JP5905734B2 - Method and apparatus for building a support vector machine - Google Patents

Description

  The present invention relates to a method and apparatus for efficiently creating a support vector machine group that performs a determination according to a plurality of determination criteria, and in particular, when each determination criterion is similar, an initial solution obtained from a learning example in which each determination criterion matches. It is related with the method and apparatus which can aim at reduction of the amount of calculations at the time of constructing | assembling the support vector machine which learns each judgment standard by learning each judgment standard using.

  A support vector machine (SVM) having the capability of nonlinear discrimination is a data classification method mainly corresponding to a binary classification problem using learning. Since SVM has high generalization performance for unlearned data, it is applied to various discrimination problems such as speech, character, figure, and pattern recognition.

  In Non-Patent Documents 1 and 2, SVM is applied to a system that finds illegal and harmful information from a large amount of content on the Internet. However, when learning SVM with a large amount of data (for example, more than 100,000 pieces of training data) in an attempt to improve the determination accuracy, a large amount of memory is consumed and the convergence speed is slowed down.

  For this reason, in an approach such as Non-Patent Document 3-6, a device for suppressing an increase in memory usage and processing time without reducing accuracy as much as possible has been proposed. These ideas focus on improving the efficiency of SVM learning for the 2-class decision problem. On the other hand, it is known that a system that performs determination based on a plurality of criteria can be realized by a combination of a plurality of SVMs that handle a 2-class determination problem.

K. Ikeda, T. Yanagihara, G. Hattori, K. Matsumoto, and Y. Takishima. Hazardous document detection based on dependency relations and thesaurus.In Australasian Conference on Artificial Intelligence, pages 455-565, 2010. Tatsuaki Matsuba, Fumito Sakurai, Ikuo Kawai, Naoki Isu, Naohiro Satomi, Harmful Information Detection in School Informal Sites, IEICE Technical Report. NLC, Language Understanding and Communication 109 (142), 93-98, 2009- 07-15 E. Osuna, R. Freund, and F. Girosi, "An improved training algorithm for support vector machines," in Neural Networks for Signal Processing VII-Proceedings of the 1997 IEEEWorkshop, NMJ Principe, L. Gile and E. Wilson, Eds ., New York, pp.276-285, 1997. T. Joachims, "Makinglarge-scale support vector machine learning practical," in Advances in Kernel Methods: Support Vector Machines, A. S. B. Scholkopf, C. Burges, Ed., MIT Press, Cambridge, MA, 1998 J. Platt, "Fast training of support vector machines using sequential minimal optimization," inAdvances in Kernel Methods-Support Vector Learning, B. Scholkopf, C. J. C. Burges, and A. J. Smola, Eds., Cambridge, MA: MIT Press, 1999. Duc Dung Nguyen; Matsumoto, K.; Takishima, Y .; Hashimoto, K .; Terabe, M., "Two-stage incremental workingset selection for fast support vector training on large datasets," Research, Innovation and Vision for the Future, 2008. RIVF 2008. IEEE International Conference on, vol., No., Pp.221-226, 13-17 July 2008

  When constructing SVM groups that can obtain judgments based on the desired judgment criteria by combining 2-class judgment SVMs corresponding to individual judgment criteria, the learning time required for system construction is the learning time for each SVM. Total time.

  As described in Non-Patent Documents 3-6, individual SVM learning can be made efficient. However, it is not possible with the conventional technology to construct a plurality of SVM groups more efficiently so that the determination can be made with a plurality of determination criteria in a time shorter than the sum of the individual SVM learning times.

  In view of the above problems, the present invention provides a method and apparatus for constructing a support vector machine capable of efficiently constructing a plurality of support vector machines so that judgments can be made with a plurality of judgment criteria when the judgment criteria are somewhat similar. The purpose is to provide.

  To achieve the above object, the present invention provides a method for constructing a support vector machine for a plurality of criteria having a common case in labeling given data, the common case being a part of the plurality of criteria. Learning a support vector machine and learning a support vector machine by using the coefficient of the decision function obtained for the common case as an initial value. A second step of constructing a support vector machine for the judgment criteria included in a part of the plurality of judgment criteria, and the first and second steps are repeatedly executed sequentially for the plurality of judgment criteria. And a third step of constructing a support vector machine corresponding to each of the plurality of determination criteria And features.

