JP2017097693A - Data prediction device, information terminal, program, and method performing learning with data of different periodic layer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a data prediction device capable of reducing computational complexity for learning processing more while securing appreciable prediction precision.SOLUTION: The present device is configured to predict object data at a certain point of time by a neural network. Input and hidden layer means of the present device has an input layer inputting, as feature quantity elements, object data values related to points of time belonging to a plurality of mutually different time periodic layers, and a hidden layer inputting a first processing signal generated by subjecting a signal output from the input layer to weighting processing with a first coupling load, and outputs a second processing signal generated by subjecting a signal output from the hidden layer to weighting processing with a second coupling load. Further, output layer means of the present device has an output layer inputting the second processing signal and outputting a value of the object data at a point of time related to prediction. Further, the input and hidden layer means and the output layer means can execute re-learning for updating the first and second coupling loads so as to reflect periodicity related to the plurality of time periodic layers.SELECTED DRAWING: Figure 2

Description

  The present invention relates to an information processing technique for predicting a user's behavior and state using machine learning based on the user's behavior history.

  In recent years, high-function communication terminals such as smartphones, tablet computers, and car navigation devices have become widespread, and users can acquire desired information regardless of time and place.

  When accessing such information, it is necessary for the user to input keywords and menu operations each time, and in some cases, it takes considerable time. Further, for example, searching for Internet information while walking or driving, or setting a navigation route / destination is not only troublesome but also involves danger.

  In order to solve such user operation problems, for example, it is possible to access information without using a finger or the like by using a voice interaction function, and further, a GPS (Global Positioning) installed in a terminal or a car. A method has been proposed in which a user's situation is grasped to some extent by using a sensor such as System) and information on the user's behavior is presented by push. However, such a method has a problem that convenience is low and it is difficult to improve prediction accuracy.

  Therefore, as another solution, a technique has been proposed in which a neural network (NN) is used to predict a user's behavior and state and present corresponding information. For example, Patent Document 1 uses a hierarchical NN to learn a nonlinear correlation between daily maximum power demand and weather conditions, etc., based on the data of actual daily maximum power demand over the past year. And a technique for generating a prediction model and predicting the daily maximum power demand. Here, an algorithm of a back propagation method (back propagation) is used for learning.

  Moreover, this patent document 1 is intended to improve prediction accuracy, performs learning processing using NN for each year based on actual data in a plurality of years, and calculates a relative error between the output value and the actual power demand amount. We have also proposed a mechanism to set an annual increase correction coefficient from among the above, and to make an annual increase correction for the predicted value.

  Furthermore, Patent Document 2 predicts a wind speed several hours ahead by inputting past continuous meteorological data and meteorological data of the day to a recurrent neural network (RNN), and wind power generation based on the information. A technology to predict the power generated by the facility is proposed.

Japanese Patent Laid-Open No. 9-22402 JP 2007-56686 A

  However, in the conventional technique using the hierarchical NN as described in Patent Document 1, for example, a high-spec server is required to execute the learning process and the prediction process, and the processing load becomes enormous. Therefore, for example, it is very difficult to generate prediction information in a terminal or an embedded system owned by a user who should present information based on the prediction result.

  In this regard, in the technique described in Patent Document 1, in order to reduce the amount of calculation, only the daily power demand for one year is used as learning data, and the annual increase correction coefficient for the predicted value by such learning is used. Is used to correct the prediction result. However, when compared with the results learned from all the actual value data retroactively, the prediction accuracy is considerably lowered.

  On the other hand, in the prior art described in Patent Document 2, RNN is used for generating prediction information. The RNN is a system suitable for learning time-series data and outputting a predicted value based on it. This RNN generally has a higher learning processing speed than the hierarchical NN as described above. However, in order to improve the prediction accuracy, it is necessary to learn a lot of past information that is close in time from the present, and the number of dimensions of the input layer and intermediate layer must be increased. It takes time to complete.

  In this regard, Patent Document 2 suggests that it is possible to improve prediction accuracy and shorten prediction calculation time by learning only known weather data in the same period of the previous year in order to reduce processing time in RNN. However, no specific criteria are given. In addition, such a learning method using partial data using seasonality is difficult to apply to prediction of data other than electric power, for example, other than electric power.

  Therefore, an object of the present invention is to provide a data prediction apparatus, program, and method that can reduce the amount of calculation for learning processing as compared with the conventional method while ensuring a considerable prediction accuracy.

According to the present invention, the apparatus includes a neural network (NN) that learns with time-series target data, and predicts the target data at a certain time point or time range.
A first combination of an input layer related to NN that inputs target data values related to time points or time intervals belonging to each of at least two different time period hierarchies as a feature quantity element, and a signal output from at least this input layer A first processing signal generated by weighting with a load is input, and a hidden layer related to NN is input. A signal output from the hidden layer is generated by weighting with a second combined weight. Input / hidden layer means for outputting a second processing signal;
An output layer means having an output layer according to NN, which inputs the second processing signal and outputs at least the value of the target data at the time point or time range related to the prediction;
The input / hidden layer means and the output layer means are capable of executing re-learning to update the first connection weight and the second connection weight to reflect the periodicity related to at least two time period layers. Is provided.

As one embodiment of the data prediction apparatus according to the present invention,
The input / hidden layer means further includes a context layer related to NN as a layer that outputs the same signal as the hidden layer.
The hidden layer inputs a third processing signal generated by weighting the signal output from the context layer with the third coupling weight as a time delay signal together with the first processing signal,
It is also preferable that the input / hidden layer means and the output layer means can execute relearning to update the third coupling weight so as to reflect the periodicity related to at least two of the time period layers.

  Moreover, it is preferable that the input / hidden layer means and the output layer means in the data prediction apparatus according to the present invention perform the update of the combined load reflecting the periodicity related to at least two of the time period layers by back propagation. The input / hidden layer means and the output layer means preferably execute relearning using data not yet used for learning among the time-series target data.

Furthermore, regarding the feature quantity element according to the data prediction apparatus according to the present invention,
A first time unit, a second time unit composed of a predetermined number of first time units, and then a predetermined number of time units one order before, sequentially incrementing the ordinal number by 1. Set the time unit, and
At least two of the time period hierarchies are a first time period hierarchy whose period is a time according to a first time unit, a second time period hierarchy whose period is a time according to a second time unit, and Thereafter, it is also preferable to include at least two time period hierarchies among time period hierarchies whose period is a time unit in which the ordinal numbers are sequentially incremented by one.

  Further, regarding the feature quantity element having the above feature, specifically, the first time unit is “day”, the first time period hierarchy is a period hierarchy related to “week”, and the second time unit. Is a "week" and the second time period hierarchy is a period hierarchy related to "month", the third time unit is "month" and the third time period hierarchy is a period hierarchy related to "year" It is also preferable.

  Furthermore, regarding the feature quantity element having the above-described features, a plurality of target data values related to time points or time intervals belonging to each of the time period layers are all the first time units constituting the first time period layer. Target data value group related to a part of time units, target data value group related to a part of time units among all second time units constituting the second time period hierarchy, and thereafter It is also preferable to include at least two of the target data value groups related to some time units among all the time units constituting the time period hierarchy in which the ordinal numbers are sequentially incremented by one.

