JP5077711B2 - Time series analysis apparatus, time series analysis method, and program - Google Patents

Description

  The present invention relates to a technique for estimating transition of time series data.

  As shown in Non-Patent Document 1, each company provides an ASP type service that collects and analyzes word-of-mouth information on a blog or a bulletin board in real time. However, with these services, it is not predicted how the number of future blog entries will change.

  Here, in other fields such as economy and traffic, a technology for predicting the transition of stock prices and traffic information is used.

  For example, the learning device described in Patent Document 1 does not predict the number of blog writes, but acquires time-series data related to traffic information and uses an AR (Auto Regression) model class to perform periodic analysis. It is possible to calculate a future predicted value to which a pattern is added.

[online], [Search September 25, 2009], Internet <URL: http://kandoreport.jp/>

JP 2006-11572 A

  However, the learning device described in Patent Literature 1 sometimes cannot accurately predict the transition of Web-related data such as the number of blog writes.

  In addition to periodicity, there is a unique tendency in fluctuations in Web-related data. On the other hand, the learning device of Patent Document 1 calculates a predicted value in consideration of periodic fluctuations, but does not take into account a specific tendency that the data has an upward trend or a downward trend.

  For this reason, when the technique described in Patent Document 1 is applied to Web-related data, there is a problem that a deviation between the predicted value and the actual value becomes large.

  An object of the present invention is to accurately obtain a predicted value of Web-related data.

In order to achieve the above-mentioned object, the time series analysis apparatus of the present invention represents an increase / decrease tendency peculiar to Web-related data, a function whose parameter is undetermined is previously defined as a trend function, and is aggregated for a specific matter. Trend parameter calculation means for calculating the parameter of the trend function as a trend parameter based on the time-series value of the Web-related data, and a periodicity peculiar to the Web-related data, and a function whose parameter is undetermined as a periodic function Based on the time series value that is defined in advance and the trend function to which the trend parameter is applied is removed, a periodic parameter calculation unit that calculates a parameter of the periodic function as a periodic parameter, and a calculation by the trend parameter calculation unit The trend function to which the trend parameter is applied, and the previous A prediction formula calculation means for calculating a prediction formula for predicting a transition of Web-related data related to the specific matter by adding the periodic function to which the periodic parameter calculated by the periodic parameter calculation means is applied; , have a,
Among the time series values, there is a correction unit that detects outliers that are suddenly increased or decreased, and corrects the outliers,
When the outlier is detected by the correcting unit, the trend parameter calculating unit calculates the trend parameter from the time series value in which the outlier is corrected by the correcting unit,
In the plurality of burst periods, each of which is a period when the time series values are aggregated, the correction means includes mintime as the first time in each burst period, and the last time in the burst period. , Maxtime is a positive real number, t is the time when the time series values are tabulated, y _{t is the} time series value at time t , m is the natural number, and BURST is the burst value in the burst period. 3 calculates the burst value of each burst period, extracts a predetermined number of burst periods in descending order of the absolute value of the burst value, and corrects the extracted time-series values within the burst period as outliers.

The time series analysis method according to the present invention represents a trend of increase or decrease peculiar to Web-related data, and the trend-parameter calculating means that holds in advance a function whose parameter is undetermined as a trend function has collected Web-related data for a specific matter. A periodic parameter calculation means for calculating a parameter of the trend function as a trend parameter based on the time series value of the time series, representing a periodicity peculiar to Web-related data, and holding a function whose parameter is undetermined in advance as a periodic function; Based on the time series value from which the trend function to which the trend parameter has been applied is removed, the parameter of the periodic function is calculated as a periodic parameter, and the prediction formula calculating means calculates the trend parameter calculated by the trend parameter calculating means. Applied trend function and periodic parameter calculation By adding the said periodic function of applying said periodic parameter calculated by the step, calculates a prediction equation for predicting a transition of data in Web relationships for the particular matter,
Detecting outliers that are suddenly increased or decreased among the time series values, correcting the outliers,
When calculating the trend parameter, if the outlier is detected, the trend parameter is calculated from the time series value in which the outlier is corrected,
In a plurality of burst periods, each period of which the time series values are aggregated being the first period, the first period in each burst period is mintime, the last period in the burst period is maxtime, and positive Assuming that the real number is α, the time when the time series values are aggregated is t, the time series value at the time t is y _t , the natural number is m, and the burst value in the burst period is BURST, each burst period is expressed by the following formulas 1 to The burst value is calculated, a predetermined number of burst periods are extracted in descending order of the absolute value of the burst values, and the time series values in the extracted burst periods are corrected as outliers .

The program of the present invention shows a time series of Web-related data aggregated with respect to a specific matter, in which a computer shows a tendency to increase or decrease specific to Web-related data, a function whose parameter is undetermined is previously defined as a trend function. A trend parameter calculation procedure for calculating a parameter of the trend function as a trend parameter based on a value, a periodicity peculiar to Web-related data, a function whose parameter is undetermined is previously defined as a periodic function, and the trend parameter Based on the time series value from which the trend function is applied, the periodic parameter calculation procedure for calculating the parameter of the periodic function as a periodic parameter, and the trend parameter calculated in the trend parameter calculation procedure is applied. Trend function and periodic parameter calculation By adding the said periodic function of applying said periodic parameter calculated in step, the prediction equation calculation step of calculating a prediction equation for predicting a transition of data in Web relationships for the particular matter,
A correction procedure for detecting outliers that are suddenly increased or decreased among the time series values and correcting the outliers,
And execute
In the trend parameter calculation procedure, when an outlier is detected by the correction procedure, the trend parameter is calculated from the time series value in which the outlier is corrected by the correction procedure,
In the correction procedure, in a plurality of burst periods in which each period of the period in which the time series values are aggregated is the first period, the first period in each burst period is mintime, and the last period in the burst period , Maxtime is a positive real number, t is the time when the time series values are tabulated, y _{t is the} time series value at time t , m is the natural number, and BURST is the burst value in the burst period. 3 is a program for calculating a burst value of each burst period, extracting a predetermined number of burst periods in descending order of the absolute value of the burst value, and correcting the extracted time series values in the burst period as outliers .

  According to the present invention, the time series analysis apparatus calculates the parameter of the function representing the increase / decrease tendency and periodicity peculiar to the Web-related data. Can accurately calculate the predicted value.

