JP2014160358A - Time series data component decomposition apparatus, method and program and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To highly accurately perform component decomposition of time series data without using a keyword.SOLUTION: Each observed value included in time series data is totalized for each predetermined time section (S106), a plurality of clusters are detected by clustering on the basis of contents of each document without using an external reference for a document set with time information and the degree which each document belongs to each cluster is calculated (S102), the degree which each document of the document set with time information belongs to each cluster for each time section is totalized (S104), a regression equation which shows relation between an explanatory variable and an objective variable is estimated by regression analysis while a document totalization result is defined as the explanatory variable and a time series data totalization result is defined as the objective variable (S108), a component ratio of each cluster is calculated for each time section on the basis of an estimation result for each time section obtained by the regression analysis and the document totalization result (S110), and each observed value included in the time series data is decomposed into a component of each cluster (S112).

Description

  The present invention relates to a time-series data component decomposition apparatus, method, program, and recording medium for performing component decomposition of time-series data.

  Microblogs, such as Twitter (registered trademark), where users mainly post their own situations and miscellaneous feelings in short sentences are widely used due to their ease of updating and their real-time nature. It continues to develop as an information infrastructure. Under these circumstances, in recent years, technology that uses microblogging as a “sensor” that captures events that occur in the real world has been developed.

  For example, the following non-patent document 1 discloses a technique for predicting the degree of influenza epidemic from the amount of tweets including influenza-related keywords. With such a technique, it is possible to know a clue as to what kind of background the time-series data observed in the real world is generated.

Aron Culotta: Towards detecting influenza epidemics by analyzing Twitter messages. SOMA Workshop 2010.

  Note that in the measurement of the number of tweets by keyword designation used by the conventional method described in Non-Patent Document 1 described above, it is necessary to comprehensively prepare keywords related to generation of time series data. However, it is difficult to comprehensively prepare keywords, and if the keywords are not prepared sufficiently, there is a problem that as a result, correct understanding about the generation process of time series data is insufficient.

  For example, when considering the number of users at an airport, considering the breakdown of the number of users, that is, the reason for the reason for using the airport, , “Depart for a honeymoon”, “play”, “dine”, “shop”, and so on. It is very difficult to prepare, and the accuracy of component decomposition of time-series data is reduced.

  In addition, the conventional keyword-matching method also has a problem that information on related documents is lost in time-series data although keywords are not included.

  The present invention has been made to solve the above-described problem, and provides a time-series data component decomposition apparatus, method, program, and recording medium capable of accurately performing time-series data component decomposition without using a keyword. The purpose is to do.

  In order to achieve the above object, the time series data component decomposition apparatus according to the present invention includes each time series data included in a series of observation values obtained by observing a temporal change of a certain phenomenon. Without using an external standard for the time-series data aggregation unit that aggregates observation values for each predetermined time interval and the document set with time information, which is a set of documents with creation time information Clustering is performed based on the contents of each document to find a plurality of clusters, and each document is calculated by the document clustering unit that calculates the degree of belonging to each of the plurality of clusters. Based on the degree of belonging to each cluster and the creation time information given to each document, each document in the document set with time information belongs to each cluster for each time interval. A document totaling unit that counts the degree of data, and for each time interval, the totaling result of the document totaling unit is used as an explanatory variable, and the totaling result of the time-series data totaling unit is used as an objective variable. A regression analysis unit that estimates a regression equation representing a relationship between the regression equation, a constant term and a regression coefficient of the regression equation estimated for each time interval by the regression analysis unit, and the time interval by the document aggregation unit The component ratio calculation unit that calculates the ratio of the components of each cluster for each time interval using the total result for each time interval, and each observation value included in the time series data corresponds to the observation time of the observation value And a component decomposing unit that decomposes into the components of each cluster using the ratio of the cluster components calculated by the component ratio calculating unit of the time interval.

  As described above, according to the time-series data component decomposition apparatus according to the present invention, the time-series data is decomposed according to the component ratio of the cluster based on the content of the document instead of the keyword unit. Decomposition can be performed with high accuracy.

