JP6174527B2 - Moving means estimation apparatus, operation method thereof, and program - Google Patents

Description

  The present invention relates to a moving means estimating apparatus, a method and a program for estimating what moving means a user has used.

  There is a technique for estimating what moving means a user has used by using GPS history information accumulated by a device such as a GPS logger (for example, Non-Patent Document 1). In order to estimate the moving means used by the user, preprocessing is performed to divide the GPS history information into sections (segments) assumed to use different moving means and extract the sections. Then, a feature is extracted from the GPS history information for the extracted segment. The features are classified by a classifier and the moving means is determined.

  Here, the classifier for moving means determination can be generated in the framework of supervised machine learning (for example, Non-Patent Document 2) using correct data prepared in advance by hand (Non-Patent Document 2). 1).

Zheng, Y., Liu, L., Wang, L., Xie, X., “Learning transportation mode from raw gps data for geographic applications on the web”, Proc. The 17th international conference on World Wide Web, pp.247 -256, 2008. Crammer, K., Dekel, O. Keshet, J., Shalev-Shwartz, S. and Singer, Y., “Online Passive-Aggressive Algorithm”, Journal of Machine Learning, Vol.7, pp.551-585, 2006 .

  However, since the cost of creating correct data is high, there is a problem that only a limited number of correct data can be prepared. Therefore, there is a problem that the prediction accuracy of the prediction model is lowered and the determination accuracy of the movement mode is deteriorated.

  The present invention has been made in view of this problem, and provides a moving means estimation device, an operation method thereof, and a program that do not deteriorate the prediction accuracy of a prediction model even when there is a limited number of correct answer data. For the purpose.

  The movement means estimation apparatus of the present invention is a movement means estimation apparatus that estimates movement means based on GPS history information in which a user's position is recorded by GPS, and includes a spatio-temporal frequency storage unit, a training case storage unit, and an unlabeled case A storage unit and a prediction model generation function unit are provided. The spatiotemporal frequency storage unit stores a user ID for identifying a user, a mesh ID representing a certain area on the map, a time zone within a day, and a frequency at which the user has entered the certain area during the time zone. The training case storage unit stores a correct answer label that correctly represents the case of the moving means and a feature vector of the case obtained from the GPS history information. The unlabeled case storage unit stores a feature vector of a case that is not given a correct answer label. The prediction model generation function unit generates a prediction model for predicting the moving means from all cases stored in the training case storage unit, and when the number of cases stored in the unlabeled case storage unit is equal to or less than a threshold value, the prediction model This is a selection that represents the degree to which the reliability of the prediction model is improved when a correct label is assigned to a case that is not given a correct answer label when the number of cases stored in the unlabeled case storage unit is larger than a threshold. The score is obtained using the data stored in the spatio-temporal frequency storage unit, the case where the selected score is larger than the score threshold is extracted from the unlabeled case storage unit, and the extracted case is presented to the outside to request an annotation. A prediction model that adds an extracted case with the input annotation as a correct answer label to the training case storage unit and predicts the moving means using all cases stored in the training case storage unit The process of recalculating is repeated until the number of cases stored in the unlabeled example storage unit falls below the threshold value.

  The operation method of the moving means estimating apparatus of the present invention is an operating method of the moving means estimating apparatus that estimates the moving means from GPS history information in which the position of the user is recorded by GPS, and the prediction model generation function unit moves When all cases stored in the training example storage unit of the means estimation device generate a prediction model for predicting the movement means, and the number of cases stored in the unlabeled case storage unit of the movement means estimation device is less than or equal to the threshold, When the number of cases stored in the unlabeled case storage unit is larger than the threshold, the degree to which the reliability of the prediction model is improved when a correct label is assigned to a case that is not given a correct answer label. A selection score to be expressed is obtained using data stored in the spatio-temporal frequency storage unit of the moving means estimation device, and a case where the selection score is greater than the score threshold is stored as an unlabeled case storage unit The extracted case is presented to the outside, an annotation is requested, the extracted case with the input annotation as the correct answer label is added to the training case storage unit, and all the cases stored in the training case storage unit are stored. Using the cases, the process of recalculating the prediction model for predicting the moving means is repeated until the number of cases stored in the unlabeled case storage unit is equal to or less than the threshold value.

