JP5802597B2 - Classification device, classification system, classification method, and classification program - Google Patents

Description

  The present invention relates to a classification device, a classification system, a classification method, and a classification program.

  As shown in Non-Patent Document 1, the inventor of the present application has previously proposed a model for estimating the user's interest by hierarchically using latent classes. This model is based on text data such as blogs on the Internet, web pages, twitters, and papers. Which community class the user (blog or web page author) belongs to, which trend class the document belongs to, and its contents The user's interest is estimated by hierarchically estimating to which topic a (word) belongs as a latent variable. This user's interest estimation result should be useful when forecasting marketing demand, presenting the optimal advertisement to the user, or recommending a new item that matches the user's interest. Can do.

Noriaki Kawamae: Latent interest-topic model: finding the causal relationships behind dyadic data.CIKM 2010: 649-658

  However, in the model shown in Non-Patent Document 1, the hierarchical structure is not reflected in the data generation process, and it cannot be classified into different hierarchies (for example, a community hierarchy and a topic hierarchy). There was a problem.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide a classification device, a classification system, a classification method, and a classification program capable of classifying data into different layers.

(1) The present invention has been made in view of the above circumstances, and one aspect of the present invention is an observation variable probability distribution acquisition unit that acquires a probability distribution of an observation variable for each hierarchy, and the observation variable probability distribution acquisition unit includes The classification apparatus includes an observation variable generation unit that switches the acquired probability distribution according to a hierarchy and generates an observation variable based on the switched probability distribution.

(2) According to one aspect of the present invention, in the above classification apparatus, the switch variable probability distribution acquisition unit that acquires a probability distribution of a switch variable for each target data including time information, and the switch variable probability distribution acquisition unit acquire A switch variable generation unit that generates a switch variable for each observation variable based on the probability distribution, and the observation variable generation unit includes a switch variable probability among the probability distributions acquired by the observation variable probability distribution acquisition unit. The distribution acquisition unit switches to a probability distribution of a hierarchy according to the switch variable acquired.

(3) According to one aspect of the present invention, in the above classification apparatus, a time frequency distribution acquisition unit that acquires a time frequency distribution that defines an appearance frequency for each time of data for each trend category that classifies a trend, and the time A trend classifying unit that classifies target data including time information into one of a plurality of trend categories based on the time frequency distribution acquired by the frequency distribution acquiring unit.

(4) One aspect of the present invention is characterized in that, in the classification device, the time frequency distribution is a probability distribution.

(5) According to one aspect of the present invention, the observation variable probability distribution acquisition unit that acquires the probability distribution of the observation variable and the probability distribution acquired by the observation variable probability distribution acquisition unit are switched and switched according to the hierarchy. And an observation variable generation unit that generates an observation variable based on the probability distribution.

(6) According to one aspect of the present invention, the observation variable probability distribution acquisition unit obtains the probability distribution of the observation variable for each hierarchy, and the observation variable generation unit acquires the probability acquired by the observation variable probability distribution acquisition unit. And a procedure for generating an observation variable based on the switched probability distribution and switching the distribution according to the hierarchy.

(7) According to one aspect of the present invention, an observation variable probability distribution acquisition step for acquiring a probability distribution of an observation variable for each layer and a probability distribution acquired by the observation variable probability distribution acquisition step according to the layer And an observation variable generation step for generating an observation variable based on the switched probability distribution.

  According to the present invention, a trend can be extracted from time series data.

It is a graphical model used to explain a preference model using a hierarchical structure of latent classes. It is a graphical model used for description of the preference model concerning the embodiment of the present invention. It is a figure which shows the correspondence of the name of each block, and the name of a function part, and the correspondence of the specific example of the probability distribution of this embodiment, and the name of probability distribution. 1 is a schematic block diagram of a classification system according to an embodiment of the present invention. It is a functional block diagram used for description of the classification system which concerns on embodiment of this invention. It is a list of parameters corresponding to random variables and random variables in the classification system according to the embodiment of the present invention. It is a flowchart used for description of the calculation process in the classification | category system which concerns on embodiment of this invention. It is explanatory drawing of the pre-processing in the pre-processing part which concerns on embodiment of this invention. It is explanatory drawing of the processing content of the calculation process in the classification | category system which concerns on embodiment of this invention. It is a related figure of the process and graphical model in the calculation process part which concerns on embodiment of this invention. It is the figure which showed the probability distribution utilized with the proposal model which concerns on embodiment of this invention. It is explanatory drawing of the algorithm which estimates the latent variable in the classification system which concerns on embodiment of this invention. The upper table shows the movie titles compiled for each period, and the lower table shows the movie titles corresponding to the trends estimated by the proposed model. It is the table which compared the precision of the recommendation of the method of this embodiment, and the conventional method.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. First, in order to understand the present invention, the preference model previously proposed by the present inventor will be described prior to the description of the embodiment of the present invention.