1 is an overall view showing a configuration example of a time-series data analysis system according to a first embodiment of the present invention. It is a block diagram which shows one structural example of the blog extraction / aggregation part of the 1st Embodiment of this invention. It is a figure which shows the information contained in the time series analysis request data of the 1st Embodiment of this invention. It is the table | surface which put together the content which the learning data of the 1st Embodiment of this invention shows. It is a table | surface which shows the burst value calculated by varying the burst period of the 1st Embodiment of this invention. (A) It is a figure which shows the aligned burst value of the 1st Embodiment of this invention. (B) It is a figure which shows the aligned burst value of the 1st Embodiment of this invention. It is the table | surface which put together the content which the prediction type | formula data of the 1st Embodiment of this invention shows. It is a flowchart which shows the time series analysis process of the analysis and visualization server of the 1st Embodiment of this invention. It is a flowchart which shows the blog extraction and totalization process of the 1st Embodiment of this invention. It is a flowchart which shows the periodic component extraction process of the 1st Embodiment of this invention. It is the line graph which plotted the time series value of the 1st Embodiment of this invention before correction | amendment of an outlier. It is the line graph which plotted the time series value of the 1st Embodiment of this invention after correction | amendment of an outlier. It is the line graph which plotted the time series value of the 1st Embodiment of this invention after a trend function calculation. It is the line graph which plotted the time series value of the 1st Embodiment of this invention after trend component removal. It is the line graph which plotted the predicted value of the 1st Embodiment of this invention. It is the line graph which plotted the predicted value of the 1st Embodiment of this invention. It is a figure which shows the information contained in the influence analysis request data of the 2nd Embodiment of this invention. It is the table | surface which put together the content which the prediction type | formula data of the 2nd Embodiment of this invention shows. It is a flowchart which shows the influence analysis process of the analysis and visualization server of the 2nd Embodiment of this invention. It is the line graph which plotted the predicted value of the 2nd Embodiment of this invention. It is a figure which shows the information contained in the growth score calculation request data of the 3rd Embodiment of this invention. It is a flowchart which shows the growth calculation process of the analysis and visualization server of the 3rd Embodiment of this invention. It is the line graph which plotted the predicted value of the 3rd Embodiment of this invention. It is a figure which shows the information contained in the simulation request data of the 4th Embodiment of this invention. It is a flowchart which shows the event effect prediction process of the 4th Embodiment of this invention.

  Embodiments for carrying out the present invention will be described in detail with reference to the drawings. FIG. 1 is an overall view showing a configuration example of a time-series data analysis system 1 according to the present embodiment. The time series data analysis system analyzes time series data indicating values aggregated in time series (hereinafter referred to as “time series values”). In the present embodiment, the number of blog writes totaled in time series is set as a time series value.

  Referring to FIG. 1, the time-series data analysis system 1 includes a user PC 10, an analysis / visualization server 20, a blog DB 30, and a prediction parameter DB 40.

  The analysis / visualization server 20 includes a blog extraction / aggregation unit 201, a trend component learning unit 202, a periodic component learning unit 203, and an analysis result creation unit 204.

  The blog DB 30 collects blog data, classifies the data into categories, and stores the data.

  The user PC 10 receives time series analysis request data for requesting analysis of time series values. When the time series analysis request data is input, the user PC 10 transmits the time series analysis request data to the analysis / visualization server 20.

  FIG. 2 is a diagram illustrating information included in the time series analysis request data. Referring to the figure, the time series analysis request data includes information indicating “category”, “keyword”, “teacher data period”, and “prediction period”.

  “Category” is an attribute of a blog to be counted. For example, keyword genres such as beverages, electrical appliances, mobile phones, beers, and celebrities are designated. A “keyword” is a phrase searched for in a blog. The “teacher data period” is a period during which blogs are tabulated. The “prediction period” is a period during which a predicted value of the number of blog writes is calculated.

  FIG. 3 is a block diagram illustrating a configuration example of the blog extraction / aggregation unit 201. Referring to the figure, the blog extraction / aggregation unit 201 includes an aggregation unit 2011 and a singularity correction unit 2012.

  The aggregation unit 2011 acquires a category, a keyword, a teacher data period, and a prediction range from the time series analysis request data illustrated in FIG. The aggregation unit 2011 extracts blog data belonging to the category in the teacher data period from the blog data stored in the blog DB 30. The tabulation unit 2011 searches the blog data for blog writings including keywords, and tabulates the number of blog writings searched for each date within the teacher data period.

  The singularity correction unit 2012 detects and corrects outliers from the counting results. An outlier is a value when the time-series value suddenly increases or decreases, such as a sudden increase in the number of writes due to irregular events or a sudden decrease or loss in the number of writes due to a system failure. Outliers are also called singularities.

  In the outlier extraction, the singularity correction unit 2012 sets the length to the minimum value for the burst period whose length is variable in the range from the minimum value to the maximum value. In this embodiment, the minimum value is set for one day, and the user sets only the maximum value.

  Then, the singularity correction unit 2012 uses the first day of the burst period as the first day of the teacher data period, and calculates the burst value using the following formula. Thereafter, the singularity correction unit 2012 calculates the burst value corresponding to each date, with each date in the burst period as the first day of the burst period. In a similar procedure, the singularity correction unit 2012 varies the length of the burst period until the maximum value is reached, and calculates a burst value corresponding to each date.

上記（式１）において、「ＢＵＲＳＴ」はバースト値であり、「α」は、長期間選択調整係数（例えば、２．５）である。「mintime」は、バースト期間内の最初の日付である。「maxtime」は、バースト期間内の最後の日付である。

In the above (Formula 1), “BURST” is a burst value, and “α” is a long-term selection adjustment coefficient (for example, 2.5). “Mintime” is the first date in the burst period. “Maxtime” is the last date in the burst period.

In the above (Formula 3), “y _i ” is the number of blog writes counted on the “i” -th date counted from “mintime”. “M” is a value (for example, 7) that is not less than the maximum value of the burst period and not more than the length of the teacher data period. In other words, “SABUN _i ” is the number of blog writes “y _i ” aggregated on the “i” th date, and “m” days before the “i” th date to the “i” th date. This is the difference between the average number of blog entries.

  The length of the burst period “W” and the value of “m” are adjusted so that outliers are properly detected. For example, when it is desired to detect outliers in units of one day, “W” is set to zero. In addition, when outliers often appear continuously for two days or more, the user may set “W” to 1 or more. “M” is set to an appropriate value according to the size of “W”.

  When there is an outlier within a certain burst period, the absolute value of the burst value during that burst period is larger than when there is no outlier within the burst period. For this reason, the singularity correction unit 2012 can efficiently determine which burst period has an outlier by calculating a burst value for each burst period.

  The singularity correction unit 2012 sorts the burst periods in descending order of the burst value. The singularity correction unit 2012 extracts a predetermined number of burst periods having a high burst value and a predetermined number of burst periods having a low burst value as burst periods including singularities. For example, the upper and lower 10 burst periods are extracted.

  However, in the burst period extraction, the singularity correction unit 2012 extracts burst periods in order by giving priority to a period in which the absolute value of the burst value is high. Then, burst periods whose dates overlap with the already extracted burst periods are excluded from extraction targets.

  The singularity correction unit 2012 uses the extracted time series values within the burst period as outliers, and corrects these outliers. For example, the singularity correction unit 2012 corrects the average value of the latest data. However, if the outlier is the last data and there is no data immediately after it, the singularity correction unit 2012 excludes the outlier and does not use the value in the time series analysis described below. The singularity correction unit 2012 outputs learning data indicating the corrected total result.

  Returning to FIG. 1, the trend component learning unit 202, the periodic component learning unit 203, and the analysis result creation unit 204 will be described.

  The trend component learning unit 202 calculates a trend parameter from the learning data output from the singular point correction unit 2012.

  The trend parameter is a parameter of a function (hereinafter referred to as “trend function”) indicating an increasing tendency or a decreasing tendency of the number of blog entries during the teacher data period. For example, a trend function coefficient or index is calculated as the trend parameter.

  The trend function is a function represented by the following formula, for example. The trend function is also referred to as a trend component because it is a component in which only an increase / decrease tendency is extracted in the prediction of the number of blog entries.