  Moreover, the program of this invention is a program for making a computer function each function part of said moving means estimation apparatus.

  According to the moving means estimation apparatus, the method, and the program of the present invention, the prediction accuracy of the prediction model can be improved even if the number of cases given labels correctly representing the moving means is small, and the moving means of the user It is possible to improve the estimation accuracy.

The figure which shows the function structural example of the moving means estimation apparatus 100 of embodiment of this invention. The figure which shows the function structural example at the time of comprising the moving means estimation apparatus 100 with a computer. The figure which shows the operation | movement flow in which the movement means estimation apparatus 100 produces | generates a prediction model. The figure which shows the operation | movement flow of the spatio-temporal frequency calculation part 10 of the movement means estimation apparatus 100. The figure which shows the example of the information memorize | stored in the GPS historical information storage part 60 of the movement means estimation apparatus 100. FIG. The figure which shows the example of the information memorize | stored in the correct annotation memory | storage part 85 of the movement means estimation apparatus 100. FIG. The figure which shows the example of the information memorize | stored in the spatiotemporal frequency memory | storage part 20 of the movement means estimation apparatus 100. FIG. The figure which shows the operation | movement flow of the segment extraction function part 70 of the movement means estimation apparatus 100. FIG. The figure which shows the example of the information memorize | stored in the segment memory | storage part 65 of the moving means estimation apparatus 100. FIG. The figure which shows the operation | movement flow of the case generation function part 60 of the moving means estimation apparatus 100. The figure which shows the example of the information memorize | stored in the training example memory | storage part 40 of the movement means estimation apparatus 100. FIG. The figure which shows the example of the information memorize | stored in the unlabeled example memory | storage part 30 of the moving means estimation apparatus 100. FIG. The figure which shows the function structural example of the prediction model production | generation function part 50 of the movement means estimation apparatus 100. FIG. The figure which shows the operation | movement flow of the prediction model production | generation function part 50. FIG. The figure which shows the example of the information memorize | stored in the prediction model memory | storage part 90 of the movement means estimation apparatus 100. FIG.

Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  In FIG. 1, the function structural example of the moving means estimation apparatus 100 of this embodiment is shown. The moving means estimation apparatus 100 estimates moving means based on GPS history information in which a user's position is recorded by GPS.

  A moving means is a walk, a car, a train, a bus, or the like used by a user to move. If the moving means can be estimated correctly, it is possible to provide an appropriate service in the device carried by the user. For example, the service to be provided can be different depending on whether the user is driving a car or traveling on a train.

  The movement means estimation device 100 includes a spatiotemporal frequency calculation unit 10, a spatiotemporal frequency storage unit 20, an unlabeled case storage unit 30, a training case storage unit 40, a prediction model generation function unit 50, a case generation function unit 60, and a segment storage unit. 65, a segment extraction function unit 70, a GPS history information storage unit 80, a correct annotation storage unit 85, and a prediction model storage unit 90. The moving means estimation apparatus 100 can be realized by a computer, for example.

  FIG. 2 shows an example of the mechanism configuration when the moving means estimating apparatus 100 is realized by a computer. In FIG. 2, each of the spatiotemporal frequency storage unit 20, the unlabeled case storage unit 30, the training case storage unit 40, the segment storage unit 65, the GPS history information storage unit 80, the correct annotation storage unit 85, and the prediction model storage unit 90. Each of the storage units is individually described for the purpose of making the correspondence with FIG. 1 easy to understand. Note that all the storage units may be consolidated in the memory 6.

  The processing performed by the functional units of the spatiotemporal frequency calculation unit 10, the prediction model generation function unit 50, and the case generation function unit 60 is realized by the CPU 5 executing a program. The input unit 7 is, for example, a keyboard. In this embodiment, an annotation is input from the input unit 7 by a user operation. The display unit 8 displays the estimated moving means together with, for example, map information.