As shown in FIG. 1, the present inventor has proposed a model for estimating user interest by using hierarchical latent classes. In this model, community classes, trend classes, and latent classes of topics are modeled in a hierarchical structure. Here, the trend is
For example, trend, trend, current or trend. The trend is, for example, a tendency of a frequency that a certain word appears within a certain period.

  In FIG. 1, a block 11 is a hyper parameter α block for obtaining a community class multinomial distribution ψ. As the hyper parameter α, a random value is used as an initial value. The block 12 is a block for obtaining a community class multinomial distribution ψ.

  A block 13 is a hyper parameter β block for obtaining a trend class multinomial distribution Ψ. As the hyper parameter β, a random value is used as an initial value. The block 14 is a block for obtaining a multinomial distribution Ψ of S types of trend classes for each community class. Here, S indicates the number of community classes.

  The block 15 is a hyper parameter γ block for obtaining the multinomial distribution θ of topics. As the hyper parameter γ, a random value is used as an initial value. The block 16 is a block for obtaining a multinomial distribution θ of C types of topics for each trend class. Here, C indicates the number of trend classes.

  The block 17 is a hyper parameter δ block for obtaining the multinomial distribution φ of the words in the token. As the hyper parameter δ, a random value is used as an initial value. The token is a delimiter (for example, a word) in a blog or web page document written on the Internet, for example. A block 18 is a block for obtaining a multinomial distribution φ of words in Z types of tokens for each topic. Here, Z indicates the number of topics.

The block 19 is a block for obtaining the latent variable s of the community class from the user ID. The probability distribution that user a is community class s is shown as latent variable s _a . Here, the user a is an author who wrote a blog or a web page on the Internet, for example, and is identified by a user ID. A is the number of users (authors). The block 19 obtains _a latent variable s _{a in} which the user a is the community class s from the user ID using the community class multinomial distribution ψ from the block 12.

A block 20 is a block for obtaining the latent variable c of the trend class from the document ID. Here, the document d is a blog or web page document written on the Internet, for example, and is identified by the document ID. Probability distribution document d belongs to the trend class c is shown as latent variables c _d. D _a is the number of documents written by the user (author) a. Block 20, the latent variable s _a user a from block 19 is community class s, is associated multinomial distribution Ψ trend classes for each community class block 14, then block 20, the corresponding community trends the multinomial distribution Ψ classes, a probability distribution document d belongs to the trend class c, as latent variables c _d, estimates.

A block 21 is a block for obtaining a latent variable z of a topic in each token. This is a block for obtaining a latent variable z of a topic from a token. The probability distribution that the i-th token belongs to topic z is shown as a latent variable z _i for the topic. Here, N _d is the number of tokens in the document d. Block 21, a latent variable c _d a document d from block 20 belongs to a trend class c, in correspondence of the multinomial distribution θ topics for each trend class block 16, then block 21, the token corresponding The probability distribution that the i-th token belongs to the topic z is estimated as the latent variable z _i by the multinomial distribution θ of the topic to be performed.

The block 22 is a block for obtaining the observation variable w. The observation variable w is information extracted from data based on a predetermined rule. For example, if the data is a book purchase history, the observation variable w is the title of the book. For example, for a magazine with data, the observation variable w is the title of a word, sentence, paragraph or chapter. For example, if the data is a movie, the observation variable w is the title of the movie.
In the present embodiment, as an example, the observation variable w is a probability distribution of tokens, and the appearance probability distribution of the i-th token w is shown as an observation variable w _i . Block 22 associates the multinomial distribution φ of the words in the token for each topic of block 18 with the latent variable z _i where the i th token belongs to topic z, and block 22 represents the occurrence of the i th token w. The probability distribution is estimated as the observation variable w _i .

  As described above, the model shown in FIG. 1 estimates the user's interest by hierarchically using the community class, the trend class, and the latent variables of the topic.

That is, the block 19 estimates the latent variable s _{a in} which the user a belongs to the community class s, and the block 20 uses the latent variable s _a to calculate the document d by the multinomial distribution Ψ of the trend class for each community class. estimating the latent variables c _d belonging to the trend class c, block 21, using the latent variables c _d, a multinomial distribution θ topics for each trend class, i-th token latent variables z _i belonging to a topic z Then, the block 22 estimates the observation variable w _i by the multinomial distribution φ of the words in the token for each topic using the latent variable z _i .

