JP5519210B2 - Model generation apparatus, model generation method, information processing system, and program - Google Patents

Description

  The present invention relates to an apparatus, a method, and a program for generating a Bayesian network model.

  In recent years, there has been a growing need for devices that estimate user preferences and intentions and provide operations and services desired by users.

  Thus, in order to estimate the operation and service desired by the user, a Bayesian network is used (for example, Patent Document 1). A Bayesian network expresses causal relationships and dependencies of various events using Bayes' theorem based on statistical data. Complex causal relationships and dependencies are represented by directed graphs, and individual qualitative It is a model of probabilistic reasoning that expresses the relationship as a conditional probability. In Patent Document 1, a model of a Bayesian network is created based on data collected by conducting a questionnaire survey on subjects, and the user's preference is estimated by checking the user's situation against the created model. The technology for recommending content is described.

JP 2007-058398 A

  However, in patent document 1, since a user's preference is estimated using the Bayesian network model created based on general statistical data, there exists a problem that a deviation from a user's individual preference will actually arise. .

  In addition, Bayesian networks derive a complex relationship between events by qualitatively capturing a wide variety of discrete data and estimate the possibility of uncertain events. Therefore, in order to estimate the operation and service desired by the user for each user, a sufficient amount of information about the user must be accumulated. Therefore, the Bayesian network is not suitable for personal use.

  Therefore, the problem to be solved by the present invention is to solve the above-mentioned problems, and to provide a technique that can be used by specializing a Bayesian network.

The model generation apparatus of the present invention that solves the above problems includes a setting unit that sets a parameter of a function that calculates a prior probability distribution according to an accumulation state of measured data, and a prior probability distribution calculated using the set parameter. And a model generation means for determining a conditional dependency between random variables based on the model and generating a Bayesian network model based on the correlation between the random variables constructed based on the determined conditional dependency. The model generation means, when the accumulated amount of the actual measurement data is a small amount for setting the conditional dependence, and when the actual measurement data is at least one older than a predetermined period, A correlation between the random variables is constructed based on a predetermined conditional dependency .

The information processing apparatus of the present invention that solves the above problems includes a setting unit that sets a parameter of a function that calculates a prior probability distribution according to an accumulation state of measured data, and a prior probability distribution calculated using the set parameter And a model generation means for determining a conditional dependency between random variables based on the model and generating a Bayesian network model based on the correlation between the random variables constructed based on the determined conditional dependency. The model generation means, when the accumulated amount of the actual measurement data is a small amount for setting the conditional dependence, and when the actual measurement data is at least one older than a predetermined period, A correlation between the random variables is constructed based on a predetermined conditional dependency .

The model generation method of the present invention that solves the above problems includes a setting step of setting a parameter of a function that calculates a prior probability distribution according to an accumulation state of measured data, and a prior probability distribution calculated using the set parameter And a model generation step of generating a Bayesian network model based on the correlation between the random variables constructed based on the determined conditional dependency. In the model generation step, when the accumulated amount of the actual measurement data is a small amount for setting the conditional dependence, and when the actual measurement data is older than a predetermined period, A correlation between the random variables is constructed based on a predetermined conditional dependency .

The present invention for solving the above-described problem is a program for an information processing apparatus, wherein the program sets a parameter of a function for calculating a prior probability distribution in accordance with an accumulation state of measured data in the information processing apparatus. Determining conditional dependence between random variables based on a prior probability distribution calculated using the set parameters, and based on a correlation between the random variables constructed based on the determined conditional dependence Te to execute the model generation process of generating a model of the Bayesian network, the model generation process, when the accumulation amount of the measured data is in an amount less to set the dependence the conditional, and, actually measured data If at least one of the older than a predetermined time period, and wherein constructing the correlation between the random variables based on the dependence with predetermined conditions That.

  According to the present invention, a Bayesian network model can be generated even if a sufficient amount of data is not accumulated.

It is a block diagram of a 1st embodiment of the present invention. It is a figure for demonstrating the algorithm of this invention. 3 is a flowchart of the present invention. It is a figure which shows the relationship between a parameter and prior probability distribution. It is a block diagram of the 2nd Embodiment of this invention.

