JP6543215B2 - Destination prediction apparatus, destination prediction method, and destination prediction program - Google Patents

Description

  The present invention relates to a destination prediction apparatus, a destination prediction method, and a destination prediction program, and in particular, to accurately estimate a user's next destination from positioning data which is a sequence of observed position information. The present invention relates to a destination prediction apparatus, a destination prediction method, and a destination prediction program for realizing the same.

  With the spread of mobile phones equipped with positioning functions such as GPS (Global Positioning System) and Wi-Fi chips, and portable terminals such as smart phones, it has become possible to easily acquire positioning data in which the position of the user is measured. Therefore, if it is possible to predict the destination of the user using positioning data from the past to the present, it can be applied to a wide range of services such as information recommendation and life support according to the prediction result.

  A method using sequence mining is known as a conventional method of predicting a user's destination using the positioning data. In the method using the sequence mining, for example, from the positioning data of the GPS by the portable terminal carried by the user, the area where the user stays is extracted by mining processing. Furthermore, by representing transition behavior among multiple stay areas as a Markov model and calculating the transition probability, it predicts a destination which is a stay area where a user staying at a certain place will stay next (See, for example, Non-Patent Document 1 and Non-Patent Document 2).

  In the conventional method described above, the destination is predicted based on a model that takes into consideration the short-term state dependency such as the first-order Markov model, as the transition behavior between the staying places. On the other hand, it is expected to predict the destination with a smaller error by considering the long-term state dependency.

  As a method of efficient destination prediction considering long-term state dependency, a sequence of positioning data is expressed in discrete grid space, and transition between grids is modeled using RNN (recurrent neural network). , And a method of predicting a destination based on the transition probability obtained from the RNN model.

Andy Yuan Xue, Rui Zhang, Yu Zheng, Xing Xie, Jin Huang, Zhenghua Xu, "Destination prediction by sub-trajectory synthesis and privacy protection against such prediction", In Proc. Of ICDE'13, pp. 254-265 (2013 ). Andy Yuan Xue, Jianzhong Qi, Xing Xie, Rui Zhang, Jin Huang, Yuan Li, "Solving the data sparsity problem in destination prediction", VLDB J. 24 (2), pp. 219-243 (2015).

  However, in the method using the RNN model described above, spatial proximity between grids is not taken into consideration when modeling transitions between discretized grids. Therefore, there is a problem that the model sometimes predicts a spatially far apart transition which can not physically occur, which worsens the prediction error.

  The present invention has been made in view of the above circumstances, and is capable of accurately predicting a user's destination in consideration of spatial proximity, a destination prediction method, and a destination prediction method. It aims to provide a destination forecasting program.

  In order to achieve the above object, the destination prediction apparatus of the present invention stores a destination prediction model parameter database storing parameters of a recurrent neural network (RNN) model that predicts a moving path from a sequence of grid IDs; A grid ID extraction unit for extracting a grid ID sequence indicating a grid which is a divided area obtained by dividing a spatial position from a sequence of positioning data indicating a parameter of the RNN model from the destination prediction model parameter database Can be visited to each destination candidate by sampling simulation using the RNN model and the distance between grids from the grid ID series extracted by the grid ID extraction unit. Calculated and calculated And D. a destination prediction unit for predicting a destination based on the possibility of visiting each destination candidate.

  The destination prediction model learning unit which learns the RNN model from the grid ID series extracted from the history of the positioning data series and stores the learned RNN model parameters in the destination prediction model parameter database You may do so.

  Further, the destination prediction unit is a grid corresponding to the final point of the series of positioning data using the RNN model based on the series of grid IDs extracted by the grid ID extraction section in the sampling simulation. The transition probability from each to the grid and the state vector are calculated, the calculated transition probability is corrected by the weight according to the distance between the grids, and the grid to be moved to the next step is sampled based on the corrected transition probability. The transition probability and the state vector may be repeated by inputting the transition probability and the state vector to the RNN model and calculating the transition probability and the state vector in the next step.

