parameters p1 * and p2 * ) as an updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * '), and the determination unit 67 Send to.
- the determining unit 67 determines whether the first condition is satisfied based on the updated parameter set ⁇ * ' (the plurality of parameters p1 * ' and p2 * '). If the first condition is not satisfied, the determination unit 67 applies the updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * ') to the model 622 and the monotonically increasing neural network 641. That is, if the first condition is not satisfied, the determining unit 67 causes the latent expression calculating unit 62, the function estimating unit 64, and the updating unit 66 to repeatedly execute a parameter updating loop.
- the determination unit 67 ends the parameter update loop, and finally updates the updated parameter set ⁇ * ' (the plurality of parameters p1 * ' and p2 * ') to the latent expression calculation unit. 63 and the function estimator 65.
- the determining unit 67 initializes the parameters applied to the latent expression calculating unit 63 and the function estimating unit 65 to this updated parameter set ⁇ * ' (the plurality of parameters p1 * ' and p2 * ').
- the configuration of the latent expression calculation unit 63 is the same as the configuration of the latent expression calculation unit 33 in FIG. 3 of the first embodiment, and includes a feature amount extraction unit 631 and a model 632. That is, the latent expression calculation unit 63 calculates the latent expression z based on the prediction target data X k * input by the user through the user interface 13 and stored in the prediction target data storage unit 70, for example. The latent expression calculation unit 63 transmits the calculated latent expression z * to the function estimation unit 65.
- the function estimation unit 65 is similar to the function estimation unit 56, and calculates the survival function S * (t, z) and the hazard function h * (t, z) based on the latent expression z * and the predicted time t. .
- the function estimation unit 65 transmits the calculated survival function S * (t,z) and hazard function h * (t,z) to the output unit 68.
- the function estimation section 65 includes a monotonically increasing neural network 651, a survival function calculation section 652, and an automatic differentiation section 653, similar to the function estimation section 56.
- the configurations of the monotonically increasing neural network 651 and the automatic differentiator 653 are equivalent to the configurations of the monotonically increasing neural network 351 and the automatic differentiator 353 in FIG. 3 of the first embodiment.
- the monotonically increasing neural network 651 to which the plurality of parameters p2 * ' is applied calculates the output f * (t,z) according to a monotonically increasing function defined by the latent representation z * and the time t.
- the monotonically increasing neural network 651 transmits the output f * (t,z) to the survival function calculation unit 652.
- the survival function calculation unit 652 is similar to the survival function calculation unit 562, and calculates the survival function S * (t,z) based on the output f * (t,z) from the monotonically increasing neural network 651.
- the survival function calculation unit 652 transmits the calculated survival function S * (t,z) to the automatic differentiation unit 653 and the output unit 68.
- the automatic differentiation section 653 is similar to the automatic differentiation section 563, and calculates the hazard function h * (t,z) by automatically differentiating the survival function S * (t,z). Automatic differentiation section 653 transmits the calculated hazard function h * (t,z) to output section 68.
- the output unit 68 sets the hazard function h * (t, z) sent from the automatic differentiation unit 653 to the hazard function h * (t
- the survival analysis device 1 can calculate the prediction target data X k stored in the prediction target data storage unit 70 based on the prediction data set D k * stored in the prediction data set storage unit 69. It has a function to calculate the hazard function h * (t
- FIGS. 8A and 8B are a series of flowcharts showing an example of the learning operation in the survival analysis device 1 as the information processing device according to the second embodiment. 8A and 8B correspond to FIGS. 4A and 4B in the first embodiment. In the examples of FIGS. 8A and 8B, it is assumed that the learning data set D k is stored in the learning data set storage unit 50 in the memory 11 in advance.
- steps S50 to S53 are executed.
- the processing in steps S50 to S53 is equivalent to the processing in steps S10 to S13 in FIG. 4A of the first embodiment. That is, the initialization unit 52 initializes the parameter set ⁇ (a plurality of parameters p1 and p2) based on an arbitrary rule R (step S50).
- the data division unit 51 randomly extracts a learning target data set for a task k from the learning data set D k stored in the learning data set storage unit 50, and extracts a support set SS and a query from the extracted learning target data set. Further extract the set QS (step S51).
- the feature extraction unit 531 extracts the feature x of each piece of data X in the support set SS from the support set SS extracted in step S51 (step S52).
