JPWO2010035690A1 - Structure learning device, structure learning system, structure learning method, program, and recording medium - Google Patents

Abstract

Provided is a structure learning device in which a plurality of agents share a network structure and can improve each learning result. A network structure learning unit that learns a network structure from a shared network and determines a candidate network structure, and a network structure mixed parameter that estimates a mixed parameter of a network structure that is a candidate by the network structure learning unit at a network structure level An estimation unit; and an agent mixed parameter estimation unit that estimates a network structure mixed parameter that is a candidate by the network structure learning unit at an agent level.

Description

  The present invention relates to a structure learning device, a structure learning system, a structure learning method, a program, and a recording medium that improve the network structure learning result of the agent by sharing and using network structure information of other agents.

  The learning method of the network structure learning system is described in Non-Patent Document 1, for example.

  The network structure learning system described in Non-Patent Document 1 includes network structure learning means and network structure mixed parameter estimation means. The network structure learning system having such a configuration operates as follows.

  That is, network structure learning and network structure mixed parameter estimation are performed by a method called an EM algorithm until the results converge.

Marina Meila, Michael I. Jordan, "Learning with Mixtures of Trees", Journal of Machine Learning Research, Vol 1, p.1-48 (2000)

  However, since the technique described in Non-Patent Document 1 does not consider sharing of the network structure by a plurality of agents, the plurality of agents share the network structure and cannot improve each learning result.

  In addition, since the network structure sharing by a plurality of agents and the similarity of the plurality of agents are not taken into consideration, the plurality of agents cannot be clustered into similar ones.

  Further, since a parametric model is used, it is necessary to determine a finite value as the number of models mixed in advance.

  Accordingly, the present invention has been made in view of the above problems, and an object of the present invention is to provide a structure learning apparatus in which a plurality of agents can share a network structure and improve each learning result.

  In order to solve the above problems, a structure learning device according to the present invention learns a network structure from a shared network and determines a candidate network structure, and a network structure made a candidate by the network structure learning means Network structure mixed parameter estimating means for estimating the mixed parameters of the network structure at the network structure level, and agent mixed parameter estimating means for estimating the network structure mixed parameters that are candidates by the network structure learning means at the agent level. It is characterized by.

  The network structure generating means for generating a new network structure when the network structure learning means determines that there is no suitable network structure as a candidate.

  The network structure learning means is characterized in that it repeatedly learns until the estimation results by the network structure mixed parameter estimation means and the agent mixed parameter estimation means converge.

  The structure learning system according to the present invention includes a plurality of the structure learning devices described above.

  The structure learning method according to the present invention is a network structure learning step for learning a network structure from a shared network and determining a network structure as a candidate, and a network structure mixed parameter determined by the network structure learning step as a network. A network structure mixed parameter estimating step for estimating at a structure level, and an agent mixed parameter estimating step for estimating a network structure mixed parameter made a candidate by the network structure learning step at an agent level.

  In the network structure learning step, a network structure generation step is provided for generating a new network structure when it is determined that there is no network structure suitable as a candidate.

  The network structure learning step is characterized by iteratively learning until the estimation results obtained by the network structure mixed parameter estimation step and the agent mixed parameter estimation step converge.

  A program according to the present invention causes a computer to execute the structure learning method described above.

  A recording medium in the present invention is a computer-readable recording medium storing the above-described program.

  According to the present invention, since a model that considers sharing of the network structure by a plurality of agents is used, a plurality of agents can share the network structure and improve each learning result.

It is a hardware block diagram which shows the structure which concerns on embodiment of this invention. It is a system block diagram which shows the structure which concerns on embodiment of this invention. It is a flowchart figure which concerns on embodiment of this invention. It is a figure which shows the specific example of the operation | movement which concerns on embodiment of this invention.

  Next, the best mode for carrying out the invention will be described in detail with reference to the drawings.

  FIG. 1 is a hardware configuration diagram of a network structure learning apparatus according to an embodiment of the present invention. The apparatus includes an operation unit 1, a CPU 2, a memory 3, a ROM 4, and a network interface 5.

  The CPU 2 controls the entire apparatus, and the ROM 4 stores programs executed by the CPU 2 and other fixed data. The memory 3 is used for temporary storage of data when the CPU 2 performs calculations. Further, a network is realized by the network interface 5, and information can be input / output by the operation unit 1.