  In addition, for each document in the document set with time information, as a value indicating what kind of emotion is described based on the emotion, further includes an emotion estimation unit that estimates the degree of each emotion included in each document, Based on the degree of belonging to each cluster calculated by the document clustering unit, the creation time information given to each document, and the estimation result of the emotion estimation unit, the document aggregation unit, for each time interval, The degree to which each document of the document set with time information belongs to each set of the cluster and the emotion is totaled, and the component ratio calculation unit is a regression equation constant estimated by the regression analysis unit for each time interval. Using a term and a regression coefficient, and a totaling result for each time interval by the document totaling unit, a ratio of components of the cluster and the emotion set for each time interval is calculated, Each observed value included in the time-series data is obtained by using the ratio of the components of the cluster calculated by the component ratio calculating unit of the time interval corresponding to the observation time of the observed value. You may make it decompose | disassemble into the component of each group.

  As described above, since the time-series data is decomposed in accordance with the component ratio of the cluster and emotion group based on the contents of the document, the time-series data can be decomposed with higher accuracy.

  A time series data component decomposition method according to the present invention includes a time series data aggregation unit, a document clustering unit, a document aggregation unit, a regression analysis unit, a component ratio calculation unit, a component ratio calculation unit, and a component decomposition unit. A time-series data component decomposing method in a decomposing apparatus, wherein each time-series data is a series of observation values obtained by observing a temporal change of a phenomenon by the time-series data totaling unit. The observed values are aggregated for each predetermined time interval, and the document clustering unit does not use an external standard for the document set with time information that is a set of documents to which creation time information is given. Clustering is performed based on the contents of each document to find a plurality of clusters, and the degree to which each document belongs to each of the plurality of clusters is calculated. Based on the degree of belonging to each cluster calculated by the document clustering unit and the creation time information given to each document, each document in the document set with time information is The degree of belonging to each cluster is tabulated, and the regression analysis unit uses the tabulation result of the document tabulation unit as an explanatory variable and the tabulation result of the time series data tabulation unit as an objective variable for each time interval. A regression equation representing a relationship between the variable and the objective variable is estimated by regression analysis, and the component ratio calculation unit calculates a constant term and a regression coefficient of the regression equation estimated by the regression analysis unit for each time interval. , By using the aggregation result for each time interval by the document aggregation unit, to calculate the ratio of the components of each cluster for each time interval, by the component decomposition unit, Each observed value included in the time-series data is converted into a component of each cluster by using the ratio of the cluster components calculated by the component ratio calculating unit in the time interval corresponding to the observation time of the observed value. Decompose.

  The document set with time information may be a set of documents in which a specific keyword is included in each document included in the document set with time information.

  Further, the document set with time information may be a set of documents created and posted by a terminal existing in a position range designated in advance.

  The document set with time information may be a set of microblog documents in which a user creates and submits text within a predetermined number of characters representing a real-time situation or miscellaneous feeling of the user.

  The program according to the present invention is a program for causing a computer to function as each unit of the time-series data component decomposition apparatus.

  The recording medium according to the present invention is a computer-readable recording medium recording a program for causing a computer to function as each means of the time-series data component decomposition apparatus.

  As described above, according to the present invention, it is possible to obtain an effect that the component decomposition of time-series data can be performed with high accuracy without using a keyword.

It is a figure which shows the functional structure of the time series data component decomposition | disassembly apparatus which concerns on 1st Embodiment. It is a schematic block diagram of the computer which functions as a time series data component decomposition | disassembly apparatus. It is a flowchart which shows the content of the processing routine which the time series data component decomposition | disassembly apparatus based on 1st Embodiment performs. It is the figure which showed typically the processing content of the time series data component decomposition | disassembly apparatus based on 1st Embodiment. It is a figure which shows the functional structure of the time series data component decomposition | disassembly apparatus which concerns on 2nd Embodiment. It is a flowchart which shows the content of the processing routine which the time series data component decomposition | disassembly apparatus based on 2nd Embodiment performs.

<First Embodiment>

  FIG. 1 is a diagram showing a functional configuration of a time-series data component decomposition apparatus 10 according to an embodiment of the present invention.

  The time-series data component decomposition apparatus 10 shown in FIG. 1 is an apparatus that decomposes the time-series data into a plurality of components from the input time-series data and the document set with time information, and outputs them. , A calculation unit 22 and an output unit 24 are provided.