  However, in the model shown in FIG. 1, the date and time (time stamp) when the document was created is not included in the model. For this reason, it is impossible to extract the trend class (the generation probability of each element and the simultaneous appearance probability of time).

  In the model shown in FIG. 1, data necessary for trend class extraction is not distinguished from data that is not. For this reason, a local trend cannot be extracted automatically and uniquely. For example, some documents, such as newspapers, are purchased by everyone at any time. Such general documentation is not suitable for reflecting trends. In addition, there are sentences such as industry papers that are of interest to people in that industry but are of little interest to general people. Such local documents are also not suitable for reflecting trends.

  FIG. 2 is a model according to the first embodiment of the present invention. In FIG. 2, blocks 111, 113, 115, and 117 are hyper parameters α, β, γ, and δ, similarly to blocks 11, 13, 15, and 17 in FIG. 1.

The blocks 112, 114, 116, 118 are similar to the blocks 12, 14, 16, 18 in FIG. 1 in that the community class multinomial distribution ψ, the trend class multinomial distribution ψ, the topic multinomial distribution θ, and the words in the token This is a block for obtaining a multinomial distribution φ. That is, the block 118 functions as an observation variable probability distribution acquisition unit that acquires the probability distribution φ of the observation variable w for each layer. In this embodiment, as an example, there are three hierarchies: a community class, a trend class, and a topic. The blocks 112, 114, and 116 are the same as the blocks 12, 14, and 16 in FIG.
In this embodiment, it is divided into three layers. However, the present invention is not limited to this, and by using observation variables such as location information and income instead of time stamps, a layer classified by region, a layer classified by income, etc. It may be divided into The number of layers may be 2 or less or 4 or more.

  The block 118 for obtaining the multinomial distribution φ of the words in the token is the multinomial distribution φ of the words in the Z types of tokens for each topic in the block 18 in FIG. The multinomial distribution φ of the words is (Z + C + S + 1) types of multinomial distributions for each topic, for each trend class, for each community class, and for the whole. This is because there are Z types of topics, C types of trend classes, and S types of community classes.

The block 119 determines the latent variable s of the community class in the same manner as the block 19 in FIG. The block 120 obtains the latent variable c of the trend class in the same manner as the block 20 in FIG. That is, the block 120 functions as a trend classifying unit that classifies target data (eg, text) including time information into one of a plurality of trend categories based on the time frequency distribution acquired by the block 125.
The block 121 determines the latent variable z of the topic in the same manner as the block 21 in FIG. Block 122 determines an observation variable w. In block 22 in FIG. 1, the observation variable w is estimated using the multinomial distribution φ of words in Z types of tokens for each topic. On the other hand, in this embodiment, the block 122 performs the estimation by switching to any one of the topic, trend class, community class, and the whole hierarchy according to the latent variable r of the switch.
That is, the block 122 functions as an observation variable generation unit that switches the probability distribution acquired by the block 118 according to the hierarchy and generates an observation variable based on the switched probability distribution. More specifically, the block 122 functions as an observation variable generation unit that switches the probability distribution acquired by the block 118 to the probability distribution of the hierarchy corresponding to the switch variable acquired by the block 129.

The block 125 obtains a beta distribution λ indicating the appearance frequency of the target data (for example, text) at each time for each trend class. Here, C indicates the number of trend classes.
That is, the block 125 functions as a time frequency distribution acquisition unit that acquires a time frequency distribution in which the appearance frequency for each time of the target data is determined for each trend category that classifies the trend. The beta distribution λ indicating the appearance frequency of the target data for each time is normalized from 0 to 1 (the oldest time is 0 and the current time is 1). In this embodiment, a beta distribution is used for the time. This is because the time is handled continuously. When the time is handled discretely, such as hourly, daily, weekly, monthly, for trend classes that change slowly, it can be treated as data with a long period like monthly, but changes frequently. The trend class must be treated as data with a short cycle, such as hourly or daily, and the data amount increases. By handling time continuously, it is possible to process only a trend class that changes slowly or a trend class that changes frequently only by the difference in the shape of the distribution curve.

A block 126 is a block for obtaining the observed variable t of the time stamp from the time stamp of the document by using the beta distribution λ of the block 125. The time stamp indicates the date and time when the document d was generated. Block 126, the latent variable c _d trend class determined in block 120, is associated with the beta distribution lambda of each trend classes from block 125, and, from the time stamp using the corresponding beta distribution lambda, document d is to estimate the probability distribution is a time stamp t as the observation variable t _d.

