JP4780670B2 - Traffic analysis model construction method, apparatus, construction program, and storage medium thereof - Google Patents

Description

  The present invention relates to a traffic analysis model construction method, apparatus, and construction program for analyzing traffic characteristics of an arbitrary user based on user attributes and actual traffic of an existing user accommodated in a network, based on the user attributes. The present invention relates to a storage medium.

  In accommodation design for accommodating a new user in an existing network or facility design for constructing a new network, it is indispensable to accurately analyze the traffic characteristics of the user newly accommodated in the network.

  Non-Patent Document 1 discloses a technique for classifying (categorizing) daily traffic fluctuations (resolution: 30 minutes) of each existing user according to the characteristics of the fluctuations using a clustering analysis technique.

  Non-Patent Document 2 discloses a technique for predicting a category to which the traffic belongs from traffic fluctuation data using the correspondence relationship between the category obtained by the classification method of Document 1 and individual traffic fluctuation as teacher data. ing.

  In Non-Patent Document 3, the traffic characteristics of dial-up users and ADSL line users are analyzed focusing on the difference of access lines, and characteristics that are hardly affected by the access line speed (TCP session duration, TCP connection rate, etc.) And the characteristics (transfer traffic volume) that are affected.

  In Patent Document 1, time series data of communication traffic statistics measured every day, and calendar information indicating calendar attributes indicating communication traffic measurement dates including month information and day information are input to a neural network, A technique for predicting communication traffic up to k days ahead of the measured communication traffic is disclosed.

  However, all of the techniques disclosed in the above-mentioned documents are techniques for classifying or predicting the traffic characteristics of users based on the actual data. However, as with new users newly accommodated in the network, It was not possible to predict the traffic characteristics of users whose traffic was completely unknown.

In response to such technical problems, the inventors of the present invention analyze the traffic characteristics of an arbitrary user based on the user attributes of the existing users accommodated in the network and the actual traffic based on the user attributes. A traffic analysis model construction method, apparatus, construction program, and storage medium thereof were invented, and a patent application (Patent Document 2) was filed.
MR de Oliveira et al, "Cluster Analysis of Internet Users Based on Hourly Traffic Utilization," In Proc. Of the First International Working Conference on Performance Modeling and Evaluation of Heterogenious Networks (HET-NETs'03), July 2003. A. Nogueira et al, "Using Neural Networks to Classify Internet Users," in Proc. Of Advanced Industrial Conference on Telecommunications / Service Assurance with Partial and Intermittent Resources Conference / E-Learning on Telecommunicatrions Workshop (AICT / SAPIR / ELETE 2005), 2005. N. Vicari et al, "The Dependence of Internet User Traffic Characteristics on Access Speed," in Proc. Of the 25th Local Computer Networks (LCN) Conference, pp.670-677, Tampa, FL, USA, Nov. 2000. JP 2004-23114 A Japanese Patent Application No. 2007-21117

  In the traffic analysis model described above, measured traffic data is divided into unit periods defined by calendar information, traffic characteristics are obtained for each divided unit, and this is associated with user information and calendar information and handled as one record. .

  When the analysis result in units of records is used as teacher data, the total number of records, that is, the number of teacher data is (measurement port number) × ((measurement data period) / (unit period)). For example, when the number of measurement ports is 5000, the measurement data period is 100 days, and the unit period is 1 day, the number of teacher data is 500,000. Since the time required for building a traffic analysis model generally increases in proportion to the number of teacher data, the above-described method for building a traffic analysis model requires a long time calculation when a large number of long-term data is used. There is a risk of need.

  The object of the present invention is to solve the above-mentioned problems of the prior art, based on the user attributes of the existing users already accommodated in the network and the actual traffic, and based on the user attributes, It is an object of the present invention to provide a traffic analysis model construction method, apparatus, construction program, and storage medium thereof that can be analyzed in a shorter time than technology.

  According to the investigation by the inventors of the present invention, the traffic characteristics of each user are attribute information specific to each user (hereinafter referred to as “contract speed, contract service type, contract area, etc.”), gender, age, etc. If the user attributes are different, the traffic characteristics are greatly different. On the other hand, if the user attributes are similar, the traffic characteristics are often similar.

