Inverse image estimation device, inverse image estimation method, and program
The inverse image estimation device addresses the challenge of estimating network quality ranges by determining hyperplane equations using a machine learning classifier and Monte Carlo method, effectively handling scenarios with many input types and few correct examples, enabling accurate user experience quality satisfaction.
Patent Information
- Application Number
- PCT/JP2023/019948
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-29
- Publication Date
- 2024-12-05
AI Technical Summary
Machine learning models that assume a large number of input types and few correct examples face challenges in estimating the inverse image, making it difficult to derive the range of network quality to satisfy user experience quality, especially when the number of output types is small compared to input types.
An inverse image estimation device that uses a machine learning classifier to determine multiple hyperplane equations for separating data sets with specified labels, employing the Monte Carlo method and margin maximization to approximate separation boundaries in multidimensional space without nonlinear mapping.
Enables the estimation of inverse images for machine learning classifiers, allowing for the determination of hyperplane equations that clearly indicate the boundaries, thereby facilitating the calculation of the range required for network quality to achieve desired user experience quality.
Smart Images

Figure JP2023019948_05122024_PF_FP_ABST
Abstract
Description
Inverse image estimation device, inverse image estimation method, and program
[0001] The present disclosure relates to techniques for estimating the inverse image of a machine learning classifier.
[0002] In recent years, machine learning has been applied in various fields, and has been particularly effective in classification problems such as image identification and handwritten character recognition. For example, in the network field, there is an index called "Quality of Experience (QoE)," which numerically expresses how comfortably a user can use a service. There is a machine learning model that uses multiple types of network quality, such as delay, packet loss rate, and throughput, as input and estimates the user's QoE when those network qualities are provided (see Non-Patent Document 1). This model can be used by providers of network-based services and network operators to check the user's QoE under the current network environment.
[0003] On the other hand, user experience quality is also used as an indicator when designing services, and network quality design is carried out so that this indicator is above a certain value. In order to implement quality design that achieves a specified user experience quality, it is necessary to derive, from the user experience quality, the range of multiple network qualities that should be realized to satisfy it, as opposed to estimating user experience quality from the multiple types of network qualities described above. The relationship between the user experience quality and multiple types of network qualities described above differs for each service, and since it would be a heavy burden for engineers to build a model every time a new service is introduced, it is desirable to automatically calculate it based on past data.
[0004] QoE visualization and prediction technology to improve user satisfaction NTT Network Technology Laboratories<https: / / journal.ntt.co.jp / backnumber2 / 1807 / files / JN20180744.pdf>
[0005] However, it is assumed that the number of types of machine learning outputs is sufficiently smaller than the number of types of inputs. Therefore, when considering applying machine learning to a case such as the above example where a range of multiple types of network quality to satisfy a user's quality of experience is derived from the user's quality of experience, it is necessary to associate many types (sets of) of output corresponding to the few types of data, so the above machine learning is not suitable for cases where multiple types of network quality to satisfy a user's quality of experience are derived from the user's quality of experience.
[0006] The present invention has been made in consideration of the above points, and aims to make it possible to estimate the inverse image of a machine learning classifier for a machine learning model that generally assumes a large number of input types and a small number of correct examples.
[0007] In order to solve the above problem, the invention of claim 1 is an inverse image estimation device that estimates an inverse image, and has an inverse image estimation unit that uses a machine learning classifier that specifies a set of multidimensional data and labels corresponding to the data, and determines a plurality of hyperplane equations for separating, in the set of data, a first set of data that is assigned a specified predetermined label and a second set of data that is assigned a label other than the specified label.
[0008] As described above, the present invention has the effect of making it possible to estimate the inverse image of a machine learning classifier for a machine learning model that is generally based on the premise that there are many types of inputs and few types of correct examples.
[0009] 1 is an overall configuration diagram of an inverse image estimation device according to an embodiment; FIG. 2 is an electrical hardware configuration diagram of an inverse image estimation device according to an embodiment; FIG. 3 is a flowchart showing processing or operations of an inverse image estimation device according to an embodiment; FIG. 4 is a detailed flowchart of cell division processing; FIG. 5 is a detailed flowchart of inverse image estimation processing; and FIG.