  In order to achieve the above object, the present invention is an apparatus for constructing a support vector machine for a plurality of judgment criteria having a common case in labeling given data, wherein a common part of the plurality of judgment criteria is shared. Learning a support vector machine in a case and finding a coefficient of a decision function corresponding to the common case, and learning a support vector machine using a coefficient of the decision function obtained for the common case as an initial value A differential case learning unit for constructing a support vector machine for a determination criterion included in a part of the plurality of determination criteria, and learning for setting a predetermined learning order of the plurality of determination criteria based on a common case The common case learning unit and the difference case sequentially according to the learning order set by the learning order setting unit. By applying 習部, characterized by constructing the support vector machine corresponding to each of the plurality of criteria.

  According to the present invention, when learning a support vector machine for a common case among a plurality of judgment criteria and obtaining a coefficient of the decision function, and constructing a support vector machine corresponding to each judgment criteria, Learning is performed using the coefficient as an initial value. At this time, since the initial value is close to the coefficient of the support vector machine to be constructed, the amount of calculation in learning is reduced, and the amount of computation as a whole when constructing a plurality of support vector machines is reduced.

It is a functional block diagram of an apparatus for constructing a support vector machine according to an embodiment of the present invention. It is a figure which illustrates the present invention conceptually using an example. 4 is a flowchart of a method for constructing a support vector machine according to an embodiment of the present invention. It is a figure for demonstrating one Example of this invention. It is a figure for demonstrating one Example of this invention. It is a figure for demonstrating one Example of this invention. It is a figure for demonstrating one Example of this invention. It is a figure which shows the actual example of the processing time reduction of SVM construction by this invention. It is a figure which shows the actual example of the processing time reduction of SVM construction by this invention.

  FIG. 1 is a functional block diagram of an apparatus for constructing a support vector machine according to an embodiment of the present invention. An apparatus 1 for constructing a support vector machine includes a common case learning unit 2, a difference case learning unit 3, and a learning order setting unit 4.

The apparatus 1 for constructing a support vector machine has T ¹ , T ² , T ³ ,..., T ^{M as} learning data for learning a predetermined plurality of M judgment criteria. Corresponding M support vector machines SVM_1, SVM_2, SVM_3,..., SVM_M are constructed. The constructed SVM group can obtain the judgment results for all the M judgment criteria for arbitrary input data. The learning data can be expressed as (Equation 1) below.

That is, each learning data ^Tu (u = 1, 2,... M) is a binary {corresponding to each judgment criterion u with respect to given data x _i (i = 1,... L) common to each other. −1, +1} are given as labels y _i ^u . For example, the common data x _i is a predetermined feature amount (document vector or the like) of each web page i (i = 1,... L) in the common web page group, and each judgment criterion u is the respective web page i. This is a criterion for determining whether (i = 1,... l) corresponds to any genre.

  In the present invention, when each criterion is similar, that is, when there is a common case at a predetermined rate for labeling each criterion (for example, whether the content of the web page is a sport and leisure) And so on), it has an effect of reducing the amount of calculation when learning all M judgment criteria.

Incidentally, T ^u and T ^v and the common case (criterion u a criterion v and the common case) The T ^u ∩T ^v, i.e. grant labels between T ^u and T ^v are common It is data for learning. The same applies in three or more common case, for example, the common case of T ^u and T ^v and T ^w is T ^u ∩T ^v ∩T ^w.

In addition, learning each criterion is an optimization for ^obtaining the following decision function f ^u (x) (= defining an identification plane for determination) from each learning data. It is a process. By determining the coefficient α ^u _i of the decision function by the optimization, the decision function is obtained.

  In (Equation 3), K is a kernel function, and C is a parameter.