  Further, when correcting the transition probability, the destination prediction unit may correct the transition probability using a Hadamard product of a weight according to the distance between grids and the transition probability.

  Further, the destination prediction unit determines the spatial position of the final point of each series in the history of the series of positioning data as the destination candidate from the series of grid IDs extracted by the grid ID extraction section. The possibility of visiting a destination candidate may be calculated by sampling simulation using the model and the distance between the grids, and the destination may be predicted based on the calculated possibility of visiting the destination candidate.

  In order to achieve the above object, the destination prediction method of the present invention is a destination prediction method performed by a destination prediction apparatus including a grid ID extraction unit and a destination prediction unit, and the grid ID extraction unit Extracting a sequence of grid IDs indicating a grid which is a divided area obtained by dividing the spatial position from the sequence of positioning data indicating the spatial position, and the destination prediction unit calculates the sequence of grid IDs from the sequence of grid IDs The parameters of the RNN model are acquired from a destination prediction model parameter database that stores parameters of a recurrent neural network (RNN) model that predicts a movement route, the RNN model is constructed, and the grid ID is extracted by the grid ID extraction unit. The RNN model from the series of grid ID and the sump using the distance between the grids Calculating a visitability to each destination candidate by ring simulation, and predicting a destination based on the calculated visitability to each destination candidate.

  The destination prediction method is performed by a destination prediction apparatus including a grid ID extraction unit, a destination prediction model learning unit, and a destination prediction unit, and the destination prediction model learning unit is a series of the positioning data. The RNN model may be learned from a series of grid IDs extracted from the history of (4), and parameters of the learned RNN model may be stored in the destination prediction model parameter database.

  In order to achieve the above object, a destination prediction program of the present invention is a program for causing a computer to function as each part of the destination prediction device of the present invention.

  According to the present invention, it is possible to accurately predict the destination of the user while taking spatial proximity into consideration.

It is a functional block diagram showing an example of composition of a destination prediction device concerning an embodiment. It is a schematic diagram which shows an example of the positioning data series historical information which concerns on embodiment. It is a schematic diagram which shows an example of the destination prediction model parameter information which concerns on embodiment. It is a schematic diagram which shows an example of the destination candidate information which concerns on embodiment. It is a schematic diagram which shows an example of the input information which concerns on embodiment. It is a flowchart which shows the flow of the process in the learning phase performed by the destination prediction apparatus which concerns on embodiment. It is a flowchart which shows the flow of the process in the prediction phase performed by the destination prediction apparatus which concerns on embodiment. It is a flowchart which shows the flow of the initial stage extraction process performed by the destination prediction apparatus which concerns on embodiment. It is a flow chart which shows a flow of extraction processing at the time of data input performed by a destination prediction device concerning an embodiment. It is a flowchart which shows the flow of the learning process performed by the destination prediction apparatus which concerns on embodiment. It is a schematic diagram which shows the structure of RNN which concerns on embodiment. It is a flow chart which shows a flow of prediction processing performed by a destination prediction device concerning an embodiment. It is a flowchart which shows the flow of the proximity | transition transition probability calculation process performed by the destination prediction apparatus which concerns on embodiment.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the present embodiment, a case will be described in which the user carries a portable terminal equipped with a GPS, a Wi-Fi chip or the like and goes to a destination.

  Further, in the present embodiment, learning of an RNN (recurrent neural network) model (see Non-Patent Document 3 and Non-Patent Document 4 below) that predicts the user's next moving destination with respect to the history of the sequence of positioning data. After that, the possibility of visiting a destination candidate is calculated by sampling simulation using the RNN model. This makes it possible to solve the data sparseness problem while considering the dependency of the state over a long period of time, and it is possible to predict the future destination of the user with higher accuracy.

      [Non-patent document 3] Sepp Hochreiter and Jurgen Schmidhuber, "Long short-term memory", Neural Computation 9 (8): pp. 1735-1780 (1997).