  Generally, various events that occur may occur independently, but there are events that are the cause and often occur due to the effects of the events that cause it.

For example, event X ¹ is “AM / PM”, event X ² is “view / do not watch today's weather forecast content”, and event X ³ is “use / do not use music tool” To do.

In creating a normal Bayesian network model,
・ We watched today's weather forecast in the morning 900 times ・ We saw today's weather forecast in 100 times in the afternoon ・ 500 times used the music tool in the morning ・ 500 times used the music tool in the afternoon Based on a conditional random variable calculated from actual measured values (sample data), it is determined whether the change in the event X ¹ is related to the occurrence of the event X ² and the occurrence of the event X ³ , it creates a dependency between the event ^{X 1} and ^{X 2} and ^{X 3.} By the way, in this example, viewing of today's weather forecast content varies greatly depending on the time zone, so it is assumed that the dependency between events X ¹ and X ² is strong, and the use of music tools varies depending on the time zone. Therefore, there is no dependency between the events X ¹ and X ³ .

However, for example,
・ We saw today's weather forecast three times in the morning ・ We saw today weather forecast twice in the afternoon ・ One time we used music tools in the morning ・ 4 times we used music tools in the afternoon When the number of actually measured values is small, actually, the dependency between the events X ¹ and X ² is strong as in the above example, and there is no dependency between the events X ¹ and X ^{3. 1} and dependencies without the X ^2, there is a possibility that can differ dependent from the actual and so there is a dependency between the event X ¹ and X ^3.

Also,
・ We saw today's weather forecast in the morning 900 times last month ・ 100 times last month we saw today's weather forecast in the afternoon ・ Measured values such as today we saw the weather forecast today in the morning 10 times this month In this example, the dependency relationship between the events X ¹ and X ² becomes strong, even though the distribution of sigma changes rapidly with time and the dependency relationship between the events X ¹ and X ² has now weakened. As a result, there is a possibility that the dependency may be different from the actual one.

  Therefore, according to the present invention, the dependency relationship between each event (probability variable) for which a Bayesian network model is to be created is set according to the accumulation state of the actual measurement value of each event.

  In order to explain the features of the present invention, it will be specifically described below with reference to the drawings.

<First Embodiment>
A first embodiment for carrying out the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a block diagram of a model generation apparatus 100 of the present invention.

  The information collection unit 101 collects observation values of various events for creating a Bayesian network model. The collected information is stored in the statistical data 102 for each event. At this time, in order to make the continuous value data meaningful, it is sampled or quantized and clustered (discretized). Remember. The observation values collected by the information collecting unit 101 are information having variables or fluctuations such as, for example, user position information, time information, weather information, log information of user actions or operations performed by the user on the apparatus. . Information may be collected when new data is observed or intermittently, and is appropriately performed according to the nature of the information to be collected.

  The statistical data 102 stores various data collected by the information collecting unit 101 for each event.

  In the reference data 103, a correlation (dependency) between events set in advance is stored as reference data. The reference data may be a Bayesian network model generated from marketing data, or may be a Bayesian network model generated by a model generation device provider.

  The Bayesian network model 105 stores a Bayesian network model generated by the model generation unit 104.

  The model generation unit 104 includes a setting unit 1041 and a construction unit 1042. The setting unit 1041 sets the dependency between each event according to the accumulation state of data stored in the statistical data 102. The construction unit 1042 constructs a correlation between each event using the dependency set by the setting unit 1041 and generates a Bayesian network model. The generated model is stored in the Bayesian network model.

  An algorithm for model creation in the present invention will be described.

An algorithm shown in FIG. 2 is used as a model generation algorithm. The currentGraph in this algorithm represents the current Bayesian network model. On the other hand, TemporaryGraph indicates the model of the immediately preceding Bayesian network. When the algorithm shown in FIG. 2 sets new data stored in the statistical data 102,
A1: Focus on one event.
A2: Build a correlation by deleting or adding an event having a correlation with the event in a state including new data.
A3: A2 is executed for each of all n events.
A4: A model is generated using the constructed correlation.
A5: Compare the generated model with the model before including new data, and generate them until they are the same, that is, for new data.
It becomes.