  The input unit 20 includes a time-series data input unit 30 and a document set input unit 32 with time information.

  The time series data input unit 30 receives input of time series data.

  Here, the time series data is a series of a series of values (observed values) obtained by observing a temporal change of a certain phenomenon. Specific examples of time-series data include stock price data and sales data. Each observation value included in the time-series data is associated with observation date / time information indicating the observation date / time when each observation value was observed. The time series data can be automatically collected from the Internet or the like, for example.

  The document set input unit 32 with time information receives input of each document in the document set with time information.

  The document set with time information is a set of document data (hereinafter simply referred to as documents), and each document has creation time information (hereinafter simply referred to as time information) indicating when each document is created. It is given as data. The document set with time information may be a set of documents in which specific keywords are included in each document. Further, the document set with time information may be a set of documents created and posted by a terminal existing in a position range designated in advance. Furthermore, the document set with time information may be a set of microblog documents in which the user creates and submits text within a predetermined number of characters representing the user's own real-time situation or feeling.

  The time information given to each document may be, for example, information indicating date and time, information indicating day of the week, or information indicating a combination of day of the week and time. Information indicating a date may be used. The document set with time information can be collected and constructed in advance from the Internet or the like, for example.

  The calculation unit 22 includes a word division unit 40, a document clustering unit 42, a document totaling unit 44, a time series data totaling unit 46, a regression analysis unit 48, a component ratio calculation unit 50, and a component decomposition unit 52.

  The word dividing unit 40 generates a word set by dividing each document included in the document set with time information into words.

  First, the document clustering unit 42 converts the document information included in the time information-added document set from the time information-added document set generated by the word dividing unit 40 without using an external reference. Based on this, a cluster set having a plurality of semantic clusters is found. Then, the document clustering unit 42 calculates and outputs the degree to which each document included in the document set with time information belongs to each cluster. Note that clustering performed without using an external criterion can use a method generally known as an unsupervised data classification method.

  The document totaling unit 44 totals the degree of belonging to each cluster of each document in the document set with time information for each time section. Here, the time section refers to each divided section when the target period is divided into a plurality of periods. To give a specific example, for example, each time zone when a day is equally divided into 24 periods (0 hour range (00:00:00 to 00:59:59), 1 hour range, ... ... the time zone can be defined as 23:00). Also, the date (365 sections), day of the week (7 sections), combination of day of the week and time zone (Monday 7 o'clock, Tuesday 10 o'clock, 7 × 24 = 168 sections) may be used. The totaling result of the document totaling unit 44 is referred to as a document totaling result.

  The time series data totaling unit 46 totals time series data for each time interval. The definition of the time interval when the time series data totaling unit 46 totals the time series data is the same as the definition of the time interval used in the totalization in the document totaling unit 44. The aggregation result of the time-series data aggregation unit 46 is referred to as a time-series data aggregation result.

  The regression analysis unit 48 performs a regression analysis for each time interval using the document aggregation result by the document aggregation unit 44 as an explanatory variable and the time series data aggregation result by the time series data aggregation unit 46 as an objective variable.

  The component ratio calculation unit 50 calculates the component ratio of each cluster for each time interval using the regression analysis result of the regression analysis unit 48.

  The component decomposing unit 52 decomposes the time series data into a plurality of components using the component ratio for each time interval calculated by the component ratio calculating unit 50.

  The output unit 24 outputs the decomposition result (component decomposition value) of the component decomposition unit 52.

  As shown in FIG. 2, the time-series data component decomposition apparatus 10 includes a CPU (Central Processing Unit) 201, a ROM (Read Only Memory) 202 that stores a program for the CPU to execute each processing routine described later, and the like. The computer 200 includes a RAM (Random Access Memory) 203. The computer 200 also includes a communication interface (IF) 204, an input / output IF 205, and a hard disk drive 206. The communication IF 204 is an interface for connecting to the network 210. The input / output IF 205 is connected to the display 208 and the keyboard 209.

  When the CPU 201 reads and executes a program stored in a recording medium such as the ROM 202 or a hard disk, the above-described functions are realized by cooperation of the hardware and the program.