- the model 532 to which the plurality of parameters p1 initialized in the process of step S50 is applied uses as input the feature quantity x of the individual data X of the support set SS extracted in the process of step S52, and generates the latent expression z. Calculate (step S53).
- the monotonically increasing neural network 551 to which the plurality of parameters p2 initialized in the process of step S50 is applied outputs according to the monotonically increasing function defined by the latent expression z calculated in the process of step S53 and the time t. f(t,z) is calculated (step S54).
- the survival function calculation unit 552 calculates the survival function S(t,z) based on the output f(t,z) calculated in the process of step S54 (step S55).
- the automatic differentiation unit 553 calculates the hazard function h(e, z) based on the survival function S(t, z) calculated in the process of step S55 (step S56).
- the updating unit 57 updates the updated parameter set ⁇ ' (a plurality of parameters p1' and p2') are calculated (step S57).
- the evaluation function estimation unit 571 calculates the evaluation function L (SS) based on the survival function S (t, z) and the hazard function h (t, z).
- the optimization unit 572 uses the error backpropagation method to optimize a plurality of parameters p1' and p2' based on the evaluation function L(SS), that is, an update parameter set ⁇ ' (a plurality of parameters p1' and p2'). ) is calculated.
- step S58 to step S62 are executed.
- the processing in steps S58 to S62 is equivalent to the processing in steps S18 to S22 in FIGS. 4A and 4B of the first embodiment. That is, the determination unit 59 determines whether the first condition is satisfied based on the updated parameter set ⁇ ' (a plurality of parameters p1' and p2') (step S58). If the first condition is not satisfied (step S58; NO), the determination unit 59 updates the parameters to be applied to the model 532 and the monotonically increasing neural network 551 from the parameter set ⁇ calculated in the process of step S57.
- the parameter set ⁇ ' (a plurality of parameters p1' and p2') is updated (step S59).
- the determination unit 59 applies the plurality of optimized parameters p1' and p2' to the model 532 and the monotonically increasing neural network 551. Then, based on the updated parameter set ⁇ ' (a plurality of parameters p1' and p2') updated in the process of step S59, the processes of steps S53 to S59 are executed. As a result, the updating process of the update parameter set ⁇ ' (the plurality of parameters p1' and p2') is repeated until it is determined in the first process of step S58 that the condition is satisfied.
- step S58 If the first condition is satisfied (step S58; YES), as shown in FIG. 8B, the determination unit 59 updates the parameters applied to the model 542 and the monotonically increasing neural network 561 in the process of step S57.
- the updated parameter set ⁇ ' (a plurality of parameters p1' and p2') is initialized (step S60).
- the feature amount extraction unit 541 extracts the feature amount x of each data X of the query set QS from the query set QS extracted in the process of step S51 (step S61).
- the model 542 to which the plurality of parameters p1' initialized in the process of step S60 is applied uses as input the feature quantity x of the individual data X of the query set QS extracted in the process of step S41, and generates the latent expression z. is calculated (step S62).
- the monotonically increasing neural network 561 to which the plurality of parameters p2' initialized in the process of step S60 is applied follows the monotonically increasing function defined by the latent expression z calculated in the process of step S62 and the time t.
- the output f(t,z) is calculated (step S23).
- the survival function calculation unit 562 calculates the survival function S(t,z) based on the output f(t,z) calculated in the process of step S63 (step S64).
- the automatic differentiation unit 563 calculates the hazard function h(t,z) based on the survival function S(t,z) calculated in the process of step S64 (step S65).
- the updating unit 58 updates the parameter set ⁇ (a plurality of parameters p1 and p2) initialized in the process of step S50, the survival function S(t,z) calculated in the process of step S64, and the survival function S(t,z) calculated in step S65.
- An updated parameter set ⁇ (a plurality of parameters p1 and p2) is calculated based on the calculated hazard function h(t, z) (step S66).
- the evaluation function estimation unit 581 calculates the evaluation function L(QS) based on the survival function S(t, z) and the hazard function h(t, z).
- the optimization unit 582 uses the error backpropagation method to calculate an optimized update parameter set ⁇ (a plurality of parameters p1 and p2) based on the evaluation function L(QS).