  FIG. 3 shows an overall operation flow in which the moving means estimation apparatus 100 generates a prediction model. The overall flow of generating a prediction model will be described with reference to FIGS.

  The spatio-temporal frequency calculation unit 10 uses the log information stored in the GPS history information storage unit 80 and the information on the moving means used in a certain time interval stored in the correct annotation storage unit 85, the user ID, the map The mesh ID representing the upper certain area, the time zone within a day, and the frequency at which the user in a certain time zone enters the certain area are calculated, and the information is stored in the spatio-temporal frequency storage unit 20 (step S10).

  The case generation function unit 60 stores the log information stored in the GPS history information storage unit 80, the information on the moving means used in a certain time interval stored in the correct answer storage unit 85, and the segment storage unit 65. A training example of a correct answer label and a feature amount that correctly represents a case of the moving means for generating a prediction model is generated from the segment information representing the same section by the moving means, and is stored in the training case storage unit 40 (step S20). . Further, the case generation function unit 60 generates a feature amount of a case that is not given a correct answer label and stores it in the unlabeled case storage unit 30 (step S30).

  The prediction model generation function unit 50 generates a prediction model for predicting the moving means from all cases stored in the training case storage unit 40, and the number of cases stored in the unlabeled case storage unit 30 is equal to or less than a threshold value. When a prediction model is output and the number of cases stored in the unlabeled case storage unit 30 is larger than the threshold, the reliability of the prediction model is improved when a correct label is assigned to a case that is not given a correct answer label. A selection score representing the degree is obtained using data stored in the spatio-temporal frequency storage unit 20, a case where the selection score is larger than the score threshold is extracted from the unlabeled case storage unit 30, and the extracted case is externally provided. Annotation is presented to request an annotation, and the extracted case with the input annotation as a correct answer label is added to the training case storage unit 40, and all cases stored in the training case storage unit 40 are used. The process of recalculating the prediction model for predicting the moving means, the number of cases stored in the unlabeled example storage unit 30 is repeated to generate a predictive model until below a threshold value (step S40).

  Hereinafter, the operation of each functional component of the moving means estimating apparatus 100 will be described in detail with reference to the drawings.

[Spatial frequency calculation part]
FIG. 4 shows an operation flow of the spatiotemporal frequency calculation unit 10. The spatio-temporal frequency calculation unit 10 acquires an unprocessed user ID from the GPS history information storage unit 80 that records the user's position by GPS and sets it as u (step S11).

  FIG. 5 shows an example of GPS history information stored in the GPS history information storage unit 60. One line in FIG. 5 represents one record of GPS history information. One record includes information of a log ID, a user ID, and an acquisition time, latitude, and longitude acquired by a GPS logger. FIG. 5 shows an example in which GPS history information is stored in a cycle of 3 seconds so that the log ID time interval is 3 seconds.

  Records can be identified by log IDs. The log ID is unique across all users. If user ID = 1 is unprocessed, user ID = 1 is set to u.

  Next, the spatio-temporal frequency calculation unit 10 acquires a record corresponding to the user ID = u from the correct annotation storage unit 85 (step S12). The correct answer annotation storage unit 85 stores information that correctly indicates what moving means (correct answer label) is used in a certain time interval for each user.

  FIG. 6 illustrates information stored in the correct annotation storage unit 85. One line in FIG. 6 represents one record of correct answer information. One record includes information on the user ID, the time interval specified by the start time and the end time, and the moving means (correct answer label) used for the time interval. The correct answer storage unit 85 need not store all information corresponding to the GPS history information stored in the GPS history information storage unit 80. Further, it is assumed that a plurality of different correct answer labels are not given to a certain time section.

  The spatio-temporal frequency calculation unit 10 obtains the positioning point of the same user corresponding to the time interval of the correct annotation information using the longitude / latitude information of the GPS history information, and obtains a set of mesh IDs corresponding to the set of positioning points. Obtain and calculate frequency information for each time zone (step S13). The frequency information represents the frequency with which the user visited the mesh ID area during the time period. Note that the frequency information is not necessarily incremented when the user visits, but may be calculated using another counting method such as adding only 1 to a set of side points for a certain day.