  10, 11, and 12 are diagrams showing the relationship between user attributes and traffic loads. Here, the ratio of low-load users with a small amount of traffic and other non-low-load users is shown for each user attribute. Show.

  FIG. 10 is a diagram illustrating a result of obtaining a ratio between a low load user and a non-low load user for each contract speed. FIG. 11 is a diagram in which the ratio between the low-load user and the non-low-load user is obtained for each contract area of each user. FIG. 12 is a diagram in which the ratio between the low-load user and the non-low-load user is obtained for each contract service type.

  In any of the figures, it can be seen that there is a significant relationship between the characteristics of actual traffic (traffic load) and user attributes. Therefore, it is understood that the traffic characteristics of an arbitrary user can be analyzed with high accuracy by obtaining the correspondence between the user attribute of the existing user and the actual traffic and applying the user attribute of the arbitrary user to this correspondence.

  Similarly, FIG. 13 is a diagram showing a relationship between a contract service type, which is one of user attributes, and a multidimensional vector which is one of traffic characteristics and represents characteristics of traffic fluctuation. Multidimensional vectors are categorized into "day type" where traffic is concentrated in the daytime and "night type" where traffic is concentrated at night. According to the figure, it is shown that whether the traffic characteristic is “day type” or “night type” depends on the contract service type.

  Further, FIG. 14 is a diagram showing the relationship between a day of the week, which is one of the calendar attributes described later, and a multidimensional vector representing the characteristics of traffic fluctuations. The “day type” in which the traffic characteristics are concentrated in the daytime and the nighttime. It is shown that which of the “night type” that concentrates on is dependent on the calendar attribute.

  The present invention has been made on the basis of the above knowledge, and based on the user attributes and actual traffic of existing users already accommodated in the network, the traffic characteristics of an arbitrary user are analyzed based on the user attributes. In the analysis model construction device, the correspondence between the user attribute of the existing user, the means of storing the actual traffic of the existing user, and the user attribute of the existing user and the actual traffic is specified by a predetermined calendar attribute. The traffic analysis means that analyzes each unit period and the correspondences that match the user attributes and calendar attributes are aggregated and segmented, and a traffic representative value that represents the traffic characteristics of each aggregated correspondence is associated with each segment Segmentation means and user and calendar attributes for each segment Based on the teacher data for traffic analysis, a teacher data table for storing traffic representative values as teacher data for traffic analysis, and a traffic analysis model for analyzing traffic characteristics for each unit period based on user attributes and calendar attributes And a traffic analysis model construction means constructed in this way.

  According to the present invention, correspondence relationships between user attributes, calendar attributes, and traffic characteristics of existing users are obtained, and correspondence relationships that match user attributes and calendar attributes are aggregated and segmented. A traffic analysis model is constructed by associating a traffic representative value representing the traffic characteristics with each segment and using the user attribute, calendar attribute, and traffic representative value of each segment as teacher data. Therefore, for a given user, if a unit period for which the traffic characteristic is to be analyzed is specified as a calendar attribute together with the user attribute, the traffic characteristic corresponding to the desired calendar attribute of the user can be analyzed in a short time.

  DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the best embodiment of the present invention will be described in detail with reference to the drawings. FIG. 1 is a diagram showing an example of a network to which an analysis result by a traffic analysis system according to the present invention is applied, and a plurality of GC stations (city exchange stations) are connected to a core network via an NC (network center). A number of user terminals are connected to each GC station via a switch / router (S / R), which are connected by a ring network (GC ring).

  FIG. 2 is a block diagram showing the configuration of the main part of the traffic analysis system according to the present invention. Here, the configuration unnecessary for the description of the present invention is omitted.

  The traffic analysis model construction unit 1 associates the user attributes of the existing users accommodated in the network with the traffic characteristics obtained by analyzing the actual traffic of the existing users for each unit period specified by a predetermined calendar attribute. Based on the above, a traffic analysis model that predicts traffic characteristics corresponding to an arbitrary calendar attribute of an arbitrary user from only the user attribute and the calendar attribute is constructed.

  For example, if “all day of the month” is designated as the calendar attribute, the traffic characteristics of “1 day”, the traffic characteristics of “2 days”,..., The traffic characteristics of “31 days” are obtained for each existing user. Each traffic characteristic is associated with a user attribute of the existing user. As the traffic characteristics, a traffic average value for one day (24 hours), a traffic standard deviation, and the like are obtained.