      [Non-patent document 4] Felix A. Gers, Nicol N. Schraudolph, and Jurgen Schmidhuber, "Learning precise timing with LSTM recurrent networks", Journal of Machine Learning Research, vol. 3, pp. 115-143 (2002).

  As shown in FIG. 1, the destination prediction apparatus 10 according to the present embodiment includes a positioning data sequence history database 12, a grid ID extraction unit 14, a destination prediction model learning unit 16, a destination prediction model parameter database 18, a destination The candidate database 20 and the destination prediction unit 22 are provided.

  As shown in FIG. 2 as an example, the positioning data sequence history database 12 is positioning data (in this embodiment, indicating the spatial position measured by the portable terminal when the user moves with the portable terminal) , A positioning number in which an identification ID of a latitude and a longitude is represented by an integer value, a latitude measured as a part of positioning data, a longitude measured as a part of positioning data, and a measured time The associated positioning data sequence history information is stored. In the positioning data series history information, a series of positioning data is configured by a plurality of positioning data each positioned at a plurality of different times in a series of movements of the user.

  The grid ID extraction unit 14 divides the spatial position into a plurality of pieces, and allocates a grid ID, which is an identification ID, to each grid, which is a divided area, with an integer value. In addition, the grid ID extraction unit 14 associates the grid ID with the set of latitude and longitude of each positioning data based on the series of positioning data stored in the positioning data sequence history database 12 to obtain a grid ID. Extract a series of

  The destination prediction model learning unit 16 learns an RNN model for predicting a grid to be a next movement destination from the sequence of grid IDs extracted by the grid ID extraction unit 14. In addition, the destination prediction model learning unit 16 stores the learned parameters of the RNN model in the destination prediction model parameter database 18. The RNN model and the learning method of the RNN model will be described later.

  The destination prediction model parameter database 18 stores, as shown in FIG. 3 as an example, destination prediction model parameter information in which parameter names of RNN model parameters and parameter values of RNN model parameters are associated with each other. Ru.

  In the destination candidate database 20, as shown in FIG. 4 as an example, the destination candidate number in which the identification ID of the destination candidate is represented by an integer value, the latitude of the destination candidate, and the longitude of the destination candidate respectively correspond The attached destination candidate information is stored.

  In this embodiment, a set of latitude and longitude of the final point of each positioning data sequence is extracted as a destination candidate from a plurality of sets of positioning data, and the extracted latitude and longitude pair is the latitude of the destination candidate The destination candidate information is constructed by setting the length and longitude.

  However, the method of constructing destination candidate information is not limited to this, and the final point of each series of positioning data is clustered by K-means clustering, Mean-shift clustering, etc., and a clustered subset is used as a destination candidate Also good. In this case, it is preferable to construct destination candidate information by setting the average value and median value of the latitude and longitude of the final point included in the subset as the latitude and longitude of the destination candidate.

  The destination prediction unit 22 is learned based on positioning data to be predicted when the destination is predicted, that is, based on input information input as positioning data indicating at least a part of the moving route to the current position of the user. Perform sampling simulation using the RNN model constructed by using the parameters of the RNN model, the destination candidate information, and the distance information between grids, and calculate the possibility of visiting each destination candidate Predict the user's destination based on the calculated purpose and the possibility of visiting a candidate.

  The input information described above is a series of positioning data to be predicted when the destination is predicted, and as shown in FIG. 5 as an example, an identification ID of positioning data for which a pair of latitude and longitude has been positioned. The positioning number represented by an integer value, the latitude which is a part of positioning data, the longitude which is a part of positioning data, and the time which was measured are each matched information. Then, based on the series of grid IDs extracted by the grid ID extraction unit 14 from the series of positioning data included in the input information, the destination prediction unit 22 is constructed by using the parameters of the RNN model as described above. The destination of the user is predicted using the RNN model thus obtained, destination candidate information, and distance information between grids.