[0010] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
[0011] [Overview of Inverse Image Estimation Device] The inverse image estimation device 10 of this embodiment utilizes a trained machine learning model to realize a method for associating a set of inputs (learning data) with the output (correct examples) of the trained model. That is, the inverse image estimation device 10 realizes an inverse image estimation method that explicitly outputs a set of input data mapped to a classification value that is the output of the original machine learning model. Specifically, the inverse image estimation device 10 considers a set of input data of the original learning model that corresponds to the classification value as being surrounded by multiple hyperplanes in a multidimensional space, and derives hyperplane equations for those hyperplanes.
[0012] Here, a method for dividing the classification space into planes is the SVM (Support Vector Machine) as described in Reference 1.
[0013] (Reference 1) Hearst, Marti A., et al., "Support vector machines", IEEE Intelligent Systems and their applications, vol.13, no.4, pp.18-28, July-Aug, 1998. However, SVM is a method that uses a nonlinear mapping φ to map the original space to another space, and separates the boundary with a plane in the original space. In order to find a boundary that divides the set to be separated in the original space, theoretically, the obtained plane can be mapped as the inverse mapping φ of the nonlinear mapping φ. -1 However, the inverse mapping φ -1 cannot be calculated analytically.
[0014] Therefore, when a trained machine learning model for mapping cases to classification types is given, the inverse image estimation device 10 of this embodiment uses a method for obtaining a set of cases belonging to each classification type from the classification type. The inverse image estimation device 10 inputs random cases (expressed as multidimensional vectors) to the trained machine learning model based on the Monte Carlo method, and plots the classification values in space. The inverse image estimation device 10 then obtains multiple hyperplanes that can efficiently approximate the boundary separating the plots of cases belonging to the desired classification value from the plots of other cases. The margin maximization method is used for the separation.
[0015] In this way, by finding multiple hyperplanes that can efficiently approximate the separating boundary, the inverse image estimation device 10 is able to determine the boundary that divides the target set in the original space without using a nonlinear mapping, and to explicitly indicate the boundary.
[0016] [Functional Configuration of Inverse Image Estimation Device] First, the functional configuration of an inverse image estimation device 10 will be described with reference to Fig. 1. Fig. 1 is a diagram showing the overall configuration of an inverse image estimation device according to an embodiment.
[0017] The inverse image estimation device 10 is a device that estimates the inverse image of a machine learning classifier for a machine learning model that is based on the assumption that there are only a few types of correct examples. In this embodiment, n-dimensional input data d is input, and a machine learning classifier M that estimates the corresponding label is given, and the estimation of the inverse image is expressed using an inverse function M.
[0018] The inverse image estimation device 10 is configured by one or more computers. When the inverse image estimation device 10 is configured by multiple computers, it may be referred to as an "inverse image estimation device" or an "inverse image estimation system."
[0019] As shown in Fig. 1, the inverse image estimation device 10 includes a data acquisition unit 11, a label separation unit 12, a cell division unit 13, and an inverse image estimation unit 14 to acquire a plurality of hyperplane equations for separating data to which a label l is assigned from data to which a label other than the label l is assigned, in an n-dimensional input data set D={d|d∈Rn}. Each of these units has a function realized by an instruction from a CPU 101 in Fig. 2 based on a program. The inverse image estimation device 10 also stores the data set D and a machine learning classifier M in a RAM 103 (described later) or the like.
[0020] The data acquisition unit 11 acquires N random pieces of multidimensional data from the "data set D" based on the idea of the Monte Carlo method by specifying that N pieces of data be acquired from the "data set D." The acquired multidimensional data set is defined as D'(⊂D).
[0021] The label separation unit 12 receives as input data belonging to the data set D' obtained by the data acquisition unit 11, and by specifying a predetermined label l to be separated, acquires a data set D'(l) to which label l is assigned and a data set nD'(l) to which labels other than label l are assigned.
[0022] When separating the data, the label corresponding to each data d∈D' can be obtained by using a machine learning classifier M to estimate the label corresponding to each multidimensional data (hereafter referred to as M(d)), making it possible to distinguish (classify) data corresponding to the specified label.
[0023] D'(l) denotes the set of all inputs to the machine learning classifier M that produce an output l {d∈D'|M(d)=l}.
[0024] nD'(l) denotes all data in D' that are not D'(l) {d∈D'|M(d)≠l}.