  Similarly, if “all month within the year” is specified as the calendar attribute, “January” traffic characteristics, “February” traffic characteristics, and “December” traffic characteristics are obtained for each existing user. Each traffic characteristic is associated with a user attribute of the existing user. As the traffic characteristics, an average value of traffic for one month, a traffic standard deviation, and the like are obtained.

  Similarly, if “daytime” is designated as the calendar attribute, the traffic characteristics of “0 hour range”, the traffic characteristic of “1 hour level”,... Each is obtained, and each traffic characteristic is associated with the user attribute of the existing user. As the traffic characteristics, an hourly traffic average value, a traffic standard deviation, and the like are obtained.

  As with the user attribute, a plurality of calendar attributes can be specified at the same time. For example, in the case of a combination of “all day in the month” and “day of the week”, “Monday of 3rd”, “25th of day” A designation such as “Friday” is also possible. In addition, in the case of a combination of “all week in the month” and “all month in the year”, “second week in January” and “first week in March” can be designated.

  However, in the present invention, in order to reduce the time required for building the traffic analysis model by reducing the total number of teacher data for building the prediction model, as will be described in detail later, the correspondence relationship in which both the user attribute and the calendar information match. They are aggregated (segmented), and a traffic representative value representative of each traffic characteristic of the aggregated correspondence is set and associated with each segment.

  The traffic prediction unit 2 uses the traffic analysis model constructed by the traffic analysis model construction unit 1 to predict the traffic characteristics of the unit period specified by the calendar attribute of a new user from only the user attribute. For example, if “25 days” is input as the calendar attribute together with the user attributes, the traffic characteristics of “25 days” of the new user corresponding to the user attributes are predicted.

  Similarly, if “25th day” and “Monday” are input as the calendar attribute together with the user attribute, the traffic characteristics of “Monday on 25th” of the new user corresponding to this user attribute are predicted.

  At this time, in the present embodiment, even when analyzing the traffic characteristics of “Monday on 25th”, it is not necessary to register the information of existing users corresponding to the combination of “25th” and “Monday”. Even if the information of the existing user corresponding to the combination is not registered, it can be analyzed based on the information of other day and day of the week.

  In the traffic analysis model construction unit 1, the existing user database (DB) 101 includes a user attribute storage unit 101 a that stores user attributes of a number of existing users, and a track record that stores track traffic of each existing user that is separately measured. In the existing user DB 101, the user attribute and the actual traffic are associated with each other on the existing user DB 101. FIG. 3 is a diagram schematically showing the configuration of the existing user DB 1, and user attributes and actual traffic are associated with each user ID that uniquely identifies each user (port number).

  In the present embodiment, various information including contract conditions such as contract speed, contract service type, contract area, and user gender and age are registered as user attributes. For example, the actual traffic is periodically transmitted over the predetermined period (for example, one month) for each upstream and downstream traffic on the port that accommodates the access line of each existing user in the switch / router (in this embodiment, every 5 minutes). ) Is time-series measurement data obtained by tapping.

  Returning to FIG. 2, the load analysis unit 102 analyzes the actual traffic of each existing user for each unit period specified by a predetermined calendar attribute, and identifies traffic having a specific load characteristic. In this embodiment, the traffic of each unit period includes “low load” in which traffic hardly flows, “temporary load” in which traffic is temporarily generated for the purpose of data backup, and “normal load” other than the above. Classify either. The traffic characteristic analysis unit 103 analyzes the correspondence between the user attributes of the existing users and the actual traffic, and extracts the correspondence for each unit period specified by a predetermined calendar attribute.

  The traffic analysis teacher data generation unit 104 segments corresponding relationships in which both the user attribute and calendar information match, sets a traffic representative value representing the traffic characteristics of each corresponding relationship in the segment, and associates each segment with the corresponding segment And a teacher data table 104b in which the correspondence relationship between the user attribute, calendar attribute, and traffic representative value of each segment is registered as teacher data for traffic analysis.

  Based on the teacher data registered in the traffic analysis teacher data table 104b, the traffic analysis model construction unit 106 analyzes a plurality of traffic characteristics corresponding to an arbitrary calendar attribute of an arbitrary user from the user attribute and the calendar attribute. Build a traffic analysis model.