In the present invention, this optimization process is performed by a known method (for example, Chih-Chung Chang and Chih-Jen Lin, LIBSVM: a library for support vector machines. ACM Transactions on Intelligent Systems and Technology, 2: 27: 1- -27: 27, 2011). In the present invention, instead of learning each learning data T ¹ , T ² , T ³ ,..., T ^M individually, advantageous initial values are set for the common case and the remaining difference cases. It is characterized by learning above.

  FIG. 2 is a diagram conceptually illustrating the present invention using an example. Here, for the training data consisting of 5 white circles, 1 gray circle (gray circle) and 5 black circles, a total of 11 learning criteria, two judgment criteria with similar labeling (standard- In this example, the learning data is abbreviated as “1” and “reference-2”, and the respective learning data are referred to as “T-1” and “T-2”.

  As shown in the figure, in the reference-1, five white circles are positive examples, and one gray circle and five black circles are negative examples. In Criterion-2, 5 white circles and 1 gray circle are positive examples, and 5 black circles are negative examples. That is, the common case of Criteria-1 and Criteria-2 (referred to as C0) is composed of 5 positive examples of white circles and 5 negative examples of black circles, with a large ratio of 10 out of 11 It is common.

  In the conventional method, the boundary surface is such that SVM-1 and SVM-2 are conceptually illustrated as a solid line and a dotted line, respectively, by learning Standard-1 and Standard-2 individually by T-1 and T-2, respectively. As obtained. In the present invention, the same SVM-1 and SVM-2 as in individual learning are obtained, but the procedure is different and the amount of calculation is reduced.

  In the present invention, SVM-0 is obtained by first learning the common case C0 and obtaining the coefficient of the decision function. Then, SVM-1 and SVM-2 are obtained by learning T-1 and T-2 using SVM-0 as an initial solution in the optimization of (Expression 2) and (Expression 3). At this time, a predetermined value is assigned to the coefficient of the decision function corresponding to the case indicated by the gray circles, that is, the difference case between the common case C0 and T-1 and T-2, and is set as the initial solution.

  Since SVM-0 as the initial solution is close to both SVM-1 and SVM-2, it functions effectively to reduce the amount of calculation. And since both T-1 and T-2 learning have the effect of reducing the amount of calculation, compared to the case of learning individually (learning 2 times of individual SVM1 + individual SVM2), the whole (common SVM-0 Learning + difference SVM-1 learning + difference SVM-2 three times of learning) is reduced.

  FIG. 3 is a flowchart of a method for constructing a support vector machine according to an embodiment of the present invention. With reference to FIG. 3, the flow will be described together with the functions of the respective parts of the apparatus 1 (FIG. 1) for constructing the support vector machine.

  When the flow is started in step S0, the learning order setting unit 4 learns in step S1, and the common case learning unit 2 and the difference case learning unit 3 sequentially learn in subsequent loops of steps S3, S4, S5, and S7. Set the target and its order. Specifically, the common case (i) and the difference case (i) whose order is determined by the counter i (i = 1, 2,...) And the end value of the counter i, that is, the common case and the difference case are set. It is set in which order i which SVM is to be constructed based on the number, common case (i), and difference case (i).

  The details will be described later with reference to FIGS. 4 to 7 and the like. Various settings are possible, and a predetermined setting is used in advance by manual selection or the like by the user. Then, the learning order setting unit 4 reads the predetermined setting. At this time, a predetermined setting based on a common case between learning data may be used. In the example of Fig. 2, the end value of i is 1 (= initial value), the common case (1) is C0, and the difference case (1) is the ash circle case (learning data). A negative example with respect to criterion-1 and a positive example with reference-2.

  In step S2, the counter i is set to an initial value of 1. In step S3, the common case learning unit 2 learns the common case (i) and obtains a coefficient of a decision function corresponding to the common case (i).

  In step S3, if i = 1, the common case (1) is learned as it is with a known method (a method for learning a single SVM), but in the second and subsequent rounds of the iterative loop, that is, i ≧ 2 The common case (i) is learned by using the coefficient in the already learned case as in step S4 described below. At this time, unlike step S4, which uses a combination of common case (i) and difference case (i), in step S3, common case (i -K) is used. Here, k> 0 and i−k <i.