[0025] The cell division unit 13 receives the data sets D'(l) and nD'(l) acquired by the label separation unit 12 as input, and divides a space represented by a virtual hyper-rectangular prism that mainly contains the data of D'(l) to obtain divided cells.
[0026] The inverse image estimation unit 14 receives the divided cells obtained by the cell division unit 13 as input and calculates a plurality of hyperplane equations (c1x1 + c2x2 + ... c n x n = 0).
[0027] When actually used, the inverse image estimation unit 14 inputs a set of input data D, a machine learning classifier M, and a label l to be separated from other data, and uses the machine learning classifier M to specify a set of multidimensional data and the label corresponding to this data, to find multiple hyperplane equations for separating, in the set of multidimensional data, a set of data D'(l) to which the specified predetermined label l is assigned and a set of data nD'(l) to which a label other than the predetermined label l is assigned.
[0028] In this case, the inverse image estimation unit 14 generates multiple hyperplanes that separate the data set D'(l) and the data set nD'(l) in the vicinity (within a predetermined distance) of each point in the multidimensional data set by maximizing the margin, thereby obtaining multiple hyperplane expressions. Furthermore, when generating hyperplanes in the vicinity of each point, the inverse image estimation unit 14 merges multiple finely separated hyperplanes to obtain the above-mentioned multiple hyperplane expressions for approximating the overall boundary.
[0029] This means that, for example, in data with n variables, the range that the nth variable should take in order to assign label l can be obtained by providing the values of variables other than the nth variable, from 1 to n-1.
[0030] [Hardware Configuration of Inverse Image Estimation Device] Next, the electrical hardware configuration of the inverse image estimation device 10 will be described with reference to Fig. 2. Fig. 2 is a diagram showing the electrical hardware configuration of the inverse image estimation device.
[0031] The inverse image estimation device 10 is a computer that includes a CPU 101, a ROM 102, a RAM 103, an SSD 104, an external device connection I / F (Interface) 105, a network I / F 106, a display 107, an input device 108, a media I / F 109, and a bus line 110.
[0032] Of these, the CPU 101 controls the overall operation of the inverse image estimation device 10. The ROM 102 stores programs such as IPL used to drive the CPU 101. The RAM 103 is used as a work area for the CPU 101.
[0033] The SSD 104 reads or writes various data under the control of the CPU 101. Note that instead of the SSD 104, a hard disk drive (HDD) may be used.
[0034] The external device connection I / F 105 is an interface for connecting various external devices, such as a display, a speaker, a keyboard, a mouse, a USB memory, and a printer.
[0035] The network I / F 106 is an interface for performing data communication via the communication network 100 .
[0036] The display 107 is a type of display means such as a liquid crystal display or organic electroluminescence (EL) display that displays various images.
[0037] The input device 108 is a keyboard, a pointing device, or the like, and is an example of an input means for accepting input operations such as characters, numbers, and various instructions.
[0038] The media I / F 109 controls reading and writing (storing) of data from and to a recording medium 109m such as a flash memory, etc. The recording medium 109m includes DVDs and Blu-ray Discs (registered trademarks).
[0039] The bus line 110 is an address bus, a data bus, etc. for electrically connecting the components such as the CPU 101 shown in FIG.
[0040] [Processing or Operation of the Embodiment] Next, processing or operation of the present embodiment will be described with reference to Fig. 3 to Fig. 6. Fig. 3 is a flowchart showing processing or operation of the inverse image estimation device according to the embodiment.
[0041] S11: The data acquisition unit 11 acquires N random pieces of data from the "data set D" based on the idea of the Monte Carlo method by specifying that N pieces of data be acquired from the "data set D".
[0042] S12: Using data belonging to the data set D' obtained by the data acquisition unit 11 as input, and specifying a predetermined label l to be separated, a data set D'(l) to which label l is assigned and a data set nD'(l) to which labels other than label l are assigned are acquired. The data set D'(l) is an example of a first data set. The data set nD'(l) is an example of a second data set.
[0043] S13: The cell division unit 13 receives the data sets D'(l) and nD'(l) acquired by the label separation unit 12 as input, and divides the space represented by a hyper-rectangular prism that mainly contains the data of D'(l) to acquire divided cells.
[0044] S14: The inverse image estimation unit 14 receives the divided cells acquired by the cell division unit 13 as input and calculates a plurality of hyperplane equations (c1x1 + c2x2 + ... c n x n = 0).