  The mesh ID is an ID corresponding to a mesh representing a certain geographical area. For example, a standard area mesh (JISX0410 area mesh code) can be used. For the time zone, for example, a unit obtained by dividing 24 hours into four can be used. Other units such as an hourly unit or morning, noon, and evening can also be used.

  The spatiotemporal frequency calculation unit 10 stores the calculated frequency information in the spatiotemporal frequency storage unit 20 (step S14). Then, the processes from step S11 to S14 are repeated until there are no unprocessed users (YES in step S15), and the above process is terminated when there are no unprocessed users (NO in step S15).

  FIG. 7 shows an example of the spatiotemporal frequency information stored in the spatiotemporal frequency storage unit 20. The frequency with which the user with the user ID = 1 visits a certain area represented by the mesh ID in the time period obtained by dividing the day into four is stored. The example shown in FIG. 7 represents that the user with the user ID = 1 visited the area represented by the mesh ID of 5438-2343 three times in the time zone from 6 am to 12 am.

[Segment Extraction Function]
FIG. 8 shows an operation flow of the segment extraction function unit 70. The segment extraction function unit 70 extracts, from the GPS history information storage unit 80, a segment that is the minimum unit of continuous trajectories moved by the same moving means.

  The segment is composed of a plurality of continuous records (one line in FIG. 5) stored in the GPS history information storage unit 60, and the plurality of segments constitutes a unit of a continuous GPS trajectory (GPS Trajectory). One session is composed of, for example, three segments: a segment moved by bus, a segment moved by foot, and a segment moved by motorcycle.

  The segment extraction function unit 70 acquires an unprocessed user ID from the GPS history information storage unit 80 and sets it as u (step S71). Then, for example, with user ID = 1 as u, the corresponding record set is acquired from the GPS history information storage unit 80, and the record set is divided into sessions (step S72). As a session division method, a division point is set when there is a difference of the threshold θs seconds or more in the time of successive records. It is assumed that the threshold value θs is set in advance. As another method of session division, for example, the method described in Non-Patent Document 1 can be used.

  Next, the segment extraction function unit 70 divides the session into segments (step S73). As a segment dividing method, a known method described in Non-Patent Document 1 can be used.

  The segment extraction function unit 70 stores the segment extracted by dividing the session in the segment storage unit 65 (step S74). Then, the processes in steps S71 to S74 are repeated until there are no unprocessed users (YES in step S75), and the above process is terminated when there are no unprocessed users (NO in step S75).

  FIG. 9 shows an example of segment information stored in the segment storage unit 65. The segment information includes a segment ID, a session ID, a user ID, a start log ID, and an end log ID. The start log ID and the end log ID correspond to the ID number of the log ID stored in the GPS history information storage unit 80.

  In this example, one session 1 has two segment IDs = 1, starting log ID = 1 to ending log ID = 10 and segment ID = 2 starting log ID = 11 to ending log ID = 204. Composed.

[Case generation function part]
FIG. 10 shows an operation flow of the case generation function unit 60. The case generation function unit 60 receives the GPS history information, the correct answer annotation information, and the segment information, and inputs the correct label that correctly represents the case of the moving means, the feature vector of the case, and the feature vector of the case that is not given the correct answer label. Generate.

  The case generation function unit 60 selects unprocessed segment information from the segment information stored in the segment storage unit 65 and sets it as s (step S61). Next, the record corresponding to the section of the start log ID and the end log ID of the selected segment information s is read from the GPS history information storage unit 80. Then, the feature of the segment information s is extracted from the read record (step S62).

As the feature, for example, an average speed between the start log ID and the end log ID of the segment information s is used. For the feature extraction, for example, a known method described in Non-Patent Document 1 can be used. The extracted feature vector _x 1 _{M-dimensional,} x _{2, ...,} expressed in _{x M.}

  Next, the case generation function unit 60 reads a label (moving means) corresponding to the segment information s from the correct annotation storage unit 85 (step S63). This will be specifically described.