In the present embodiment, the teacher data registered in the traffic analysis teacher data table 104b is applied to an appropriate analysis method such as Neural network, Naive Bayes, Decision Tree, Support Vector Machine, and the like from user attributes and calendar attributes. Build a traffic analysis model to analyze traffic characteristics.

  Here, if the Neural Network is applied to construct an analysis model, Network parameters (coupling weights) are optimized so as to minimize the error between the teacher data and the output. Similarly, when other analysis methods are applied, the parameters are optimized so that the error between the teacher data and the output is minimized.

  The load determination teacher data generation unit 105 determines a low load determination for the correspondence between user attributes, calendar attributes, and load characteristics (low load or normal load) for each unit period in which the traffic volume is classified as low load or normal load. Is registered in the teacher data table 105a as teacher data for use.

  FIG. 4 is a diagram schematically showing an example of the teacher data table 105a, in which correspondence relationships between user attributes, calendar attributes, and load characteristics (low load or normal load) are registered.

The low load discrimination model construction unit 107 applies the teacher data registered in the low load discrimination teacher data table 105a to an appropriate analysis method such as Neural network, Naive Bayes, Decision Tree, Support Vector Machine, Based on the user attribute and calendar attribute of an arbitrary user, a low load determination model is determined for determining whether or not the unit period characterized by the calendar attribute of the user has a low load.

  Returning to FIG. 2, in the traffic prediction unit 2, the low load port prediction unit 201 applies the user attribute and calendar attribute of the new user to the low load determination model constructed by the low load determination model construction unit 107, It is determined whether or not the traffic of the unit period specified by the calendar attribute of the new user is light. The determination result is temporarily stored in the storage unit 203 and then output.

  The traffic characteristic prediction unit 202 applies a new user to the traffic analysis model constructed by the traffic analysis model construction unit 106 for only a normal load unit period determined by the low load port prediction unit 201 as other than a low load. Are applied to predict the traffic characteristics of the unit period specified by the calendar attribute of the new user. The prediction result is output after being temporarily stored in the storage unit 203.

  Next, the operation of the traffic analysis model construction unit 1 will be described in detail with reference to the flowchart of FIG.

  In step S <b> 1, one of the existing users is selected as the current user of attention from the existing user DB 101, and the user attribute of the target user and the actual traffic thereof are taken into the load analysis unit 102. In step S2, the actual traffic is extracted by the load analysis unit 102 for each unit period specified by the calendar attribute. In the present embodiment, since the actual traffic is analyzed using the day and day of the week as a calendar attribute, the actual traffic is extracted every 24 hours.

  In step S3, the extracted actual traffic for 24 hours is analyzed, and an average value of the traffic volume is obtained. In step S4, the average value of the traffic volume is compared with a predetermined reference value, and based on the comparison result, it is determined whether or not the traffic of the current unit period is lightly loaded.

  In this embodiment, if the logic determination is set to “average traffic volume is 1 Mbps or less” and the average traffic volume for 24 hours is 1 Mbps or less, the current unit period is classified as low load, and the process proceeds to step S5. . In step S5, the load determination teacher data generation unit 105 stores the user attribute, calendar attribute, and load characteristic (low load) in the load determination teacher data table 105a as low load determination teacher data for the current unit period. be registered. As the calendar attribute, if the current unit period is [Monday of 1st], “1 day” as the whole day and “Monday” as the day of the week are registered as the calendar attributes.

  In step S10, it is determined whether or not the analysis of all the unit periods has been completed for the current user of interest, and if not completed yet, the process returns to step S2, and the actual traffic for 24 hours of the next day is set as the next unit period. Are extracted and the analysis is performed in the same manner as described above.

  If it is not determined as “low load” in step S4 for the current unit period, the process proceeds to step S6, and the load analysis unit 102 determines whether the current unit period is a “temporary load” in which the traffic is bursty. It is determined whether or not. If it is determined as “temporary load”, the process proceeds to step S9, and information regarding the current unit period is discarded.