- step S67 to S69 is equivalent to the processing in steps S18 to S22 in FIGS. 4A and 4B of the first embodiment. That is, the determination unit 60 determines whether the second condition is satisfied based on the updated parameter set ⁇ (a plurality of parameters p1 and p2) (step S67). If the second condition is not satisfied (step S67; NO), the determination unit 60 selects the parameters to be applied to the model 232 and the monotonically increasing neural network 251 from the parameter set ⁇ ' (the plurality of parameters p1' and p2'). is updated to the updated parameter set ⁇ (a plurality of parameters p1 and p2) calculated in the process of step S26 (step S68).
- the determination unit 60 applies the plurality of optimized parameters p1 and p2 to the model 532 and the monotonically increasing neural network 551. Thereafter, the processes of steps S51 to S66 are executed based on the updated parameter set ⁇ (a plurality of parameters p1 and p2) updated in the process of step S68. As a result, the updated parameter set ⁇ (the plurality of parameters p1 and p2) is calculated again.
- step S67 the determination unit 30 converts the updated parameter set ⁇ (the plurality of parameters p1 and p2) calculated in the above step S66 into the learned parameter set ⁇ * (the plurality of parameters p1 * and p2 * ) in the learned parameter storage unit 61 (step S69).
- step S69 ends, the learning operation in the survival analysis device 1 ends (end).
- FIG. 9 is a flowchart showing an example of prediction operation in the survival analysis device 1 as an information processing device according to the second embodiment.
- FIG. 9 corresponds to FIG. 5 in the first embodiment.
- the prediction data set D k * is stored in the prediction data set storage unit 69 in the memory 11 due to a learning operation performed in advance.
- prediction target data X k * is stored in the prediction target data storage section 70 in the memory 11.
- steps S70 to S72 are executed.
- the processing in steps S70 to S72 is equivalent to the processing in steps S30 to S32 in FIG. 5 of the first embodiment. That is, the parameters applied to the model 622 and the monotonically increasing neural network 641 are initialized to the learned parameter set ⁇ * (a plurality of parameters p1 * and p2 * ) stored in the learned parameter storage unit 61 (step S70).
- the feature extraction unit 621 extracts the feature x * of each data X of the prediction data set D k * from the prediction data set D k * stored in the prediction data set storage unit 69 (step S71 ).
- the model 622 to which the plurality of parameters p1 initialized in the process of step S70 is applied calculates a latent expression z * by inputting the feature amount x * extracted in the process of step S71 (step S72).
- the monotonically increasing neural network 641 to which the plurality of parameters p2 initialized in the process of step S70 is applied follows the monotonically increasing function defined by the latent expression z * calculated in the process of step S72 and the time t.
- the output f * (t,z) is calculated (step S73).
- the survival function calculating unit 642 calculates the survival function S * (t,z) based on the output f * (t,z) calculated in the process of step S73 (step S74).
- the automatic differentiation unit 643 calculates the hazard function h * (t,z) based on the survival function S * (t,z) calculated in the process of step S74 (step S75).
- the update unit 66 updates the update parameter set ⁇ * (multiple parameters p1 * and p2 * ) are calculated (step S76). Specifically, the evaluation function estimation unit 661 calculates the evaluation function L * (D) based on the survival function S * (t,z) and the hazard function h * (t,z). The optimization unit 662 uses the error backpropagation method to generate a plurality of optimized parameters p1 * and p2 *' based on the evaluation function L(D), that is, an update parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * ').
- step S77 to step S81 are executed.
- the processing from step S77 to step S81 is equivalent to the processing from step S37 to step S41 in FIG. 5 of the first embodiment. That is, the determination unit 37 determines whether the first condition is satisfied based on the updated parameter set ⁇ * ' (the plurality of parameters p1 * ' and p2 * ') (step S77). If the first condition is not satisfied (step S77; NO), the determination unit 67 determines the parameters to be applied to the model 622 and the monotonically increasing neural network 641 from the parameter set ⁇ * calculated in the process of step S76.
- the updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * ') is updated (step S78).
- the determination unit 67 applies the plurality of optimized parameters p1 * ′ and p2 * ′ to the model 622 and the monotonically increasing neural network 641. Then, based on the updated parameter set ⁇ * ' (a plurality of parameters p1 * ' and p2 * ') updated in the process of step S78, the processes of steps S73 to S78 are executed. As a result, the updating process of the update parameter set ⁇ * ' (the plurality of parameters p1 * ' and p2 * ') is repeated until it is determined in the process of step S77 that the first condition is satisfied.