上記（式４）において、関数「ｙ（ｘ）」は、トレンド関数である。この「ｙ（ｘ）」の値は、教師データ期間における最初の日付から「ｘ」日目におけるブログ書き込み数となる。「α」、「β」は、トレンドパラメータである。

In the above (Formula 4), the function “y (x)” is a trend function. The value of “y (x)” is the number of blog writes on the “x” day from the first date in the teacher data period. “Α” and “β” are trend parameters.

  The trend component learning unit 202 approximates the measurement data to the above (formula 4) by using, for example, the least-squares method, and calculates the trend parameter.

  Specifically, the trend parameter is calculated by substituting the measured value into the following equation.

上記（式５）、（式６）において、「ｔ」は教師データ期間内の最初の日からの日数であり、「Σ」は、教師データ期間における総和をとることを示している。「α」
、「β」は、トレンドパラメータである。「ｙ」は、集計されたブログの書き込み数である。

In the above (Expression 5) and (Expression 6), “t” is the number of days from the first day in the teacher data period, and “Σ” indicates that the sum in the teacher data period is taken. "Α"
, “Β” is a trend parameter. “Y” is the total number of written blogs.

  The trend component learning unit 202 derives a trend function to which the calculated trend parameter is applied, and obtains the trend function by substituting the number of days (“x”) corresponding to the date from the number of written learning data for each date. The written number (“y (x)”) is subtracted.

  The periodic component learning unit 203 calculates a periodic parameter by performing Fourier analysis on the learning data after subtracting the calculated value of the trend function.

  The periodic parameter is a parameter of a function (hereinafter referred to as “periodic function”) indicating the number of blog writings in the teacher data period, that is, a periodic variation of the time series value. For example, a periodic function coefficient or index is calculated as the periodic parameter.

  The periodic function is a function represented by the following formula, for example. Since the periodic function is a component in which only periodicity is extracted in the prediction of the number of blog writes, it is also called a periodic component.

上記（式７）〜（式１３）において、「Ｙ（ｔ）」は周期関数である。この「Ｙ（ｔ）」の値は、教師データ期間における最初の日付から「ｔ」日目におけるブログ書き込み数となる。「An」、および「Bn」は、１年を周期とする周期パラメータである。「an」、および「bn」は、１週間を周期とする周期パラメータである。「ｎ」は、フーリエ係数である。例えば、周期成分学習部２０３は、１〜５の「ｎ」について、周期パラメータを算出し、それぞれを加算する。

In the above (Expression 7) to (Expression 13), “Y (t)” is a periodic function. The value of “Y (t)” is the number of blog writes on the “t” day from the first date in the teacher data period. “An” and “Bn” are periodic parameters having a period of one year. “An” and “bn” are periodic parameters having a period of one week. “N” is a Fourier coefficient. For example, the periodic component learning unit 203 calculates periodic parameters for “n” of 1 to 5 and adds them.

  The analysis result creation unit 204 adds a trend function and a periodic function to generate a prediction formula for predicting the transition of the number of blog writes.

  The analysis result creation unit 204 also stores, in the prediction parameter DB 40, prediction formula data indicating each parameter of the prediction formula, that is, a trend parameter and a periodic parameter.

  Then, the analysis result creation unit 204 calculates the number of blog writes for each date within the prediction period indicated by the time-series analysis request data using a prediction formula. The analysis result creation unit 204 transmits analysis result data indicating each predicted value to the user PC 10 using the value calculated within the prediction period as the predicted value.

  In addition, the analysis result creation unit 204 predicts using the previously obtained prediction formula if the category and keyword indicated by the time series analysis request data are the same as the category and keyword for which the trend component and periodic component have been calculated previously. Calculate the value.

  FIG. 4 is a table summarizing the contents indicated by the learning data created by the blog extraction / aggregation unit 201. Referring to the figure, the learning data includes data indicating a plurality of “dates” and “number of writings” in association with “category” and “keyword”. “Category” is a category to which the blog data to be extracted belongs. “Keyword” is a keyword searched in the blog data. “Date” is each date within the aggregation period. The “number of writes” is the number of blogs containing keywords.

  For example, consider a case where the category indicated by the time series analysis request data is “food” and “price increase” is specified as a keyword. The blog extraction / aggregation unit 201 searches for specified keywords for blog data belonging to the category of “groceries”. The blog extraction / aggregation unit 201 extracts blog data including the keyword “price increase”, and totals the number of data on each date within the teacher data period.

  A method for detecting and correcting a singular point from learning data will be described with reference to FIGS. Here, it is assumed that the length of the burst period is variable from 1 day to 7 days. Further, “m” in the above (Expression 3) is set to 7.

  As shown in FIG. 5, the blog extraction / aggregation unit 201 changes the burst period length (burst range) from 1 to 7, and calculates a burst value with each date in the teacher data period as the first day of the burst period. To do.

  Referring to FIG. 5, for example, when the burst range is 1 day and “mintime” is August 25, 2009, the burst value is “0”. When the burst range is 2 days and “mintime” is August 25, 2009, “maxtime” is August 26, 2009, and the burst value is “22.5”.

  The blog extraction / aggregation unit 201 sorts the burst values calculated in each period in descending order as shown in FIG. The blog extraction / aggregation unit 201 extracts burst periods in order from the top of the burst value.

  Here, when there are a plurality of periods in which the burst values are the same, the longer burst range is extracted with priority. Even if there are a plurality of periods with the same burst value, the blog extraction / aggregation unit 201 may extract all of them.

  Referring to FIG. 6A, the highest burst value is a burst value calculated with the first day of the burst period being August 28, 2009 and the burst range being one day. The second highest burst value is a burst value calculated with the first day of the burst period being September 14, 2009 and the burst range being two days. The third highest burst value is the burst value corresponding to September 14, 2009. First, these three periods are extracted as burst periods that can include outliers.

  The fourth highest burst value was a burst value calculated with the first day of the burst period being August 27, 2009 and the burst range being two days. This burst period includes August 28, the date within the already extracted burst period. For this reason, the burst period in which the first day of the burst period is August 27, 2009 and the burst range is two days is excluded from the extraction target. The shaded portion in FIG. 6A is a burst period excluded from the extraction target. The result of excluding these burst periods is shown in FIG.

  In FIG. 5, the number surrounded by a thick solid line is a burst value corresponding to the extracted burst period. The numbers surrounded by diagonal lines are burst values corresponding to burst periods in which the extracted burst periods and dates overlap. Burst periods corresponding to these hatched parts are excluded from extraction targets.

  The blog extraction / aggregation unit 201 extracts the top 10 burst periods after removing the burst periods corresponding to the shaded portions in FIGS. 5 and 6A. Next, in order from the burst value with the lowest burst value, the lower 10 burst periods are extracted in the same manner.

  In addition, it is good also as a structure extracted sequentially from the burst period corresponding to the burst value with the largest absolute value irrespective of a high-order and a low-order.

  Thus, the blog extraction / aggregation unit 201 can detect outliers with high accuracy by not extracting burst periods with overlapping dates.

  Then, the blog extraction / aggregation unit 201 corrects the extracted value in the burst period to an average value before and after the extracted value.

  FIG. 7 is a table summarizing the contents of the prediction formula data. As shown in the figure, the prediction formula data includes information indicating “category”, “keyword”, “teacher data period”, “trend parameter”, and “periodic parameter”.