  Among the correct annotation information, the start log ID of the segment information s for the time range of the start time and end time of each record (one line in FIG. 6) having the same user ID as the user ID of the segment information s is stored in the GPS history information. The record having the maximum coverage in the time range determined by the time (start time) held by the section 80 and the time (end time) whose end log ID is held by the GPS history information storage section 80 is selected, and the movement that the record has The means is adopted as the correct answer label.

  Whether or not to give the correct annotation information moving means (correct label) to the feature obtained from the segment information s is determined by the coverage. The feature of the segment information s with the coverage ratio equal to or greater than the threshold value θc and correctly representing the case of the moving means is given a correct answer label. The feature given the label is stored in the training example storage unit 40 as a training example (step S65).

  The segment information s whose coverage is less than the threshold θc has no correct answer label, and the user ID, session ID, feature, mesh ID, and time zone of the segment information s are stored in the unlabeled case storage unit 30 as unlabeled cases. (Step S66). As the mesh ID here, a mesh ID corresponding to the start position, end position, or intermediate position of the segment can be used. Also, the time zone corresponding to the start position, the end position, or the intermediate position can be used for the time zone.

  The case generation function unit 60 repeats the processing from step S61 to S66 until there is no unprocessed segment information s (YES in step S67), and ends the above processing when there is no unprocessed segment information s ( NO of step S67).

  In FIG. 11, the example of the training example which the example generation function part 60 produced | generated is shown. A feature vector corresponding to the moving means (correct answer label) is stored. FIG. 12 shows an example of an unlabeled case generated by the case generation function unit 60. A feature vector, a mesh ID, and a time zone are stored for each user ID for each session ID.

[Prediction model generation function]
FIG. 13 shows a more specific functional configuration example of the prediction model generation function unit 50. The prediction model generation unit function 50 includes a prediction model generation unit 52, a prediction model output unit 53, a selection score calculation unit 54, an annotation request unit 55, an unlabeled case storage update unit 56, and a prediction model recalculation unit 57. It comprises.

  FIG. 14 shows an operation flow of the prediction model generation function unit 50. The operation of the prediction model generation function unit 50 will be described with reference to FIGS. 13 and 14. The prediction model generation unit 52 reads all the training cases stored in the training case storage unit 40 and sets it as a training case set T (step S51). And the prediction model C is produced | generated using the training example set T (step S52). The prediction model C is generated by multi-class classification using all cases stored in the training case storage unit 40.

  Multi-class classification is a classification method that predicts one of a plurality of classes. The prediction model is constructed using a known method described in Non-Patent Document 2, for example. Other than this method, any method can be used as long as it is a supervised machine learning method capable of multi-class classification. In the present embodiment, an example of a linear identification model that performs prediction based on an inner product of weight vectors of the same dimension as a feature vector will be described. However, an algorithm that can be used in the present embodiment is not limited to a linear identification model. When a nonlinear model is applied, a schema that matches the algorithm is used for the data structure of the storage unit that stores the prediction model.

  The prediction model output unit 53 outputs the prediction model C to the outside when the number of elements (number of cases) of the unlabeled case set U stored in the unlabeled case storage unit 30 is equal to or less than the threshold θ (YES in step S53). And it memorize | stores in the prediction model memory | storage part 90 (step S58), and operation | movement is complete | finished. When the number of cases is larger than the threshold θ, the prediction model C is output to the selection score calculation unit 54 (NO in step S53). The threshold value θ is set in advance.

  The selection score calculation unit 54 performs prediction of the moving means for each case stored in the unlabeled case storage unit 30 using the prediction model C, and a certainty factor F indicating the degree of certainty of the prediction and a spatio-temporal frequency storage. The geographic score G representing the degree of spatiotemporal rarity is obtained from the mesh ID and time zone stored in the unit 20, and the selection score S representing the degree to which the prediction model performance is improved from the certainty factor F and the geographic score G. Is calculated (step S54). For the certainty factor F, in the case of the linear identification model, for example, the maximum value of the inner product of the feature vector and the weight vector can be used. In addition, the well-known method described in the nonpatent literature 2, for example can be used. When such a prediction model that cannot calculate the certainty factor F is used, the certainty factor F is assumed to be 1.