  In step S4, the difference case learning unit 3 combines the common case (i) and the difference case (i) as a union "common case (i) ∪ difference case (i)" (= SVM construction target) 1) learning data of each criterion, and constructing one of the construction target SVMs. At this time, the SVM is learned by setting an initial value using the coefficient of the decision function corresponding to the common case (i) calculated in step S3.

Here, the coefficient of the decision function calculated for the common case (i) in step S3 is {α1, α2,..., Αi _m }, and the initial value of the coefficient of the decision function corresponding to the common case (i) is Uses the calculated coefficient. Further, in the initial value of the combined case, zero is set as a predetermined value for the coefficient of the decision function corresponding to the difference case (i). That is, when the coefficient corresponding to the difference example (i) is written backward, the initial value is {α1, α2,..., Αi _m , 0,. A predetermined value other than zero may be used as the initial value corresponding to the difference example (i).

  In addition, although there is one common case (i), since there may generally be one or more difference cases (i) for each i, the SVM learning in step S4 is the same as the common case (i). This is performed for all combinations with the difference example (i), and the SVM of the corresponding criterion is constructed. The order of learning by combining the difference example (i) may be arbitrary. Depending on the learning setting in the learning order setting unit 4, step S4 may be skipped (omitted) for a predetermined i.

  In general, the difference case (i) = “learning data of the judgment criteria for constructing the SVM at the step S4 \ common case (i)”, so the common case (i) and the judgment criteria for the SVM construction target are The difference example (i) is automatically determined by setting. “\” Represents a difference set as described later.

  In step S3 in the case of i ≧ 2 described above, the common case learning unit 2 performs learning by setting initial values as in the following [1] or [2]. [1] When the common case (i) to be learned is configured as a combination of the past common case (i−k) and some case, that is, the union “common case (i−k) ∪some case” Similarly to the above, the coefficient of the calculated common case (i−k) and the coefficient corresponding to the case are set to zero or the like as the initial value.

  [2] Further, in step S3 when i ≧ 2, the common case (i) of the learning target is configured as a part of the common case (i−k) using the initial value (“common case (i− k) ⊃ common case (i) ”), the extracted coefficients of the common case (i−k) are extracted from the coefficients corresponding to the common case (i). Use as a value.

  In the learning in steps S3 and S4, as described above, optimization by a known method implemented by libSVM or the like may be performed.

  Details of steps S3 and S4 will be described later, but in the example of FIG. In step S3, common case (1) = C0 is learned, and SVM-0 is obtained as a coefficient of the decision function. In step S4, as one of the difference cases (1), a gray case as a negative example in criterion-1 and a common case (1) = C0 calculated as SVM-0 are combined (= T-1 SVM-1 can be obtained by setting and learning an initial value for). In addition, another example of difference (1) is a combination of gray circle as a positive example in Standard-2 and common case (1) = C0 calculated as SVM-0 (= T-2) SVM-2 can be obtained by setting and learning the initial value of.

  In the present invention, as described with reference to FIG. 2, the coefficient of the decision function calculated for the common case (i) in step S4 is used as an initial value, and the past in step S3 where i ≧ 2. By using the coefficient of the decision function that has already been calculated for the common case (i−k) as the initial value, the amount of calculation can be reduced.

  In step S5, it is confirmed whether or not SVMs corresponding to all the judgment criteria have been constructed. If they have been constructed, the process proceeds to step S6, and the flow ends. If there is an unconstructed SVM, the process proceeds to step S7. i is incremented by 1, and steps S3 and S4 are repeated for the subsequent common case (i) and difference case (i).

  In the example of FIG. 2, since all the SVM-1 and SVM-2 to be constructed have been constructed in step S4 when i = 1, the flow ends without repeating.

  FIGS. 4 to 7 are diagrams for supplementary explanation of the details of the flow of FIG. 3, and predetermined setting of the common case (i) and the difference case (i) by the learning order setting unit 4 in step S1, Each example of learning by the common case learning unit 2 in step S3 and predetermined learning by the difference case learning unit 3 in step S4 under the setting is shown.