  On the other hand, if it is not “temporary load”, it is classified as “normal load” and the process proceeds to step S7, and the user attribute, calendar attribute, and load characteristic (normal load) for the current unit period are determined for load determination. It is registered in the load determination teacher data table 105a as teacher data. In other words, in the present embodiment, both the correspondence between unit periods with a low load characteristic and the correspondence with unit periods with a normal load characteristic are registered as teacher data for load determination.

  In step S8, the traffic analysis teacher data generation unit 104 records the correspondence between the user attribute, calendar attribute, and traffic characteristic for the current unit period, and temporarily stores it as a correspondence record.

  As described above, the actual traffic analysis of the current user of interest proceeds, and if it is determined in step S10 that the analysis has been completed for all unit periods, the process proceeds to step S11. In step S <b> 11, it is determined whether or not the actual traffic of the existing user remains. If the actual traffic remains, the process returns to step S1, the target user is switched to the next existing user, and the above-described processes are repeated.

  When the analysis of the actual traffic is completed for all existing users or a predetermined number of existing users, the process proceeds to step S12. In step S12, as will be described in detail below, the correspondence record of the user attribute, calendar attribute, and traffic characteristic is segmented by the segmentation unit 104a of the teacher data generation unit 104, and traffic is used as the teacher data for traffic analysis. It is registered in the analysis teacher data table 104b.

  FIG. 6 is a flowchart showing the operation of the segmentation unit 104a. In step S120, all the registered correspondence records are compared with each other in their user attributes and calendar attributes, and the correspondence records whose user attributes and calendar attributes match each other are segmented. In step S121, a segment with a small number of correspondence record aggregations (frequency) is discarded.

  In step S122, with respect to the traffic characteristics of the correspondence records L of all the segments, the time resolution of the traffic data that was 5 minutes apart is converted to 1 hour, and 24 traffic data x (L, t), t = 0. , 1,2 ... 23 are obtained. This time resolution conversion is performed, for example, by obtaining an average value of measurement results (12 pieces) every 5 minutes within each hour.

  In step S123, the average value μ (L) and standard deviation σ (L) of the 24 traffic data x (L, t) are calculated for each correspondence record L. In step S124, the 24 traffic data x (L, t) are normalized by the average value μ (L) and the standard deviation σ (L) to obtain 24 normalized traffic data R [x (L, t )] Is calculated.

  In step S125, the average value μg of the average value μ (L), standard deviation σ (L), and 24 normalized traffic data R [x (L, t)] of all correspondence records L for each segment. , Σg, and 24 R [xg (t)] are calculated as traffic representative values representing the traffic characteristics of all correspondence records L aggregated in the segment. In step S126, the user attribute, calendar attribute, and traffic representative value (μg, σg and 24 R [xg (t)]) of each segment are registered in the teacher data table 104b as teacher data.

  FIG. 7 is a diagram schematically showing an example of the teacher data table 104b, in which the correspondence relationship between the user attribute, calendar attribute, each traffic representative value, and frequency is registered for each segment ID.

  Returning to FIG. 5, in step S13, the traffic analysis model construction unit 106 applies the teacher data registered in the traffic analysis teacher data table 104b to an algorithm such as a neural network, An error from the algorithm output is obtained. Then, a traffic analysis model is constructed by optimizing various parameters so that the sum of errors in all segments is minimized.

  In this embodiment, 26 items (μg, σg, R [xg (t)]) of traffic representative values are used as objective variables, and user attributes (contract speed, contract area, contract service type, etc.) and calendar attributes of existing users are set. As explanatory variables, the traffic characteristics of each user are obtained for the 26 items.

  If it is desired to preferentially obtain traffic characteristics corresponding to a segment with a high frequency with high accuracy, the error may be amplified by multiplying the error between the teacher data and the algorithm output by a coefficient corresponding to the frequency. good. In this way, even if the correspondence between user attributes, calendar attributes, and traffic characteristics is segmented and used as teacher data, the learning data is similar to the case of teaching data in units of records without segmentation as in the prior art. Therefore, the analysis time can be shortened while maintaining the same accuracy.