  The geographic score G is calculated using the mesh ID of the unlabeled case and the time zone information. If the user ID of a case included in the unlabeled case set U is u, the mesh ID is m, and the time zone is t, the geographic score G is calculated by the equation (1).

ここでc(u,m,t)は、時空間頻度記憶部２０におけるユーザｕ，メッシュＩＤ_ｍ，時間帯
ｔの頻度を表す。同様にc(u,m)はユーザｕ，メッシュＩＤ_ｍの頻度、c(u,t)はユーザｕ,時間帯ｔの頻度、c(m,t)はメッシュＩＤ_ｍ, 時間帯ｔの頻度、c(u)はユーザｕの頻度、c(m)はメッシュＩＤ_ｍの頻度、c(t)は時間帯ｔの頻度を表す。また、λ_１，…，λ_７は各頻度の重み係数でありλ_１＋…＋λ_７＝１となるように予め設定されているものとする。

Here, c (u, m, t) represents the frequency of the user u, mesh ID _m , and time zone t in the spatio-temporal frequency storage unit 20. Similarly, c (u, m) is the frequency of user u and mesh ID _m , c (u, t) is the frequency of user u and time zone t, and c (m, t) is the frequency of mesh ID _m and time zone t. , C (u) represents the frequency of user u, c (m) represents the frequency of mesh ID _m , and c (t) represents the frequency of time zone t. In addition, λ ₁ ,..., Λ ₇ are weighting coefficients for each frequency, and are set in advance so that λ ₁ +... + Λ ₇ = 1.

  Using the certainty factor F and the geographic score G, the selection score S can be calculated by, for example, Expression (2).

ここでαは重みパラメータであり予め設定されているものとする。選択スコアＳが大きいほど、現在の予測モデルＣでは正確な移動手段の予測が困難であることを表す。

Here, α is a weight parameter and is set in advance. The larger the selection score S, the more difficult it is to accurately predict the moving means in the current prediction model C.

  For the selection, for example, a method of selecting only cases having a threshold value θs or more with respect to the selection score S can be used. It is assumed that the threshold value θs is set in advance. Further, a threshold value θs may be provided for the minimum value of the selection score S for the unlabeled case set U having the same user ID. The selected case set is set as an additional case set V and removed from the unlabeled case set U (formula (3)).

更新したラベルなし事例集合Ｕをラベルなし事例記憶部３０に出力する（式（４））。

The updated unlabeled case set U is output to the unlabeled case storage unit 30 (formula (4)).

アノテーション要求部５５は、選択スコアＳがスコア閾値θｓ以上のラベルなし事例記憶部３０に記憶した各事例を抽出し、当該抽出した事例を外部に提示してアノテーションを要求する（ステップＳ５５）。アノテーションを要求するとは、例えばユーザに事例を提示して、当該事例でユーザが利用した移動手段（注釈：正解ラベル）の入力を要求することである。選択された追加事例集合Ｖをユーザに提示し、アノテーションを要求する。追加事例集合Ｖの提示方法としては、セグメントに対応する事例を地図上に可視化するなどの方法を用いることができる。ユーザによって入力されたアノテーションを正解ラベルとして採用する。

The annotation request unit 55 extracts each case stored in the unlabeled case storage unit 30 with the selection score S equal to or greater than the score threshold θs, presents the extracted case to the outside, and requests an annotation (step S55). Requesting an annotation means, for example, presenting a case to the user and requesting input of moving means (annotation: correct answer label) used by the user in the case. The selected additional case set V is presented to the user and an annotation is requested. As a method for presenting the additional case set V, a method of visualizing a case corresponding to a segment on a map can be used. The annotation entered by the user is adopted as the correct answer label.