  “Category” is a category to which the blog data to be extracted belongs. “Keyword” is a keyword searched in the blog data. The “teacher data period” is a period in which time series values are tabulated to calculate a prediction formula. The “trend parameter” is a parameter of a trend function calculated by the least square method. “Periodic parameter” is a parameter of a periodic function calculated by Fourier analysis.

  Next, the operation of the analysis / visualization server 20 will be described. FIG. 8 is a flowchart showing the operation of the analysis / visualization server 20. This operation starts when the analysis / visualization server 20 receives time-series analysis request data.

  The blog extraction / aggregation unit 201 extracts blog data based on the time series analysis request data, and executes blog extraction / aggregation processing for aggregation (step S1).

  The trend component learning unit 202 extracts a trend component by approximating the counting result to the above equation (4) (step S2). The periodic component learning unit 203 executes a periodic component extraction process for extracting a periodic component (step S3).

  The analysis result creation unit 204 stores the trend parameter and the periodic parameter in the prediction parameter DB 40, and creates a prediction formula by adding the trend function and the periodic function (step S4). The analysis result creation unit 204 calculates the predicted value of the time series value in the designated prediction period from the prediction formula, and outputs the analysis result data to the user PC 10 (step S5). After step S5, the analysis / visualization server 20 ends the operation.

  FIG. 9 is a flowchart showing blog extraction / aggregation processing. Referring to the figure, the blog extraction / aggregation unit 201 extracts and aggregates blogs that include words that match all or partially match the category and keyword specified in the time series analysis request data within the teacher data period ( Step S11).

  The blog extraction / aggregation unit 201 divides the teacher data period for each burst period, and calculates a burst value in each burst period. Then, the blog extraction / aggregation unit 201 detects and corrects a singular point by extracting a predetermined number of burst periods having a high burst value (step S12). After step S12, the blog extraction / counting unit 201 ends the blog extraction / counting process.

  FIG. 10 is a flowchart showing the periodic component extraction process. Referring to the figure, the periodic component learning unit 203 calculates yearly periodic parameters An and Bn using the above (Expression 8) and (Expression 9) (step S31). Then, the periodic component learning unit 203 calculates the yearly periodic parameters an and bn using the above (formula 10) and (formula 11) (step S32). After step S32, the periodic component learning unit 203 ends the periodic component extraction process.

  An example of the operation result of the analysis / visualization server 20 will be described with reference to FIGS. FIG. 11 is a line graph in which the analysis / visualization server 20 plots the result of totaling the number of blog writes during the teacher data period based on the time-series analysis request data. In the figure, the horizontal axis represents each counting date within the teacher data period, and the vertical axis represents the number of blog writes counted on the counting date.

  The analysis / visualization server 20 divides the teacher data period into a plurality of burst periods, and calculates a burst value in each burst period using the above (Equation 1). As shown in the figure, the number of writes suddenly increases at B1 and the number of writes suddenly decreases at B2, so that the absolute value of the burst value is high in the burst period including these B1 and B2. Become. The analysis / visualization server 20 corrects outliers within these burst periods.

  FIG. 12 is a line graph showing the counting result after correcting the singular points. In the figure, the horizontal axis represents each counting date within the teacher data period, and the vertical axis represents the number of blog writes counted on the counting date. As shown in the figure, the singular points B1 and B2 are both corrected.

  FIG. 13 is a diagram illustrating a result of the trend component being obtained by the analysis / visualization server 20 from the aggregation result. In the figure, the horizontal axis represents each counting date within the teacher data period, and the vertical axis represents the number of blog writes counted on the counting date. A logarithmic function represented by a one-dot chain line in FIG. As shown in the figure, the transition of the number of blog entries has a tendency to gradually increase with time, that is, a trend.

  FIG. 14 is a diagram illustrating a result of the trend component obtained by the analysis / visualization server 20 removing the trend component from the counting result. In the figure, the horizontal axis represents each counting date within the teacher data period, and the vertical axis represents the number of blog writes counted on the counting date. From the total result, a periodic function represented by a two-dot chain line in FIG. As shown in the figure, the number of blog entries changes periodically on a yearly basis and also periodically on a weekly basis.

  FIG. 15 is a line graph showing the result of the prediction value calculated by the analysis / visualization server 20 using the prediction formula. In the figure, the horizontal axis represents each counting date within the teacher data period, and the vertical axis represents the number of blog writes counted on the counting date. The solid line part in the figure is the total result obtained by correcting the singular points, and the two-dot chain line part in the figure is the result of plotting the predicted values within the prediction period calculated by the prediction formula. As shown in the figure, the analysis / visualization server 20 can predict the future transition of the number of blog writes by using the prediction formula.

  The analysis / visualization server 20 can also simultaneously display the aggregation results calculated under different conditions and the analysis results calculated under different conditions for comparison.

  FIG. 16 is a line graph showing a totaled result obtained by counting similar products “product A” and “product B” as keywords in the same category and a calculated analysis result. In the figure, the solid line portion is the total result for these keywords. The broken line is a plot of the predicted values for these keywords.

  The user can compare how the number of writings related to each product changes, and can consider how to execute sales or promotion of the product.

  In this embodiment, the analysis / visualization server 20 collects the number of blog writes as a time-series value. However, if the value is collected in time-series, it collects and analyzes other time-series values. You can also For example, the time series value may be Web-related data other than a blog. The web-related data includes the number of articles on the web page, the number of times the article has been searched, an access log to the website, and the like.

  Further, the time series value may be a stock price or an exchange rate. Alternatively, the analysis / visualization server 20 can also analyze system log data such as a CPU usage rate as a time series value.

  In the present embodiment, the analysis / visualization server 20 totals the number of blog writes every day, but the totaling time is not limited to one day as long as it is within the teacher data period. For example, the analysis / visualization server 20 may count the number of blog writes every hour.

  In this embodiment, the analysis / visualization server 20 detects outliers using the above (Expression 1) to (Expression 3), but if these values can be detected suddenly, these An outlier may be detected using an expression other than an expression.

For example, in the above (Equation 3), the difference between the average value and “y _i ” is “SABUN _i ”, but the absolute value of the difference may be calculated as “SABUN _i ”. In this case, the analysis / visualization server 20 extracts a high-order burst period and a predetermined number of burst periods.

  In the present embodiment, the analysis / visualization server 20 approximates the aggregation result to a logarithmic function as in the above (Equation 4), but it goes without saying that the analysis / visualization server 20 may approximate a function other than the logarithmic function. For example, the analysis / visualization server 20 may perform linear approximation using an expression as shown below.

または、分析・可視化サーバ２０は、下記に示すような式を用いて累乗近似を行ってもよい。

Alternatively, the analysis / visualization server 20 may perform power approximation using an expression as shown below.

もしくは、分析・可視化サーバ２０は、下記に示すような式を用いて指数近似を行ってもよい。

Alternatively, the analysis / visualization server 20 may perform exponential approximation using an expression as shown below.

上記（式１４）〜（式１６）において、「α」、「β」はトレンドパラメータ、「ｙ」はブログ書き込み数、「ｘ」は集計日である。

In the above (Expression 14) to (Expression 16), “α” and “β” are trend parameters, “y” is the number of blog writes, and “x” is a counting date.