  FIG. 3 is a flowchart showing this algorithm. The following operation may be started every time a predetermined amount of new data is stored, may be started every time new data is stored, or may be started at a predetermined time. Further, a plurality of setting units 1041 may be prepared and operated in parallel for different events.

  First, it is determined whether new data stored in the statistical data 102 is added to the model (step S1). New data can be added depending on the model to be generated and the nature or characteristics of the information to be measured, for example, for each event, or a plurality of correlated events may be grouped, and data may be added for each group. All new data may be added.

  The model stored in the Bayesian model 105 is stored in the variable g as the current Bayesian network model currentGraph (step S2).

  It is determined whether the variable g matches the variable g ′ in which TemporaryGraph, which is a model being created, is stored (step S3). If there is no model being created, it is determined that the models do not match because the model is not stored in the variable g ′.

  If it is determined in step S3 that they do not match, the variable g 'is stored in the variable g (step S4).

  An event correlated with the event of interest is added and deleted, and a correlation of the event of interest is constructed (step S5).

  Based on the constructed correlation, a Bayesian network model is generated, which is used as TemporaryGraph, and TemporaryGraph is stored in variable g '(step S6).

  Steps S2 and S3 are executed for all events (step S7).

  If it is determined in step S3 that they do not match, that is, if a Bayesian network model that takes into account the added new data is created, the process returns to step S1 to determine whether to add new data. If not, TemporaryGraph is set to currentGraph. And stored in the Bayesian model 105 (step S8).

  Here, step S5 will be described in detail.

  Bayesian network, as its name suggests, is a technique that uses Bayes' theorem to express the causal relationship (dependency) between events representing each observation, that is, between random variables, as a network consisting of directed edges, and performs probability inference. It is.

と表すことができ、これを単一サンプルデータと呼ぶ。そして、ｎは、単一サンプルデータのサンプルサイズを表す。Ｖはデータの分散を示し、

で定義される集合である。 If there are N events in the statistical data 102, the entire statistical data, which is a set of random variables, can be expressed as X ^N = X ⁽¹⁾ X ⁽²⁾ ... X ^(N) . Each random variable is a collection of data. The set of data for each random variable is

This is called single sample data. N represents the sample size of single sample data. V indicates the distribution of data,

Is a set defined by

Some random variable and ^X i, ^{X j, the} correlation between ^the variables ^{X j} and X ⁱ is the variable ^{X j} are affected by ^{X i,} that, conditional between ^{X i} and ^{X j} The dependency is expressed as X ⁱ → X ^j . When this relationship is established, X ⁱ is called a parent node and X ^j is called a child node. Events that occur in general do not have a single causal relationship of obtaining one consequent event for one causal event, but have a causal relationship in which various events are intertwined in a complex manner. Therefore, a plurality of parent nodes exist for one child node, which is expressed as follows. In step S5, attention is paid to each event that can be a child node, and an event that can be a parent node is deleted and added to be searched.

と表すが、これらは子ノードに相関関係があると決定した親ノードである。親ノードを決定するには、各事象を親ノードの候補として追加及び削除することにより、最適なモデルの構造を探索しなければならない。そのために、親ノードの候補を追加したり削除したりした際に、追加又は削除した状態のモデルを評価して構造学習を行い、ヒューリスティックにモデルを生成する。この構造学習には、ベイズ情報量基準を近似した記述長基準（Ｍｉｎｉｍｕｍ ｄｅｓｃｒｉｐｔｉｏｎ ｌｅｎｇｔｈ：ＭＤＬ）を用いる。ある確率変数のサンプルデータＸ^Ｎが与えられたときの、その確率変数のモデルｇの記述長基準Ｌ(ｇ｜Ｘ^Ｎ)は、