- the determination unit 67 sets the parameters to be applied to the model 632 and the monotonically increasing neural network 651 to the updated parameter set ⁇ *' that was last updated in the process of step S76. (a plurality of parameters p1 *' and p2 *' ) (step S79).
- the feature amount extraction unit 631 extracts the feature amount x * from the prediction target data Xk * stored in the prediction target data storage unit 70 (step S80).
- the model 632 to which the plurality of parameters p1 *' initialized in the process of step S79 is applied calculates a latent expression z * by inputting the feature amount x * extracted in the process of step S78 (step S81 ).
- the monotonically increasing neural network 651 to which the plurality of parameters p2 *' initialized in the process of step S79 is applied has a monotonically increasing function defined by the latent expression z * calculated in the process of step S81 and the time t.
- the output f * (t,z) is calculated according to (step S82).
- the survival function calculating unit 652 calculates the survival function S * (t,z) based on the output f * (t,z) calculated in the process of step S82 (step S83).
- the automatic differentiator 653 calculates the hazard function h * (t,z) based on the survival function S * (t,z) calculated in the process of step S83 (step S84).
- the output unit 68 sets the hazard function h * (t,z) calculated in the process of step S84 as the hazard function h*(t
- step S85 the prediction operation in the survival analysis device 1 ends (end).
- the monotonically increasing neural network 651 performs latent expression calculation that calculates a latent expression representing a feature amount from processing target data including a feature amount related to a prediction target event. It is configured to calculate as an output a monotonically increasing function defined by the latent expression calculated by the unit 63 and the time.
- the survival function calculation unit 652 and automatic differentiation unit 653 of the function estimation unit 65 estimate the survival function and the hazard function based on the monotonically increasing function output from the monotonically increasing neural network 651. In this way, by modeling using the monotonically increasing neural network 651, integral calculations by approximation can be avoided. Therefore, it is possible to calculate the hazard function and survival function for the prediction target data without making any assumptions.
- a learning function configuration (for learning It further includes a data set storage unit 50 to a determination unit 60). Therefore, even if there is not enough prediction data in which an event to be predicted has occurred, it is possible to calculate the hazard function and survival function for the prediction target data.
- parameters that update the parameters learned by the learning function configuration based on a plurality of prediction data including feature amounts related to the prediction target event stored in the prediction data set storage unit 69 It includes a latent expression calculating section 62, a function estimating section 64, and an updating section 66, which function as an updating section. Therefore, by updating the parameters learned through meta-learning using MAML to parameters corresponding to the prediction target data, it is possible to more accurately calculate the hazard function and survival function.
- the survival function calculation unit 652 of the function estimation unit 65 calculates the survival function based on the scalar value output from the monotonically increasing neural network 651, and the automatic differentiation unit 653 of the function estimation unit 65 calculates the survival function based on the scalar value output from the monotonically increasing neural network 651.
- the hazard function is calculated by automatically differentiating the survival function calculated by . In this way, the hazard function and survival function can be calculated based on the scalar values output from the monotonically increasing neural network 651. Furthermore, since the survival function S(t) satisfies 0 ⁇ S(t) ⁇ 1, unlike the cumulative hazard function, no scale adjustment is required. Therefore, it can be expected that the learning configuration will be easier than in the first embodiment.
- z ⁇ is z when (x0, . . . , x ⁇ ) is input to the model 232 or the like.
- the feature quantity x includes both time-series data (x0,..., x ⁇ ,..., xe) and static data xs, the data xs is used to calculate the data z ⁇ .
- the survival analysis device 1 as an information processing device according to the first and second embodiments, an example has been described in which parameter sets are learned by meta-learning using MAML, but it goes without saying that the meta-learning method is not limited to MAML. It is. A wide variety of advanced versions of MAML have been proposed, and meta-learning may be performed using such advanced versions of MAML. Furthermore, the parameter set may be learned using a meta-learning method other than MAML.
- the survival analysis device 1 as an information processing device according to the first and second embodiments receives a learning program and a prediction program from a program server on the cloud through the communication module 12, stores them in the memory 11, and stores them in the memory 11. It may also be one that performs operations according to a program. Furthermore, instead of providing the learning data set storage units 20 and 50 and the prediction data set storage units 40 and 69 in the memory 11, data sets on the cloud may be used.
- the cumulative hazard function may be calculated by converting it from the survival function.
- learning operations and the predicted operation are executed by the program stored in the survival analysis device 1 as the information processing device according to the embodiment, but this is not limited to this. I can't.