  4 to 7 are all examples in which the criterion is five, and the learning data is Ta, Tb, Tc, Td and Te, but the general number n (n ≧ 2) is exactly the same. The present invention can be applied to a predetermined method. The support vector machines constructed from the learning data Ta, Tb, Tc, Td, and Te are referred to as SVM-a, SVM-b, SVM-c, SVM-d, and SVM-e.

  In FIGS. 4 to 7, the relationship constituting the common case from each learning data is represented as a combination of lines corresponding to each learning data. Further, the learning process of the common case in step S3 is indicated by a dotted line frame block, and the construction process of the judgment criterion in step S4 is indicated by a solid line frame block. The processes in steps S3 and S4 are both “learning” of the SVM, but step S4 is called “construction” in the sense that the SVMs of the respective determination criteria that are finally obtained are sequentially completed.

  In the embodiment of FIG. 4, as shown in the figure, four common cases (i) (i = 1 to 4) are sequentially learned in step S3 in each round. Common case (1) = Ta∩Tb, Common case (2) = Ta∩Tb∩Tc, Common case (3) = Ta∩Tb∩Tc∩Td, Common case (4) = Ta∩Tb∩Tc∩ Td∩Te. Learning by combining each common case (i) and the corresponding difference case (i) in step S4 is as follows.

  In step S3 with i = 1, the common case (1) is learned according to the known single SVM method, and the coefficient is obtained. In step S4 of i = 1, learning is performed by combining the common case (1) and the two difference cases (1) to construct SVM-a and SVM-b. Here, SVM-a uses "Ta \ common case (1)" which is one of the differential cases (1), and SVM-b uses "Tb \ common case" which is another of the differential cases (1). (1) ".

  Here, the symbol “\” represents a difference set. For example, “Ta \ common case (1)” is obtained by excluding data belonging to the common case (1) from the learning data Ta. Hereinafter, the symbol notation is used in the same manner.

  In step S3 of i = 2, the common case (2) is learned by using the common case (1) whose coefficient has been calculated as an initial value. At this time, the common case (2) is configured as a part of the common case (1) ("common case (1) ⊃ common case (2)"). The initial value is extracted from the corresponding part. In step S4 with i = 2, learning is performed by combining the common case (2) and one difference case (2), and SVM-c is constructed. The difference case (2) is “Tc \ common case (2)”.

  In steps S3 and S4 where i = 3, 4, learning and construction similar to the case where i = 2 are repeated. In step S3 where i = 3, the common case (3) is learned using the common case (2) as an initial value (using a part of the coefficient). In step S4 where i = 3, SVM-d is constructed by learning by combining the common case (3) and one difference case (3) = “Td \ common case (3)”. In step S3 of i = 4, the common case (4) is learned using the common case (3) as an initial value (using a part of the coefficient). In step S4 with i = 4, SVM-e is constructed by learning by combining the common case (4) and one difference case (4) = "Te \ common case (4)", and all SVMs are built. Complete.

  In the embodiment of FIG. 5, as shown in the figure, only one common case (1) is learned, and the loop ends when i = 1. Common case (1) = “Ta∩Tb∩Tc∩Td∩Te”. That is, the common case (1) is a common case of all learning data.

  In step S3 with i = 1, the common case (1) is learned as a single SVM. In step S4 with i = 1, learning is performed by combining the common case (1) and the five difference cases (1), and all SVMs to be built are built, and the flow after i = 2 is executed. Quit without For example, SVM-a is constructed by learning by combining the common case (1) and one difference case (1) = “Ta \ common case (1)”. Similarly, the remaining SVM-b to SVM-e are constructed by learning by combining each difference case (1) obtained by subtracting the common case (1) from each learning data Tb to Te and the common case (1). The

  In the embodiment of FIG. 6, first, bottom-up clustering is performed on the five determination criteria based on their similarities, and the determination criteria are connected by a similarity relationship. The similarity is determined by the size of the ratio of the common cases to the total data. After clustering (integration), the ratio of the common cases is determined using the total number of data when the initial integration is not performed at all as the denominator. Further, the similarity may be determined based on other criteria.