  In addition, with respect to the feature quantity element having the above-described feature, the target data value related to a part of the time units among all the first time units constituting the first time period hierarchy, the second time period hierarchy is constituted. A target data value related to a part of time units in all the second time units, and a part of all time units constituting the time period hierarchy in which the ordinal number is sequentially incremented by one The target data value relating to the time unit preferably takes a value in the basic time unit constituting the first time unit in the time unit constituting the time period hierarchy relating to itself.

Furthermore, regarding the feature quantity element having the above characteristics, specifically,
The first time unit is “day”, the first time period hierarchy is a period hierarchy related to “week”, the second time unit is “week”, and the second time period hierarchy is “month”. , The third time unit is “month”, and the third time period hierarchy is a period hierarchy related to “year”,
It is also preferable that the target data values related to the first time unit, the second time unit, and the third time unit take a value at a predetermined “time” or “time zone” in the time unit.

Furthermore, as another embodiment of the data prediction apparatus according to the present invention,
An interface capable of inputting an evaluation result for the value of the target data as a prediction result output from the output layer means;
The input / hidden layer means and the output layer means preferably execute relearning using teacher data generated based on the acquired evaluation result.

In addition, regarding the target data in the data prediction apparatus according to the present invention,
The target data value related to the time point or time interval belonging to each of the time period hierarchy is a data value related to the reference position of the movement prediction and the position of the movement destination in the prediction target at the time point or time interval,
The value of the target data output at the time point or time range related to the prediction output by the output layer means is preferably a predicted value of the position of the movement destination in the prediction target.

  Further, the data device that handles such target data further has a positioning unit that can acquire information related to the location by GPS (Global Positioning System), and based on the acquired information related to the location It is also preferable that the information terminal be capable of generating at least one of the feature amount input by the layer and the teacher data used for relearning.

According to the present invention, there is also provided a program for causing a computer mounted on an apparatus for predicting the target data at a certain time point or time range to have a NN that learns with time-series target data,
A first combination of an input layer related to NN that inputs target data values related to time points or time intervals belonging to each of at least two different time period hierarchies as a feature quantity element, and a signal output from at least this input layer A first processing signal generated by weighting with a load is input, and a hidden layer related to NN is input. A signal output from the hidden layer is generated by weighting with a second combined weight. Input / hidden layer means for outputting a second processing signal;
The second processing signal is input, and at least the value of the target data at the time point or time range related to the prediction is output, and the computer functions as output layer means having an output layer related to NN,
The input / hidden layer means and the output layer means are capable of executing re-learning to update the first connection weight and the second connection weight to reflect the periodicity related to at least two time period layers. A data prediction program is provided.

According to the present invention, there is further provided a data prediction method in an apparatus that includes an NN that learns with time-series target data and predicts the target data at a certain time point or time range,
Inputting a target data value related to a time point or a time interval belonging to each of at least two different time period layers as a feature quantity element to an input layer related to NN;
The first processed signal generated by weighting at least the signal output from the input layer with the first coupling weight is input to the hidden layer related to NN, and the signal output from the hidden layer is set to the second Outputting a second processing signal generated by weighting with a combined weight;
Inputting the second processed signal to the output layer related to the NN, and outputting at least the value of the target data in the time point or time range related to the prediction;
Performing a relearning to update the first and second coupling weights to reflect the periodicity associated with the at least two time period hierarchies.

  According to the data prediction apparatus, information terminal, program, and method of the present invention, it is possible to further reduce the amount of calculation for the learning process as compared with the related art while ensuring considerable prediction accuracy.

It is a schematic diagram which shows one usage pattern of the data prediction apparatus by this invention. It is a functional block diagram which shows the function structure in one Embodiment of the data prediction apparatus by this invention. It is a functional block diagram which shows the function structure in other embodiment of the data prediction apparatus by this invention. It is a schematic diagram which shows one Embodiment of the learning and data prediction process in a deep learning part. It is a schematic diagram which shows one Embodiment of the production | generation of the feature-value vector based on this invention. It is a schematic diagram which shows the program code for performing one Embodiment of each step shown by FIG. It is a schematic diagram which shows the program code for performing one Embodiment of each step shown by FIG. It is a table which shows one Example and the comparative example of the data prediction method by this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a schematic diagram showing one usage pattern of a data prediction apparatus according to the present invention.

  According to FIG. 1, a smartphone 1 which is a data prediction device of the present invention is attached to a cradle 8 installed in a car as a moving body, and is used as a terminal for providing information including a car navigation function. In FIG. 1, the smartphone 1 is arranged at a position that is well visible from the driver's seat in the middle front position between the driver's seat and the passenger seat. In this way, the smartphone 1 is installed in a simple form without having to be electrically connected to the automobile.

  Here, the data prediction apparatus according to the present invention is not limited to a smartphone as a matter of course. For example, a mobile portable terminal such as a tablet computer or a wearable terminal may be used. Furthermore, it may be an information device that can be loaded into a car such as a car navigation unit or a personal computer (PC). In any case, such a device exhibits the function as the data prediction device of the present invention by causing an in-device computer to execute an application program for operating the main functions of the present invention described below, for example. Can be.

  The smartphone 1 includes a neural network (NN) that learns with time-series target data, and predicts the value of the target data at a certain time point or time range. In the present embodiment, the smartphone 1 has a positioning unit 102 that can acquire information related to a location by GPS (Global Positioning System), and based on the information related to the location acquired by the positioning unit 102, a predetermined time It is possible to predict a destination after the passage (a time point to be predicted or a location in a time range). It is also preferable that content related to the result can be acquired or selected based on the generated prediction result of the destination and provided to the user.

Here, if the specific structure in the smart phone 1 is given, first, as a feature quantity vector input to the provided NN,
(A) A feature amount vector having a feature value element as a target data value related to a time point or a time interval belonging to each of “at least two time period layers” different from each other is adopted.

This feature vector will be described in detail later with reference to FIG. 5. For example, “1 week” and “1 month” are adopted as two different “time period hierarchies”,
(A) Adopting “1 day ago”, “2 days ago” and “3 days ago” as three time points (time intervals) belonging to “1 week”
(A) “1 week ago”, “2 weeks ago”, and “3 weeks ago” are adopted as three time points (time intervals) belonging to “1 month”.
Here, it is possible to generate a feature quantity vector whose elements are data values at each time point.

In addition, the smartphone 1
(B) An “input layer” relating to NN to which the feature vector of (A) is input and a signal output from at least this “input layer” are weighted with a first combined weight and generated. A second processing signal generated by weighting the signal output from the “hidden layer” with a second coupling weight. Input / hidden layer part to output,
(C) The second processing signal is input, and at least the value of the target data in the time point or time range related to the prediction is output, and the output layer unit having the “output layer” related to the NN,
(D) The input / hidden layer unit and the output layer unit perform relearning to update the first coupling weight and the second coupling weight to reflect the periodicity related to “at least two time period layers”. It is characterized by being possible.

  As in the configuration (C), the “output layer” outputs the value of the target data at the time point or time range related to the prediction. For example, when a feature vector is generated based on information related to the location of the smartphone 1 acquired by the positioning unit 102 and input to the “input layer”, the “output layer” is the time point or time range related to the prediction. It is possible to output a predicted value of the location (movement destination).