と定義づけられる。式（１）の項目は、

と定義される量である。ここで、親ノード候補の集合φ(j)は

と表され、これは{1, 2, …, j-1}の部分集合である。

は、ｉ番目のサンプルＸ^(ｉ)における親ノードの集合

の実現値の組となる。Ｓ（ｊ；ｇ）は、モデルｇが与えられたときの変数Ｘ^ｊに関する親ノードの集合Πｊ実現値全体の集合を表す。1｛ ｝は、指示関数を表し、ある事象が真である場合に「１」を、偽の場合に「０」を返す関数である。n[s,j;g]は、モデルｇの事象ｓ及びｊの周辺頻度関数を表し、n[q,s,j;g]はモデルｇの事象ｑ、ｓ及びｊの同時頻度関数を表す。周辺頻度関数と同時頻度関数との間には、

の関係が成り立つ。これをもとに条件付頻度関数n[q,j|s;g]を定義すると、

と定義することが出来る。この条件付頻度関数は、

という性質を持つ。同様に、同時関数及び周辺頻度関数についても、

という性質を持つ。 As above, the parent node is

These are parent nodes that have been determined to be correlated with child nodes. In order to determine a parent node, an optimal model structure must be searched by adding and deleting each event as a parent node candidate. Therefore, when a candidate for a parent node is added or deleted, the model in the added or deleted state is evaluated to perform structure learning, and a model is generated heuristically. This structure learning uses a description length criterion (MDL) that approximates a Bayesian information criterion. Given sample data X ^N of a random variable, the description length criterion L (g | X ^N ) of the model g of the random variable is

It is defined as The item of formula (1) is

It is an amount defined as Here, the parent node candidate set φ (j) is

Which is a subset of {1, 2,…, j-1}.

Is the set of parent nodes in the i th sample X ⁽ⁱ⁾

Is a set of realization values. S (j; g) represents a set of parent nodes Π ^j realization value for the variable X ^j given the model g. 1 {} represents an instruction function, and is a function that returns “1” when an event is true and returns “0” when it is false. n [s, j; g] represents the peripheral frequency function of events s and j of model g, and n [q, s, j; g] represents the simultaneous frequency function of events q, s and j of model g . Between the peripheral frequency function and the simultaneous frequency function,

The relationship holds. If we define a conditional frequency function n [q, j | s; g] based on this,

Can be defined as This conditional frequency function is

It has the nature of Similarly, for simultaneous functions and marginal frequency functions,

It has the nature of

と書き直すことが出来る。このことから、各変数に対して

を最小化することとなる。 The expression (1) of the description length criterion L (g | X ^N ) evaluates a term representing an empirical entropy (a function given by H) and a degree of freedom of a Bayesian network (a function given by k) serving as a penalty term. It is shown that. Therefore, this formula eliminates or adds parent node variables to the child nodes of interest to create a temporary model structure, thereby minimizing empirical entropy and freedom of the joint probability distribution of the entire Bayesian network model. Indicates that a small model structure is selected. Furthermore, when attention is given to each variable of the description length criterion L (g | X ^N ),

Can be rewritten. From this, for each variable

Will be minimized.

The joint probability distribution of the entire Bayesian network model is a multinomial distribution. In the present invention, a Bayesian network model is constructed from the observation data collected in the statistical data 102 using the reference data stored in the reference data 103, that is, the prior assumptions. Since the information criterion is used, a Dirichlet distribution that is a conjugate prior distribution of a multinomial distribution is assumed as a prior probability distribution. This Dirichlet distribution is a multivariate distribution, and assumes the entire prior probability distribution of a CPT (Conditional Probability Table) in which conditional probabilities between parent nodes and child nodes are tabulated. Since the Dirichlet distribution is a mixed distribution of β distribution, in order to pay attention to one entry of the CPT, that is, a state with a parent node, the β distribution may be considered. Prior probability distribution f (x) of each entry of CPT is given by the following β distribution.