  In step S14, in the low load discrimination model construction unit 107, the teacher data registered in the load discrimination teacher data table 105a is applied to an algorithm such as Neural network, and the error between the teacher data and the algorithm output is minimized. A load discrimination model is constructed by optimizing various parameters so that

  FIG. 8 is a flowchart illustrating a prediction operation by the traffic prediction unit 2, in which the low load analysis model is registered in the low load port prediction unit 201, and the traffic analysis is performed in the traffic characteristic prediction unit 202. The model is registered. Here, a description will be given focusing on the operation of the traffic characteristic prediction unit 202.

  In step S21, contract conditions (contract speed and contract service type), gender, age, and the like are taken in as user attributes of the new user, and further, in this embodiment, as calendar attributes that specify the prediction period of traffic characteristics, in this embodiment. The day and day of the week are captured. That is, if it is desired to predict the traffic characteristics of “Monday of 3rd”, “3rd day” is designated as the day of the month and “Monday” is designated as the day of the week. Similarly, if the traffic characteristics of “Sunday on the 20th” are to be predicted, “20th day” is designated as the day of the month and “Sunday” is designated as the day of the week.

  The calendar attribute that can be specified here is not limited to the combination of the calendar attributes registered as the teacher data (that is, the correspondence relationship of the existing user). Even if the correspondence relationship relating to the combination of “20th day” and “Sunday” of the day of the week is unregistered, the traffic characteristics of “Sunday on 20th” can be predicted based on another correspondence relationship.

  Similarly, the user attribute of the new user need not match the combination of user attributes registered as teacher data. For example, the user attribute of the new user is “contract speed is 10 Mbps”, “Service type is A”, “contract area is 2”, etc., and even if such correspondence is not registered as teacher data, traffic characteristics can be predicted based on other correspondence.

  Furthermore, even if a combination of each user attribute and each calendar attribute of a new user is not registered as teacher data, traffic characteristics can be predicted based on other combinations.

  In step S22, the user attribute and the calendar attribute are applied to the traffic characteristic prediction model, and the average value, standard deviation, and 24 normalized traffic data of the unit period specified by the calendar attribute are included in the period of the new user. It is predicted as the traffic characteristics. In step S23, the prediction results of the 24 normalized traffic data are denormalized with the average value and the standard deviation prediction result to obtain 24 traffic data. In step S24, this traffic characteristic is output as an analysis result.

  In the above-described embodiment, the average value of the traffic characteristics of a number of correspondence records aggregated in each segment has been described as the traffic representative value of each segment. However, the present invention is not limited to this. Instead, a probability value such as the maximum value, minimum value, mode value, or 95% value of the traffic characteristics of a large number of correspondence records aggregated in each segment may be used as the traffic representative value of the segment.

  Further, in the above-described embodiment, it has been described that “calendar day of the month” and “day of the week” are designated as calendar attributes. However, the present invention is not limited to this, and the distinction between holidays and weekdays is made. It may be specified or “all week of the month (the week number in the month)” may be specified.

  At this time, if the new user's traffic is analyzed separately for holidays and weekdays, the actual traffic is analyzed for each day in the same manner as described above, and the “calendar attribute” of the teacher data is set to the holidays / weekdays. Another is registered. If the traffic of a new user is analyzed by “all week”, the performance of the actual traffic is analyzed every week, and “all week” is registered in the “calendar attribute” of the teacher data.

  FIG. 9 is a block diagram showing the configuration of the main part of another embodiment of the present invention. The same reference numerals as those described above represent the same or equivalent parts.

  In the present embodiment, a segmentation unit 105b is provided in the load determination teacher data generation unit 105 so that the load determination teacher data described with reference to FIG. 4 is segmented in the same manner as the traffic analysis teacher data. There is a feature in the point.

  The load characteristic analysis unit 108 extracts a correspondence relationship between the user attribute, calendar attribute, and load characteristic (low load or normal load) for each unit period in which the traffic amount is classified as low load or normal load. The segmenting unit 105b segments correspondence records that match both user attributes and calendar information, sets a load representative value that represents the load characteristics of each correspondence in the segment, and associates them with each segment. Are registered in the low load determination teacher data table 105a as load determination teacher data.

  The load representative value representing the load characteristic may be, for example, “low load” in a segment where the low load is higher than the normal load, and “normal load” in a segment where the normal load is higher than the low load. Alternatively, if the segment is 70% low load and 30% normal load, the representative load value may be expressed by a probability value such as “70% low load”.