  FIG. 2 is a system block diagram of the network structure learning apparatus according to the embodiment of the present invention. The network structure learning apparatus includes an input unit 11, a network structure learning unit 12, a network structure generation unit 13, a network structure mixed parameter estimation unit 14, an agent mixed parameter estimation unit 15, and an output unit 16. Configured.

  The input unit 11 takes in the data to be observed, and the network structure learning unit 12 learns the network structure that generates the observation data.

  The network structure generation unit 13 generates a new network structure when the network structure learning unit 12 does not match the existing network structure well with the observation value.

  The network structure mixed parameter estimation means 14 estimates the mixing probability at the network structure level, and the agent mixing parameter estimation means 15 estimates the mixing probability at the agent level.

  On the other hand, the output means 16 outputs the learned network structure and the estimation result in the mixing probability.

  With reference to the flowchart of FIG. 3, the overall operation in the embodiment of the present invention will be described in detail.

  First, data to be observed is captured by the input means 11 (step S1). Next, the network structure learning unit 12 learns the most suitable structure as a network structure for generating observation data (step S2). It is determined whether or not there is (step S3).

  If none of the network structures obtained by learning so far fits well, a new most suitable network structure is generated from the network structure base distribution (step S4).

  When there is a compatible network structure or after a new network structure is generated, the network structure level is calculated to calculate what mixing probability the network structure has, and the mixing probability is estimated (step S5). It is determined whether or not it has converged (step S6). Here, when it does not converge, it returns to step S2 and learns again.

  On the other hand, if it converges, it calculates what mixing probability the agent has at the agent level, estimates the mixing probability (step S7), and determines whether the estimation result has converged (step S8). . Here, when it does not converge, it returns to step S2 and learns again.

  It repeats until the result of learning / estimation converges. Finally, the network structure learned in step S2 and the mixed probability estimated in steps S5 and S7 are output (step S9).

  Hereinafter, the operation of the best mode for carrying out the present invention will be described more specifically with reference to FIG.

  Given a network structure G = (V, E) consisting of a vertex set {vεV} representing a certain attribute and a branch set {(u, v) εE} connecting the vertices, the network observation data x Consider a generation model.

  A parameter set representing the behavior of a vertex pair and a branch (u, v) ∈ E between them is θE = {θuv} (u, v) ∈E, and θT = {E, θE} is defined including the network structure. The probability distribution T of the network is

It is expressed.

  Next, let us consider a case where agents j, (j = 1,..., J) share information related to each other's network structure when they have vertices representing common attributes among the agents. Let xji, (i = 1,..., Nj) be an observed value from agent j, (j =,..., J). Here, Nj represents the number of observations of data at agent j. The observation values are assumed to be exchangeable within the agent, and are generated according to the probability distributions xji to T (· | θT: ji). Here the network structure and its parameter θT: ji are generated from Gj,

Are linked via a mixed model represented by the following mathematical expression.

  Where δ is the Dirac delta function. The mixing probability πk, ωkl is

Follow. This is one of the construction methods of Dirichlet process and is called stick-breaking process. This is expressed as π (= (π1,..., Πk)) to Stick (α), ωk (= (ωk1,..., Ωkl)) to Stick (β). H is the prior distribution for the network structure and its parameters.

  {G1,..., Gj} used as a mixture distribution in the above process is said to follow a nested dirichlet process (nDP). The parameters are α, β, H and are expressed as nDP (α, β, H). Since nDP uses a mixed distribution at both the network structure level and the agent level, network structure and agent clustering can be performed simultaneously.

  Here, the following index is introduced. The index cj represents the mixed network structure that generates the network structure of agent j. The index sji represents the network structure for generating the data point xji.

  Learning and estimating nDP and shared network structure can be done based on Gibbs sampler, and the following steps are repeated until convergence.

i). For all agents j = 1,..., J, the index cj is sampled according to the posterior probability p (cj = k | {xji}).

ii). For all agents j = 1, ..., J and all data i = 1, ..., Nj, index sji is sampled according to posterior probability p (sji = 1 | {xji}) To do.

iii). The agent level mixture weight πk is generated according to the posterior probability p (πk = {xji}, {cj}).

iv). The network structure level mixture weight ωkl is generated according to the posterior probability p (ωkl = {xji}, {sji}).

v). Network structure and parameters

The posterior probability

According to the sample. If no observation is assigned, it is generated from the prior distribution H.