  Here, the item given by p in Equation (3) is a parameter for the prior probability distribution. When p is sufficiently close to 0, a prior probability distribution having a dependency relationship between a child node and a parent node of interest is assumed. On the other hand, by putting a large value in p, an a priori probability distribution having no dependency relationship between the child node and the parent node of interest is assumed. FIG. 4 shows the relationship between the parameters and the prior probability distribution. The horizontal axis in FIG. 4 indicates the prior probability value, and the vertical axis indicates the distribution of the value. When a distribution close to 0 or 1 is set as a prior probability distribution, a causal relationship is assumed between the child node and the parent node of interest (graph A: p = 0.5). Conversely, when a distribution that is convex at the center is used, it is assumed that there is no dependency between the child node and the parent node of interest (graph B: p = 3). That is, when p is small, a U-shaped distribution biased to 0 and 1 is assumed, and a dependency relationship is assumed as a prior probability distribution between the child node and the parent node of interest. When P = 1, the distribution is uniform and nothing is assumed (graph C). If the value of p increases, the shape approaches a normal distribution with an average of 1/2, and it is assumed that there is no dependency between the child node and parent node of interest. As a result, the value of p acts to weaken the influence of the degree of freedom of the probability distribution of the entire model of the Bayesian network acting as a penalty term.

  In the present invention, the setting unit 1041 adjusts the parameter p shown in FIG. 4 according to the accumulation state of the data collected in the statistical data 102, and the dependency between the child node and the parent node of interest. Strengthen the relationship. For example, if the setting unit 1041 determines that the accumulated amount of data collected in the statistical data 102 is a small amount for setting the dependency using the data of the statistical data 102, the setting unit 1041 stores the data in the reference data 103. Set with the dependencies that are present. If the setting unit 1041 determines that the data of a certain event collected in the statistical data 102 is data older than a certain period, the setting unit 1041 sets using the dependency stored in the reference data 103, For example, the dependency derived from the data of the statistical data 102 and the dependency stored in the reference data 103 are set in consideration. In addition, a dependency that is derived in consideration of characteristics of data transition such as abrupt change in data of an event collected in the statistical data 102 and repetition at a certain period is set. More specifically, since the information is not yet accumulated in the statistical data 102 in the initial state, if the Bayes information criterion is used as it is, a result in which there is no dependency between the child node and the parent node of interest is output. I often do. Therefore, the prior probability distribution p shown in FIG. 4 is adjusted to strengthen the dependency between the child node and the parent node of interest. That is, the parameter term represented by p is brought close to 0. Changing this parameter will positively strengthen the dependency with the parent node that is likely to be related to the child node of interest. This is because the information observed by the information collecting unit 101, that is, the probability variable is known in advance, so that the influence relationship between the influence random variables is set in advance. On the other hand, the dependency may be set to be weak.

  According to the present invention, since the dependency between events can be set dynamically, a model of a Bayesian network can be generated even if a sufficient amount of data is not accumulated. Moreover, since the dependency between each event can be set dynamically according to the transition of data, it is possible to cope with the change in dependency between each event due to a change in environment.

<Second Embodiment>
Next, a second embodiment of the present invention will be described. In addition, the same number is attached | subjected about the structure similar to the said embodiment, and detailed description is abbreviate | omitted.

The model generation device of the present invention can be used for various embedded devices. For example,
・ Dynamic presentation of menus such as mobile phones according to user TPO and behavior patterns ・ Distribution information and advertisement according to user TPO and behavior patterns ・ Network services according to user TPO, behavior patterns and usage patterns It can be used for the operation of various electric products and the user assist operation according to the TPO, the behavior pattern, and the usage pattern of the user.

  FIG. 5 is a block diagram of the information processing apparatus 600 according to this embodiment. In the following explanation, a mobile phone will be described as an example, but it should be executed for a certain event such as providing or presenting contents, various information and services, or performing various operations of the own device. Any device that can determine the process may be used.

  When the information collection unit 101 collects the new measurement value, the information collection unit 101 outputs the measurement value to the estimation unit 602.

  The application unit 601 stores various functions such as a telephone function, a music player, mail, a camera, a game, and the Internet, and tools for executing content.

