JP2012253445A - Traffic prediction method, device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable a traffic prediction within a practical time with increased traffic prediction accuracy and a reduced processing scale by sampling.SOLUTION: A packet flow time and a packet size are obtained through a communication network, and stored into packet data storage means. External cause information is extracted from external cause information storage means that stores external cause information including a cause affecting a traffic volume. Packet data is extracted from the packet data storage means. A traffic volume prediction formula is obtained using given time granularity to be predicted. From the external cause prediction information storage means that stores prediction information of a future external cause, external cause data for a period to be predicted is obtained and applied to the traffic volume prediction formula to calculate the prediction traffic volume.

Description

  The present invention relates to a traffic prediction method, apparatus, and program, and in particular, by using an external factor (such as weather information and event information) that is data that cannot be obtained from the communication network, and the amount of traffic that flows on the communication network. The present invention relates to a traffic prediction method, apparatus and program for estimating future traffic volume.

  The amount of traffic flowing on the communication network varies greatly, including the factors that determine whether each user uses the communication network (calendar information (weekdays / holidays and time zones), weather, events, etc.), It is determined by various factors such as the spread of applications and is very complex.

  As a conventional technique, there is a study in which these factors are not taken into consideration and a future traffic amount is simply predicted from time series data of the traffic amount by regression analysis or the like. For example, there is a method in which time-series fluctuation data of the latest traffic is given as an input, and a traffic amount at the next time is estimated using a support vector machine (see, for example, Non-Patent Document 1). The input of this method is only the traffic amount, but the above-described factors that change the traffic amount are indirectly included.

  On the other hand, the future traffic volume is predicted by determining in which range the traffic fluctuates from time-series traffic data using the concept of a steady state called an attractor. At this time, as data for obtaining the attractor, the data to be given is determined based on the calendar information (for example, see Patent Document 1).

  In addition, calendar information and weather information are input as input, how these information affect the traffic volume, and the traffic volume is estimated by multiple regression analysis (for example, non-patent Reference 2).

JP-A-11-96135

Bermolen, P. et al. "Support vector regression for link load prediction" in proceedings of the Telecommunication Networking Workshop on QoS in Multiservice IP Networks, 2008, pp268-273, 2008. Kazuaki Akinaga, Shigeru Kaneda, Junteru Shinagawa, Akira Miura, "Proposal of Traffic Prediction Method Using Call Characteristics and Surrounding Environment Characteristics" IEICE Technical Report, NS2005-6, pp.21-24, April 2005.

  Many factors that determine whether or not a user uses a communication network are non-linear (for example, it can be easily imagined that the amount of increase in traffic when the temperature rises once varies with temperature). In addition, each element has a complex influence. For example, in the case of “the amount of traffic increases on a rainy day at night”, the case where two elements occur at the same time is more likely than when each (“night” and “rain”) occurs independently. The increase in traffic volume increases. As described above, there is a problem in that the prediction accuracy cannot be improved unless a plurality of non-linear elements are considered at the same time. Therefore, since multiple regression analysis such as that used in many prediction methods including the method of Non-Patent Document 2 assumes linearity, accurate prediction is difficult.

  The present invention has been made in view of the above points, and provides a traffic prediction method, apparatus, and program capable of improving the accuracy of traffic prediction, reducing the processing scale by sampling, and enabling traffic prediction within a realistic time. The purpose is to provide.

In order to solve the above problems, the present invention (Claim 1) is a traffic prediction method for estimating a future traffic volume,
A traffic data collecting means for acquiring the time and the size of the packet that flowed through the communication network, and storing the packet data in the packet data storage means;
The learning means extracts the external factor information from the external factor information storage means storing the external factor information including the factors affecting the traffic volume, extracts the packet data from the packet data storage means, and A learning step for obtaining a traffic volume prediction formula using a time granularity to be predicted;
The prediction means obtains the data of the external factor in the period to be predicted from the external factor prediction information storage means storing the prediction information of the external factor in the future, and applies it to the traffic amount prediction formula, thereby predicting the traffic volume And a prediction step of calculating.

In the present invention (Claim 2), in the learning step,
For the packet data, a feature amount calculation step using a traffic amount determined in units of the time granularity and a total number of external factors determined from the external factor information as a feature vector;
A sample reduction step of applying a K-means method to the feature vector to reduce the number of samples;
Applying a kernel regression analysis to the sample reduced in the sample reduction step by using the external factor information and the traffic volume to create a traffic prediction formula to obtain the traffic volume prediction formula.