- learning operations and prediction operations may be performed on computational resources on the cloud.
- the method described in the embodiments uses a program (software means) that can be executed by a computer (computer), such as a magnetic disk (floppy (registered trademark) disk, hard disk, etc.), an optical disk (CD-ROM, DVD, etc.). It can also be distributed by being stored in a recording medium such as a semiconductor memory (ROM, RAM, flash memory, etc.), or by being transmitted via a communication medium.
- a computer such as a magnetic disk (floppy (registered trademark) disk, hard disk, etc.), an optical disk (CD-ROM, DVD, etc.
- a recording medium such as a semiconductor memory (ROM, RAM, flash memory, etc.)
- the programs stored on the medium side also include a setting program for configuring software means (including not only execution programs but also tables and data structures) in the computer to be executed by the computer.
- the computer that realizes this device reads a program recorded on a recording medium, and if necessary, constructs software means using a setting program, and executes the above-described processing by controlling the operation of the software means.
- the recording medium referred to in this specification is not limited to one for distribution, and includes storage media such as a magnetic disk and a semiconductor memory provided inside a computer or in a device connected via a network.
- the present invention is not limited to the above-described embodiments, and various modifications can be made at the implementation stage without departing from the spirit thereof.
- each embodiment may be implemented in combination as appropriate as possible, and in that case, the combined effects can be obtained.
- the embodiments described above include inventions at various stages, and various inventions can be extracted by appropriately combining the plurality of disclosed constituent elements. For example, if a problem can be solved and an effect can be obtained even if some constituent features are deleted from all the constituent features shown in the embodiment, the configuration from which these constituent features are deleted can be extracted as an invention.
- Prediction data set storage section 41, 70 ...Prediction target data storage unit 231, 241, 321, 331, 531, 541, 621, 631...Feature amount extraction unit 232, 242, 322, 332, 532, 542, 622, 632...Model 251, 261, 341, 351 , 551, 561, 641, 651... Monotonically increasing neural network 252, 262, 342, 352... Cumulative hazard function calculation unit 253, 263, 343, 353, 553, 563, 643, 653... Automatic differentiation unit 271, 281, 361 , 571, 581, 661... Evaluation function estimation section 272, 282, 362, 572, 582, 662... Optimization section 552, 562, 642, 652... Survival function calculation section