A block 127 is a hyper parameter ε block for obtaining the multinomial distribution μ. As the hyper parameter ε, a random value is used as an initial value. Block 128 is seeking _{D a} type of multinomial mu. Here, D _a is the number of documents written by the user (author) a. That is, the block 128 functions as a switch variable probability distribution acquisition unit that acquires a probability distribution of switch variables for each target data (for example, text).

A block 129 is a block for obtaining a latent variable r of the switch. Block 129 determines the switch variable r _i of the i th token from the multinomial distribution μ _d corresponding to document d. The multinomial distribution used for estimation of the observation variable w in block 122 is switched by this switch variable. The block 129 functions as a switch variable generation unit that generates a switch variable for each observation variable based on the probability distribution acquired by the block 128.

  In the above model, the time of the observation variable is not included in the model. For this reason, the appearance probability of the sentence d for each time cannot be extracted for each trend class. On the other hand, in the model according to this embodiment shown in FIG. 2, time t is introduced as one of the observation variables. When the latent variable c of the trend class changes in block 120, the block 126 switches the temporal distribution of the appearance probability of the sentence d provided from the block 125, so the appearance probability of the sentence d at each time t changes. Thereby, the appearance probability of the sentence d for every time can be extracted for every trend class.

  In the above-described model, the observation variable w is obtained using the multinomial distribution φ for each topic. On the other hand, in the present embodiment, the multinomial distribution for obtaining the observation variable w is converted into a multinomial distribution for each topic, a multinomial distribution for each trend class, a multinomial distribution for each community class, And switching to the multinomial distribution. Since the number of topics is Z, the number of trend classes is C, the number of community classes is S, and the number handled as a whole is “1”, the number of multinomial distributions φ in block 118 is (Z + C + S + 1). )

  If the switch variable r is (r = 0), the entire multinomial distribution is selected, and the observation variable w is generated from the simultaneous multinomial distribution. The entire multinomial distribution is a general distribution regardless of its contents and time.

  If the switch variable r is (r = 1), a multinomial distribution for each topic is selected. The multinomial distribution of topics is sustained over a long period.

  If the switch variable r is (r = 2), a multinomial distribution for each trend class is selected. The trend class multinomial distribution has a short duration in which the trend changes with time.

  If the switch variable r is (r = 3), a multinomial distribution for each community class is selected. The multinomial distribution of community classes is a local distribution that is specialized for the community class.

  Thus, in the model according to the present embodiment, by introducing the switch variable r, it is possible to distinguish between those that change with time and those that do not. Thereby, not only the simultaneous probability of the combination of time and the element but also the probability of only the element can be expressed.

FIG. 3 is a diagram illustrating a correspondence relationship between the names of the blocks and the names of the functional units, and a correspondence relationship between a specific example of the probability distribution of the present embodiment and the names of the probability distributions. The table T31 is a table showing the correspondence between the name of each block and the name of the functional unit. The table T32 is a table showing a correspondence relationship between a specific example of the probability distribution of the present embodiment and the name of the probability distribution.
Next, the process for estimating the user's interest by executing the process shown by the model as shown in FIG. 2 will be specifically described.

  FIG. 4 is a block diagram showing a configuration of the classification system according to the embodiment of the present invention, and FIG. 5 shows a functional block diagram of each part. As shown in FIG. 4, the system according to the embodiment of the present invention includes a file server 501, a calculation server (classification device) 502, a database 503, and a service server 504. Hereinafter, a word will be described as an example of a token.

  As shown in FIG. 5, the file server 501 includes a data file storage unit 511 that stores document data serving as a processing unit. As the document data to be stored in the data file storage unit 511, blog documents on the Internet, document data of web pages, tweeters, papers, and the like are used. Note that document data as a processing unit is not limited to documents on the Internet. Further, the data file storage unit 511 includes each document data, a document ID for identifying a document as a processing unit, an author ID indicating the author of the document, and a time stamp indicating the date and time when the document was generated. Are stored in association with each other.

  The calculation server 502 takes out the document data from the file server 501, performs a calculation process as shown by the model shown in FIG. 2, and outputs a calculation result. The calculation server 502 includes a pre-processing unit 521 and a calculation processing unit 522 as illustrated in FIG.

  For example, the pre-processing unit 521 receives a document data file to be subjected to calculation processing from the file server 501, extracts a document ID, an author ID, and a time stamp for each document data from the document data file, Extract words that are elements of the document. Then, the preprocessing unit 521 assigns a processing document ID to the document, assigns a processing author ID to the author, and assigns a processing word ID to each extracted word.