  Further, for example, teacher data may be generated based on information related to the location of the smartphone 1 acquired by the positioning unit 102, and the relearning in the configuration (D) may be executed. In this case, periodicity related to “at least two time period hierarchies” seen at the destination of the user of the smartphone 1, such as going to a predetermined place every week on a predetermined day of the week, or meeting in a predetermined place at the end of every month Thus, the periodicity seen in the moving behavior can be reflected in the first combined load and the second combined load. As a result, it is possible to ensure sufficient prediction accuracy while suppressing the amount of calculation for the learning process.

Furthermore, in the smartphone 1,
(E) The input / hidden layer unit further includes a “context layer” related to the NN as a layer that outputs the same signal as the “hidden layer”.
(F) The “hidden layer” uses the third processing signal generated by weighting the signal output from the “context layer” with the third coupling weight as a time-delayed signal, together with the first processing signal. type in,
(G) It is also preferable that the input / hidden layer unit and the output layer unit be able to execute relearning for updating the third coupling weight to reflect the periodicity related to “at least two time period layers”.

  In this case, the machine learning function unit corresponds to a recurrent neural network (RNN). In particular, since the connection load is given to the output from the “context layer” of the first-order lag, this function unit has an internal storage portion that can appropriately reflect a plurality of “time period hierarchies” of the data. Will have. In this way, a part of the output of the “hidden layer” feeds back to the input layer side, and the functional unit with the asymmetric coupling state accurately determines the periodicity of the time-series data that is input and its change. Suitable for identification.

  Here, in particular, the smartphone 1 of the present invention employs a feature quantity vector having feature data elements as target data values related to each of “at least two time period hierarchies” different from each other in the configuration (A). . As a result, even if the target data is not continuous in time series and is discrete and has a smaller amount of data, by capturing data values in each “time period hierarchy”, various data possessed by the target data can be obtained. Periodicity and its change can be learned efficiently. As a result, it is possible to further reduce the amount of calculation for the learning process as compared with the prior art while ensuring a considerable prediction accuracy.

  Further, when learning is performed using the RNN having the above-described configurations (E) to (G), for example, this HMM is also compared with a hidden Markov model (HMM) that has been conventionally used when handling time-series data. Does not require a priori knowledge. That is, without using such a priori knowledge, for example, based on only the user's movement history data, a deep layer using a feedback coupling capable of performing information analysis combining past time series data and current time series data Learning can be carried out. This makes it possible to achieve both high prediction accuracy and a small amount of calculation (short calculation processing time) that are normally in a trade-off relationship in the NN. Therefore, instead of high-spec devices such as servers, information terminals such as smartphones and in-vehicle car navigation devices can be used as the data prediction device of the present invention.

  In particular, in the past, deep learning such as RNN was mostly executed by a server on the Internet, but as described above, the feature amount is based on the data value in each “time period hierarchy”. By generating, it is possible to speed up the convergence of the computation and suppress the amount of calculation required for learning, so that it can be executed in a mobile information terminal that has been increasing in specifications in recent years.

  Incidentally, the above-mentioned HMM requires a priori knowledge for the configuration design of the state transition topology, and cannot cover all state transitions comprehensively at the design stage (for example, included in the topology at the design stage). Unrecognized state transition behavior). Therefore, for example, when predicting the destination of a user, it is impossible to know in advance how many types of movement patterns are included in the time series, so the transition topology between hidden states should be determined in advance. Becomes very difficult. Furthermore, in the HMM, as a new hidden state is added, it becomes necessary to re-learn the continuous-time model from the beginning, and this leads to an enormous amount of calculations. It is also extremely difficult to construct a system that performs adaptive learning processing online using a terminal.

  Here, when the specific usage example of the smart phone 1 is shown, for example, based on the prediction result of the destination after a predetermined time predicted from the past movement history, not only the predicted destination but also the destination Or you may provide the user in a vehicle with the outing spot information on the path | route until it arrives there. In addition, various guidance information such as a parking lot near the destination can be provided to users in the vehicle. Furthermore, when the predicted destination is the company of the commuting destination, information relating to the company, for example, a work schedule for the day may be presented. Further, when the destination is a store, it is possible to provide time service information at the store.

  In this way, information such as preset content that matches the destination as the prediction result can be presented to the user (driver and / or passenger). Here, this presentation information may be personalized information unique to each user. When presenting the above information, current location information by GPS (Global Positioning System), traffic congestion and traffic regulations by VICS (registered trademark) (Vehicle Information and Communication System), which will be described in detail later, It is also possible to acquire and use road traffic information.

  In addition, naturally the mobile body in which the data prediction apparatus by this invention, such as the smart phone 1, is installed is not limited to a motor vehicle. For example, a bicycle, a pedestrian, a passenger car other than an automobile, a truck, a bus, an elevator, etc. Can be a target.

  Furthermore, the prediction target data predicted by the data prediction apparatus according to the present invention is not limited to the movement destination (location). For example, financial data such as stock prices and exchange rates, usage data such as electricity, gas, and water, as well as data related to social phenomena such as holding events and the occurrence of incidents, etc. Can be predicted data according to the present invention.

[Functional configuration of data prediction device]
FIG. 2 is a functional block diagram showing a functional configuration in an embodiment of the data prediction apparatus according to the present invention.

  According to FIG. 2, the smartphone 1 as a data prediction device includes a communication interface unit 101, a positioning unit 102, a touch panel display (TP / DP) 103 as a user interface, a speaker 104 and a microphone 105, and a log storage unit. 106, a combined load storage unit 107, a content storage unit 108, and a processor memory. Here, the processor memory realizes a data prediction function by executing a program that causes a computer mounted on the smartphone 1 to function.

  Further, the processor memory includes a position information acquisition unit 111, a situation acquisition unit 112, a feature amount generation unit 113, a deep learning unit 12, a prediction result processing unit 114, and a user interface (IF) control unit 115. Have. Here, the user IF control unit 115 preferably includes a dialogue processing unit 115c. In addition, according to FIG. 2, the flow of the process which connected each function structure part with the arrow is understood also as one Embodiment of the data prediction method by this invention.

In FIG. 2, the communication interface unit 101
(A) Receive weather information from the weather information server 6 provided by the private or public sector at present or after a predetermined time (period),
(B) From the VICS (registered trademark) server 7, traffic or operation information (for example, VICS (registered trademark) information) for distributing traffic or operation information can be received. Also,
(D) transmitting the prediction result of the destination determined by the prediction result processing unit 114 or the presentation information determined according to the result to an external device, for example, a display device or a speaker device installed in the vehicle. Is also preferable.

  The positioning unit 102 captures a positioning radio wave from the GPS satellite 5 constituting the GPS system, and measures the current position of the smartphone 1 (automobile). Further, the position information acquisition unit 111 receives the measurement result from the positioning unit 102 and acquires the current position information, for example, the latitude and longitude information of the current position.

  The situation acquisition unit 112 acquires the weather information, traffic, or operation information received by the communication interface unit 101, determines the weather state, traffic, or operation state at the current position (current time), and the feature value generation unit 113 Provides information for feature generation.