  Next, the operation of the time-series data component decomposition apparatus 10 in the present embodiment will be described in detail. When the time-series data input unit 30 and the time-information-added document set input unit 32 accept input of the time-series data and the time-information-added document set, the time-series data component decomposition apparatus 10 executes the processing routine shown in FIG.

In step 100, the word dividing unit 40 divides each document d _i included in the document set with time information accepted by the document set input unit 32 with time information into word units to generate a word set. At this time, only nouns may be extracted by using a morphological analyzer to obtain word sets, or only noun / verb / adjective words may be extracted to form word sets. Moreover, you may add the word of another part of speech to the said word set. Here, an example has been described for implementing the morphological analysis by the morphological analyzer, instead of carrying out the morphological analysis, continuous all characters n g (contained in each document d _i of the document set with time information n Character) may be the word set.

In step 102, the document clustering unit 42 selects a cluster having a plurality of semantic groups based on the contents of the document from the document set with time information in which the word set of each document is generated without using an external criterion. A set is found, and the degree P (c | d _i ) belonging to each cluster is output for each document d _i included in the document set with time information. C is a parameter indicating each cluster, and takes a value from 1 to the number C of clusters.

  Here, the relationship shown in the following formula (1) is established.

  Here, an algorithm (Fuzzy c-means, for example) called soft clustering (see Reference 1 “Hiroyuki Shinno, Cluster Analysis Learned by R, Ohmsha, 2007”) that can output the degree to which each data belongs to each cluster. Any mixture distribution model, pLSI, LDA, NMF, etc. can be used. The number C of clusters is a parameter given in advance, such as C = 20.

In step 104, the document counting unit 44, a value indicating the degree of belonging to the cluster of each document d _i of the time information attached document set (hereinafter, a value indicating the degree of convenience referred to as document number) Time Aggregate by interval h. Specifically, when the time when the document d _i is posted is t _i , the number of documents n (c, h) belonging to the cluster c in the time interval h is defined as shown in the following formula (2) Then, all combinations of <c, h> are calculated.

Here, D is the number of documents included in the document set with time information, and δ (t _i , h) is a function that returns 1 when the time t _i is included in the time interval h, and returns 0 otherwise. It is.

  The document totaling unit 44 outputs the calculated document number n (c, h) as a document totaling result.

  As described above, the time interval h may be a time zone in units of one hour, a date (365 intervals), a day of the week (7 intervals), a combination of day of the week and time zone (Monday 7 o'clock, Tuesday 10 Time zone, 7 × 24 = 168 sections).

In step 106, the time-series data totaling unit 46 totals the average value for each time interval h for the _j-th observed value x _j of the time-series data.

Specifically, when the time (observation time) at which the observed value x _j is observed is t _j , the average value m (h) of the observed values x _j in the time interval h is expressed by the following equation (3). Thus, Equation (3) is calculated for all time intervals h.

Here, J is the number of observation values included in the time-series data, and N _h is the number of observation values x _j for which δ (t _j , h) = 1. In addition to the average value, the aggregation method may be a maximum value, a minimum value, a total, or the like.

Here, based on the definition of the time interval h, the time series data is aggregated according to the observation time t _j of the observation value x _j , but depending on the definition of the time interval h, the observation date of the observation value x _j It can also be tabulated according to the day of observation.

  The time series data totaling unit 46 outputs the calculated m (h) as a time series data totaling result.

  In step 108, the regression analysis unit 48 uses the document aggregation result n (c, h) of the document aggregation unit 44 as an explanatory variable, and the time series data aggregation result m (h) of the time series data aggregation unit 46 as an objective variable. Perform regression analysis. A linear regression model representing the relationship between m (h) and n (c, h) is defined as the following equation (4).

Constant terms and regression coefficients (β ₀ , β ₁ ,… β _C ) can be found in Reference 2 (Shuichi Kawano, Satoshi Hirose, Shohei Tateishi, Sadanori Konishi “Recent Developments in Regression Modeling and L1 Type Regularization” It can be derived by linear multiple regression, lasso regression, ridge regression, etc. described in the Journal of Statistical Society, 39, 2, pp.211-242, 2010).