  The present invention proposes a method for predicting the traffic volume with high accuracy. By using external factors such as weather and calendar information, the traffic volume from several days to several weeks ahead is predicted. If the amount of traffic can be predicted, the possibility of quality degradation due to congestion in the communication network can be predicted. At that time, the influence of congestion can be suppressed in advance by taking measures such as adding traffic engineering technology and servers. In addition, it will contribute to future facility design plans by clarifying under what conditions traffic will increase.

It is a block diagram of the traffic prediction apparatus in one Embodiment of this invention. It is a block diagram of the traffic data collection part in one Embodiment of this invention. It is a block diagram of the learning part of the traffic estimation part in one Embodiment of this invention. It is a flowchart of the feature-value calculation part of the learning part in one Embodiment of this invention. It is an image of traffic volume totaling with the designated granularity in one embodiment of the present invention. It is an image of a dummy variable in one embodiment of the present invention. It is a flowchart of the sample number reduction part of the learning part in one Embodiment of this invention. It is a flowchart of the prediction formula preparation part by the kernel regression of the learning part in one Embodiment of this invention. It is a block diagram of the prediction part of the traffic prediction part in one Embodiment of this invention. It is a mode of distribution in the feature space in one embodiment of the present invention.

  Embodiments of the present invention will be described below with reference to the drawings.

  In the present invention, highly accurate traffic prediction is performed by combining the following three factors.

(1) Input data for prediction:
Improve the accuracy of traffic prediction by setting factors that affect traffic volume, such as weather information, calendar information, and event information, from inside and outside the network, as well as traffic volume information obtained within the communication network in the input data be able to. In the traffic prediction device to be described later, this will be described as being stored in the external factor DB 101.

(2) Reduction of processing scale by sampling:
By performing prediction using only representative points from the obtained plurality of sample points, it is possible to cope with nonlinear processing that requires a large amount of calculation as shown in (3). Specifically, “Reference 1: Masanori Ichino, Akira Sakano, Naohisa Komatsu,“ A method for reducing the throughput of the nuclear nonlinear mutual subspace method using clustering, ”IEICE Transactions D, J90-D, no. 8, pp.2168-2181, (2007.8) ”, it is possible to significantly reduce the amount of processing in other methods using the kernel method. This method is applied to the sample number reduction unit 132 of the learning unit 130 described later.

(3) Traffic prediction by kernel regression:
High-dimensional mapping is performed by the kernel method in order to formulate prediction formulas considering the nonlinearity of traffic and the correlation of multiple elements. By performing linear regression such as multiple regression analysis on a high-dimensional space and returning to the original dimension, nonlinear processing becomes possible. Although processing in a high-dimensional space is very computationally intensive, prediction within a realistic time becomes possible by performing the sample reduction shown in (2). This process is applied to a prediction formula creation unit 133 by kernel regression of the learning unit 130 described later.

  This will be specifically described below.

  In the following embodiment, description will be made assuming that Internet traffic prediction is realized from weather information, calendar information, and traffic volume. In the following, weather information and calendar information are referred to as external factors.

  FIG. 1 shows the configuration of a traffic prediction apparatus according to an embodiment of the present invention.

  The traffic prediction apparatus shown in FIG. 1 includes a traffic data collection unit 110, a traffic amount prediction unit 120, a predicted traffic amount display unit 150, an external factor DB 101, and a future external factor prediction DB 103.

  The traffic data collection unit 110 has a packet data DB 102 and stores traffic data acquired via a communication network. The external factor DB 101 is created by, for example, acquiring weather information from the Japan Meteorological Agency or the like.

  Below, each component of a traffic prediction apparatus is demonstrated.

・ Traffic data acquisition unit:
The traffic data acquisition unit 110 is a functional unit for measuring in advance the amount of traffic flowing to a location to be predicted in order to learn the relationship between the traffic amount and external factors. A block diagram of the traffic data acquisition unit 110 is shown in FIG. In order to acquire the traffic volume, the packet data acquisition unit 112 records the packet flow time and the packet size for each packet (data division unit) flowing through the communication network (Internet) 111, and The data is stored in the data DB 102 sequentially.

・ Traffic prediction part:
The traffic prediction unit 120 is divided into a learning unit 130 that designs a formula for predicting the traffic volume from existing data, and a prediction unit 140 that actually predicts the future traffic volume using the prediction formula.