  Here, clustering is performed in the following order. [1] Ta and Tb are integrated 90% in common to make Ta∩Tb. [2] Tc and Td are integrated 80% in common to make Tc∩Td. [3] “Ta∩Tb” and “Tc∩Td” will be integrated by 70% in common and become “Ta∩Tb∩Tc∩Td”. [4] "Ta ∩ Tb ∩ Tc ∩ Td" and Te are integrated 60% in common and become "Ta ∩ Tb ∩ Tc ∩ Td ∩ Te" As shown in the figure, each common case (i) is determined in the reverse order of the integrated cluster, and the learning order is also determined in the same manner.

  In step S3 where i = 1, the common case (1) = “Ta∩Tb∩Tc∩Td∩Te” is learned as a single SVM. In step S4 with i = 1, SVM-e is constructed by learning by combining the common case (1) and one difference case (1) = “Te \ common case (1)”.

  In step S3 of i = 2, the common case (2) = “Ta∩Tb∩Tc∩Td” is learned using the common case (1) as an initial value. Here, since it is configured as common case (2) = “common case (1) 事例 some case”, the corresponding coefficient is set to zero and an initial value is obtained. Step S4 with i = 2 is skipped because the construction target SVM is not set. Therefore, it is not necessary to set the differential case (2).

  In step S3 with i = 3, common case (3) = "common case (2) ∪ some case", so set the corresponding coefficient to zero and use common case (2) as the initial value. (3) = “Tc∩Td” is learned. In step S4 with i = 3, the common case (3) and the two difference cases (3) are learned in combination to construct SVM-c and SVM-d. SVM-c is constructed with one difference case (3) = "Tc \ common case (3)", SVM-d is built with one difference case (3) = "Td \ common case (3)" .

  In step S4 with i = 4, common case (4) = "common case (2) ∪ some case", so set the corresponding coefficient to zero and use the common case (2) as the initial value. (4) = “Ta∩Tb” is learned. In step S4 where i = 4, the common case (4) and the two difference cases (4) are learned in combination, SVM-a and SVM-b are built, and the construction of all SVMs is completed. SVM-a is constructed with one difference case (4) = "Ta \ common case (4)", and SVM-b is built with one difference case (4) = "Tb \ common case (4)" .

  The embodiment of FIG. 7 can be applied in combination with the embodiments of FIGS. FIG. 7 shows the result of bottom-up clustering similar to FIG. Here, when a predetermined threshold is set for the similarity, and clustering is terminated for a common ratio of 75% or less, for example, a cluster 1 (Ta and Tb) and a cluster 2 (Tc and Td) are obtained as a plurality of clusters as illustrated. Cluster 3 (Te) is obtained. In this embodiment, the embodiments of FIGS. 4 to 6 are applied to each cluster individually.

  In addition, instead of the embodiment of FIG. 7 by bottom-up clustering, more generally, a plurality of judgment criteria for SVM construction are divided into predetermined groups based on the similarity between judgment criteria, and each group is individually The embodiments of FIGS. 4 to 6 may be applied. At this time, a predetermined upper limit may be provided for the number of determination criteria belonging to each group. Since the present invention assumes that judgment criteria are somewhat similar, applying the present invention individually within a group of judgment criteria having a certain degree of similarity reduces the amount of calculation as a whole more reliably. Is done.

  In the example of FIG. 4, the order of Ta to Te was not particularly limited, but the highest similarity among the five judgment criteria is Ta and Tb, and the similarity to “Ta∩Tb” Tc is the highest of the remaining three criteria, and Td is the highest of the remaining two criteria, which is similar to Ta∩Tb∩Tc. Also good.

  Moreover, you may combine the Example of FIGS. 4-6 suitably. For example, the common case (1) and the difference case (1) in the embodiment of FIG. 4 may be set, and the method of FIG. 5 may be adopted for the common case (2) and the difference case (2). That is, common case (2) = "Ta∩Tb∩Tc∩Td∩Te", difference case (2) = "Tc \ common case (2)", "Td \ common case (2)" and "Te \ common As example (2), all SVMs may be constructed with i = 1,2.