  The feature quantity generation unit 113 generates a feature quantity vector having a target data value related to a time point or time interval belonging to each of “at least two time period hierarchies” that are different from each other, and outputs the feature quantity vector to the deep learning section 12. . A specific configuration of the feature vector will be described later in detail with reference to FIG.

The deep learning unit 12
(A) an input / hidden layer unit 121 having an input layer 121i, a hidden layer 121h, and a context layer (CON layer) 121c;
(B) It has the output layer part 122 which has the output layer 122o.

Here, the input layer 121i is a layer for inputting target data values related to time points or time intervals belonging to each of at least two different time period layers as feature elements. The hidden layer 121h is
(A1) A first processing signal generated by weighting a signal output from the input layer 121i with a first coupling weight is input, and
(A2) A third processing signal generated by weighting the signal output from the context layer 121c with the third coupling weight is input as a time delay signal together with the first processing signal.

  Further, the input / hidden layer unit 121 outputs, to the output layer unit 122, a second processing signal generated by weighting the signal output from the hidden layer 121h with the second coupling weight. The output layer 122o of the output layer unit 122 receives the second processing signal, and outputs at least the value of the target data at the time point or time range related to the prediction.

  Furthermore, the input / hidden layer unit 121 and the output layer unit 122 update the first coupling load, the second coupling load, and the third coupling load to reflect the periodicity related to at least two time period layers. The learning is configured to be executable by an error back propagation method (back propagation).

  Here, when the target data value related to the time point or time interval belonging to each of the time period hierarchy is the data value related to the reference position of the movement prediction and the position of the movement destination in the prediction target at these time points or time intervals, The value of the target data at the time point or the time range related to the prediction output by the output layer unit 122 is a predicted value of the position of the movement destination in the prediction target. The learning / prediction processing in the deep learning unit 12 will be described later in detail with reference to FIG.

  The prediction result processing unit 114 processes the value of the target data at the time point or time range related to the prediction output from the output layer unit 122 into a prediction result, and outputs information on the prediction result to the user IF control unit 115. For example, when the predicted output value from the output layer 122o (the output layer unit 122) is a numerical value normalized by a value from 0 to 1, this numerical value is converted into a prediction result (such as a destination). Here, the prediction result processing unit 114 sets the number of predictions, causes the deep learning unit 12 to repeat the prediction process for the number of predictions, and sets the predetermined time (period) as one unit, and only one unit from the present. It is also preferable to generate not only the predicted value at the previous time but also the predicted value at a plurality of units ahead and generate a predicted result according to this.

  Furthermore, the prediction result processing unit 114 may extract content corresponding to the generated prediction result from the content storage unit 108 and present the content to the outside via the user IF control unit 115. For example, the presentation information related to the content may be displayed on the TP / DP 103, and the presentation information as sound may be output to the speaker 104. Here, the content storage unit 108 may obtain presentation information (content) from an external content server via the communication interface 101. Moreover, you may input predetermined | prescribed presentation information by input operation via TP / DP103 by a user.

  The user IF control unit 115 displays the prediction result information and content input from the prediction result processing unit 114 on the TP / DP 103 as presentation information, or outputs a voice corresponding to the presentation information from the speaker 104 via the dialogue processing unit 115c. I will let you. In addition, the user IF control unit 115 sends response information, which is an evaluation result such as affirmation or denial by the user, to the presented prediction result via voice input to the microphone 105 and the dialogue processing unit 115c through an operation on the TP / DP 103. Or via an external device such as a remote controller, and outputs the result to the prediction result processing unit 114.

  At this time, the prediction result processing unit 114 generates teacher data for relearning based on the evaluation result (response information) by the user, and outputs it to the deep learning unit 12 to execute relearning. That is, the prediction result processing unit 114 also functions as a means for updating the learning data by using the user's affirmative / negative response in using the real environment in order to increase the reliability of the NN system. If no response is input from the user after the predicted value is output and the predicted result is presented, the feature amount and the predicted value that are currently used are stored in the log storage unit 106. On the other hand, for example, when a negative (objection) is input from the user, the predicted value is replaced with correction information generated based on a response from the user, and is written in the log storage unit 106 together with the used feature vector.

  The log storage unit 106 stores user movement history data (for example, time-series data in which a date and time and a location at that time (period) are associated) for generating a feature vector. Moreover, the function which updates the movement history data which should be learned based on the evaluation result (response information) by a user is also provided.

  The combined load storage unit 107 includes the combined loads (first, second, and third combined loads) calculated and updated during the learning process in the deep learning unit 12 and the learned learning stored in the log storage unit 106. The record ID of a certain last record is stored. Further, at the time of re-learning, in response to a read instruction from the deep learning unit 12, the stored connection load and record ID are output to the deep learning unit 12. Here, this relearning is preferably performed from a new record having an ID next to the read record ID. In this case, in other words, relearning is executed using a record that is not yet used for learning. Thus, by performing relearning only on the history difference, it is possible to further reduce the amount of calculation required for the learning process.

  FIG. 3 is a functional block diagram showing a functional configuration in another embodiment of the data prediction apparatus according to the present invention.

  According to FIG. 3, the data prediction system is constructed | assembled by the data prediction server 2 as the data prediction apparatus of this invention, and the smart phone 3 being connected for communication via an access network and the internet, for example.

  Among these, the data prediction server 2 is a communication interface unit 201, a communication control unit 211, a log storage unit 206, a combined load storage unit 207, and a means corresponding to a corresponding functional unit in the smartphone 1 (FIG. 2). And a deep learning unit 22.

  Further, the smartphone 3 is also a communication interface unit 301 and a communication control unit 312, a positioning unit 302 and a position information acquisition unit 311, and a user interface, which are equivalent to corresponding functional units in the smartphone 1 (FIG. 2). User IF unit 303 and user IF control unit 315, a feature amount generation unit 313, and a prediction result processing unit 314.

  That is, the present data prediction system is an embodiment in which a deep learning function in the smartphone 1 shown in FIG. 2 is provided in an external server, and this server is used as a data prediction device. Here, the feature vector generated by the smartphone 3 is transmitted to the data prediction server 2. On the other hand, the prediction value generated by the data prediction server 2 is transmitted to the smartphone 3, converted into a prediction result by the prediction result processing unit 314 of the smartphone 3, and the prediction result is presented to the user via the user IF. . Furthermore, the prediction result processing unit 314 of the smartphone 3 generates teacher data based on a positive or negative evaluation result by the user input via the user IF, and the teacher data is used for re-learning in the data prediction server 2. May be.

  Here, a form in which the feature prediction unit 313 and / or the prediction result processing unit 314 is provided in the data prediction server 2 instead of the smartphone 3 is also possible. In any case, in the data prediction system configured as described above, a feature vector having elements of target data values related to time points or time intervals belonging to different “at least two time period hierarchies” is adopted. By doing so, it is possible to further reduce the amount of calculation for the learning process as compared with the conventional one while ensuring a considerable prediction accuracy.

  In addition, for example, the present data prediction system is applied to a rental car service business, and a rental car user can use a smartphone owned by the rental car as a car navigation apparatus familiar to him / her. At this time, a deep learning unit 22 constructed for each user that performs learning based on the movement history and teacher data for each user is prepared in the data prediction server 2 installed in the call center, and the calculated predicted movement destination (predicted value) ) Can be downloaded to the rental car used by the corresponding user, and this prediction result can be used.