  Also in this embodiment, the low-load discrimination model construction unit 107 applies the teacher data registered in the low-load discrimination teacher data table 105a to an algorithm such as Neural network, and outputs the teacher data and algorithm output for each segment. The error is required. Then, a low-load discrimination model is constructed by optimizing various parameters so that the sum of errors in all segments is minimized.

  Note that in this embodiment, if it is desired to obtain a load characteristic corresponding to a segment having a high frequency with high accuracy, the error is amplified by multiplying the error between the teacher data and the algorithm output by a coefficient corresponding to the frequency. You may do it.

It is a figure of the network to which the prediction result of the traffic characteristic using the traffic analysis model constructed | assembled with the traffic analysis model construction | assembly apparatus of this invention is applied. It is the block diagram which showed the structure of the traffic analysis system to which this invention is applied. It is the figure which showed the structure of the existing user database typically. It is the figure which showed an example of the load discrimination teacher data table. It is the flowchart which showed operation | movement of the traffic analysis model construction part. It is the flowchart which showed the segmentation procedure. It is the figure which showed an example of the traffic analysis teacher data table. It is the flowchart which showed operation | movement of the traffic estimation part. It is a block diagram of other embodiment of the traffic analysis system to which this invention is applied. It is the figure which showed the relationship between the ratio of a low load / non-low load user, and contract speed. It is the figure which showed the relationship between the ratio of a low load / non-low load user, and a contract area. It is the figure which showed the relationship between the ratio of a low load / non-low load user, and contract service classification. It is the figure which showed the relationship between a contract service classification and the multidimensional vector of a traffic characteristic. It is the figure which showed the relationship between the day of the week of a calendar characteristic and the multidimensional vector of a traffic characteristic.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Traffic analysis model construction part, 2 ... Traffic prediction part, 101 ... Existing user DB, 101a ... User attribute memory | storage part, 101b ... Performance traffic memory | storage part, 102 ... Load analysis part, 103 ... Special characteristic analysis part, 104 ... Traffic Analysis teacher data generation unit, 104a ... segmentation unit, 104b ... teacher data table, 105 ... load discrimination teacher data generation unit, 105a ... teacher data table, 105b ... segmentation unit, 106 ... traffic analysis model construction unit, 107 ... low Load discriminating model construction unit 108, load characteristic analysis unit

Claims (21)