  The unlabeled case storage update unit 56 deletes the extracted case using the input annotation as a label from the unlabeled case storage unit 30 and adds the extracted case to the training case storage unit 40 (step S56). For the case to be added, a case to be used for additional learning is selected from the case set U of the unlabeled case storage unit 30 using the selection score S.

  The prediction model recalculation unit 57 recalculates the prediction model C that predicts the moving means from all cases (training case set T) stored in the training case storage unit 40 and outputs the prediction model C to the prediction model output unit 53 (step). S57). The prediction model recalculation unit 57 generates a prediction model C using the newly set training case set T. Here, a method for recalculating the prediction model C uses the same algorithm as in step S52.

  The prediction model output unit 53 repeats the processes of steps S53 to S57 until the number of cases stored in the unlabeled case storage unit 30 becomes equal to or less than the threshold θ. Therefore, the prediction model C is output after the number of cases stored in the unlabeled case storage unit 30 is less than the predetermined number. That is, the prediction model C generated from more training cases than the predetermined number is output.

  As described above, according to the moving unit estimation apparatus 100 of the present embodiment, in a situation where a small amount of correct answer labels are given, from the viewpoint of improving the performance of the prediction model from the unlabeled case set to which no labels are given. A case can be selected. The prediction model performance can be improved by requesting an annotation for the selected unlabeled case and adding a correct label. As a result, the estimation accuracy of the user's moving means can be improved.

  The actions of the present embodiment exhibiting these effects are summarized as follows. The prediction model generation function unit 50 of the moving means estimation apparatus 100 has a certainty factor F representing the degree of certainty of the current prediction model for a case not included in the current training case, and a spatio-temporal frequency storage unit 20. Use this to calculate a geographic score that represents whether the same case has already been used to build a prediction model from a spatio-temporal perspective, and use these two perspectives to indicate the degree to which the performance of the prediction model is improved Calculate the score. This selection score represents the degree to which the performance of the prediction model for estimating the moving means is improved.

  The annotation request unit 55 of the prediction model generation function unit 50 requests an annotation by presenting a user with a high degree of improvement in the performance of the prediction model. Then, the unlabeled case storage update unit 56 adds a case in which the input annotation is a correct label to the training case storage unit 40. This process of adding training cases is repeated until the number of unlabeled cases is less than a predetermined number. The prediction model recalculation unit 57 recalculates the prediction model from the training case set added in this way. Therefore, the prediction accuracy of the prediction model for estimating the moving means can be improved.

  In this embodiment, an example is described in which an annotation is requested from the user. However, if the correct answer label is stored in, for example, the action log, the correct answer label is automatically acquired with reference to the action log. You may comprise. Thus, the present invention can be modified within the scope of the gist thereof.

  When the processing unit in the above apparatus is realized by a computer, the processing contents of the functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing unit in each device is realized on the computer.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of the server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

10: Spatiotemporal frequency calculation unit 20: Spatiotemporal frequency storage unit 30: Unlabeled case storage unit 40: Training case storage unit 50: Prediction model generation function unit 52: Prediction model generation unit 53: Prediction model output unit 54: Selection score Calculation unit 55: Annotation request unit 56: Unlabeled case storage update unit 57: Prediction model recalculation unit 60: Case generation function unit 65: Segment storage unit 70: Segment extraction function unit 80: GPS history information storage unit 85: Correct answer annotation Storage unit 100: moving means estimation device

Claims (5)