The destination prediction unit 22 is configured to include the proximity transition probability calculation unit 4.
The proximity transition probability calculation unit 24 calculates a proximity transition probability according to the distance between grids for each grid using a weight vector based on the distance between grids. Further, the proximity transition probability calculation unit 24 updates the transition probability to each grid so as to correct the calculated proximity transition probability.

  Also, when performing the sampling simulation, the destination prediction unit 22 uses the RNN model based on the grid ID sequence extracted from the positioning data sequence, and uses the RNN model to generate each grid from the grid corresponding to the final point of the positioning data sequence. Calculate transition probability to grid and state vector. Then, the destination prediction unit 22 corrects the calculated transition probability with a weight according to the distance between the grids, and based on the corrected transition probability, samples the grid to be moved to the next step, and the transition probability and the state vector Are input to the RNN model to calculate the transition probability and the state vector in the next step.

  The destination prediction device 10 according to the present embodiment is configured by, for example, a computer device provided with a central processing unit (CPU), a random access memory (RAM), and a read only memory (ROM) for storing various programs. . Moreover, the computer which comprises the destination prediction apparatus 10 may be equipped with memory | storage parts, such as a hard disk drive and non-volatile memory. In the present embodiment, when the CPU reads and executes the program stored in the storage unit such as the ROM and the hard disk, the above-described hardware resource and the program cooperate with each other to realize the above-described function.

  In the present embodiment, the positioning data sequence history database 12, the destination prediction model parameter database 18, and the destination candidate database 20 are provided inside the destination prediction apparatus 10. However, the present invention is not limited to this. It may be provided outside the destination prediction device 10.

  The processing executed by the destination prediction apparatus 10 having the above functions includes a learning phase for storing parameters of the RNN model, and a prediction phase for predicting the destination based on the input information inputted. It is divided into

  First, an overview of the flow of processing in the learning phase by the destination prediction apparatus 10 according to the present embodiment will be described using the flowchart shown in FIG.

  In step S101, the grid ID extraction unit 14 acquires positioning data sequence history information from the positioning data sequence history database 12, and based on the sequence of positioning data included in the acquired positioning data sequence history information, the grid ID sequence Perform extraction processing at the time of data input to be extracted. The extraction processing at the time of data input will be described later using the flowchart shown in FIG.

  In step S103, the destination prediction model learning unit 16 performs a learning process of learning the RNN model from the extracted grid ID sequence. The learning process will be described later using the flowchart shown in FIG.

  In step S105, the destination prediction model learning unit 16 stores the learned parameters of the RNN model in the destination prediction model parameter database 18.

  Next, the outline of the flow of processing in the prediction phase by the destination prediction device 10 according to the present embodiment will be described using the flowchart shown in FIG. 7.

  In step S201, the grid ID extraction unit 14 receives input information, and performs extraction processing at the time of data input for extracting a grid ID sequence based on the sequence of positioning data included in the input information. The extraction processing at the time of data input will be described later using the flowchart shown in FIG.

  In step S203, the destination prediction unit 22 acquires the parameters of the RNN model from the destination prediction model parameter database 18, and acquires destination candidate information from the destination candidate database 20. In addition, the destination prediction unit 22 uses the RNN model constructed using the parameters of the acquired RNN model from the sequence of the extracted grid ID, and the acquired destination candidate information to use the destination of the user. Perform prediction processing to predict. The prediction process will be described later using a flowchart shown in FIG.

  Here, the process of the grid ID extraction unit 14 in the present embodiment will be described in detail. The grid ID extraction unit 14 performs initial extraction processing performed at the start of prediction of a destination and extraction processing at the time of data input performed when a sequence of positioning data is input.

  First, the flow of the initial extraction process of the grid ID extraction unit 14 will be described using the flowchart shown in FIG. The initial extraction process is a process performed prior to the extraction process at the time of data input.

  In step S301, the grid ID extraction unit 14 divides the latitude and longitude space into grids at regular intervals to create a plurality of grids that are rectangular areas.