  In step 110, the component ratio calculation unit 50 uses the regression coefficient estimation result obtained by the regression analysis unit 48 and the document aggregation result to calculate the cluster component ratio r (c, h) for each time interval h by the following formula ( Calculate and output according to 5).

  Here, the following expression (6) is satisfied.

In step 112, the component decomposing unit 52 uses the value x _j observed at time t _j to calculate the cluster component ratio r (c, h for the time interval h corresponding to time t _j calculated by the component ratio calculating unit 50. ) To decompose into C components as shown in the following equation (7).

  The output unit 24 outputs the component decomposition result of the component decomposition unit 52.

  FIG. 4 schematically shows the processing contents of the time-series data component decomposition apparatus 10 described above. As shown in FIG. 4, the time-series data component decomposition apparatus 10 according to the present embodiment has a plurality of meanings of time-series data, which is a series of values obtained by observing a temporal change of a certain phenomenon. It can be decomposed into specific components and can be used for data mining and marketing.

  As described above, according to the time-series data component decomposition apparatus according to the first embodiment, the observation values included in the time-series data are aggregated for each predetermined time interval, and the document set with time information is collected. From the above, a plurality of clusters are found based on the contents of each document without using external criteria, and the degree to which each document belongs to each of the plurality of clusters is calculated and belongs to each cluster. Based on the degree and the creation time information given to each document, the degree to which each document of the document set with time information belongs to each cluster for each time interval is totaled, and the document set with time information for each time interval The summary results of the time series data are used as explanatory variables, and the summary results of the time series data are used as objective variables. A regression expression representing the relationship between the explanatory variables and the objective variables is estimated by regression analysis. The cluster component ratio for each time interval is calculated based on the aggregation result of the document set, and each observation value included in the time series data is decomposed into cluster components based on the calculation result. The component decomposition of time series data can be performed with high accuracy without using it.

  In the first embodiment, an example is described in which steps 100 to 104 are performed before step 106 (see FIG. 3), but steps 100 to 104 are performed between step 106 and step 108. Alternatively, each of steps 100 to 104 and step 106 may be performed in parallel.

<Second Embodiment>

  In the second embodiment, an example will be described in which time-series data component decomposition is performed by a time-series data component decomposition apparatus provided with an emotion estimation unit that estimates emotion for each document in a document set with time information.

  FIG. 5 is a diagram illustrating a functional configuration of the time-series data component decomposition apparatus 60 according to the second embodiment. In FIG. 5, the same or equivalent parts as in FIG.

  The time series data component decomposition device 60 includes an input unit 20, a calculation unit 26, and an output unit 24. The calculation unit 26 according to the second embodiment includes a word division unit 40, an emotion estimation unit 41, a document clustering unit 42, a document totaling unit 62, a time series data totaling unit 46, a regression analysis unit 64, and a component ratio calculation unit 66. , And a component decomposition unit 68.

  The emotion estimation unit 41 uses an emotion estimator to set each document as a value indicating what kind of emotion the document has been described for each document that has been generated in the word set included in the document set with time information. Estimate and output the degree of each emotion component included in. Here, the emotion estimation unit 41 estimates the degree of each emotion of positive (positive), negative (negative), and neutral (neutral).

  The document totaling unit 62 totalizes the degree to which each document in the document set with time information belongs to each cluster and emotion group for each time interval. In the present embodiment, the totaling result of the document totaling unit 62 is referred to as a document totaling result.

  The regression analysis unit 64 performs regression analysis using the document aggregation result of the document aggregation unit 62 as an explanatory variable and the time series data aggregation result of the time series data aggregation unit 46 as an objective variable.

  The component ratio calculation unit 66 calculates the ratio of each component (a set of clusters and emotions) for each time interval using the regression analysis result of the regression analysis unit 64.

  The component decomposition unit 68 decomposes the time series data into a plurality of components using the component ratio calculated by the component ratio calculation unit 66.

  Note that the time-series data component decomposition apparatus 60 of the second embodiment is also configured by the computer illustrated in FIG. 2, and the CPU 201 reads out and executes a program stored in a recording medium such as the ROM 202 or a hard disk. The above-described functions are realized by cooperating the hardware and the program.