  The calculation processing unit 522 receives the data processed by the preprocessing unit 521 and performs calculation processing corresponding to the model shown in FIG. As will be described later, in this embodiment, Gibbs sampling is used for estimating a latent variable.

  As shown in FIG. 5, the database 503 has a calculation result storage unit 531. The calculation result of the calculation processing unit 522 of the calculation server 502 is sent to the database 503 and stored in the calculation result storage unit 531. As shown in FIG. 6, the calculation results include community, trend class, time stamp, switch, topic, word random variables, parameters of each probability distribution, and their types. In the calculation result of FIG. 6, the time stamp and the word are observation variables, and the others are latent variables.

  The service server 504 is a server for providing a calculation result for using the service. As illustrated in FIG. 5, the service server 504 includes a calling unit 541.

  In response to a call from the user terminal 505, the calling unit 541 sends a calculation result to the user terminal 505. This calculation result can be used for various services such as marketing, demand prediction, advertisement, and recommendation.

  FIG. 7 is a flowchart showing processing in the calculation server 502. In FIG. 7, first, the pre-processing unit 521 of the calculation server 502 performs processing for assigning a processing document ID, a processing author ID, and a processing word ID for the calculation target document data.

  That is, as shown in FIG. 8A, each document is given a unique author ID or document ID. In FIG. 8A, the document data of the first record has “A” as the author ID and “001” as the document ID. Tokens (here, words) that are elements of the document include “Rome”, “History”,... The document data of the last record has “Z” as the author ID and “087” as the document ID. There are "ancient", "art", ..., "culture" as tokens that are elements of the document.

  FIG. 8B shows an example of processing for assigning a processing author ID, a processing document ID, and a processing word ID to such data in step S1. As shown in FIG. 8B, the document data of the first record is assigned “0” as the processing author ID and “0” as the processing document ID. A processing word ID “22”, a processing word ID “0”,..., And a processing word ID “1212” are assigned to tokens 1 to N. The document data of the last record is assigned “100” as the processing author ID and “223” as the processing document ID. Then, the processing word ID “4”, the processing word ID “1”,..., And the processing word ID “557” are allocated to the tokens 1 to N.

  Next, the calculation processing unit 522 of the calculation server 502 sets the number of random variables (C, S, Z) and initial values of the hyper parameters (α, β, γ, δ, ε), and performs calculation processing. The number of repetitions is set (step S2). Then, the calculation processing unit 522 of the calculation server 502 generates a random number and gives the value to the random variable (C, S, Z) (step S3).

  That is, FIG. 9A shows a state in which the processing document ID, the processing author ID, and the processing word ID are assigned to the calculation target document data in the processing of step S1. For such data, random numbers are inserted into the community class S, trend class C, and topic Z as shown in FIG. Here, for example, “0” to “20” are arbitrarily inserted as community class random numbers, for example, “0” to “40” are arbitrarily inserted as trend class random numbers, and topic random numbers, for example, “ 0 "to" 99 "are arbitrarily inserted. In the example of FIG. 9B, the random number “7” is inserted as the community class S, the random number “20” is inserted as the trend class C, and the tokens “1” to “N” are inserted into the data of the first record. Random numbers “7”, “5”,... “8” are inserted as topics. In the data of the last record, the random number “12” is inserted as the community class S, the random number “11” is inserted as the trend class C, and the random numbers “8”, “ 8 ”,...,“ 3 ”are inserted.

  Next, the calculation processing unit 522 of the calculation server 502 estimates a latent variable by Gibbs sampling (step S4). When the number of Gibbs sampling iterations reaches a predetermined number of computation iterations, the computation process ends (step S5).

  Thus, in this embodiment, the latent variable is estimated by Gibbs sampling. FIG. 10 shows a correspondence between values estimated by Gibbs sampling and latent variables obtained thereby.

  Next, in this embodiment, it will be described that each latent variable is obtained by Gibbs sampling. In the present embodiment, as shown in FIG. 11, a Dirichlet distribution is introduced in order to estimate each latent variable by statistical processing. The Dirichlet distribution is a continuous probability distribution, but can be converted to a discrete type by integration.

  By introducing such a probability distribution, the joint probability of the entire data is expressed as the following equation.

Here, the joint probability p (|) of the entire data indicates a conditional probability.
When the above equation is integrated, it can be expressed as the following equation.

Here, Γ is a gamma function. n _s is the number of authors belonging to the community s. It is. n _sc is the number of documents in which the trend class c is selected from the community class s. n _cz is the number of times the community class z is selected from the trend class c. n _dr is the number of tokens for which the switch variable r is selected from the community class z. n _zw is the number of times the observation variable w is selected from the community class z.
In this way, by introducing the Dirichlet distribution and integrating it, the parameters of the multinomial distribution disappear, and the hyper parameters and frequency information remain. In this way, by introducing and integrating the Dirichlet distribution, the continuous probability distribution becomes a discrete probability distribution.