  In the present embodiment, the analysis / visualization server 20 obtains a periodic component having a cycle of one year and a periodic component having a cycle of one week. However, a periodic component having a different period may be obtained. Of course. For example, when the number of blog writes is counted every hour, the analysis / visualization server 20 can also calculate a periodic component with a period of one day, and the period is not one year but two years or three. It may be a year.

  The analysis / visualization server 20 of the present embodiment uses the least-squares method in calculating trend components, but if the regression results can be analyzed, the trend components are calculated by a method other than the least-squares method. May be.

  The analysis / visualization server 20 of this embodiment calculates a periodic component by performing Fourier analysis. However, if a periodic variation pattern can be obtained, the periodic component is obtained by a method other than Fourier analysis. It may be calculated. For example, the analysis / visualization server 20 can obtain a periodic component by using an AR (Autoregressive model) model.

  As described above, according to the present embodiment, the time series analysis apparatus calculates the function parameters representing the increase / decrease tendency and the periodicity peculiar to the Web-related data. Even if there is a tendency to increase or decrease, the predicted value can be calculated accurately.

(Second Embodiment)
A second embodiment of the present invention will be described. The time series analysis system according to the present embodiment is different from the time series analysis system according to the first embodiment in that an event influence degree can be further calculated. In addition, about the time series analysis system of this embodiment, the detailed description of the structure similar to the structure demonstrated in FIGS. 1-16 is abbreviate | omitted, and a different point from 1st Embodiment is demonstrated in detail.

  The user PC 10 of the present embodiment accepts input of influence degree analysis request data for requesting calculation of the influence degree of an event in addition to time series analysis request data. The event influence level is a value indicating the degree to which the actual number of blog writes after the event is executed is increased or decreased with respect to the predicted value. For example, the analysis / visualization server 20 uses the value obtained by dividing the actual number of writes by the predicted value as the degree of influence.

  Specifically, when the user wants to determine the degree of influence of an event when a certain period has elapsed since the start of the event, the user sets the event type, event start time, prediction period, and teacher data period to the user PC 10. input. Here, a predetermined period before the start time of the event is set as the teacher data period. Further, a predetermined period after the event start time is set as the prediction period. In the present embodiment, it is desirable that the length of the teacher data period is set to be equal to or longer than the length of the prediction period.

  For example, when a sense of one month has passed since the start of the event and it is desired to calculate the degree of influence of the event during that one month, the user can use the one year before the start of the event for the teacher data period and the first month of the event. Enter the period as the forecast period.

  The user PC 10 outputs impact analysis request data including information indicating an event type, an event start time, a prediction period, and a teacher data period to the analysis / visualization server 20.

  FIG. 17 is a diagram illustrating information included in the impact analysis request data. Referring to the figure, the impact analysis request data includes information indicating “category”, “keyword”, “event type”, “teacher data period”, and “prediction period”.

  “Category” and “Keyword” are the same as those in the first embodiment. The “teacher data period” is a predetermined period before the event start time. The “prediction period” is a predetermined period after the event start time.

  When the impact analysis request data is input, the analysis / visualization server 20 obtains a prediction formula by the same method as in the first embodiment, and calculates a prediction value for each date within the prediction period from the prediction formula. .

  On the other hand, the analysis / visualization server 20 reads the blog data in the prediction period from the blog DB 30 and totals the actual number of blog writes in the prediction period to obtain an actual measurement value. Then, the analysis / visualization server 20 sets the value obtained by dividing the total of the actual measurement values in the prediction period by the total of the prediction values in the prediction period as the event influence level.

  The analysis / visualization server 20 transmits analysis result data for displaying the predicted value and the actual measurement value to the user PC 10. The analysis / visualization server 20 stores prediction formula data further including information indicating the type of event and the degree of influence in the prediction parameter DB 40.

  FIG. 18 is a table summarizing the contents indicated by the prediction formula data. As shown in the figure, in addition to information indicating “category”, “keyword”, “teacher data period”, “trend parameter”, and “periodic parameter”, the prediction formula data includes “event” and “influence degree”. Contains information indicating.

  “Event” is the type of event to be analyzed. The “influence degree” is the influence degree of the event obtained from the actual measurement value and the predicted value.

  FIG. 19 is a flowchart showing the operation of the analysis / visualization server 20 of the present embodiment. This operation starts when the analysis / visualization server 20 receives the impact analysis request data.

  The analysis / visualization server 20 executes a time series analysis process for calculating a predicted value in the prediction period from the total number of blog writings in the teacher data period (step T1). The time series analysis process in FIG. 19 is the same process as the time series analysis process shown in FIG.

  The analysis / visualization server 20 calculates the influence degree of the event by dividing the total of the actual measurement values in the prediction period by the total of the prediction values in the prediction period (step T2).

  With reference to FIG. 20, an example of the calculation result of the event influence level will be described. The figure is a line graph showing predicted values and actual measured values. The solid line in the figure is a plot of the number of blog writes actually tabulated. The alternate long and short dash line plots the predicted values. The “P1” time is the start time of the event. The hatched portion is a portion indicating a difference between the predicted value and the actually measured value.

  As shown in FIG. 20, from the event start time “P1”, the actual number of blog writes increases with respect to the predicted value within the prediction period. This increase is estimated to be the number of writes increased due to the event.

  In the present embodiment, the analysis / visualization server 20 uses the value obtained by dividing the total of the actual measurement values by the total of the prediction values as the degree of influence. The degree of influence may be calculated by other methods. For example, the difference itself between the actually measured value and the predicted value on a certain date within the prediction period may be used as the degree of influence.

  As described above, according to the present embodiment, the analysis / visualization server 20 can quantitatively measure the effect of the event.

(Third embodiment)
A third embodiment of the present invention will be described. The time series analysis system of this embodiment is different from the time series analysis system of the first embodiment in that a growth score can be further calculated. In addition, about the time series analysis system of this embodiment, the detailed description of the structure similar to the structure demonstrated in FIGS. 1-16 is abbreviate | omitted, and a different point from 1st Embodiment is demonstrated in detail.

  The user PC 10 of this embodiment accepts input of growth score calculation request data for requesting calculation of a growth score in addition to time series analysis request data.

  The growth degree score is a value indicating the degree of increase or decrease of the predicted value calculated with the time after that as the all day of the teacher data period with respect to the predicted value calculated with the certain period as the all day of the teacher data period. For example, the growth score is obtained by dividing the sum of the predicted values for the later day of the teacher data period by the sum of the predicted values for the earlier day of the teacher data period.

  The analysis / visualization server 20 calculates the growth score based on the growth score calculation request data.

  Here, the date when the number of blog writes at the end is counted is “present”. The user sets the date before the current time as the “comparison target time”. The user sets a predetermined period after the present as the prediction period based on the present, and sets a predetermined period before the present as the teacher data period.

  For example, the user sets the date one week before the current time as the comparison target time, sets the two months after the current time as the prediction period, and sets the previous year as the teacher data period.

  The user PC 10 outputs growth score calculation request data including information indicating the comparison target time, the prediction period, and the teacher data period to the analysis / visualization server 20.

  FIG. 21 is a diagram illustrating information included in the growth score calculation request data. Referring to the figure, the impact analysis request data includes information indicating “category”, “keyword”, “teacher data period”, “prediction period”, and “comparison target time”.