  Next, the operation of the time series data component decomposition apparatus 60 in the present embodiment will be described in detail. When the time-series data component decomposition apparatus 60 receives the input of the time-series data and the document set with time information by the time-series data input unit 30 and the document set input unit 32 with time information, the time-series data component decomposition apparatus 60 executes the processing routine shown in FIG.

In step 300, the word dividing unit 40 generates a word set for each document d _i included in the document set with time information accepted by the document set input unit 32 with time information. Since the word set generation method is as described in the first embodiment (see step 100 in FIG. 3), detailed description thereof is omitted here.

In step 302, using the emotion estimator, the degree of each emotion of positive (positive), negative (negative), and neutral (neutral) for each document in the document set with time information for which a word set has been generated for each document. P (s = 1 | d _i ), P (s = 2 | d _i ), and P (s = 3 | d _i ) are estimated. The emotion estimator is input to the time-series data component decomposition apparatus 60 by the method described in Reference 3 (Alexander Pak and Patrick Paroubek, “Twitter as a Corpus for Sentiment Analysis and Opinion Mining”, LREC 2010.), for example. It is possible to prepare and construct a corpus different from the input data. Specifically, the corpus can be constructed by learning using emotion data to which correct data is given in advance.

  Here, the relationship shown in the following formula (8) is established.

  As long as the emotion estimator can estimate the degree of each of positive (positive), negative (negative), and neutral (neutral) emotions, a method other than the above Reference 3 may be used.

In step 304, as described in the first embodiment (see also step 102 in FIG. 3), the document clustering unit 42 externally extracts the document set with time information from which the word set of each document has been generated. Without a reference, a cluster set having a plurality of semantic groups is found based on the content of the document, and for each document d _i included in the document set with time information, the degree P (c | d _i belonging to each cluster ) Is output.

In step 306, the document counting unit 62, the cluster and the value indicating the degree of belonging to a set of feelings of each document d _i time information with the set of documents (hereinafter, referred to as document number value indicating the degree of) the Aggregate by time interval h. Specifically, when the time at which the document d _i is posted t _i, the time interval h, cluster and emotional set <s, c> document number n belonging to the (s, c, h) the following The calculation is performed for all combinations of <s, c, h>.

Here, D is the number of documents included in the document set with time information, as in Expression (2) of the first embodiment, and δ (t _i , h) is the time interval between time t _i This function returns 1 when included in h and 0 otherwise.

  The document totaling unit 64 outputs the calculated document number n (s, c, h) as a document totaling result.

In step 308, the time-series data totaling unit 46 calculates the average value m for each time interval h for the _j-th observed value x _j (observation time t _j ) of the time-series data as described in the first embodiment. Aggregate (h). The time series data totaling unit 46 outputs the calculated m (h) as the time series data totaling result. Here, m (h) is an average value, but may be a minimum value, a maximum value, or a total value. The time-series data tabulation method is the same as that described with reference to step 106 in FIG. 3 in the first embodiment, and a detailed description thereof is omitted here.

  In step 310, the regression analysis unit 64 uses the document total result n (s, c, h) of the document total unit 62 as an explanatory variable, and the time series data total result m (h) of the time series data total unit 46 as an objective variable. As a regression analysis. A linear regression model representing the relationship between m (h) and n (s, c, h) is defined as the following equation (10).

The constant term and the regression coefficient (β ₀ , β _1,1 ,..., Β _{3, C} ) are as described in the first embodiment, as described in Reference Document 2 (Shuichi Kawano, Satoshi Hirose, Shohei Tateishi, Sadanori Konishi “ It can be derived by linear multiple regression, lasso regression, ridge regression, etc. described in "Recent Modeling of Regression Modeling and L1 Type Regularization", Journal of the Japan Statistical Society, 39, 2, pp.211-242, 2010).

  In step 312, the component ratio calculation unit 66 uses the estimation result of the regression coefficient by the regression analysis unit 64 to calculate the ratio r (s, c, h) of each component (cluster / emotion pair) for each time interval h. Calculate and output according to the following equation (11).

  Here, the following expression (12) is satisfied.

In step 314, the component decomposition unit 68 uses the value x _j observed at time t _j to calculate the ratio r (s, cluster component ratio of the time interval h corresponding to time t _j calculated by the component ratio calculation unit 66. c, h) to be decomposed into 3 × C components as shown in the following equation (13).