  When the equation (1) obtained by integrating the Dirichlet distribution described above is modified, an equation for estimating each latent variable is obtained by Gibbs sampling. That is, the probability that the community class z is g can be derived as in the following equation.

  The probability that the trend class c is j can be derived as follows.

  The probability of the variable r can be derived as follows:

Equation (4) is the probability of variable r = 0. Equation (5) is the probability of variable r = 1. k is a topic identifier. Equation (6) is the probability of variable r = 2. Equation (7) is the probability of variable r = 3.
FIG. 12 shows an algorithm for estimating a latent variable by Gibbs sampling.
The calculation processing unit 522 performs initialization processing according to this algorithm, and sets the number of _iterations of Gibbs sampling N _iteration .
Then, the calculation processing unit 522 repeats the following processing for the number of authors A. Calculation processing unit 522 estimates the latent variable _{s a} community class by equation (2) for each author (number Author A), a updates the variable _{n s.} Calculation processing unit 522, every time the updating of the estimated and variables n _s of the latent variables s _a, the following processing is performed.

The calculation processing unit 522 repeats the following processing for the number of documents D. Calculation processing unit 522 estimates the latent variable _{c d} trend classes for each document (document count D) by the equation (3), and updates the variable _{n c} and variable lambda _c. Calculation processing unit 522, every time the update processing of the latent variables c _d estimated and variables n _c, the following process is performed.

The calculation processing unit 522 repeats the following processing for the number of words determined for each document d. The calculation processing unit 522 estimates the latent variable r _di of the switch and the latent variable z _{di of the} topic by the equations (4), (5), (6), and (7), and the variables n _dr , n _cz , and n _zw. Update.

  The calculation processing unit 522 estimates the parameters ψ, ψ, θ, φ, and μ of the multinomial distribution when it is desired to know each probability distribution after the above iterative processing is completed. In FIG. 12, the symbols (hats) attached to the parameters ψ, ψ, θ, φ, and μ of the multinomial distribution mean estimated values.

  Next, the effects obtained by the present invention will be described. FIG. 13 shows the ranking of popularity of movies by dividing them into the whole (6 years from 2000 to 2005), 2000 to 2001, 2002 to 2003, and 2004 to 2005. . FIG. 13A shows a simple tabulation result. Of the counting results shown in FIG. 13A, a movie with a title as shown by the underline retains a certain popularity throughout the entire period. That is, they do not change much in time.

  FIG. 13B shows the result obtained by this embodiment. Here, the community class number C is 75, the movie trend class number S is 75, and the movie topic number Z is 100. φb is an overall trend class, and is displayed in descending order of probability from the overall probability distribution φ. φc (2000-2001) extracts the trend class having the trend class distribution whose beta distribution peak is closest to t = 0, and displays the probability distribution φ of the extracted trend class in descending order of probability. It is a thing. φc (2002-2003) extracts a trend class having a trend class distribution whose beta distribution peak is closest to t = 0.5, and from the probability distribution φ of the extracted trend class, the probability is high They are displayed in order. φc (2004-2005) extracts the trend class having the trend class distribution whose beta distribution peak is closest to t = 1, and displays the probability distribution φ of the extracted trend class in descending order of probability. It is a thing.

  In simple aggregation, movie titles that maintain a certain level of popularity over the entire period are included in the ranking, as shown by the underlined titles in FIG. On the other hand, as shown in FIG. 13B, in this embodiment, the entire period as indicated by the underline in each period from 2000 to 2001, 2002 to 2003, and 2004 to 2005. Films with titles that have a certain popularity over time are excluded (or fall out of the above ranking), and movies with titles reflecting the trend class for each period are ranked. As described above, in the present embodiment, it is possible to obtain a trend class by distinguishing between those that change with time and those that do not.

  FIG. 14 compares the method according to the present embodiment with the conventional method. In FIG. 14, TOT, DTMs, gPLSA, and LIT are conventional techniques. POT is a method according to the present embodiment, and POT r = {0, 1, 2, 3} is a case where r is selectable by 0, 1, 2, 3. POT r = {1, 2, 3} is a case where r is selectable with 1, 2, 3, and r cannot be selected with 0. POT r = {2, 3} is a case where r is selectable with 2, 3, and r is not selectable between 0 and 1. Thus, when POT r = {2, 3}, the observation variable w is calculated from the trend class probability distribution or the topic probability distribution.