  Note that the method of dividing the latitude and longitude space is not limited to this, as long as each set of latitude and longitude is divided so as to be included in any division area, for example, disclosed in Non-Patent Document 5 below. It may be divided using a geo-hash which is a known technique. Further, in the present embodiment, a plurality of grids each having the same shape are created by dividing the latitude and longitude space into grids at regular intervals, but the shape of the grid is not limited to this, and a plurality of grids having different sizes or shapes You may create

[Non-patent document 5] Kisung Lee, Raghu K. Ganti, Mudhakar Srivatsa, Ling Liu, "Efficient spatial query processing for big data", In Proc. Of SIGSPATIAL'14, pp. 469-472 (2014).

  In step S303, the grid ID extraction unit 14 assigns integer values as grid IDs to each grid obtained by division. In the present embodiment, the grid ID extraction unit 14 avoids overlapping of grid IDs in each of a plurality of grids by assigning an integer value to each grid while increasing the value one by one from 0.

  Next, the flow of extraction processing at the time of data input of the grid ID extraction unit 14 will be described using the flowchart shown in FIG.

  In step S401, the grid ID extraction unit 14 determines a corresponding grid for each piece of positioning data of the series of positioning data included in the input information input.

  In step S403, the grid ID extraction unit 14 outputs the sequence of integer values assigned to the corresponding grid as the grid ID sequence to the destination prediction unit 22.

  Next, the learning process performed by the destination prediction model learning unit 16 in the present embodiment will be described in detail using the flowchart shown in FIG.

  Hereinafter, in the mathematical expression, a vector is represented in bold, and in the text, the symbol of the vector is represented by adding “^”.

  In step S501, the grid ID extraction unit 14 acquires positioning data sequence history information from the positioning data sequence history database 12, and based on the sequence of positioning data included in the acquired positioning data sequence history information, at the time of learning described above A series of grid IDs obtained by extraction processing is extracted. Further, the grid ID extraction unit 14 calculates a series of feature vectors of one-hot expression including dimensions of all grid numbers. Here, the feature vector of the one-hot expression is a binary vector in which the element of the corresponding grid ID is 1 and the other elements are 0.

  In step S503, the destination prediction model learning unit 16 learns an RNN model that predicts the feature vector of the next positioning timing with respect to the feature vector of any positioning timing, using the calculated sequence of feature vectors as training data. In addition, as a method of RNN, RNN provided with LSTM (long short term memory) which is a known technique disclosed in the above-mentioned Non-Patent Document 3, oblivion gate which is a known technique disclosed in the above-mentioned Non-Patent Document 4 It is possible to use well-known techniques, such as RNN provided with attached LSTM.

As shown in FIG. 11, in the prediction of the feature vector of the next positioning timing using the RNN model, first, the feature vector of the one-hot expression at the current positioning timing is determined based on the parameters (S30) Embed vectors in low dimensions (S32). Specifically, the feature vector embedded in another vector space is obtained by taking the product of w _embed ^, which is one of the model parameters, and the feature vector of the one-hot expression.

  Next, the obtained feature vector and the potential state vector (S34) obtained using the RNN model at the previous positioning timing are input, and the LsTM disclosed in Non-Patent Document 3 described above is used. A state vector at the current positioning timing is calculated (S36).

  Finally, from the calculated state vector, the soft max function is used to predict the feature vector at the next positioning timing, that is, the movement probability to each grid (S38).

  In addition, it is possible to use methods, such as Truncated BPTT (Truncated back propagation through time) currently disclosed by the nonpatent literature 3 mentioned above, for learning of RNN.

In step S505, the learned RNN model parameters are stored in the destination prediction model parameter database 18. The destination prediction model parameter information shown in FIG. 3 is an example of the parameters of the RNN model provided with the above-described LSTM. That is, w _embed ^ is a parameter for embedding the input feature vector of the one-hot expression in another low-dimensional vector space. Further, w _in ^ is a parameter in the input gate of LSTM, w _out ^ is a parameter in the output gate, and w _c ^ is a parameter in the memory cell. Further, w _softmax ^ is a parameter for calculating a feature vector at the next positioning timing from the RNN state vector using the soft max function.