[0045] <Detailed Description of Cell Division Processing> Next, processing S13 will be described in detail with reference to Fig. 4. Fig. 4 is a detailed flowchart of the cell division processing.
[0046] S31: The cell division unit 13 receives two types of data obtained by the label separation unit 12: D'(l), which is a set of elements d such that M(d)=l, and nD'(l), which is a set of elements d such that M(d)≠l.
[0047] S32: The cell division unit 13 generates the smallest virtual hyper-rectangular parallelepiped Rect (n-dimensional) that includes all points of the set D'(l) from the two types of data input in step S31. Rect = {(x1,..., x n )∈R ≧0 n |a i ≦x i ≦b i}. Also, the i-th (1≦i≦n) dimension data x i The lower limit a i and upper limit b i The data d that is an element of D'(l) is d = [d1, d2, … ,d n ], a i = min{d i | d i ∈D'(l)}, b i = max{d i | d i ∈D'(l)}.
[0048] S33: Finally, the cell dividing unit 13 divides each of the n sides of the hyper-rectangle Rect generated in step S32 into p sides to obtain p n Outputs cells (p is an arbitrary integer). Here, each cell is represented by the x i The axial edge is divided into p parts, and k parts are i th (1≦k i ≦p), we can generate cells by selecting cell[k1, k2,…k n ]
[0049] <Detailed Description of Inverse Image Estimation Processing> Next, processing S14 will be described in detail with reference to Fig. 5 and Fig. 6. Fig. 5 and Fig. 6 are detailed flowcharts of the inverse image estimation processing. Note that all of the processing in Fig. 5 and Fig. 6 is performed by the inverse image estimation unit 14.
[0050] S41: The inverse image estimation unit 14 calculates the p n The input data is a set of cells Cell.
[0051] S42: Among Cells, cells that are not on the boundary do not need to be taken into consideration when estimating the inverse image, so the inverse image estimation unit 14 excludes cell c that is not on the boundary between D'(l) and nD'(l). "Cell c is not on the boundary" means the following case, in which the data contained in cell c is biased toward D'(l) or nD'(l).
[0052] n(D'(l)∩c) / n(D'(l)∩c)+n(nD'(l)∩c) < r, or n(nD'(l)∩c) / n(D'(l)∩c)+n(nD'(l)∩c) < r Here, n(D'(l)∩c) means the number of data elements belonging to D'(l) that exist in cell c. Also, n(nD'(l)∩c) means the number of data elements belonging to nD'(l) that exist in cell c.
[0053] Furthermore, r is a small parameter close to 0, which excludes cells that are not on the boundary surface.
[0054] S43: Next, the inverse image estimation unit 14 generates a variable cCell for storing connected cells generated by subsequent processing. A "connected cell" is a set of cells c, and any cell in the set includes another adjacent cell, and cCell is a set of adjacent boundary cells extracted in processing S42 that are considered connected cells. Note that a set consisting of a single cell is also considered a connected cell, and this is set as the initial state of cCell (i.e., cCell = {{c} | c∈Cell}). For example, if two connected cells cCell_1 and cCell_2 each include a cell c1∈cCell_1 and c2∈cCell_2, and the cells c1 and c2 are adjacent to each other, then the connected cells cCell_1 and cCell_2 are said to be adjacent to each other. Here, "cells c1 and c2 are adjacent" means that c1 = cell(k1, k2, ..., k n ), c2 = cell(k'1, k'2, ..., k' n ), then for some i, |k i - k' i |=1, and the j-th dimension element is k j = k' j This shows the case where (1≦j≦n, j≠i).
[0055] S44: Next, the inverse image estimation unit 14 extracts pairs of adjacent connected cells from the set of connected cells belonging to cCell and creates a set p_cCell of such pairs. p_cCell comprehensively includes pairs of adjacent connected cells.
[0056] S45: After that, the inverse image estimation unit 14 pops (takes out and deletes from the original set p_cCell) the original (cCell_1, cCell_2) belonging to p_cCell.