  The estimation unit 602 verifies suitability for each function and content of the application unit 601 based on the data collected by the information collection unit 101 and the Bayesian network model stored in the Bayesian model 105. Then, based on the verification result, the operation to be executed is estimated, and the corresponding function or content is executed. The verification method is not specified in the present invention, and may be executed using a general technique.

  In the above description, the case where the model generation apparatus is configured in the information processing apparatus has been described. However, the present invention is not limited thereto. For example, the home server that manages a plurality of electrical devices installed in the user's home is configured with a model generation device and the home server measures the user's operation and information that affects the user's operation, or each electric device To create and store a Bayesian network model. Then, in response to a command from the electric device, the model or a part of the model is notified, and the estimation unit provided on the electric device side estimates the operation to be executed by the own device and executes the corresponding operation. It may be.

  According to the present invention, it is possible to create a Bayesian network model for individuals by dynamically setting dependencies between events, and to estimate events that can occur from the current situation based on this model. And functions, contents, and services that match the environment.

  The above-described model generation apparatus and information providing apparatus of the present invention can be configured by hardware as is apparent from the above description, but can also be realized by a computer program. In this case, functions and operations similar to those of the above-described embodiment are realized by a processor that operates according to a program stored in the program memory. Note that some of the functions of the above-described embodiments can be realized by a computer program.

  The present invention has been described with reference to the embodiment and examples. However, the present invention is not necessarily limited to the above embodiment, and various modifications can be made within the scope of the technical idea. I can do it.

DESCRIPTION OF SYMBOLS 100 Model production | generation apparatus 101 Information collection part 102 Statistical data 103 Reference | standard data 104 Model production | generation part 1041 Setting part 1042 Construction part 105 Bayesian model 600 Information processing apparatus 601 Application part 602 Estimation part

Claims (4)

A setting means for setting a parameter of a function for calculating a prior probability distribution according to an accumulation state of measured data;
Based on the prior probability distribution calculated using the set parameters, the conditional dependency between random variables is determined, and based on the correlation between the random variables constructed based on the determined conditional dependency have a model and generating means for generating a model of Bayesian network,
The model generation unit is configured to generate a predetermined amount when the accumulated amount of the measured data is a small amount for setting the conditional dependence and when the measured data is at least one older than a predetermined period. A model generation apparatus characterized in that a correlation between the random variables is constructed based on conditional dependence . A setting means for setting a parameter of a function for calculating a prior probability distribution according to an accumulation state of measured data;
Based on the prior probability distribution calculated using the set parameters, the conditional dependency between random variables is determined, and based on the correlation between the random variables constructed based on the determined conditional dependency have a model and generating means for generating a model of Bayesian network,
The model generation unit is configured to generate a predetermined amount when the accumulated amount of the measured data is a small amount for setting the conditional dependence and when the measured data is at least one older than a predetermined period. An information processing system characterized by constructing a correlation between the random variables based on conditional dependency . A setting step for setting a parameter of a function for calculating a prior probability distribution according to the accumulation state of the measured data;
Based on the prior probability distribution calculated using the set parameters, the conditional dependency between random variables is determined, and based on the correlation between the random variables constructed based on the determined conditional dependency possess a model generating step of generating a model of the Bayesian network,
The model generation step is performed when the accumulated amount of the measured data is a small amount for setting the conditional dependence, and when the measured data is at least one older than a predetermined period. A model generation method characterized by constructing a correlation between the random variables based on conditional dependence . A program for an information processing apparatus, wherein the program sets a parameter of a function for calculating a prior probability distribution according to an accumulation state of measured data in the information processing apparatus ;
Based on the prior probability distribution calculated using the set parameters, the conditional dependency between random variables is determined, and based on the correlation between the random variables constructed based on the determined conditional dependency a model generation process of generating a model of the Bayesian network to execute the <br/>,
The model generation process is performed when the accumulated amount of the measured data is a small amount for setting the conditional dependence and when the measured data is at least one older than a predetermined period. A program for constructing a correlation between the random variables based on conditional dependence .

Images