First, the learning unit 130 will be described. A block diagram of the learning unit 130 is shown in FIG. The learning unit 130 includes a feature amount calculation unit 131, a sample number reduction unit 132, and a prediction formula creation unit 133. By giving a time granularity to be predicted, the learning unit 130 extracts data from the packet data DB 101 and the external factor DB 102, and traffic volume. Design prediction formulas.
The feature amount calculation unit 131 calculates a feature amount from traffic data. The feature amount calculation unit 131 performs aggregation for predicting the traffic amount with the prediction granularity designated from the external factor DB 101 and the packet data DB 102. The operation is shown in FIG.

  Step 101) First, the feature quantity calculation unit 131 obtains data for the past year to be used for prediction from the external factor DB 101 and the packet data DB 102 (in this example, it is assumed to be one year, but the calculation quantity is actually calculated). -It is necessary to set an appropriate period from the accuracy).

  Step 102) Next, the traffic volume and external factors are tabulated in the unit of the predicted granularity specifying the data for one year. For example, as shown in FIGS. 5A, 5B, and 5C, when the daily traffic load is predicted, the traffic volume is summed up every day. For external factors that can be expressed in quantity, the one corresponding to the extraction interval is calculated. External factors that cannot be expressed in quantity are expressed by dummy variables as shown in FIG.

  Step 103) Thereby, the obtained traffic amount and external factors are output as feature vectors. The total amount N of traffic volume and external factors is the number of dimensions of each vector. The number of vectors is calculated by dividing the year by the prediction granularity (365 N-dimensional vectors are calculated when the traffic volume in a day unit is predicted).

  Next, the sample number reduction unit 132 of the learning unit 130 uses the K-average amount in order to reduce feature vectors (samples). Thereby, prediction within a realistic time becomes possible.

  The operation of the sample number reduction unit 132 is shown in FIG.

Let K be the number of cluster centers (this number will eventually be the number of samples reduced). Each sample is classified into one of K classes. The number of samples classified into each class C _j is N _j (j = 1,..., K), and the cluster center of each class C _j (j = 1,.

とする。以下の処理を行うことで、与えたサンプルを代表するのに適切なK個のクラスタ中心を得ることが出来る。ただし、

And By performing the following processing, K cluster centers suitable for representing a given sample can be obtained. However,

はi番目のサンプルである．
ステップ２０１） m個のサンプル集合

Is the i th sample.
Step 201) m sample sets

からj個のクラスタ中心をランダムに選ぶ。

Choose j cluster centers at random.

  Step 202) m samples

それぞれに対して、

For each

（あるサンプルデータとクラス中心の距離の2乗)が最小になるクラスタＣ_jへクラス分けする。

Classification is made into clusters C _j where (some sample data and the square of the distance between class centers) are minimized.

  Step 203) Indicator of whether the entire sample data is divided into appropriate classes (= the square of the distance from the class center of all sample data. The smaller the better))

を計算する。

Calculate

  Step 204) If the square value of the sum of the distances obtained in Step 203 is equal to or smaller than the threshold ε, it is determined that an appropriate cluster center is obtained. To get to step 205.

  Step 205) Cluster center for each cluster

を再計算し、クラスタ中心を更新してステップ２０３に移行する。

Is recalculated, the cluster center is updated, and the process proceeds to step 203.

  Step 206) Cluster center obtained in Step 203

を学習データとしてメモリ(図示せず)に保存する。

Are stored in a memory (not shown) as learning data.

  Step 207) An N-dimensional vector at the center of the cluster, which is sample-reduced to K samples specified by the input from the user, is output.

  When the above process is completed, K points representing m data

が求められるのでこれを削減されたサンプルとして利用する．
次に、学習部１３０のカーネル回帰によるトラヒック予測式作成部１３３について説明する。

Is used as a reduced sample.
Next, the traffic prediction formula creation unit 133 by kernel regression of the learning unit 130 will be described.

  In step 102, using the traffic volume and external factors measured in advance, a regression equation representing the relationship between the traffic volume (target variable) and the external factors (explanatory variable) is obtained.

  In a space (hereinafter referred to as a “feature space”) in which the traffic amount and external factors are taken on each axis (hereinafter referred to as “feature space”), the traffic distribution may be nonlinear. This is because there are many states that can be taken in the feature space depending on external factors, and the degree of freedom for increasing the number of external factors used for prediction also increases. An example of the distribution in the feature space is shown in FIG.

  Therefore, the present invention proposes a traffic prediction method combined with kernel regression analysis. Kernel regression analysis is a technique that enables nonlinear regression analysis in the original dimensional space by mapping each point to a higher-order space using a nonlinear function and performing regression analysis.

  FIG. 8 shows a flow of the traffic prediction formula creation unit 133 by kernel regression.