  As described above, if (r = 0), the entire multinomial distribution which is a general distribution is selected. If (r = 1), a multinomial distribution for each topic that is a long and persistent distribution is selected. If (r = 2), a multinomial distribution is selected for each trend class whose tendency changes with time. If (r = 3), a multinomial distribution for each local community class specialized for the community class is selected.

Next, each index in FIG. 14 will be described in detail. Top-10 is a pre-determined period immediately before the user's recommended movie titles ranked in the top 10 from the data in the test period other than the pre-determined period (for example, the latest one month). It is the probability of watching during the period.
Further, UC (User Coverage: coverage rate of recommended users) is the ratio of the number of users who can recommend each recommendation method to the number of users who watched the movie title during the test period. Since the movie title can be recommended to many users as the UC is higher, the system is more valuable for the entire user.

  IC (Item Coverage: coverage of recommended items) is the ratio of the number of titles that can be recommended by each recommendation method to the number of movie titles viewed during the test period. The IC is one index indicating the size of the movie title domain in the system that can be recommended by the system. Therefore, a system with a low IC can present only a limited number of movie titles, and is therefore a low-value system for the user.

  The Gini coefficient is an index indicating a recommended statistical variance when a movie title is recommended to the user. The Gini coefficient takes a value from 0 to 1, and the closer the value is to 0, the smaller the difference in the number of recommended users for each movie title, and the closer the value is to 1, the larger the difference.

  AE (Average Elapsed time) is an average of elapsed time from the release of a movie title to viewing. If this value is small, the novelty of the movie title is high for the user.

  AD (Average Difference time) is the average of the difference between the test period start time and the movie title viewing time. If this value is large, the movie title becomes harder to notice.

As shown in FIG. 14, in the case of POT r = {1, 2, 3}, the value of Top-10 is statistically larger than any conventional method, so the top 10 movies that are better than the conventional ones Can recommend a title.
Further, when POT r = {2, 3}, the IC value is statistically larger than any conventional method, so that the recommended movie title is different for each user than before. Thereby, since the classification system of this embodiment can present a wide range of movie titles, it is highly valuable for the user.

  In addition, when POT r = {2, 3}, the value of Gini is statistically smaller than any conventional method, so that there is less disparity in the number of recommended users for each movie title than before. Thereby, since the classification system of this embodiment can recommend a movie title widely, its value is high for a user.

  In addition, when POT r = {2, 3}, since the value of AE is statistically smaller than any conventional method, the elapsed time from the release to viewing of the movie title is shorter than before. Thereby, since the classification system of this embodiment can recommend a new movie title than before, it is highly valuable for the user.

  As described above, the calculation server 502 in the present embodiment introduces the time t as an observation variable, and extracts the appearance probability of the sentence d for each trend class and for each time. As a result, the calculation server 502 can simultaneously extract the trend period and tokens (for example, words) constituting each trend from the time-series data. As a result, for example, the calculation server 502 can simultaneously extract a community (a set of users with similar preferences).

  Further, the calculation server 502 according to the present embodiment uses a switch variable r to calculate a multinomial distribution for obtaining an observation variable w, a multinomial distribution for each topic, a multinomial distribution for each trend class, a multinomial distribution for each community class, Switch between overall multinomial distributions. Thereby, what changes with time can be separated from what does not, and can be expressed not only by the simultaneous probability of the combination of time and element but also by the probability of only the element.

  In the conventional analysis, since the change with respect to the time series was not continuously taken into consideration, it was impossible to predict the change with the passage of time. In contrast, in this embodiment, the time t of the observation variable is introduced into the model. As a result, it is possible to perform a prediction that incorporates an element that continuously fluctuates with respect to the time series, so that a trend can be extracted.

In the conventional analysis, the hierarchical structure is not reflected in the data generation process, and the data cannot be classified by hierarchy (for example, community or topic). As a result, since the classification of the data to be analyzed is not hierarchical, analysis based on the data attributes (related to the size of the corresponding community, etc.) could not be performed.
On the other hand, in the present embodiment, a probability distribution for generating the observation variable w is generated for each hierarchy (for example, community or topic), and the probability distribution to be used is switched according to the value of the switch variable r. Thereby, a constituent factor can be classified for every hierarchy (for example, community, trend, or topic). As a result, for example, when an author ID, a document ID, and a word ID are assigned to the analysis data, it is possible to analyze whether the analysis data is specific or universal for the specific author ID, for example.