  Furthermore, an embodiment in which a data prediction device equivalent to the data prediction server 2 is incorporated in an automobile (moving body) is also possible. For example, when a control CPU mounted on an automobile implements the data prediction function of the present invention in parallel with other standard functions, a unit in the automobile including the CPU is included in the data prediction apparatus of the present invention. That will be true. In this case, it is also preferable that the in-vehicle unit and a smartphone or the like brought into the vehicle by the user are connected for communication by wire or wirelessly.

  FIG. 4 is a schematic diagram illustrating an embodiment of learning / data prediction processing in the deep learning unit 12.

  As shown in FIG. 4, the deep learning unit 12 configures an RNN by an input layer 121i, a hidden layer 121h, a context layer 121c, and an output layer 121o. The RNN is a very suitable network for handling time series data. Hereinafter, for the purpose of easily explaining the functions of the deep learning performed, the deep learning unit 12 will be described as a single-layered Elman network that is a kind of RNN.

  Originally, a neural network (NN) is a model that mimics the mechanism of a human brain neural circuit, and uses an input layer, a hidden layer (intermediate layer), and an output layer as model components. Each of these layers includes a plurality of neurons (Neuron) corresponding to nerve cells in the brain. The basic function of each neuron is signal input and output. However, when a signal is transmitted from a neuron in one layer to a neuron in the next layer, the signal output from the neuron in one layer is not directly input to the neuron in the next layer, but is output to the output signal. Weighting is performed with connection weights, and a signal is transmitted to a neuron in the next layer only when the sum of the weighted output signals exceeds a threshold set for each neuron. Here, by calculating, updating, and applying the connection weight between these neurons from learning using teacher data, it is possible to output predicted values when new data to be predicted is input. is there.

  The Elman network (deep learning unit 12) illustrated in FIG. 4 is a deep layer NN including a loop structure, and includes an input layer 121i, a hidden layer 121h, a context layer 121c, and an output layer 121o. The input layer 121i has n input layer neurons ni that process a feature vector that is an input signal. The number n of the input layer neurons ni matches the number of dimensions (number of elements) of the feature vector. The context layer 121c includes m context layer neurons nc that replicate the state of the immediately preceding hidden layer 121h. The number m of the context layer neurons nc matches the number of hidden layer neurons nh described below.

  The hidden layer 121h includes m hidden layer neurons nh that process input signals from the input layer neuron ni and the context layer neuron nc. The number m of hidden layer neurons nh is arbitrary and is one of the network design parameters. The output layer 121o includes an output layer neuron no that processes an input signal from the hidden layer neuron nh. The number k of the output layer neurons no coincides with the number of dimensions of the prediction vector constituting the prediction value. For example, if the predicted value is only the position information of the movement destination (predicted location), k = 1 can be set.

  The neurons described above have their activation functions, and hold different parameters for each neuron. The neuron, for example, performs a learning process by backpropagation, adjusts the parameters to be held, determines the expression of the activation function, and calculates / updates the above-described connection weight. This learning process is repeated many times. That is, the error calculated from the predicted data (learned value) and the teacher data (actually measured value) output as a result of data input is propagated backward to fine-tune the above parameters, and the next data is input As a result, the parameter fine adjustment process is continued until the error is minimized (minimum) and converged, such that the error due to the output data is propagated backward again and the parameter is finely adjusted again. Next, in the prediction process, for example, a predicted value, for example, a destination (predicted location position) is obtained by inputting an immediate status value such as the current position to be measured and the date and time thereof to the NN (deep learning unit 12). It is output.

  Since the RNN (Elman network) shown in FIG. 4 has the above-described configuration, it has a feature that is relatively close to a living body (brain), and has a behavior pattern of a human (brain) that has a strong correlation with the past. Very suitable for prediction. Specifically, the hidden layer neuron nh updates the state while interacting with the hidden layer neuron nh at the adjacent stage via the context layer neuron nc, so that the behavior pattern to be predicted is further increased. It is reflected in the combined load comprehensively and accurately. As described above, since the deep learning unit 12 of FIG. 4 is a recurrent type, it is possible to perform multi-layer deep learning by expanding and circulating the hidden layer and the context layer to function. Further, by repeating this development / circulation process, it is possible to recognize a periodic pattern other than the expected time period.

  Note that the number of hidden layer neurons nh (number of units of the hidden layer 121h) m, the number of repetitions of the learning process, and the adjustment range (learning rate) in parameter fine adjustment are network design items and are determined as empirical values. May be.

  Hereinafter, specific contents of Step 1 to Step 7 in the data prediction method shown in FIG. 4 will be described in detail with reference to FIGS.

[Data prediction method]
FIG. 5 is a schematic diagram showing an embodiment of generating a feature vector according to the present invention.

  FIG. 5A shows an embodiment of the feature quantity vector input to the input layer 121i in step 1 (FIG. 4). This feature vector is obtained by using time-series data of the user's past movement history (for example, time-series data in which the date / time and the location (starting place) at the date / time are associated with the destination) It is generated by analyzing where it moved from where.

Specifically, the feature vector of this embodiment is:
1. 1. Date of departure of the target travel history 2. Day of departure of target travel history 3. Weekdays and holidays at the departure of the target travel history 4. Time zone classification at the time of departure of the target movement history 5. Location classification at the start of the target movement history Destination of time series based on multiple time period hierarchies
6a ・ Destination on the day (immediately before)
6b · Destination of the same time zone one day ago (time zone 4 above)
6c · The destination of the same time zone 2 days ago
6d 3 days before the same time zone
6e · Destination of the same time zone one week ago
6f · Travel destination in the same time zone 2 weeks ago
6g · Travel destination in the same time zone 3 weeks ago
6h · 1 month before the same time zone
6i · Destination of the same time zone 2 months ago
6j · The destination of the same time zone 3 months ago
6k · 6 months before the same time zone
6l · It has a vector element with the destination of the same time zone one year ago. The number of these vector elements coincides with the number n of input layer neurons ni. Specifically, the number n of vector elements is the above items 1. to 5. And the above item 6. P (= 12) in the sum,
n = 5 + p = 17
It has become. Incidentally, the value of the vector element is preferably normalized to a value from 0 to 1.

(A) Here, in the feature vector elements 6a to 6d, the first time unit is “day”, and the first time period hierarchy is composed of a predetermined number (seven) “days”. It is a periodic hierarchy related to “week”.
(B) In the feature vector elements 6e to 6g, the second time unit is “week”, and the second time period hierarchy is composed of a predetermined number of “weeks” (including fractions). It is a periodic hierarchy related to “Month”.
(C) Further, in the feature vector elements 6h to 6k, the third time unit is “month”, and the third time period hierarchy is composed of a predetermined number (12) of “months”. It is a periodic hierarchy related to “year”.

Thus, for the time unit and time period hierarchy set when generating the feature vector,
(A) A first ordinal number composed of a first time unit, a second time unit composed of a predetermined number of first time units, and then a predetermined number of time units one order before. A time unit incremented by 1 is set,
(A) As the time period hierarchy, a first time period hierarchy whose period is the time according to the first time unit, a second time period hierarchy whose period is the time according to the second time unit, and thereafter At least two time period hierarchies are adopted among the time period hierarchies whose period is a time unit in which the ordinal number is incremented by 1 sequentially.