  “Category” and “Keyword” are the same as those in the first embodiment. “Teacher data period” and “prediction period” are a teacher data period and a prediction period set based on the current time. The “comparison time” is a predetermined date before the present time.

  FIG. 22 is a flowchart showing the operation of the analysis / visualization server 20 of this embodiment. This operation starts when the analysis / visualization server 20 receives the growth score calculation request data.

  The analysis / visualization server 20 sets a “teacher data period” and a “prediction period” based on the current time indicated by the growth score calculation request data (step T3).

  The analysis / visualization server 20 executes a time series analysis process for calculating a predicted value in the prediction period from the result of counting the number of blog writes in the teacher data period (step T4). The time series analysis process in FIG. 22 is the same process as the time series analysis process shown in FIG.

  Next, the analysis / visualization server 20 sets the “teacher data period” so that the entire day becomes the “comparison time” without changing the length of the “teacher data period” indicated by the growth score calculation request data. Try again.

  Also, the analysis / visualization server 20 resets the “prediction period” so that the first day becomes the “comparison period” without changing the length of the “prediction period” indicated by the growth score calculation request data. .

  That is, the analysis / visualization server 20 sets a teacher data period and a prediction period based on the comparison target time (step T5).

  The analysis / visualization server 20 executes time series analysis processing based on the newly set teacher data period and prediction period (step T6).

  The analysis / visualization server 20 divides the total of the predicted values within the prediction period calculated based on the current time by the total of the predicted values within the prediction period calculated based on the comparison target time, thereby obtaining a growth score. Is calculated (step T7). After step T7, the analysis / visualization server 20 ends the operation.

  With reference to FIG. 23, an example of the calculation result of the growth degree score will be described. This figure is a line graph in which the predicted values calculated based on the respective comparison target periods are plotted. The thin solid line in the figure plots a part of the total result of the teacher data period based on the present. The thick solid line is a plot of the predicted value within the prediction period based on the current time. The alternate long and short dash line plots the predicted values within the prediction period with the comparison target time as a reference. The comparison target time is set one week ago.

  As shown in FIG. 23, the predicted value calculated on the basis of the present is higher than the predicted value calculated on the basis of one week ago. An increase rate of the predicted value is calculated as a growth score.

  In the present embodiment, the analysis / visualization server 20 divides the total of the prediction values within the prediction period calculated based on the current time by the total of the prediction values within the prediction period calculated based on the comparison target time. By doing so, the growth score is calculated.

  However, as long as the value indicates the degree of increase / decrease in the predicted value thereafter with respect to the predicted value at the past time point, the value calculated by another calculation method may be used as the growth score. For example, the analysis / visualization server 20 calculates the difference between the total of the predicted values within the prediction period calculated based on the current time and the total of the predicted values within the prediction period calculated based on the comparison target time as a growth score. It is good.

  In this embodiment, only one comparison target time is set, but it is needless to say that a plurality of times may be set. For example, the analysis / visualization server 20 calculates the current growth degree for one week ago and the current growth degree for one month ago when three time points one week ago and one month ago are set.

  As described above, according to the present embodiment, the difference between the past predicted value and the predicted value thereafter can be quantified as the degree of growth.

(Fourth embodiment)
A fourth embodiment of the present invention will be described. The time series analysis system of this embodiment is different from the time series analysis system of the second embodiment in that the effect of an event can be further predicted. In addition, about the time series analysis system of this embodiment, the detailed description of the structure similar to the structure demonstrated in FIGS. 17-20 is abbreviate | omitted, and a different point from 2nd Embodiment is demonstrated in detail.

  The user PC 10 of this embodiment accepts input of simulation request data for requesting prediction of the effect of an event in addition to time series analysis request data.

  The user sets the category, keyword, event type, event start time, and prediction period. The user PC 10 transmits simulation request data including information indicating these.

  The analysis / visualization server 20 searches the data stored in the prediction parameter DB 40 for events that belong to the same category as the category indicated by the simulation request data and are similar. Similar events are events in which some or all of the keywords match and the event type is the same. The analysis / visualization server 20 reads the component of the predicted expression of the searched event. Further, the analysis / visualization server 20 acquires the actual measurement values after the event start time of the searched event from the blog data of the blog DB 30.

  The analysis / visualization server 20 transmits data for displaying the predicted value calculated from the prediction formula and the actual measurement value to the user PC 10.

  FIG. 24 is a diagram showing information included in the simulation request data. Referring to the figure, the simulation request data includes information indicating “category”, “keyword”, “event type”, “event start time”, and “prediction period”. The contents of “category”, “keyword”, “event type”, “event start time”, and “prediction period” are the same as those in the second embodiment.

  FIG. 25 is a flowchart showing event effect prediction processing executed by the analysis / visualization server 20 of this embodiment. The event effect prediction process starts when the analysis / visualization server 20 receives simulation request data.

  The analysis / visualization server 20 searches for similar events that belong to the same category as the category indicated by the simulation request data. The analysis / visualization server 20 reads the component of the predicted expression of the searched event. When a plurality of events are searched, the average value of the components of those events is calculated (step T11).

  The analysis / visualization server 20 calculates a predicted value in the prediction period indicated by the simulation request data using a prediction formula obtained from the acquired component (step T12).

  The analysis / visualization server 20 reads the blog data of the searched event from the blog DB 30, and totals the measured values of the number of blog writes after the start time of the event. When a plurality of events are searched, an average value of actually measured values of those events is calculated. The analysis / visualization server 20 totals the period from the event start time until the period of the same length as the prediction period indicated by the simulation request data elapses.

  The analysis / visualization server 20 transmits data for displaying the predicted value and the actual measurement value to the user PC 10 (step T13). After step T13, the analysis / visualization server 20 ends the event effect prediction process.

  As described above, the analysis / visualization server 20 obtains the predicted value and the actual measurement value of the past similar event, and therefore can simulate the effect when the event is executed.

1 Time-series data analysis system 10 User PC
20 Analysis / Visualization Server 30 Blog DB
40 Prediction parameter DB
201 Blog extraction / aggregation unit 202 Trend component learning unit 203 Periodic component learning unit 204 Analysis result creation unit 2011 Measurement unit 2012 Singularity detection / correction unit

Claims (27)

A function that represents an increase / decrease characteristic peculiar to Web-related data, a function whose parameter is undetermined is defined in advance as a trend function, and the parameter of the trend function is based on the time-series values of Web-related data aggregated for a specific matter Trend parameter calculation means for calculating as a trend parameter,
Based on the time series value representing the periodicity peculiar to Web-related data, defining a function whose parameter is undetermined in advance as a periodic function, and removing the trend function to which the trend parameter is applied, A periodic parameter calculating means for calculating the parameter of
By adding the trend function to which the trend parameter calculated by the trend parameter calculation means is applied and the periodic function to which the periodic parameter calculated by the periodic parameter calculation means is added, the particular function A prediction formula calculation means for calculating a prediction formula for predicting the transition of Web-related data;
Among the time series values, there is a correction unit that detects outliers that are suddenly increased or decreased, and corrects the outliers,
When the outlier is detected by the correcting unit, the trend parameter calculating unit calculates the trend parameter from the time series value in which the outlier is corrected by the correcting unit,
In the plurality of burst periods, each of which is a period when the time series values are aggregated, the correction means includes mintime as the first time in each burst period, and the last time in the burst period. Maxtime, α is a positive real number, t is a time when time series values are tabulated, y _{t is the} time series value at time t , m is a natural number, and BURST is a burst value in the burst period.