A moving means estimating device for estimating a moving means based on GPS history information in which a user's position is recorded by GPS,
Spatio-temporal frequency storage unit that stores a user ID for identifying a user, a mesh ID representing a certain area on the map, a time zone within a day, and a frequency at which the user has entered the certain area during the time zone When,
A training case storage unit that stores a correct answer label that correctly represents a case of the moving means, and a feature vector of the case obtained from the GPS history information;
An unlabeled case storage unit that stores feature vectors of the case that are not given the correct answer label;
A prediction model for predicting the moving means is generated from all the cases stored in the training case storage unit, and when the number of cases stored in the unlabeled case storage unit is equal to or less than a threshold, the prediction model is When the number of cases stored in the unlabeled case storage unit is greater than the threshold, the reliability of the prediction model is determined when the correct label is assigned to the case that is not given the correct label. A selection score representing the degree of improvement is obtained using data stored in the spatio-temporal frequency storage unit, the case where the selection score is larger than the score threshold is extracted from the unlabeled case storage unit, and the extracted case Is added to the training case storage unit, and the extracted case with the input annotation as the correct answer label is added to the training case storage unit. A prediction model generation function that repeats the process of recalculating the prediction model for predicting the moving means using all the stored cases until the number of cases stored in the unlabeled case storage unit is equal to or less than the threshold value A moving means estimation device. In the movement means estimation apparatus according to claim 1,
The prediction model generation function unit
A prediction model generation unit that generates a prediction model by performing multi-class classification using all the cases stored in the training case storage unit;
When the number of cases stored in the unlabeled case storage unit is less than or equal to a threshold, the prediction model is output to the outside, and when the number of cases is greater than the threshold, the prediction model is output to a selection score calculation unit. A prediction model output unit,
Predicting the moving means for each case stored in the unlabeled case storage unit using the prediction model, a certainty factor representing the degree of certainty of the prediction, and the mesh stored in the spatio-temporal frequency storage unit Selection score calculation that obtains a geographic score representing the degree of spatiotemporal rarity from the ID and the time zone, and calculates a selection score that represents the degree of improvement in the performance of the prediction model from the certainty factor and the geographic score And
Extracting each case stored in the unlabeled case storage unit having the selection score equal to or higher than a score threshold, and presenting the extracted case to the outside and requesting an annotation;
The extracted case with the input annotation as a correct label is deleted from the unlabeled case storage unit, and the unlabeled case storage update unit adds the extracted case to the training case storage unit;
And a prediction model recalculation unit that recalculates a prediction model for predicting the moving means from all the cases stored in the training case storage unit and outputs the prediction model to the prediction model generation unit. Means estimation device. In the movement means estimation apparatus according to claim 2,
The geographic score is calculated by the following formula:

Here, G is the geographic score, c (u, m, t) is the frequency of user u, mesh ID _m and time zone t in the spatio-temporal frequency storage unit, and c (u, m) is the user u and mesh ID _m . Frequency, c (u, t) is frequency of user u, time zone t, c (m, t) is mesh ID _m , frequency of time zone t, c (u) is frequency of user u, c (m) is The frequency of the mesh ID _m , c (t) represents the frequency of the time zone t, λ is a weighting factor that becomes λ ₁ +... + Λ ₇ = 1,
The selection score is calculated by the following formula:

Here, S is a selection score, F is the certainty factor, and α is a weight parameter. An operation method of a moving means estimating device for estimating a moving means from GPS history information in which a user's position is recorded by GPS,
The prediction model generation function unit of the moving unit estimation device generates a prediction model for predicting the moving unit from all the cases stored in the training case storage unit of the moving unit estimation device, and the label of the moving unit estimation device When the number of cases stored in the none case storage unit is less than or equal to a threshold, the prediction model is output, and when the number of cases stored in the unlabeled case storage unit is greater than the threshold, the moving means A selection score representing the degree of improvement in reliability of the prediction model when the correct label is assigned to the case that is not given a correct label that correctly represents the case is stored in the spatio-temporal frequency storage unit of the moving means estimating device. Using the stored data, find the case where the selection score is greater than the score threshold from the unlabeled case storage unit, and present the extracted case to the outside Requesting an annotation, adding the extracted case with the input annotation as the correct answer label to the training case storage unit, and using all the cases stored in the training case storage unit, the moving means The operation method of the moving means estimating apparatus is characterized in that the process of recalculating the prediction model for predicting is repeated until the number of cases stored in the unlabeled case storage unit becomes equal to or less than the threshold value.   The program for functioning a computer as a moving means estimation apparatus in any one of Claims 1 thru | or 3.

Images