  Of course, the mode of the time unit and time period hierarchy set when generating the feature vector is not limited to the above (a) to (c). For example, “season” (spring / summer / autumn / winter) may be adopted as a time unit, and the corresponding time period hierarchy may be “year” composed of “season”. It is also possible to adopt “half year” as the time unit and set the corresponding time period hierarchy to “year” or “10 years” composed of “half year”.

  Further, the target data related to the same time zone (time zone 4 above) belonging to the first time period hierarchy “week” in the specific feature vector described above is four (6a, 6b, 6c and 6d) ( Multiple) is set. These are data related to a part of “day” (the previous day, the day before two days, and the day before three days) among all (seven) “days” constituting the first time period hierarchy “week”. . That is, as the input data for reflecting the periodicity related to the first time period hierarchy “week” in the prediction, the target data for all seven days in one week is not handled, but a part of the data Is adopted.

  Similarly, three (plural) target data items 6e, 6f, and 6g are set for the same time zone (time zone 4 above) belonging to the second time period hierarchy “month”. These are data relating to a part of “weeks” (weeks 1, 2, and 3 weeks before) of all “weeks” constituting the second time period hierarchy “month”. That is, as input data for reflecting the periodicity related to the second time period hierarchy “month” in the prediction, not all the target data for 4 to 5 weeks in one month are handled, but a part thereof. The data is adopted.

Summarizing the adoption of such target data, a plurality of target data values related to time points or time intervals belonging to each of the time period hierarchy are as follows:
(A) a target data value group related to a part of time units among all the first time units constituting the first time period hierarchy;
(A) a target data value group related to a part of time units among all the second time units constituting the second time period hierarchy, and thereafter
(C) It is supposed to include at least two of the target data value groups related to some time units among all the time units constituting the time period hierarchy in which the ordinal number is incremented by 1. . As a result, the feature amount vector has elements composed of discrete time history data with nonuniform time intervals.

  In this way, by limiting the number of target data belonging to each time period hierarchy discretely, the period related to each time period hierarchy with a smaller amount of calculation than when all target data is continuously adopted It becomes possible to learn sex. That is, instead of learning using all continuous time-series data, for example, by learning with discrete data of the same period of the previous month or the month after the previous month, the calculation time for prediction can be shortened. Practical processing can be realized.

  In addition, the target data values relating to the first time unit “day”, the second time unit “week”, and the third time unit “month” in the specific feature vector described above are predetermined in the time unit. The value in “time” or “time zone” of the above (in the present embodiment, the time zone of 4. above) is taken. When these settings are generalized and summarized, the target data values relating to some time units among all the first time units constituting the first time period hierarchy, the second time period hierarchy is constituted. The target data value related to a part of time units in all the second time units, and a part of all time units constituting the time period hierarchy in which the ordinal number is sequentially incremented by one thereafter. The target data value relating to the unit takes a value in the basic time unit (in this embodiment, “time zone”) constituting the first time unit in the time unit constituting the time period hierarchy relating to the unit. .

  In this way, for the target data belonging to each time period hierarchy, data that can be effectively compared can be selected discretely by unifying the values in basic time units. It becomes possible to learn the periodicity related to the periodic hierarchy.

  In addition, the temporal discrete degree (discrete interval) and the number of vector elements n (number of dimensions) in each data value that is an element of the feature vector can be set in accordance with the use of the prediction process and the usage target. In addition, in order to realize a more accurate prediction, as part of the vector element, weather information (for example, distinction between sunny, cloudy, rainy, etc.), traffic or operation information (for example, road congestion at the current location) It is also possible to employ the presence or absence or degree).

  As described above, in the generation of feature vectors, it is possible to effectively learn (especially human) behavior patterns by adopting discrete target data values having a predetermined time interval for each of a plurality of time period hierarchies. Thus, it is possible to achieve more accurate prediction of data values. Here, human behavior often occurs at predetermined time intervals or has a pattern including a plurality of periodicities. For example, it is common to have a tendency to take a predetermined action at a specific day of the week or a specific day of the month, or at a predetermined number of days or a predetermined monthly interval, for example, to go to a predetermined place. Therefore, by inputting the feature vector composed of the target data values as described above to the RNN that performs feedback coupling processing that mutually influences the past input and the current input, the time not considered in the feature vector. It is also possible to identify behavior patterns related to intervals and periodicity and to perform more accurate prediction.

Here, with reference to FIG. 5B, a specific mode for determining the elements of the feature quantity vectors related to a plurality of time period layers will be described.
(S51, S52) It is determined whether or not the time H related to the time-series data is closer in time than the time point 7 days ago when viewed from the prediction time point (date and time) when the prediction is performed. If a true determination is made here, of these time-series data, data at the same time zone (destination time) on the current day, 1 day before, 2 days before and 3 days before the time of prediction (destination data) Is used to generate a time-series feature vector.

(S53, S54) On the other hand, when a false determination is made in step S51, the time H related to the time-series data is closer in time than the time point one month ago when viewed from the prediction time point (date and time) when the prediction is performed. It is determined whether or not. If the true determination is made here, data in the same time zone (as the prediction time point) one week before, two weeks before and three weeks before the time point of the prediction among these time series data (data of the destination) ) To generate a time-series feature vector.
(S55, S56) On the other hand, when a false determination is made in step S53, the time H related to the time-series data is 12 months (one year) ahead of the time point 12 months (one year) before the prediction time point (date and time) for prediction. It is determined whether it is close in time. If true judgment is made here, data in the same time zone (as the forecast time) one month before, two months before, and three months before the time of forecasting among these time series data (data of the destination) ) To generate a time-series feature vector.

(S57) On the other hand, if a false determination is made in step S55, among these time-series data, at the same time zone (predicted time point) one year before, two years ago and three years before the time of prediction Data (destination data) is adopted to generate a time-series feature vector.

  6 and 7 are schematic views showing program codes for executing one embodiment of each step shown in FIG.

  Hereinafter, each step of the deep learning process shown in FIG. 4 will be described using FIGS. 6 (A) and 6 (B) and FIGS. 7 (A) and 7 (B) in sequence. An outline of the learning process will be described. First, learning is performed by inputting a feature amount input as a learning target from the input layer 121i to the output layer 121c via the hidden layer 121h. In each layer, the input feature value data is multiplied by a connection weight (weighting factor) in the neurons in the layer and output to the neurons in the subsequent layers. A mechanism for predictive learning of an input obtained at the next time as a result of such selective processing is generally called feedforward learning.

  In contrast to this feedforward learning, backpropagation learning is a mechanism for updating the connection weight by the error back propagation method. Here, the error back propagation method refers to an error between a value output from the output layer 121c (for example, a predicted value of the destination) and an actually obtained actual value from the output layer 121c via the hidden layer 121h. In this method, feedback is made toward the input layer 121i and adjustment is performed so as to minimize. Specifically, as shown in FIG. 4, in feedforward, data moves in this direction sequentially in the input layer, hidden layer, and output layer, and in backpropagation, the data is in the opposite direction to feedforward. Moving.