  Next, the prediction processing performed by the destination prediction unit 22 in the present embodiment will be described in detail using the flowchart illustrated in FIG. 12.

  Essentially, the RNN model is a model that estimates the state of the transition destination when transitioning from the current positioning timing to the positioning timing one ahead. However, in the case of estimating the destination, it does not necessarily reach from the current location to the destination in one step. Therefore, with the general RNN model usage method, it is not possible to realize destination prediction.

  Therefore, in the present embodiment, the destination prediction unit 22 inputs a grid ID sequence up to the current positioning timing, simulates transitions of a plurality of steps by random sampling using the RNN model, and can be used as destination candidates for each destination. Count the number of times to reach The counted value can be regarded as the possibility of visiting each next destination candidate for the input. Therefore, the destination is determined based on the counted value.

  Hereinafter, a specific method of determining the destination will be described in detail.

  In step S601, the destination prediction unit 22 initializes a visit probability vector P ^ representing the possibility of visiting each destination candidate with a zero vector.

  In step S603, the destination prediction unit 22 initializes an initial state vector h ^ with a zero vector.

  In step S605, the destination prediction unit 22 acquires the parameters of the RNN model from the destination prediction model parameter database 18, and constructs the RNN model using the acquired parameters of the RNN model. At this time, as described above, the destination prediction unit 22 uses the feature vector of the one-hot expression generated from the grid ID of the current positioning timing and the state vector obtained from the RNN model at the immediately preceding positioning timing as the RNN model. , And outputs the feature vector and the state vector of the prediction result at the next positioning timing.

  Here, the feature vector of the prediction result can be regarded as the transition probability to each grid. Hereinafter, the feature vector of the one-hot expression generated from the grid ID of the grid g to be processed and the state vector h ^ obtained from the RNN model at the positioning timing immediately before the grid g to be processed , RNN (grid g, state vector h ^) obtained by using the RNN model.

  In step S607, the destination prediction unit 22 starts loop A which repeats the processing in the loop until the variable i, which is the number of simulation trials, changes from 1 to N. For example, N = 1000.

  In step S609, the destination prediction unit 22 causes the grid ID extraction unit 14 to execute the above-described extraction process at the time of prediction on the series of positioning data included in the input information, Assuming that G, a loop B is started which repeats the processing in the loop for each grid g of the grid ID of each element in the series G of grid IDs.

  In step S611, the destination prediction unit 22 uses the RNN model based on the following equation (1) to determine the transition probability p ^ to each grid from the grid g and the state vector h ^ of the previous positioning timing. Are calculated, and the state vector h ^ is further updated.

  In step S613, the destination prediction unit 22 determines whether or not the processing in the loop has been performed for all the grids g in the series G of grid IDs. If a negative determination is made, the processing in the loop is repeated. If the determination is affirmative, the loop B is ended, and the process proceeds to step S615.

  In step S615, the destination prediction unit 22 starts loop C, which repeats the processing in the loop until the variable j, which is the number of times of sampling, changes from 1 to M. Here, M is the upper limit number of samplings, and for example, M = 100.

  In step S616, the proximity transition probability calculation unit 24 of the destination prediction unit 22 executes proximity transition probability calculation processing for updating the transition probability p ^ to each grid. The flow of the proximity transition probability calculation process will be described below with reference to the flowchart showing the flow of the proximity transition probability calculation process shown in FIG.

  In step S701, the proximity transition probability calculation unit 24 calculates a weight vector s ^ based on the distances from the grid g to all | G | grids based on the following equation (2).

  Here, Dist (..., ...) represents the distance [meters] between two grids. σ is a fixed value that determines the degree of variation of the weight of each element, and in this embodiment, σ = 200 [meters].