  The output unit 24 outputs the component decomposition result of the component decomposition unit 68.

  The time-series data component decomposition apparatus 60 described in the present embodiment is also obtained by observing a temporal change of a certain phenomenon, similarly to the time-series data component decomposition apparatus 10 described in the first embodiment. It is a technique that can decompose time-series data, which is a series of values, into a plurality of semantic components, and can be used for data mining, marketing, and the like.

  As described above, according to the time-series data component decomposition apparatus according to the second embodiment, the observation values included in the time-series data are aggregated for each predetermined time interval, and the document with time information is added. Estimate the degree of each emotion included in each document of the set, find a plurality of clusters based on the contents of each document from the document set with time information without using external criteria, and for each document, The degree of belonging to each of a plurality of clusters is calculated, and each document of the document set with time information is clustered for each time interval based on the degree of belonging to each cluster and the creation time information given to each document. And the degree of belonging to each set of emotions, and for each time interval, the summary result of the document set with time information is the explanatory variable, and the summary result of the time series data is the objective variable. Seki Is calculated by regression analysis, and based on the estimation result of the regression analysis and the total result of the document set with time information, the component ratio of each set of clusters and emotions for each time interval is calculated, and the calculation result Based on this, each observation value included in the time-series data is decomposed into each set of cluster and emotion components, so that the time-series data can be accurately decomposed without using keywords.

  In this embodiment, an example in which steps 300 to 306 are performed before step 308 has been described (see FIG. 6). However, steps 300 to 306 may be performed between step 308 and step 310. Alternatively, each of steps 300 to 306 and step 308 may be performed in parallel. Further, the processing of step 302 and the order of step 304 are not limited to the example shown in FIG. 6, and either may be performed first or may be performed in parallel.

  Note that the time-series data component decomposition apparatus described in the first embodiment and the second embodiment can be applied to any document set with time information. This is particularly effective for a microblog document set having a high real-time property such as a trademark. Further, the document set with time information may be limited to those including specific keywords such as place names and product names, and those including the position of the terminal that created and posted the document within the specified range.

  Further, as described in the first embodiment and the second embodiment, the operation of the components shown in FIG. 1 or FIG. 5 is constructed as a program and used as a time-series data component decomposition apparatus. It can be installed and executed on the network, or distributed via a network.

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium. For example, the program can be stored not only in the above-described ROM and hard disk but also in a portable storage medium such as a flexible disk or a CD-ROM, and installed in a computer or distributed. Alternatively, the program may be stored in another storage device on the network, and the program may be downloaded and executed via the network.

  Further, the above-described time-series data component decomposition apparatus has a computer system therein, but if the “computer system” uses a WWW system, a homepage providing environment (or display environment) is also provided. Shall be included.

  The time-series data component decomposition apparatus described in the first embodiment and the second embodiment may be configured by hardware such as an ASIC.

10 time-series data component decomposition apparatus 20 input unit 22 calculation unit 24 microblog user attribute estimator construction unit 24 output unit 26 calculation unit 30 time-series data input unit 32 document set input unit with time information 40 word division unit 41 emotion estimation unit 42 Document Clustering Unit 44 Document Aggregation Unit 46 Time Series Data Aggregation Unit 48 Regression Analysis Unit 50 Component Ratio Calculation Unit 52 Component Decomposition Unit 60 Time Series Data Component Decomposition Device 62 Document Aggregation Unit 64 Regression Analysis Unit 64 Document Aggregation Unit 66 Component Ratio Calculation Unit 68 Component decomposition unit 200 Computer 201 CPU
202 ROM
203 RAM
204 Communication IF
205 I / O IF
206 Hard disk drive

Claims (8)