  Hereinafter, each step shown in FIG. 4 will be described in detail with reference to FIGS. In step 1, a feature vector generated according to the state (immediate state) at the time of prediction is input to the input layer 121i.

  FIG. 6A shows an example of program code for executing the feedforward process (input layer + context layer → hidden layer) in step 2 (FIG. 4). Here, the output inputs [inp] of the input layer neuron are input to the hidden layer neuron together with the output context [con] of the context layer neuron which is a duplicate of the immediately preceding hidden layer neuron. Here, the input layer neuron output inputs [inp] is weighted and combined with the connection load (first connection load) wih [inp] [hid] between the input layer and the hidden layer. Furthermore, the context layer neuron output context [con] is weighted with a connection weight (third connection weight) wch [con] [hid] between the context layer and the hidden layer, and is summed. The value obtained by subtracting the threshold value set for each unit (neuron) from the sum of these inputs is nonlinearly transformed by the sigmoid function (f (x) = 1 / (1 + exp (x)), and the next unit ( The initial values before learning in the connection weights wih [inp] [hid] and wch [con] [hid] can be set to random values, for example.

  Next, FIG. 6B shows a program for executing the feedforward process (hidden layer → output layer) in step 3 (FIG. 4) and the copy process (hidden layer → context layer) in step 4 (FIG. 4). An example code is shown. Here, the connection weight (second connection weight) who [hid] [out] between the hidden layer and the output layer is weighted to the output hidden [hid] of the hidden layer neuron, and the sum is obtained. The value obtained by subtracting the threshold value set for each unit (neuron) from the sum of the inputs is nonlinearly transformed by the sigmoid function, and the output actual [out] of the output layer neuron is generated. Note that the initial value before learning in the connection weight who [hid] [out] can also be a random value, for example.

  Further, the hidden [hid] of the hidden layer neuron is output to the output layer neuron as described above, while being duplicated as the output context [con] of the context layer neuron.

  Next, FIG. 7A shows an example of program code for executing the error calculation process in step 5 (FIG. 4). Here, the error actual erro [out] is calculated by comparing the output actual [out] (predicted value) derived in steps 1 to 4 obtained by feed-forwarding the input with the actually measured value (correct answer) target [out]. ing. Here, for example, in the case of prediction of the destination of a car (user), the value of the set destination input and set by the user at the time of departure may be used as the actual measurement value target [out]. Moreover, it is good also as a value at the time of deriving the location position (arrival position) when an engine is stopped by the positioning part.

  Next, FIG. 7B shows an example of program code for executing the back-propagation process in steps 6 and 7 (FIG. 4). Here, the pack propagation process is performed on the error erro [out] acquired in step 5 to perform learning. Specifically, the error erro [out] calculated from the predicted value actual [out] and the correct target [out] is traced back from the output layer to the input layer and propagated to be smaller. The connection weights who [hid] [out], wih [inp] [hid], and wch [inp] [hid] are updated. This back-propagation process is repeatedly executed, and this update is also repeatedly performed.

  Note that the amount of calculation can be greatly reduced by performing relearning only on the error erro [out] in this way.

  Here, in the back-propagation process, the minimum value of the error is searched by the steepest descent method, but in general, the learning speed becomes higher as the learning coefficient (learnRate in FIG. 7B) is set larger. . However, when the error between the predicted value and the correct answer is viewed as a function of the combined weight, if the learning coefficient is too large, the error increases or vibrates. On the other hand, as the learning coefficient is set smaller, problems such as vibrations are less likely to occur, but the correction amount is reduced and the learning speed is further reduced. As described above, the learning speed, that is, the speed of convergence and the prediction accuracy are greatly affected by how the learning coefficient is set. In contrast, the present invention inputs a combination of relatively close time-series data and distant time-series data when viewed from the prediction time point when the prediction is performed, and these time-series data act differently on learning. Using this phenomenon, the trade-off between learning speed and prediction accuracy is improved.

  Further, in the backpropagation process according to the present invention, it is possible to further improve the prediction accuracy by searching for and determining the optimum value of the learning coefficient for each time period hierarchy using the test data.

  Further, the above-described embodiments of Steps 1 to 7 reflect the information of both the input layer and the output layer in the prediction by duplicating (mapping) the context layer peculiar to the RNN as the intermediate layer of the first-order lag. The pattern of the target data is identified covering the history. As a result, for example, an unknown pattern including a defect can be accurately recognized.

  Moreover, in embodiment of step 1-7 mentioned above, for example, by inputting the predicted moving destination which is the predicted value output from the output layer into the input layer as new input data, and executing learning again, It is possible to predict the destination at a point further ahead of the point at which the destination is predicted. Similarly, by repeating this re-input process, it is possible to predict a continuous movement destination within a predetermined time, for example, a movement route.

[Example of data prediction]
FIG. 8 is a table showing an embodiment and a comparative example of the data prediction method according to the present invention.

  As shown in FIG. 8, in the experiment including this example, in order to verify the effect (prediction accuracy and learning time) of this example, the conventional method (time period hierarchy is introduced for the history data for the experiment. The data prediction process which employ | adopted the method using no time series data) was implemented together, and it was set as Comparative Examples 1-3. In this example and Comparative Examples 1 to 3, a MacBook Pro Mid 2012 (built-in 2.3 GHz Intel Core i7), which is a PC terminal, was used as a personal computer (PC) 4 as a data prediction device.

  In this example and Comparative Examples 1 to 3, the learning object was 1231 records in a predetermined three years, and the number of learnings was 20000. The number of dimensions of the feature vector, that is, the number of neurons (units) in the input layer was 24 in this example, and 24, 48, and 96 in Comparative Examples 1 to 3, respectively. In addition, the number of neurons (units) in the hidden layer was 16 in this example, and 16, 32, and 64 in Comparative Examples 1 to 3, respectively. Furthermore, the learning coefficient (learnRate) is set to 0.2 in all the time period layers in the present embodiment, and is also set to 0.2 in Comparative Examples 1 to 3. Moreover, the latest 100 records were used for the input for evaluating the error estimation rate.

  As shown in the table of FIG. 8, in this embodiment, a very low false estimation rate (4%) equivalent to Comparative Example 3 in which the number of neurons in the input layer (number of feature dimensions) is four times is realized. is made of. Moreover, the learning time (76.1 cpu sec) of this example is very short, about 1/15 of that of Comparative Example 3. This is a high learning speed equivalent to that of Comparative Example 1 that has the same number of input layer neurons (number of feature dimensions) as that of the present embodiment, but the error estimation rate has deteriorated to 33%. From this, it can be understood that the present embodiment achieves both higher prediction accuracy and higher learning speed as compared with the prior art.

  As described above in detail, according to the present invention, since the feature value vector having the feature data element as the feature value element for each of the “at least two time period hierarchies” different from each other is adopted, the target data It is possible to efficiently learn various periodicities and changes thereof. As a result, it is possible to further reduce the amount of calculation for the learning process as compared with the prior art while ensuring a considerable prediction accuracy.