[0057] S46: Based on the idea of maximizing margins, the inverse image estimation unit 14 calculates (generates) a hyperplane s1 that separates each point of D'(l) contained in the original cell cCell_1 and a hyperplane s2 that separates each point of nD'(l) contained in the original cell cCell_2, and calculates the coefficients of the hyperplane equations. In this case, the inverse image estimation unit 14 merges hyperplanes that achieve separation in multiple finely decomposed spaces to calculate the coefficients of multiple hyperplane equations that approximate the overall boundary. Note that hyperplane s1 is an example of a first hyperplane. Hyperplane s2 is an example of a second hyperplane.
[0058] S47: The inverse image estimation unit 14 determines whether the angle θ between the generated hyperplanes s1 and s2 is equal to or less than an arbitrary threshold θ (0≦θ′≦90°). If θ<θ′ is not true, and p_cCell is not an empty set (S52; NO), the process returns to step S45. If it is an empty set (S52; YES), there are no more cells to be separated, and step S14 ends. If θ<θ′, the process proceeds to step S47.
[0059] S48: The inverse image estimation unit 14 combines the original (c_cell_1, c_cell_2) to generate a new cell c_cell_tmp.
[0060] S49: The inverse image estimation unit 14 calculates hyperplanes that separate each point of D'(l) and each point of nD'(l) contained in each cell of c_cell_tmp in the same manner as in process S46, and finds the coefficients of the hyperplane equation.
[0061] S50: Here, the inverse image estimation unit 14 determines whether the separation accuracy at the time of separation is 90% or more. If the separation accuracy is not 90% or more (NO), and if p_cCell is not an empty set (S52; NO), the process returns to step S52. If it is an empty set (S52; YES), there are no more cells to be separated, and process S14 ends. On the other hand, if the separation accuracy is 90% or more (YES), the process proceeds to step S51.
[0062] S51: The inverse image estimation unit 14 deletes c_cell_1 and c_cell_2 from cCell, adds a temporary c_cell_tmp, and returns to process S44. Thereafter, the inverse image estimation unit 14 repeats processes S44 to S51, and ends process S14 when p_cCell finally becomes an empty set.
[0063] As described above, the inverse image estimation device 10 according to this embodiment can calculate a plurality of hyperplane equations that separate each point of D'(l) from each point of nD'(l), which leads to specific calculation and understanding of the possible range of each feature value of the n-dimensional data d that becomes the label l.
[0064] [Major Effects of the Embodiments] As described above, the embodiments have the effect of making it possible to estimate the inverse image of a machine learning classifier for a machine learning model that is generally based on the premise that there are many types of inputs and few types of correct examples.
[0065] [Supplementary Note] The present invention is not limited to the above-described embodiment, and may have the following configurations or processes (operations).
[0066] (1) The inverse image estimation device 10 can be realized by a computer and a program, but this program can also be recorded on a (non-transitory) recording medium or provided via the communication network 100.
[0067] (2) The CPU 101 may be a single CPU or multiple CPUs.
[0068] 10 Inverse image estimation device 11 Data acquisition unit 12 Label separation unit 13 Cell division unit 14 Inverse image estimation unit M Machine learning classifier
Claims
1. An inverse image estimation device that estimates an inverse image, comprising an inverse image estimation unit that uses a machine learning classifier that specifies a set of multidimensional data and a label corresponding to the data, and determines a number of hyperplane equations for separating, in the set of data, a first set of data to which a specified label is assigned and a second set of data to which a label other than the specified label is assigned.
2. The inverse image estimation device according to claim 1, wherein the inverse image estimation unit determines the multiple hyperplane equations by generating multiple hyperplanes that separate the first data set and the second data set within a predetermined distance of each point of the multidimensional data set through margin maximization.
3. The inverse image estimation device according to claim 2, wherein, in generating the hyperplane, the inverse image estimation section merges the multiple hyperplanes to obtain the multiple hyperplane equations that approximate an entire boundary.
4. An inverse image estimation method executed by an inverse image estimation device that estimates an inverse image, wherein the inverse image estimation device uses a machine learning classifier that specifies a set of multidimensional data and a label corresponding to the data, and executes an inverse image estimation process to find multiple hyperplane equations for separating, in the set of data, a first set of data to which a specified predetermined label is assigned and a second set of data to which a label other than the specified label is assigned.
5. A program for causing a computer to execute the method according to claim 4.
Citation Information
Patent Citations
Biological information processor, biological information processing method, and biological information processing program
JP2007052507A
Detection method, detection program, and information processing device
WO2021079459A1