  Step 301) The feature vector sampled by the sample number reduction unit 132

を入力する。

Enter.

  Step 302) Store the input sample-reduced feature vector in a memory (not shown).

  Step 303) Select a kernel function k to be used.

  Step 304) Set the value of the kernel parameter in the kernel function.

  Step 305) A prediction equation is generated and output using the sample-reduced feature vector and kernel parameter held in Step 302.

  Specifically, the regression model can be expressed as follows.

ただし、

However,

は

Is

における予測値、

Predicted value at,

はカーネル関数、

Is the kernel function,

は事前に測定した各サンプルのうち、それぞれのクラスタに所属しているトラヒック量の平均値である。

Is an average value of traffic volume belonging to each cluster among samples measured in advance.

  Next, the prediction unit 140 of the traffic prediction unit 120 will be described.

  FIG. 9 shows the configuration of the prediction unit 140.

  The prediction unit 140 includes a feature amount calculation unit 141 and a prediction processing unit 142.

  The feature amount calculation unit 141 acquires data of external factors for a period to be predicted from the conventional external factor prediction DB 103. After acquisition, the feature quantity corresponding to the extraction section is calculated for those that can be expressed in quantity. External factors that cannot be expressed in quantity are expressed by dummy variables as shown in FIG. Then, the prediction processing unit 142 calculates the predicted value by substituting the feature amount into the prediction formula (1) obtained by the learning unit 130.

  In addition, the operation of the traffic data collection unit 110 and the traffic prediction unit 120 in the above embodiment is constructed as a program and installed in a computer used as a traffic prediction device to be executed or distributed via a network. Is also possible.

  The present invention is not limited to the above-described embodiments, and various modifications and applications can be made within the scope of the claims.

100 Traffic prediction device 101 External factor DB
102 packet data DB
103 Future External Factor Prediction DB
DESCRIPTION OF SYMBOLS 110 Traffic data collection part 120 Traffic prediction part 130 Learning part 131 Feature-value calculation part 132 Sample number reduction part 133 Prediction formula preparation part 140 by kernel regression 140 Prediction part 141 Feature-value calculation part 142 Prediction processing part 150 Predictive traffic amount display part

Claims (5)

A traffic prediction method for estimating future traffic volume,
A traffic data collecting means for acquiring the time and the size of the packet that flowed through the communication network, and storing the packet data in the packet data storage means;
The learning means extracts the external factor information from the external factor information storage means storing the external factor information including the factors affecting the traffic volume, extracts the packet data from the packet data storage means, and A learning step for obtaining a traffic volume prediction formula using a time granularity to be predicted;
The prediction means obtains the data of the external factor in the period to be predicted from the external factor prediction information storage means storing the prediction information of the external factor in the future, and applies it to the traffic amount prediction formula, thereby predicting the traffic volume A prediction step of calculating
The traffic prediction method characterized by performing. In the learning step,
For the packet data, a feature amount calculation step using a traffic amount determined in units of the time granularity and a total number of external factors determined from the external factor information as a feature vector;
A sample reduction step of applying a K-means method to the feature vector to reduce the number of samples;
Applying a kernel regression analysis to the sample reduced in the sample reduction step using the external factor information and the traffic volume, a traffic prediction formula creating step for obtaining the traffic volume prediction formula;
The traffic prediction method according to claim 1, comprising: A traffic prediction device for estimating future traffic volume,
External factor information storage means for storing external factor information including factors that affect traffic volume;
An external factor prediction information storage means storing prediction information of future external factors;
A traffic data collecting means for acquiring a time when the packet has flowed through the communication network and a size of the packet and storing the packet data in the packet data storage means;
Learning means for extracting the external factor information from the external factor information storage means, extracting the packet data from the packet data storage means, and obtaining a traffic amount prediction formula using a given time granularity to be predicted;
Predicting means for obtaining data of external factors for a period to be predicted from the external factor prediction information storage means, and calculating a predicted traffic amount by applying to the traffic amount prediction formula;
A traffic prediction apparatus comprising: The learning means includes
For the packet data, a feature amount calculation means having a feature vector that is a traffic amount determined in units of the time granularity and a total number of external factors determined from the external factor information;
Sample reduction means for reducing the number of samples by applying a K-means method to the feature vector;
Traffic prediction formula creation means for obtaining the traffic amount prediction formula by applying kernel regression analysis to the sample reduced by the sample reduction means using the external factor information and the traffic volume;
The traffic prediction apparatus according to claim 3, comprising: Computer
The traffic prediction program for functioning as each means of the traffic prediction apparatus of Claim 3 or 4.

Images