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
Description
1.イベントが発生したデータと発生していないデータを用いてモデルの学習(或いは人手による設計)を行う。
2.モデルを利用して、実際に予測したいデータに対して、イベントの発生し易さを表すハザード関数、及び/又は、イベントが或る時間を超えて発生する確率を表す生存関数を求める。
第1実施形態に係る情報処理装置について説明する。以下では、第1実施形態に係る情報処理装置の一例として、生存分析装置について説明する。
第1実施形態に係る情報処理装置としての生存分析装置の構成について説明する。
図1は、第1実施形態に係る情報処理装置としての生存分析装置1のハードウェア構成の一例を示すブロック図である。図1に示すように、生存分析装置1は、制御回路10、メモリ11、通信モジュール12、ユーザインタフェース13、及びドライブ14を含む。
図2は、第1実施形態に係る情報処理装置としての生存分析装置1の学習機能の構成の一例を示すブロック図である。
指示変数δは、1又は0の値を取る。δ=1はイベント発生を示し、δ=0は打ち切りを示す。打ち切りの場合、当該生存時間データXは、イベント発生前までの特徴量xしか含まないことを示す。
評価関数推定部271は、累積ハザード関数H(t,z)及びハザード関数h(t,z)に基づいて、評価関数L(SS)を算出する。評価関数L(SS)は、例えば、以下のような負の対数尤度である。
評価関数推定部281は、累積ハザード関数H(t,z)及びハザード関数h(t,z)に基づいて、評価関数L(QS)を算出する。評価関数L(QS)は、例えば、以下のような負の対数尤度である。
図3は、第1実施形態に係る情報処理装置としての生存分析装置1の予測機能の構成の一例を示すブロック図である。
評価関数推定部361は、累積ハザード関数H*(t,z)及びハザード関数h*(t,z)に基づいて、評価関数L*(D)を算出する。評価関数L*(D)は、例えば、以下のような負の対数尤度である。
次に、第1実施形態に係る情報処理装置としての生存分析装置1の動作について説明する。
図4A及び図4Bは、第1実施形態に係る情報処理装置としての生存分析装置1における学習動作の一例を示す一連のフローチャートである。図4A及び図4Bの例では、予めメモリ11内の学習用データセット記憶部20に学習用データセットDkが記憶されているものとする。
図5は、第1実施形態に係る情報処理装置としての生存分析装置1における予測動作の一例を示すフローチャートである。図5の例では、予め実行された学習動作によって、メモリ11内の予測用データセット記憶部40に予測用データセットDk*が記憶されているものとする。また、図5の例では、メモリ11内の予測対象データ記憶部41に予測対象データXk*が記憶されているものとする。
第1実施形態によれば、単調増加ニューラルネットワーク351は、予測対象イベントに関する特徴量を含む処理対象データから、特徴量を表す潜在表現を算出する潜在表現算出部33によって算出された潜在表現及び時刻によって規定される単調増加関数に従ってスカラ値を出力するように構成される。関数推定部35の累積ハザード関数算出部352及び自動微分部353は、単調増加ニューラルネットワーク351から出力されたスカラ値に基づいて、ハザード関数を推定する。このように、単調増加ニューラルネットワーク351を用いてモデル化することにより近似による積分計算を回避することができる。よって、仮定無しに、予測対象データについてのハザード関数の算出が可能となる。
次に、第2実施形態に係る情報処理装置について説明する。
第2実施形態に係る情報処理装置としての生存分析装置1の構成について説明する。
図6は、第2実施形態に係る情報処理装置としての生存分析装置1の学習機能の構成の一例を示すブロック図である。図6は、第1実施形態における図2に対応する。
評価関数推定部571の構成は、累積ハザード関数H(t,z)に代えて生存関数S(t,z)を使用する点を除いて、第1実施形態の図2における評価関数推定部271の構成と同等である。評価関数推定部571は、生存関数S(t,z)及びハザード関数h(t,z)に基づいて、評価関数L(SS)を算出する。評価関数推定部571は、算出された評価関数L(SS)を最適化部572に送信する。
評価関数推定部581は、累積ハザード関数H(t,z)に代えて生存関数S(t,z)を使用する点を除いて、第1実施形態の図2における評価関数推定部281の構成と同等である。評価関数推定部581は、生存関数S(t,z)及びハザード関数h(t,z)に基づいて、評価関数L(QS)を算出する。評価関数推定部581は、算出された評価関数L(QS)を最適化部582に送信する。
図7は、第2実施形態に係る情報処理装置としての生存分析装置1の予測機能の構成の一例を示すブロック図である。図7は、第1実施形態における図3に対応する。
評価関数推定部661は、評価関数推定部571と同様のものであり、生存関数S*(t,z)及びハザード関数h*(t,z)に基づいて、評価関数L*(D)を算出する。評価関数推定部661は、算出された評価関数L*(D)を最適化部662に送信する。
次に、第2実施形態に係る情報処理装置としての生存分析装置1の動作について説明する。
図8A及び図8Bは、第2実施形態に係る情報処理装置としての生存分析装置1における学習動作の一例を示す一連のフローチャートである。図8A及び図8Bは、第1実施形態における図4A及び図4Bに対応する。図8A及び図8Bの例では、予めメモリ11内の学習用データセット記憶部50に学習用データセットDkが記憶されているものとする。
図9は、第2実施形態に係る情報処理装置としての生存分析装置1における予測動作の一例を示すフローチャートである。図9は、第1実施形態における図5に対応する。図9の例では、予め実行された学習動作によって、メモリ11内の予測用データセット記憶部69に予測用データセットDk*が記憶されているものとする。また、図9の例では、メモリ11内の予測対象データ記憶部70に予測対象データXk*が記憶されているものとする。
第2実施形態によれば、単調増加ニューラルネットワーク651は、予測対象イベントに関する特徴量を含む処理対象データから、特徴量を表す潜在表現を算出する潜在表現算出部63によって算出された潜在表現及び時刻によって規定される単調増加関数を出力として算出するように構成される。関数推定部65の生存関数算出部652及び自動微分部653は、単調増加ニューラルネットワーク651から出力された単調増加関数に基づいて、生存関数及びハザード関数を推定する。このように、単調増加ニューラルネットワーク651を用いてモデル化することにより近似による積分計算を回避することができる。よって、仮定無しに、予測対象データについてのハザード関数及び生存関数の算出が可能となる。