  Further, in the present embodiment, the structure in which the time t of the observation variable is introduced into the model, the probability distribution for generating the observation variable w for each hierarchy is generated, and the probability distribution used according to the value of the switch variable r is switched. For example, a trend for each community or topic) can be uniquely extracted without human intervention. Here, unique means that the same result is always obtained no matter who does it.

  Furthermore, in this embodiment, in order to handle the time t as a continuous value, a trend probability distribution (for example, a beta distribution) is introduced. With this configuration, by treating time with continuous values, it is possible to simultaneously extract trends having different periods and elements constituting each trend. Here, the period is a time scale of time series change. At the same time, it is easy to compare trends by expressing the trend distribution as a probability distribution. Moreover, since it is not necessary to discretize the frequency for every certain time width, it is not necessary to adjust the time width for each data, and the time and effort for adjusting the time width can be reduced.

  Further, in the present embodiment, the time t of the observation variable is introduced into the model, the probability distribution for generating the observation variable w for each hierarchy is generated, the probability distribution used according to the value of the switch variable r is switched, and the time t is set. In order to deal with continuous values, a probability distribution of trends was introduced. With this configuration, a plurality of time-series data having different periods can be classified hierarchically for each period of the time-series change.

The probability distribution at time t is not limited to the beta distribution λ, but may be a probability distribution or a gamma distribution.
In addition, a system including a plurality of devices may process each processing of the calculation server 502 of this embodiment in a distributed manner by the plurality of devices.
Further, by recording a program for executing each process of the calculation server 502 of the present embodiment on a computer-readable recording medium, causing the computer system to read and execute the program recorded on the recording medium, The various processes described above related to the calculation server 502 may be performed.

  Here, the “computer system” may include an OS and hardware such as peripheral devices. Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. The “computer-readable recording medium” means a flexible disk, a magneto-optical disk, a ROM, a writable nonvolatile memory such as a flash memory, a portable medium such as a CD-ROM, a hard disk built in a computer system, etc. This is a storage device.

  Further, the “computer-readable recording medium” refers to a volatile memory (for example, DRAM (Dynamic) in a computer system serving as a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. Random Access Memory)) that holds a program for a certain period of time is also included. The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, what is called a difference file (difference program) may be sufficient.

  As mentioned above, although embodiment of this invention was explained in full detail with reference to drawings, the concrete structure is not restricted to this embodiment, The design etc. of the range which does not deviate from the summary of this invention are included.

501 File server 502 Calculation server (classification device)
503 Database 504 Service server 505 User terminal 511 Data file storage unit 521 Pre-processing unit 522 Calculation processing unit 531 Calculation result storage unit 541 Calling unit

Claims (7)

Observation variable probability distribution acquisition unit for acquiring probability distribution of observation variables for each hierarchy,
Switching the probability distribution acquired by the observation variable probability distribution acquisition unit according to the hierarchy, and generating an observation variable based on the switched probability distribution; and
A classification apparatus comprising: A switch variable probability distribution acquisition unit that acquires a probability distribution of switch variables for each target data including time information,
Based on the probability distribution acquired by the switch variable probability distribution acquisition unit, a switch variable generation unit that generates a switch variable for each observation variable;
With
The observation variable generation unit switches the probability distribution acquired by the observation variable probability distribution acquisition unit to a probability distribution of a hierarchy corresponding to the switch variable acquired by the switch variable probability distribution acquisition unit. Classification device according to. A time frequency distribution acquisition unit that acquires a time frequency distribution that defines an appearance frequency for each time of data for each trend category that classifies a trend;
Based on the time frequency distribution acquired by the time frequency distribution acquisition unit, a trend classification unit that classifies target data including time information into one of a plurality of trend categories,
Classification apparatus according to 請 Motomeko 2 comprising: a.   The classification device according to claim 3, wherein the time frequency distribution is a probability distribution. Observation variable probability distribution acquisition unit for acquiring probability distribution of observation variables for each hierarchy,
Switching the probability distribution acquired by the observation variable probability distribution acquisition unit according to the hierarchy, and generating an observation variable based on the switched probability distribution; and
A classification system comprising: The observation variable probability distribution acquisition unit acquires the probability distribution of observation variables for each hierarchy, and
The observation variable generation unit switches the probability distribution acquired by the observation variable probability distribution acquisition unit according to the hierarchy, and generates an observation variable based on the switched probability distribution;
A classification method characterized by comprising: On the computer,
Observation variable probability distribution acquisition step for acquiring probability distribution of observation variables for each hierarchy,
Switching the probability distribution acquired by the observation variable probability distribution acquisition step according to the hierarchy, and generating an observation variable based on the switched probability distribution; and
Classification program for running

Images