  In step S703, the proximity transition probability calculation unit 24 calculates the proximity transition probability to each grid according to the following equation (3) based on the weight vector s ^ and the transition probability p to each grid, The transition probability p ^ to each grid is updated, and the process proceeds to step S617 in FIG.

  Here, “o” in the equation (3) represents the Hadamard product of the vector.

  Thus, the transition probability to each grid is corrected according to the distance between the grids, and the grid to be moved to the next step is sampled based on the corrected transition probability to each grid as described later. By doing so, it is possible to avoid predicting spatially far apart transitions which can not occur physically, thereby making it possible to predict the destination with smaller error.

  In step S617 of FIG. 12, the destination prediction unit 22 randomly samples the grid g, which is a grid in the case of moving to the next step, in accordance with the transition probability p to each grid.

  In step S619, the destination prediction unit 22 determines whether the distance to the destination candidate d closest to the grid g is closer than the threshold ε. In this embodiment, the average value of the distance traveled by the user in one step is set as the threshold value ε, and for example, ε = 100 m (meters). When an affirmative determination is made in step S619, the process proceeds to step S625, and when a negative determination is made, the process proceeds to step S621.

  In step S621, the destination prediction unit 22 calculates the transition probability p ^ to each grid from the grid g and the state vector h ^ one step earlier using the RNN model based on the equation (1). And further update the state vector h ^.

  In step S623, the destination prediction unit 22 determines whether or not the variable j has become M. If the determination is negative, the process in the loop is repeated, and if the determination is positive, the loop C is ended. Then, the process proceeds to step S627.

  In step S625, the destination prediction unit 22 adds 1 to the element Pd of the visit probability vector P ^ corresponding to the destination candidate d based on the following equation (4).

  In step S627, the destination prediction unit 22 determines whether or not the variable i has become N. If negative, the process in loop A is repeated, and if positive, loop A is ended. Then, the process proceeds to step S629.

  In step S629, the destination prediction unit 22 calculates the grid ID of the destination candidate with the highest possibility of visiting from the visit probability vector P ^ indicating the possibility of visiting each destination candidate.

  In step S631, the destination prediction unit 22 acquires the latitude and longitude of the destination candidate that matches the calculated grid ID from the destination candidate information acquired from the destination candidate database 20, and displays the latitude and longitude on the display means. It outputs by storing it in the storage means.

  As described above, the destination prediction apparatus 10 according to the present embodiment receives the history of positioning data sequences, which is a set of latitude and longitude, from positioning data sequence history information, and is included in the history of received positioning data sequences. A grid ID sequence is extracted based on the sequence of positioning data. In addition, the destination prediction apparatus 10 learns the RNN model from the series of grid IDs, and stores the learned parameters of the RNN model as destination prediction model parameter information.

  The destination prediction apparatus 10 receives a sequence of positioning data as input information, and extracts a sequence of grid IDs based on the received sequence of positioning data. In addition, the destination prediction apparatus 10 acquires parameters of the destination prediction model, constructs an RNN model, and visits each destination candidate by sampling simulation using the RNN model from the extracted grid ID series The calculated visitability is corrected according to the distance between the grids, and the latitude and longitude of the destination are output as the prediction result of the destination based on the corrected visitability.

  Thus, the RNN model is learned using the history of the series of positioning data of the user, and a sampling simulation using the RNN model is performed from the series of positioning data of the user newly acquired. In this way, the dependence of the state over time can be taken into account. Furthermore, since a destination can be predicted based on learning data that includes only a part of a new user's positioning data sequence by the same method, it is possible to extend the conventional Markov model to a third or higher order. It is sufficient to prepare learning data of a small movement pattern, which solves the problem of data sparsity. In addition, it is possible to avoid predicting a spatially far apart transition which can not occur physically, which makes it possible to predict a destination with a smaller error. Since the possibility of visiting a destination candidate is calculated based on the above effects, it is possible to accurately predict the destination of the user.