上述した第1及び第2実施形態には、種々の変形を適用され得る。
例えば、特徴量xが時系列データ(x0,….xτ,…,xe)の場合、尤度を各時刻について計算可能であるため、第1及び第2実施形態における評価関数である負の対数尤度についても、それに応じて変更することができる。例えば、第1実施形態における評価関数L(SS)は、以下のように変更することができる。
他の評価関数についても同様である。
なお、特徴量xが時系列データ(x0,…,xτ,…,xe)と静的なデータxsとの双方を含む場合には、データxsはデータzτの計算に用いられる。
10…制御回路
11…メモリ
12…通信モジュール
13…ユーザインタフェース
14…ドライブ
15…記憶媒体
20,50…学習用データセット記憶部
21,51…データ分割部
22,52…初期化部
23,24,32,33,53,54,62,63…潜在表現算出部
25,26,34,35,55,56,64,65…関数推定部
27,28,36,57,58,66…更新部
29,30,37,59,60,67…判定部
31,61…学習済みパラメータ記憶部
38…変換部
39,68…出力部
40,69…予測用データセット記憶部
41,70…予測対象データ記憶部
231,241,321,331,531,541,621,631…特徴量抽出部
232,242,322,332,532,542,622,632…モデル
251,261,341,351,551,561,641,651…単調増加ニューラルネットワーク
252,262,342,352…累積ハザード関数算出部
253,263,343,353,553,563,643,653…自動微分部
271,281,361,571,581,661…評価関数推定部
272,282,362,572,582,662…最適化部
552,562,642,652…生存関数算出部
Claims (8)
- 予測対象イベントに関する特徴量を含む処理対象データから、前記特徴量を表す潜在表現を算出する潜在表現算出部と、
前記潜在表現算出部によって算出された前記潜在表現及び時刻によって規定される単調増加関数に従ってスカラ値を出力するようにモデル化された単調増加ニューラルネットワークと、
前記単調増加ニューラルネットワークから出力された前記スカラ値に基づいて、ハザード関数及び生存関数の少なくとも一方を推定する関数推定部と、
を備える、情報処理装置。 - 前記潜在表現算出部及び前記単調増加ニューラルネットワークのパラメータをメタ学習により学習する学習部を更に備える、
請求項1に記載の情報処理装置。 - 前記予測対象イベントに関する前記特徴量を含む複数の予測用データに基づいて、前記学習部によって学習した前記パラメータを更新するパラメータ更新部を更に備える、
請求項2に記載の情報処理装置。 - 前記関数推定部は、
前記単調増加ニューラルネットワークから出力された前記スカラ値に基づいて、累積ハザード関数を算出する関数算出部と、
前記関数算出部が算出した前記累積ハザード関数を自動微分することにより前記ハザード関数を算出する自動微分部と、
を含む、
請求項1乃至3の何れか1項に記載の情報処理装置。 - 前記関数算出部が算出した前記累積ハザード関数を前記生存関数に変換する変換部を更に備える、
請求項4に記載の情報処理装置。 - 前記関数推定部は、
前記単調増加ニューラルネットワークから出力された前記スカラ値に基づいて、前記生存関数を算出する関数算出部と、
前記関数算出部が算出した前記生存関数を自動微分することにより前記ハザード関数を算出する自動微分部と、
を含む、
請求項1乃至3の何れか1項に記載の情報処理装置。 - 予測対象イベントに関する特徴量を含む処理対象データから、前記特徴量を表す潜在表現を算出することと、
前記潜在表現及び時刻によって規定される単調増加関数に従ってスカラ値を出力するようにモデル化された単調増加ニューラルネットワークに、前記算出した前記潜在表現を入力して前記スカラ値を出力することと、
前記単調増加ニューラルネットワークから出力された前記スカラ値に基づいて、ハザード関数及び生存関数の少なくとも一方を推定することと、
を備える、情報処理方法。 - コンピュータを、請求項1に記載の情報処理装置が備える各部として機能させるためのプログラム。
Priority Applications (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2024520123A JP7754297B2 (ja) | 2022-05-10 | 2022-05-10 | 情報処理装置、情報処理方法及びプログラム |
| US18/862,979 US20250315669A1 (en) | 2022-05-10 | 2022-05-10 | Information processing apparatus, information processing method, and program |
| PCT/JP2022/019840 WO2023218535A1 (ja) | 2022-05-10 | 2022-05-10 | 情報処理装置、情報処理方法及びプログラム |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| PCT/JP2022/019840 WO2023218535A1 (ja) | 2022-05-10 | 2022-05-10 | 情報処理装置、情報処理方法及びプログラム |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2023218535A1 true WO2023218535A1 (ja) | 2023-11-16 |
Family
ID=88730007
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/JP2022/019840 Ceased WO2023218535A1 (ja) | 2022-05-10 | 2022-05-10 | 情報処理装置、情報処理方法及びプログラム |
Country Status (3)
| Country | Link |
|---|---|
| US (1) | US20250315669A1 (ja) |
| JP (1) | JP7754297B2 (ja) |