A time series data totaling unit that totals each observation value included in time series data that is a series of observation values obtained by observing a temporal change of a phenomenon, for each predetermined time interval;
A document set with time information, which is a set of documents to which creation time information is assigned, is clustered based on the contents of each document without using external criteria, and a plurality of clusters are found. A document clustering unit for calculating the degree of belonging to each of the plurality of clusters;
Based on the degree of belonging to each cluster calculated by the document clustering unit and the creation time information assigned to each document, each document in the document set with time information is added to each cluster for each time interval. A document aggregation section that aggregates the degree of belonging to
For each time interval, the regression result representing the relationship between the explanatory variable and the objective variable is regressed using the aggregation result of the document aggregation unit as an explanatory variable and the aggregation result of the time series data aggregation unit as an objective variable. A regression analysis unit that estimates by analysis,
Using the constant term and regression coefficient of the regression equation estimated for each time interval by the regression analysis unit and the aggregation result for each time interval by the document aggregation unit, the components of each cluster for each time interval A component ratio calculation unit for calculating the ratio of
Each observed value included in the time-series data is converted into a component of each cluster by using the ratio of the cluster components calculated by the component ratio calculating unit in the time interval corresponding to the observation time of the observed value. A component decomposition part to be decomposed;
A time-series data component decomposition apparatus. For each document in the document set with time information, as a value indicating what kind of emotion the document is described based on, it further includes an emotion estimation unit that estimates the degree of each emotion included in each document,
The document totaling unit, for each time interval, based on the degree of belonging to each cluster calculated by the document clustering unit, the creation time information given to each document, and the estimation result of the emotion estimation unit , Totalizing the degree to which each document of the document set with time information belongs to each set of the cluster and the emotion,
The component ratio calculation unit uses the constant term and regression coefficient of the regression equation estimated for each time interval by the regression analysis unit and the aggregation result for each time interval by the document aggregation unit, and uses the time interval. Calculating the ratio of the components of each cluster and the set of emotions for each
Each observation value included in the time series data is obtained by using the ratio of the cluster components calculated by the component ratio calculation unit of the time interval corresponding to the observation time of the observation value, and the cluster and the emotion. Break down into each set of components,
The time-series data component decomposition apparatus according to claim 1. The document set with time information is
Each document included in the document set with time information is a set of documents including a specific keyword.
The time series data component decomposition | disassembly apparatus of Claim 1 or Claim 2. The document set with time information is
A set of documents created and posted on a terminal that exists within a pre-specified location range.
The time series data component decomposition | disassembly apparatus of any one of Claims 1-3. The document set with time information is
It is a set of microblog documents that users create and post sentences within a predetermined number of characters representing the user's own real-time situation or miscellaneous feeling.
The time series data component decomposition | disassembly apparatus of any one of Claims 1-4. A time-series data component decomposition method in a time-series data component decomposition apparatus including a time-series data aggregation unit, a document clustering unit, a document aggregation unit, a regression analysis unit, a component ratio calculation unit, a component ratio calculation unit, and a component decomposition unit ,
The time series data totaling unit aggregates each observation value included in the time series data, which is a series of observation values obtained by observing a temporal change in a phenomenon, for each predetermined time interval. And
A plurality of clusters obtained by performing clustering based on the contents of each document without using an external reference for the document set with time information, which is a set of documents to which creation time information is given, by the document clustering unit. And for each document, calculate the degree of belonging to each of the plurality of clusters,
Based on the degree of belonging to each cluster calculated by the document clustering unit and the creation time information given to each document by the document aggregation unit, the document set with time information is set for each time interval. Aggregate the degree to which each document belongs to each cluster,
By the regression analysis unit, for each time interval, the aggregation result of the document aggregation unit is an explanatory variable, and the aggregation result of the time series data aggregation unit is an objective variable, and the relationship between the explanatory variable and the objective variable The regression equation representing is estimated by regression analysis,
By using the constant term and regression coefficient of the regression equation estimated by the regression analysis unit for each time interval by the component ratio calculation unit, and the aggregation result for each time interval by the document aggregation unit, the time interval Calculating the proportion of each cluster component for each
By using the component ratio of the cluster calculated by the component ratio calculation unit of the time interval corresponding to the observation time of the observation value, each observation value included in the time series data by the component decomposition unit, A time-series data component decomposition method for decomposing into components of each cluster.   The program for functioning a computer as each means of the time series data component decomposition | disassembly apparatus of any one of Claims 1-5. A computer-readable recording medium recording a program for causing a computer to function as each means of the time-series data component decomposition apparatus according to any one of claims 1 to 5.

Images