  In the processing described above, (ii) and (iv) correspond to the network structure mixed parameter estimating means 14 of FIG. 2 and step S5 of FIG. 3, and (i) and (iii) are the agent mixed parameters of FIG. 3 corresponds to the estimation means 15 and step S7 in FIG. 3, and (v) corresponds to the network structure learning means 12, the network structure generation means 13 in FIG. 2, and steps S2 to S4 in FIG.

  Next, the effect of this embodiment will be described.

  In the present embodiment, since a plurality of agents share a network structure, a plurality of agents can share a network structure and improve each learning result.

  Further, in the present embodiment, since a mixed model is used for each agent model, a plurality of agents can be clustered into similar ones.

  Further, in the present embodiment, modeling, learning, and estimation are further performed using a nonparametric model, so that it is not necessary to set a finite value as the number of models mixed in advance.

  The present invention can be used as follows.

  For example, in a mechanical system such as an automobile, a correlation structure is observed in the behavior between specific parts of various vehicle types. These parts correlation structures are the result of realizing a specific function, and almost the same parts correlation structure is shared among vehicle types.

  When considering multiple vehicle types (light, truck, diesel, hybrid, etc.), the basic part correlation structure such as engine control shows similar behavior, but for example, in the case of hybrid vehicles, electronic motor control, etc. It has a parts correlation structure unique to the model.

  In such a case, with regard to the part correlation structure that is common among the vehicle types, the learning accuracy is improved by sharing information on other vehicle types, and the part correlation structure that can be found only in a specific vehicle type is learned independently for each vehicle type. Furthermore, the present invention can be applied to such a case that vehicle types that share a part correlation structure with high uniqueness in the vehicle type are clustered as similar vehicle types.

  Also, in a data consortium where financial institutions share internal loss data, when considering internal loss events of each bank, it is said that there is usually a correlation structure between occurrences of specific loss events.

  When considered among multiple banks, loss events related to a specific business such as deposit business and loan business are considered to have a similar loss event correlation structure.

  In such a case, regarding the loss event correlation structure common to the banks, the learning accuracy can be improved by sharing information of other banks.

  However, when considering the characteristics of each bank (the number of specific operations, regional characteristics such as the occurrence of earthquakes and typhoons), there is also a loss event correlation structure that can be found only in specific banks. The present invention can also be applied to applications such as sharing information between banks having the specific loss event correlation structure and clustering them as banks having a similar loss event correlation structure.

  The present invention has been described above by the preferred embodiments of the present invention. While the invention has been described with reference to specific embodiments thereof, various modifications and changes can be made to these embodiments without departing from the broader spirit and scope of the invention as defined in the claims. is there.

  This application claims priority based on Japanese Patent Application No. 2008-244168, filed on September 24, 2008, the entire disclosure of which is incorporated herein.

DESCRIPTION OF SYMBOLS 11 Input means 12 Network structure learning means 13 Network structure generation means 14 Network structure mixed parameter estimation means 15 Agent mixed parameter estimation means 16 Output means

Claims (9)

A network structure learning means for learning a network structure from a shared network and determining a candidate network structure;
Network structure mixed parameter estimating means for estimating the network structure mixed parameters that are candidates by the network structure learning means at a network structure level;
A structure learning apparatus comprising: an agent mixed parameter estimation unit that estimates a mixed parameter of the network structure that is a candidate by the network structure learning unit at an agent level.   2. The structure learning apparatus according to claim 1, further comprising a network structure generation unit that generates a new network structure when it is determined by the network structure learning unit that there is no network structure suitable as a candidate.   The structure learning apparatus according to claim 1, wherein the network structure learning unit repeatedly learns until estimation results obtained by the network structure mixed parameter estimation unit and the agent mixed parameter estimation unit converge.   A structure learning system comprising a plurality of the structure learning devices according to claim 1. A network structure learning step for learning a network structure from a shared network and determining a candidate network structure;
A network structure mixed parameter estimation step for estimating a network structure mixed parameter that is a candidate in the network structure learning step at a network structure level;
A structure learning method comprising: an agent mixed parameter estimation step that estimates a mixed parameter of the network structure that is a candidate in the network structure learning step at an agent level.   6. The structure learning method according to claim 5, further comprising a network structure generation step of generating a new network structure when it is determined that there is no network structure suitable as a candidate in the network structure learning step.   7. The structure learning method according to claim 5 or 6, wherein the network structure learning step repeatedly learns until estimation results obtained by the network structure mixed parameter estimation step and the agent mixture parameter estimation step converge.   The program which makes a computer perform the structure learning method of any one of Claim 5 to 7.   A computer-readable recording medium storing the program according to claim 8.

Images