In a traffic analysis model construction apparatus that analyzes the traffic characteristics of a new user based on the user attributes and actual traffic of existing users accommodated in the network, based on the user attributes,
Means for storing user attributes of existing users;
Means for storing the actual traffic of existing users;
A traffic characteristic analyzing means for analyzing a correspondence relationship between a user attribute of an existing user and actual traffic for each unit period specified by a predetermined calendar attribute;
Segmenting means for associating and segmenting corresponding relationships in which user attributes and calendar attributes of existing users match, and associating traffic representative values representing the existing traffic characteristics of each aggregated corresponding relationship with each segment;
A teacher data table that stores user attributes, calendar attributes and traffic representative values of each segment as teacher data for traffic analysis;
And a traffic analysis model construction means for constructing a traffic analysis model for analyzing the traffic characteristics of the new user for each unit period based on the user attributes and calendar attributes of the new user based on the teacher data for the traffic analysis. A device for building a traffic analysis model characterized by   The apparatus for constructing a traffic analysis model according to claim 1, further comprising means for discarding a segment with a small number of correspondences.   The traffic representative value includes at least one of traffic data measured at a plurality of timings within the unit period, and an average value and a standard deviation of the traffic data. Traffic analysis model construction device.   The traffic representative value is obtained by normalizing at least one of an average value and a standard deviation of traffic data measured at a plurality of timings within the unit period, and normalizing the traffic data with at least one of the average value and the standard deviation. The apparatus for constructing a traffic analysis model according to claim 3, comprising the obtained normalized traffic data.   The traffic analysis model construction means obtains an error between the teacher data and the algorithm output by applying the teacher data to a predetermined analysis algorithm for each segment, and sets various parameters so that the sum of errors of all the segments is minimized. 5. The traffic analysis according to claim 1, further comprising means for constructing an analysis model by optimization and multiplying the error by a coefficient corresponding to the number of correspondences aggregated for each segment. Model building device. Load analysis means for identifying load characteristics of the unit period,
6. The apparatus for constructing a traffic analysis model according to claim 1, wherein the segmenting means does not set a correspondence relationship between unit periods having a specific traffic load as a segmentation target.   The apparatus for constructing a traffic analysis model according to claim 6, wherein the load analysis unit identifies a unit period of low load with little traffic usage.   7. The apparatus for constructing a traffic analysis model according to claim 6, wherein the load analysis means identifies a unit period in which traffic usage is bursty. Load characteristic analysis means for analyzing the correspondence between user attributes of existing users and load characteristics for each unit period specified by a predetermined calendar attribute;
Segmentation means for associating and segmenting the correspondences that match the user attributes and calendar attributes of existing users, and associating each segment with a load representative value that represents the load characteristics of each aggregated correspondence,
A teacher data table that stores user attributes, calendar attributes, and load representative values of each segment as teacher data for load determination;
The load characteristic determination model to determine the load characteristics of the new user for each unit period based on the user attributes and calendar attribute of the new user, and a load characteristic determination model construction means for constructing on the basis of the training data for the load determination The apparatus for constructing a traffic analysis model according to claim 6, comprising:   10. The traffic analysis model according to claim 9, wherein the load representative value is any one of information indicating whether the traffic load is specific and a probability value indicating that the traffic load is specific. apparatus. The load characteristic discrimination model construction means obtains an error between the teacher data and the algorithm output by applying the teacher data to a predetermined discrimination algorithm for each segment, and sets various parameters so that the sum of errors of all segments is minimized. By constructing a discriminant model by optimizing
11. The apparatus for constructing a traffic analysis model according to claim 9, further comprising means for multiplying the error by a coefficient corresponding to the number of correspondences aggregated for each segment. In a method for constructing a traffic analysis model for analyzing the traffic characteristics of a new user based on the user attributes and actual traffic of existing users accommodated in the network, based on the user attributes,
A procedure for analyzing the correspondence between user attributes of existing users and actual traffic for each unit period specified by a predetermined calendar attribute;
A segmentation procedure for associating and segmenting corresponding relationships in which user attributes and calendar attributes of existing users match, and associating traffic representative values representing the existing traffic characteristics of each aggregated correspondence relationship with each segment,
A procedure for storing the user attribute, calendar attribute and traffic representative value of each segment as teacher data for traffic analysis,
Including a traffic analysis model construction procedure for constructing a traffic analysis model for analyzing a traffic characteristic of a new user for each unit period based on a user attribute and a calendar attribute of the new user based on the teacher data for the traffic analysis. A method for building a traffic analysis model characterized by   The method of constructing a traffic analysis model according to claim 12, further comprising a procedure of discarding a segment having a small number of correspondences.   14. The traffic representative value includes at least one of traffic data measured at a plurality of timings within the unit period, and an average value and a standard deviation of the traffic data. How to build a traffic analysis model.   The traffic representative value is obtained by normalizing at least one of an average value and a standard deviation of traffic data measured at a plurality of timings within the unit period, and normalizing the traffic data with at least one of the average value and the standard deviation. 15. The method for constructing a traffic analysis model according to claim 14, comprising the obtained normalized traffic data. In the traffic analysis model construction procedure, teacher data is applied to a predetermined analysis algorithm for each segment to obtain an error between the teacher data and the algorithm output, and various parameters are set so that the sum of errors of all segments is minimized. By optimizing, an analytical model is built,
16. The method for constructing a traffic analysis model according to claim 12, further comprising a step of multiplying the error by a coefficient corresponding to the number of correspondences aggregated for each segment. Including a procedure for identifying a load characteristic of the unit period;
17. The method for constructing a traffic analysis model according to claim 12, wherein in the segmentation procedure, a correspondence relationship between unit periods having a specific traffic load is not targeted for segmentation.   18. The method for constructing a traffic analysis model according to claim 17, wherein a unit period of low load with little traffic use is identified.   18. The method for constructing a traffic analysis model according to claim 17, wherein the traffic use identifies a burst unit period.   20. A traffic analysis model construction program for causing a computer to execute the traffic analysis model construction method according to any one of claims 12 to 19.   A storage medium storing the traffic analysis model construction program according to claim 20 in a computer-readable manner.

Images