Calculating a burst value of each burst period by the following formula, extracting a predetermined number of burst periods in descending order of the absolute value of the burst value, and correcting the extracted time series value in the burst period as an outlier Analysis equipment.   The time series analysis apparatus according to claim 1, wherein the periodic parameter calculation unit defines a plurality of periodic functions having different periods and calculates a plurality of periodic parameters corresponding to the respective periodic functions. The time series analysis apparatus according to claim 1 , wherein the correction unit calculates each burst value by changing a length of the burst period from a minimum value to a maximum value. The time series analysis apparatus according to claim 3 , wherein the correcting unit preferentially extracts burst periods having a high absolute value of burst values in order, and does not extract burst periods that overlap with the already extracted burst periods. When a prediction time for predicting the value of the time series value is input, the time series value at the prediction time is converted into a prediction value by substituting the prediction time into the prediction formula calculated by the prediction formula calculation means. The time series analysis apparatus according to claim 1 , further comprising: a predicted value calculation unit that calculates as follows. A function that represents an increase / decrease characteristic peculiar to Web-related data, a function whose parameter is undetermined is defined in advance as a trend function, and the parameter of the trend function is based on the time-series values of Web-related data aggregated for a specific matter Trend parameter calculation means for calculating as a trend parameter,
Based on the time series value representing the periodicity peculiar to Web-related data, defining a function whose parameter is undetermined in advance as a periodic function, and removing the trend function to which the trend parameter is applied, A periodic parameter calculating means for calculating the parameter of
By adding the trend function to which the trend parameter calculated by the trend parameter calculation means is applied and the periodic function to which the periodic parameter calculated by the periodic parameter calculation means is added, the particular function A prediction formula calculation means for calculating a prediction formula for predicting the transition of Web-related data;
When a prediction time for predicting the value of the time series value is input, the time series value at the prediction time is converted into a prediction value by substituting the prediction time into the prediction formula calculated by the prediction formula calculation means. Predicted value calculation means for calculating as
The degree of influence of the event is calculated on the basis of the time-series values collected from the time when the event is held until the predetermined period elapses and the predicted value calculated within the period by the predicted value calculation means. A time-series analysis apparatus having an influence degree calculation means. The time series analysis apparatus according to claim 6 , wherein the periodic parameter calculation unit defines a plurality of periodic functions having different periods and calculates a plurality of periodic parameters corresponding to the respective periodic functions. Among the time series values, it further includes a correcting unit that detects an outlier that is a value that suddenly increases or decreases, and corrects the outlier.
8. The trend parameter calculation unit according to claim 6 , wherein, when an outlier is detected by the correction unit, the trend parameter calculation unit calculates the trend parameter from the time-series value in which the outlier is corrected by the correction unit. Time series analyzer. In the plurality of burst periods, each of which is a period when the time series values are aggregated, the correction means includes mintime as the first time in each burst period, and the last time in the burst period. Maxtime, α is a positive real number, t is a time when time series values are tabulated, y _{t is the} time series value at time t, m is a natural number, and BURST is a burst value in the burst period.

Calculates a burst value for each burst period by the formula, extracting a predetermined number of burst periods absolute value in descending order of the burst value, corrects the series values as outliers when in extracted the burst period, claim 8. The time series analysis apparatus according to 8 . The time series analyzer according to claim 9 , wherein the correction unit calculates each burst value by changing a length of the burst period from a minimum value to a maximum value. The time series analysis apparatus according to claim 10 , wherein the correction unit extracts burst periods with a high absolute value of the burst value in priority order, and does not extract burst periods that overlap with the already extracted burst periods. A function that represents an increase / decrease characteristic peculiar to Web-related data, a function whose parameter is undetermined is defined in advance as a trend function, and the parameter of the trend function is based on the time-series values of Web-related data aggregated for a specific matter Trend parameter calculation means for calculating as a trend parameter,
Based on the time series value representing the periodicity peculiar to Web-related data, defining a function whose parameter is undetermined in advance as a periodic function, and removing the trend function to which the trend parameter is applied, A periodic parameter calculating means for calculating the parameter of
By adding the trend function to which the trend parameter calculated by the trend parameter calculation means is applied and the periodic function to which the periodic parameter calculated by the periodic parameter calculation means is added, the particular function A prediction formula calculation means for calculating a prediction formula for predicting the transition of Web-related data;
When a prediction time for predicting the value of the time series value is input, the time series value at the prediction time is converted into a prediction value by substituting the prediction time into the prediction formula calculated by the prediction formula calculation means. Predictive value calculation means for calculating as
The predicted value calculation means is configured to input a first aggregation period, a second aggregation period, and a prediction time based on a time series value aggregated within the first aggregation period when a prediction period is input. A prediction value obtained by calculating a prediction formula, calculating a second prediction formula based on the time series values aggregated within the second aggregation period, and substituting the prediction timing into the first prediction formula for, we calculate the degree of the predicted value obtained by substituting the predicted time to the prediction equation of the second as growth degree, time series analysis apparatus. The time series analysis apparatus according to claim 12 , wherein the periodic parameter calculation unit defines a plurality of periodic functions having different periods and calculates a plurality of periodic parameters corresponding to the respective periodic functions. Among the time series values, it further includes a correcting unit that detects an outlier that is a value that suddenly increases or decreases, and corrects the outlier.
14. The trend parameter calculation unit according to claim 12 or 13 , wherein, when an outlier is detected by the correction unit, the trend parameter is calculated from the time series value in which the outlier is corrected by the correction unit. Time series analyzer. In the plurality of burst periods, each of which is a period when the time series values are aggregated, the correction means includes mintime as the first time in each burst period, and the last time in the burst period. Maxtime, α is a positive real number, t is a time when time series values are tabulated, y _{t is the} time series value at time t, m is a natural number, and BURST is a burst value in the burst period.

Calculates a burst value for each burst period by the formula, extracting a predetermined number of burst periods absolute value in descending order of the burst value, corrects the series values as outliers when in extracted the burst period, claim 14. A time-series analysis apparatus according to 14 . The time series analysis apparatus according to claim 15 , wherein the correction unit calculates each burst value by changing a length of the burst period from a minimum value to a maximum value. 17. The time series analysis apparatus according to claim 16 , wherein the correction unit sequentially extracts burst periods having a high absolute value of the burst value in order, and does not extract a burst period that overlaps with the already extracted burst period. The time series analysis apparatus according to claim 1 , wherein the trend function is any one of a logarithmic function, a linear function, a power function, and an exponential function. The trend parameter calculation means uses the least power method to calculate the trend parameter that minimizes the difference between the time series value and a value obtained by substituting the time when the time series value is aggregated into a trend function. The time series analysis apparatus according to claim 1 , wherein the time series analysis apparatus calculates the time series analysis apparatus. The time series analysis apparatus according to claim 1 , wherein the periodic parameter calculation unit calculates the periodic parameter by performing a Fourier analysis on the time series value after the subtraction. 21. The time series analysis apparatus according to claim 1 , wherein the periodic parameter calculation unit calculates the periodic parameter by performing an AR model analysis on the time series value after the subtraction. A trend parameter calculation means that represents a trend of increase / decrease peculiar to Web-related data and previously stores a function whose parameter is undetermined as a trend function is based on the time series value of Web-related data aggregated for a specific matter. Calculate function parameters as trend parameters,
The periodic parameter calculation means that represents the periodicity peculiar to the Web-related data and previously holds a function whose parameter is undetermined as a periodic function is based on the time series value obtained by removing the trend function to which the trend parameter is applied. Calculating the parameters of the periodic function as periodic parameters;
The prediction formula calculation means adds the trend function to which the trend parameter calculated by the trend parameter calculation means is applied and the periodic function to which the periodic parameter calculated by the periodic parameter calculation means is added. , Calculating a prediction formula for predicting the transition of the web-related data related to the specific matter ,
Detecting outliers that are suddenly increased or decreased among the time series values, correcting the outliers,
When calculating the trend parameter, if the outlier is detected, the trend parameter is calculated from the time series value in which the outlier is corrected,
In a plurality of burst periods, each period of which the time series values are aggregated being the first period, the first period in each burst period is mintime, the last period in the burst period is maxtime, and positive The real number is α, the time when the time series values are tabulated, t, the time series value at time t is y _t , the natural number is m, and the burst value in the burst period is BURST,