  In addition, the data prediction process according to the present invention is naturally not limited to execution by a smartphone or a car navigation device brought into the automobile. For example, even when a smartphone user moves by means other than an automobile such as walking, bicycle, train, etc., the present invention can be applied to predict the destination. Furthermore, the prediction target of this data prediction process is naturally not limited to the destination, and can be used in fields such as information pushes such as news that the user expects or desires, and prediction of exchange rates and stock prices. Is possible. In any case, the mechanism for handling feature quantities based on different time period hierarchies, which is a feature of the present invention, can be applied and extended to data prediction in various fields related to time-series data.

  Various changes, modifications, and omissions of the above-described various embodiments of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

1, 3 Smartphone (data prediction device)
101, 201, 301 Communication interface unit 102, 302 Positioning unit 103 Touch panel display (TP / DP)
104 Speaker 105 Microphone 106, 206 Log storage unit 107, 207 Bond load storage unit 108 Content storage unit 111, 311 Position information acquisition unit 112 Situation acquisition unit 113, 313 Feature value generation unit 114, 314 Prediction result processing unit 115, 315 User Interface (IF) control unit 115c Dialog processing unit 12, 22 Deep learning unit 121 Input / hidden layer unit 121i Input layer 121h Hidden layer 121c Context layer 122 Output layer unit 122o Output layer 2 Data prediction server (data prediction device)
211, 312 Communication control unit 303 User IF
4 PC (data prediction device)
5 GPS satellite 6 Weather information server 7 VICS (registered trademark) server 8 Cradle

Claims (14)

A device comprising a neural network (NN) for learning with time-series target data, and predicting the target data at a certain time point or time range,
A first combination of an input layer related to NN, which inputs a target data value related to a time point or time interval belonging to each of at least two different time period layers as a feature quantity element, and a signal output from at least the input layer A first processing signal generated by weighting with a load is input and a hidden layer related to NN is input, and a signal output from the hidden layer is generated by weighting with a second combined weight. Input / hidden layer means for outputting a second processing signal;
An output layer means having an output layer according to NN, which inputs the second processing signal and outputs at least the value of the target data at the time point or time range related to the prediction;
The input / hidden layer means and the output layer means can execute relearning to update the first coupling weight and the second coupling weight to reflect the periodicity related to at least two of the time period layers. A data prediction apparatus characterized by the above. The input / hidden layer means further includes a context layer related to NN as a layer for outputting the same signal as the hidden layer,
The hidden layer inputs a third processing signal generated by weighting a signal output from the context layer with a third coupling weight as a time-delayed signal together with the first processing signal,
The input / hidden layer means and the output layer means are capable of executing re-learning to update a third coupling weight to reflect periodicity related to at least two time period layers. The data prediction apparatus according to 1.   3. The input / hidden layer means and the output layer means update the coupling load reflecting the periodicity related to at least two time period layers by backpropagation. The data prediction apparatus described.   4. The input / hidden layer means and the output layer means execute re-learning using data not yet used for learning among the time-series target data. The data prediction apparatus according to item 1. A first time unit, a second time unit composed of a predetermined number of first time units, and then a predetermined number of time units one order before, sequentially incrementing the ordinal number by 1. Set the time unit, and
At least two of the time period hierarchies are a first time period hierarchy whose period is a time according to a first time unit, a second time period hierarchy whose period is a time according to a second time unit, and 5. The method according to claim 1, further comprising: at least two time period hierarchies among time period hierarchies having time periods according to time units obtained by sequentially incrementing the ordinal number by one. 6. The data prediction apparatus described.   The first time unit is “day”, the first time period hierarchy is a period hierarchy related to “week”, the second time unit is “week”, and the second time period hierarchy is “month”. The data prediction apparatus according to claim 5, wherein the third time unit is “month” and the third time period hierarchy is a period hierarchy related to “year”. .   A plurality of target data values related to time points or time intervals belonging to each of the time period hierarchies are target data values related to some time units among all first time units constituting the first time period hierarchy. A group, a target data value group related to a part of time units among all the second time units constituting the second time period hierarchy, and a time period hierarchy in which the ordinal number is sequentially incremented by 1 The data prediction apparatus according to claim 5 or 6, comprising at least two groups among target data value groups related to a part of time units among all the time units to be configured.   Target data values related to some time units in all first time units constituting the first time period hierarchy, one of all second time units constituting the second time period hierarchy The target data value related to the time unit of the part, and the target data value related to some time units among all the time units constituting the time period hierarchy in which the ordinal number is sequentially incremented by 1 are related to itself. The data prediction apparatus according to claim 7, wherein a value in a basic time unit constituting a first time unit in a time unit constituting a time period hierarchy is taken. The first time unit is “day”, the first time period hierarchy is a period hierarchy related to “week”, the second time unit is “week”, and the second time period hierarchy is “month”. , The third time unit is “month”, and the third time period hierarchy is a period hierarchy related to “year”,
The target data value relating to the first time unit, the second time unit, and the third time unit takes a value in a predetermined “time” or “time zone” in the time unit. 8. The data prediction apparatus according to 8. An interface capable of inputting an evaluation result for the value of the target data as a prediction result output from the output layer means;
10. The input / hidden layer means and the output layer means execute re-learning using teacher data generated based on the acquired evaluation result. The data prediction apparatus described in 1. The target data value related to the time point or time interval belonging to each of the time period hierarchy is a data value related to the reference position of the movement prediction and the position of the movement destination in the prediction target at the time point or time interval,
11. The value of the target data at the time point or time range related to prediction output by the output layer means is a predicted value of the position of the movement destination in the prediction target. The data prediction apparatus according to item 1.   A positioning unit that can acquire information related to the location by GPS (Global Positioning System), and a feature amount input by the input layer based on the acquired information related to the location and the relearning The data terminal according to claim 11, wherein the information terminal is capable of generating at least one of teacher data used for the communication. A program comprising a neural network (NN) for learning with time-series target data and functioning a computer mounted on a device for predicting the target data at a certain time point or time range,
A first combination of an input layer related to NN, which inputs a target data value related to a time point or time interval belonging to each of at least two different time period layers as a feature quantity element, and a signal output from at least the input layer A first processing signal generated by weighting with a load is input and a hidden layer related to NN is input, and a signal output from the hidden layer is generated by weighting with a second combined weight. Input / hidden layer means for outputting a second processing signal;
The second processing signal is input, and at least the value of the target data at the time point or time range related to the prediction is output, and the computer functions as output layer means having an output layer related to NN,
The input / hidden layer means and the output layer means can execute relearning to update the first coupling weight and the second coupling weight to reflect the periodicity related to at least two of the time period layers. A data prediction program characterized by A data prediction method in an apparatus that includes a neural network (NN) that learns with time-series target data and predicts the target data at a certain time point or time range,
Inputting a target data value related to a time point or a time interval belonging to each of at least two different time period layers as a feature quantity element to an input layer related to NN;
A first processed signal generated by weighting at least a signal output from the input layer with a first coupling weight is input to a hidden layer related to the NN, and a signal output from the hidden layer is set to a second value. Outputting a second processing signal generated by weighting with the combined weight of:
Inputting the second processed signal to the output layer related to the NN, and outputting at least the value of the target data in the time point or time range related to the prediction;
And re-learning to update the first connection weight and the second connection weight to reflect the periodicity related to at least two of the time period layers.

Images