  8 and 9 are graphs showing an example of the effect of applying the present invention in comparison with the prior art. Both show the case of constructing an SVM of learning data private1, private2, and private3 according to three criteria, and the present invention is applied by the method of FIG. 4 and is pre-learned as a common case at i = 1. Private1 and private2 are constructed using “private1∩private2”, and private3 is constructed using common2 = “private1∩private2∩private3” learned in advance using common1 as a common case in i = 2.

  8 and 9, the individual learning time (private1 + private2 + private3) in the prior art is indicated by the item name of LibSVM, and the learning time according to the present invention (common1 + common2 + private1 + private2 + private3) is indicated by the item name of PerSVM (Personal SVM). The degree of reduction in processing time (PerSVM processing time ÷ LibSVM processing time) is also shown.

  8 and 9, the items of LibSVM and PerSVM are arranged on the left and right and the learning data number is changed from 50,000 to 250,000. As an index showing the degree of similarity of each judgment criterion, when the ratio of different labels is adopted due to the difference in judgment criteria in all learning data, the example when the degree of difference is 5% and the example when it is 10%, respectively FIG. 8 and FIG.

  8 and 9, it can be seen that the method of the present invention (PerSVM) always requires less time for learning than the conventional method on the left (LibSVM). In addition, it can be seen that the degree of reduction in processing time is larger when the degree of difference is 5% than when it is 10%. From this, it can be seen that the present invention is effective if the judgment criteria are similar.

  As described above, according to the present invention, when learning a plurality of similar determination criteria, all of the learning of a plurality of SVMs can be completed in a shorter time than the conventional method.

1 ... Device for constructing a support vector machine, 2 ... Common case learning unit, 3 ... Difference case learning unit, 4 ... Learning order setting unit

Claims (5)

A method of constructing a plurality of criteria per support vector machine respectively correspond to a plurality of learning data having mutually common case in the label application to the given data,
A first step of learning a support vector machine with a common case in a part of a plurality of learning data corresponding to each of the plurality of determination criteria to obtain a coefficient of a decision function corresponding to the common case;
A second support vector machine is constructed by learning a support vector machine by using a coefficient of a decision function obtained for the common case as an initial value, so as to be included in a part of the plurality of judgment criteria. And the steps
And a third step of constructing a support vector machine corresponding to each of the plurality of determination criteria by repeatedly executing the first and second steps sequentially for the plurality of determination criteria. How to build a vector machine.   In the third step, when the support vector machine repeatedly executes the first and second steps sequentially for the unestablished criterion, the common example in which the coefficient of the decision function has already been obtained in the first step 2. The coefficient of the decision function is obtained as an initial value by using a coefficient in step (2), and a support vector machine corresponding to the unbuilt criterion is constructed in the second step. How to build a support vector machine.   3. The method according to claim 1, wherein, in the third step, the first and second steps are repeatedly executed sequentially in a predetermined order based on the similarity between the determination criteria for the plurality of determination criteria. How to build the described support vector machine.   4. The method according to claim 1, wherein the third step is individually executed for each predetermined group obtained by dividing the plurality of determination criteria based on the similarity between the determination criteria. 5. How to build a support vector machine. An apparatus for constructing a support vector machine for a plurality of judgment criteria respectively corresponding to a plurality of learning data having mutually common cases in labeling for given data,
A common case learning unit that learns a support vector machine in a common case in a part of a plurality of learning data respectively corresponding to the plurality of determination criteria and obtains a coefficient of a decision function corresponding to the common case;
A differential case that constructs a support vector machine of a judgment criterion included in a part of the plurality of judgment criteria by learning a support vector machine by using a coefficient of a decision function obtained for the common case as an initial value The learning department,
A learning order setting unit that sets a predetermined learning order of the plurality of determination criteria based on a common case;
A support vector machine corresponding to each of the plurality of determination criteria is constructed by sequentially applying the common case learning unit and the difference case learning unit according to the learning order set by the learning order setting unit. A device to build a support vector machine.