Calculating a burst value of each burst period by the following formula, extracting a predetermined number of burst periods in descending order of the absolute value of the burst value, and correcting the extracted time series value in the burst period as an outlier Analysis method. A function that represents an increase / decrease characteristic peculiar to Web-related data, a function whose parameter is undetermined is defined in advance as a trend function, and the parameter of the trend function is based on the time-series values of Web-related data aggregated for a specific matter As a trend parameter,
Based on the time series value representing the periodicity peculiar to Web-related data, defining a function whose parameter is undetermined in advance as a periodic function, and removing the trend function to which the trend parameter is applied, Is calculated as a periodic parameter,
By adding the trend function to which the trend parameter is applied and the periodic function to which the periodic parameter is applied, a prediction formula for predicting the transition of Web-related data related to the specific matter is calculated,
By substituting into the prediction formula the prediction time when the value of the time series value should be predicted, the time series value at the prediction time is calculated as the prediction value,
A time-series analysis method for calculating an influence degree of an event based on a time-series value totaled until a predetermined period elapses from a time when the event is held and a predicted value calculated within the period. A function that represents an increase / decrease characteristic peculiar to Web-related data, a function whose parameter is undetermined is defined in advance as a trend function, and the parameter of the trend function is based on the time-series values of Web-related data aggregated for a specific matter As a trend parameter,
Based on the time series value representing the periodicity peculiar to Web-related data, defining a function whose parameter is undetermined in advance as a periodic function, and removing the trend function to which the trend parameter is applied, Is calculated as a periodic parameter,
By adding the trend function to which the trend parameter is applied and the periodic function to which the periodic parameter is applied, a prediction formula for predicting the transition of Web-related data related to the specific matter is calculated,
A time series analysis method for calculating a time series value at a prediction time as a prediction value by substituting a prediction time at which a value of a time series value is to be predicted into the prediction formula,
Calculating a first prediction formula based on a time series value aggregated within a first aggregation period, calculating a second prediction formula based on a time series value aggregated within a second aggregation period; A time-series analysis method for calculating a degree of growth of a prediction value obtained by substituting the prediction time into the second prediction formula with respect to a prediction value obtained by substituting the prediction time into the first prediction formula . On the computer,
A function that represents an increase / decrease characteristic peculiar to Web-related data, a function whose parameter is undetermined is defined in advance as a trend function, and the parameter of the trend function is based on the time-series values of Web-related data aggregated for a specific matter Trend parameter calculation procedure to calculate as a trend parameter,
Based on the time series value representing the periodicity peculiar to Web-related data, defining a function whose parameter is undetermined in advance as a periodic function, and removing the trend function to which the trend parameter is applied, A periodic parameter calculation procedure for calculating the parameter of the above as a periodic parameter, the trend function to which the trend parameter calculated by the trend parameter calculation procedure is applied, and the periodic parameter calculated by the periodic parameter calculation procedure is applied. A prediction formula calculation procedure for calculating a prediction formula for predicting a transition of Web-related data related to the specific matter by adding a periodic function;
A correction procedure for detecting outliers that are suddenly increased or decreased among the time series values and correcting the outliers,
And execute
In the trend parameter calculation procedure, when an outlier is detected by the correction procedure, the trend parameter is calculated from the time series value in which the outlier is corrected by the correction procedure,
In the correction procedure, in a plurality of burst periods in which each period of the time series values is counted as the first period, the first period in each burst period is mintime, and the last period in the burst period Maxtime, α is a positive real number, t is a time when time series values are tabulated, y _{t is the} time series value at time t , m is a natural number, and BURST is a burst value in the burst period.

A program for calculating a burst value of each burst period by the following formula, extracting a predetermined number of burst periods in descending order of the absolute value of the burst value, and correcting the extracted time-series values within the burst period as outliers . On the computer,
A function that represents an increase / decrease characteristic peculiar to Web-related data, a function whose parameter is undetermined is defined in advance as a trend function, and the parameter of the trend function is based on the time-series values of Web-related data aggregated for a specific matter Trend parameter calculation procedure to calculate as a trend parameter,
Based on the time series value representing the periodicity peculiar to Web-related data, defining a function whose parameter is undetermined in advance as a periodic function, and removing the trend function to which the trend parameter is applied, A periodic parameter calculation procedure for calculating the parameter of as a periodic parameter,
A prediction formula for calculating a prediction formula for predicting a transition of Web-related data related to the specific matter by adding the trend function to which the trend parameter is applied and the periodic function to which the periodic parameter is applied Calculation procedure,
A prediction value calculation procedure for calculating a time series value at the prediction time as a prediction value by substituting a prediction time at which the value of the time series value is to be predicted into the prediction formula;
An influence calculation procedure for calculating the influence degree of the event based on the time-series values collected until the predetermined period elapses from the time when the event was held and the predicted value calculated within the period; The program to be executed. On the computer,
A function that represents an increase / decrease characteristic peculiar to Web-related data, a function whose parameter is undetermined is defined in advance as a trend function, and the parameter of the trend function is based on the time-series values of Web-related data aggregated for a specific matter Trend parameter calculation procedure to calculate as a trend parameter,
Based on the time series value representing the periodicity peculiar to Web-related data, defining a function whose parameter is undetermined in advance as a periodic function, and removing the trend function to which the trend parameter is applied, A periodic parameter calculation procedure for calculating the parameter of as a periodic parameter,
A prediction formula for calculating a prediction formula for predicting a transition of Web-related data related to the specific matter by adding the trend function to which the trend parameter is applied and the periodic function to which the periodic parameter is applied Calculation procedure,
By substituting the prediction time when the value of the time series value should be predicted into the prediction formula, a prediction value calculation procedure for calculating the time series value at the prediction time as a prediction value is executed,
In the predicted value calculation procedure, the first prediction formula is calculated based on the time series values aggregated within the first aggregation period, and the second prediction value is calculated based on the time series values aggregated within the second aggregation period. When the prediction formula is calculated, the degree of growth is the degree of the prediction value obtained by substituting the prediction time into the second prediction formula with respect to the prediction value obtained by substituting the prediction time into the first prediction formula. The program to calculate.

Images