  In the present embodiment, although the case of using a set of latitude and longitude as positioning data indicating a spatial position has been described, the present invention is not limited to this, and in a space for predicting a destination as positioning data indicating a spatial position Information that can identify the position may be used.

  Further, in the present embodiment, the operation of the component of the function shown in FIG. 1 is constructed as a program, installed in a computer used as the destination prediction apparatus 10 and executed, but the present invention is not limited thereto. You may distribute it.

  Further, the constructed program may be stored in a portable storage medium such as a hard disk, a flexible disk, a CD-ROM, etc., and may be installed in a computer or distributed.

10 Destination prediction apparatus 12 Positioning data sequence history database 14 Grid ID extraction unit 16 Destination prediction model learning unit 18 Destination prediction model parameter database 20 Destination candidate database 22 Destination prediction unit

Claims (8)

A destination prediction model parameter database storing parameters of a recurrent neural network (RNN) model that predicts a movement path from a sequence of grid IDs;
A grid ID extraction unit that extracts a grid ID sequence indicating a grid that is a divided area obtained by dividing the spatial position from the sequence of positioning data indicating the spatial position;
The parameters of the RNN model are acquired from the destination prediction model parameter database to construct the RNN model, and the RNN model and the distance between the grids are used from the series of grid IDs extracted by the grid ID extraction unit. A destination prediction unit which calculates visitability to each destination candidate by sampling sampling, and predicts a destination based on the calculated visitability to each destination candidate;
Destination prediction device equipped with. It further comprises a destination prediction model learning unit for learning the RNN model from the grid ID sequence extracted from the history of the positioning data sequence and storing the learned RNN model parameters in the destination prediction model parameter database The destination prediction apparatus according to claim 1. The destination prediction unit uses the RNN model on the basis of the grid ID sequence extracted by the grid ID extraction unit in the sampling simulation, and uses each of the grids corresponding to the final point of the positioning data sequence. Calculate transition probability to grid and state vector,
The calculated transition probability is corrected by a weight according to the distance between the grids, and the grid to be moved to the next step is sampled based on the corrected transition probability, and the transition probability and the state vector are converted to the RNN model. The destination prediction device according to claim 1, wherein the destination prediction device is repeatedly input to calculate the transition probability and the state vector in the next step. The destination prediction unit corrects the transition probability by using a Hadamard product of a weight according to a distance between grids and the transition probability when correcting the transition probability. apparatus. The destination prediction unit sets the spatial position of the final point of each series in the history of the series of positioning data as the destination candidate, and the RNN model from the series of grid IDs extracted by the grid ID extraction section, The destination possibility is calculated by sampling simulation using the distance between the grid and the grid, and the destination is predicted based on the calculated possibility of visiting the destination candidate. The destination prediction device according to item 1. A destination prediction method performed by a destination prediction apparatus including a grid ID extraction unit and a destination prediction unit,
The grid ID extraction unit extracts a grid ID sequence indicating a grid which is a divided area obtained by dividing the spatial position from the sequence of positioning data indicating the spatial position;
The RNN model is constructed by acquiring the parameters of the RNN model from a destination prediction model parameter database storing the parameters of a recurrent neural network (RNN) model that predicts a movement path from a series of grid IDs by the destination prediction unit. The possibility of visit to each destination candidate is calculated by sampling simulation using the RNN model and the distance between the grids from the grid ID series extracted by the grid ID extraction unit, Predicting a destination based on the possibility of visiting a destination candidate;
Destination prediction method with. 7. The destination prediction method according to claim 6, performed by a destination prediction apparatus provided with a grid ID extraction unit, a destination prediction model learning unit, and a destination prediction unit,
The destination prediction model learning unit learns the RNN model from the grid ID sequence extracted from the history of the positioning data sequence, and stores the learned RNN model parameters in the destination prediction model parameter database The destination prediction method according to claim 6, further comprising a step.   The program for functioning a computer as each part of the destination prediction apparatus in any one of Claims 1-5.

Images