| WO (1) | WO2023218535A1 (ja) |
-
2022
- 2022-05-10 WO PCT/JP2022/019840 patent/WO2023218535A1/ja not_active Ceased
- 2022-05-10 US US18/862,979 patent/US20250315669A1/en active Pending
- 2022-05-10 JP JP2024520123A patent/JP7754297B2/ja active Active
Non-Patent Citations (1)
| Title |
|---|
| OMI TAKAHIRO, UEDA NAONORI, AIHARA KAZUYUKI: "Fully Neural Network based Model for General Temporal Point Processes", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, ARXIV.ORG, ITHACA, 23 May 2019 (2019-05-23), Ithaca, XP093107262, [retrieved on 20231130], DOI: 10.48550/arxiv.1905.09690 * |
Also Published As
| Publication number | Publication date |
|---|---|
| US20250315669A1 (en) | 2025-10-09 |
| JPWO2023218535A1 (ja) | 2023-11-16 |
| JP7754297B2 (ja) | 2025-10-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP7344149B2 (ja) | 最適化装置及び最適化方法 | |
| CN107909427A (zh) | 一种提升推荐模型时序数据挖掘能力的循环神经网络方法 | |
| EP3596662B1 (en) | Imagination-based agent neural networks | |
| US20230186092A1 (en) | Learning device, learning method, computer program product, and learning system | |
| CN114897173A (zh) | 基于变分量子线路确定PageRank的方法及装置 | |
| US20200066377A1 (en) | Method and apparatus for generating chemical structure using neural network | |
| CN106022517A (zh) | 一种基于核极限学习机的风险预测的方法和装置 | |
| CN111178986A (zh) | 用户-商品偏好的预测方法及系统 | |
| WO2021106111A1 (ja) | 学習装置、推定装置、学習方法、推定方法、及びプログラム | |
| JP2024170949A (ja) | 機械学習プログラム、機械学習方法および情報処理装置 | |
| US12248534B2 (en) | Automated feature engineering for predictive modeling using deep reinforcement learning | |
| US20220405640A1 (en) | Learning apparatus, classification apparatus, learning method, classification method and program | |
| CN112905809B (zh) | 知识图谱学习方法和系统 | |
| CN115718740A (zh) | 用于稀疏时间序列数据集的数据插补的方法和装置 | |
| JP7754297B2 (ja) | 情報処理装置、情報処理方法及びプログラム | |
| CN117893274A (zh) | 基于高维多目标优化的商品推荐方法、装置、设备及介质 | |
| KR20200107347A (ko) | 클릭율 최대화를 위한 이미지 디자인 파라미터 최적화 방법 및 시스템 | |
| JP7489275B2 (ja) | 情報処理装置、情報処理システムおよび情報処理方法 | |
| CN116185568A (zh) | 一种容器扩容方法、装置、电子设备及存储介质 | |
| JP7283548B2 (ja) | 学習装置、予測システム、方法およびプログラム | |
| CN115423186A (zh) | 基于神经网络模型的成本预测方法、装置、介质及设备 | |
| JP7718004B1 (ja) | 学習装置、学習方法、推論装置、推論方法及びプログラム | |
| JP7625404B2 (ja) | 計画作成装置、計画作成方法、およびプログラム | |
| JP7568090B2 (ja) | 予測装置、予測方法およびプログラム | |
| WO2021250753A1 (ja) | 学習方法、学習装置及びプログラム |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 22941617 Country of ref document: EP Kind code of ref document: A1 |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 2024520123 Country of ref document: JP |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 18862979 Country of ref document: US |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 22941617 Country of ref document: EP Kind code of ref document: A1 |
|
| WWP | Wipo information: published in national office |
Ref document number: 18862979 Country of ref document: US |


