Prediction device, program, and method for predicting user movement trajectory using sequence prediction model

Abstract

To provide a prediction device or the like capable of predicting a user's movement locus by considering temporal periodicity based on human social activities and using a series prediction model.SOLUTION: Series data is composed of a plurality of pieces of period data, and each piece of period data is composed of a plurality of tokens. A prediction device sets the maximum number of tokens including the plurality of pieces of period data which include cycle period data including a plurality of pieces of period data for one cycle, and subsequent period data following the cycle period data as a prescribed length of series data. A prediction device includes: padding means for filling a rear part that is less than a prescribed length with in padding for each series data; and prediction control means for inputting a token string of a prescribed length including the cycle period data and at least one beginning token of the subsequent period data for each series data filled with padding to a series prediction model, and predicting a token string of a prescribed length obtained by incrementing a token of the subsequent period data following the cycle period data by 1 and removing the last padding.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a sequence prediction model technology, and is particularly suited to predicting a user's movement trajectory. [Background technology]

[0002] Conventionally, there is a technology that uses a deep learning neural network to train a large number of location movement trajectories collected from smartphones owned by many users (for example, see Non-Patent Document 1). According to this technology, a position sequence for each day is trained using LSTM (Long Short Term Memory) to predict movement trajectories. This makes it possible to predict the movement trajectory of an arbitrary user for a day from the movement trajectory in the morning, for example. This technology can also be used to forecast traffic demand. For example, based on the movement trajectories of a sample user in a given area, it is possible to predict and aggregate the movement trajectories of subsequent users and predict road traffic volume. For example, it is possible to predict evening traffic congestion based on the movement trajectories of many smartphones in the morning.

[0003] There is also a technique for estimating the length of stay of a user for each region (see, for example, Non-Patent Document 2). According to this technique, a location is represented by BERT (Bidirectional Encoder Representations from Transformers) and prediction is performed using LSTM. BERT is based on a Transformer for pre-training of natural language processing. By representing a location using BERT incorporating time, it is possible to predict how often and for how long a user will stay at each location.

[0004] Furthermore, there is a technology that can estimate a user's significant location without activating the positioning function of a mobile terminal (see, for example, Patent Document 1). According to this technology, base station location information with coarse spatial granularity and non-constant time intervals that can be obtained by a communication carrier's equipment is used. A connection history including a large amount of data on date, time, and base station identifiers is constantly collected from the base station under which the mobile terminal is located, and is divided for each mobile terminal. Next, the connection history is divided into a predetermined time window (time interval), and the number of occurrences of each base station identifier is counted for each time window. Next, for each time window, a vector is generated with multiple base station identifiers as elements and the number of occurrences as values, and the vectors are aggregated into clusters. Then, for each cluster, the significant zone of the user is estimated based on the number of occurrences of each base station identifier in the vector included in the cluster. This technology regards a sequence of location information as text, and regards a location that appears within a time frame as a document. Whether or not a location exists in a document is trained by expression learning of unsupervised learning. However, location prediction (i.e., generation of a sentence) is not taken into consideration, and it is limited to expressing the significant zone of the user.

[0005] Another conventional technique is the GPT2 (Generative Pre-trained Transformer - 2) model as a sequence prediction model (see, for example, Non-Patent Document 3). This is a supervised learning language generation model that predicts the next word (next word prediction) of a word sequence (sentence) extracted from a large corpus. When an arbitrary word sequence is input, the GPT2 model can sequentially predict the next word, and can generate sentences that sound like they were written by a human. For example, when the word sequence "today" and "good" are input, the word "weather" with the highest probability as the next word is output. Next, when "today", "good", and "weather" are input, the word "is" with the highest probability as the next word is output. Sentence: "Today is good weather."

[0006] FIG. 1 is an explanatory diagram showing sequence data of the GPT2 model in the prior art. We assume that there are multiple texts (sentences) as training data, and the sequence data is the concatenation of these multiple texts.

[0007] FIG. 2 is an explanatory diagram showing the process of training the GPT2 model in the prior art.

[0008] GPT2 is based on Transformer and has an attention mechanism (scoring the relevance of tokens to each other). The attention of each token is always applied to the token before it. This allows each token to learn which part of the previous token to focus on to infer the next token.

[0009] According to FIG. 2, the GPT2 model divides the sequence data constructed in FIG. 1 into a predetermined length (e.g., 40) that can be input, and trains by associating the input data with the correct answer data. Input data sequence: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 0, 1, 19, 0 Correct data sequence: 1,2,3,4,5,6,7,8,9,0,1,2,·······19,0,1 In this way, in the GPT2 model, the correct answer data is obtained by deleting the first token and adding the last token to the input data. The sequence data is input to the GPT2 model while shifting it, and the resulting data sequence is predicted. The difference between the resulting data sequence and the correct answer data sequence is expressed as a loss function (generally called cross entropy loss) and training is carried out.

[0010] According to the above-mentioned example of input data and correct answer data, in order to generate token 1, it is sufficient to focus only on token 0 (to begin with, there is only token 0 to focus on before token 1). Also, in order to generate token 3, it is sufficient to focus on token 1 by the attention mechanism (tokens 0 and 2 can be ignored), and so on.

[0011] In this way, for the sequence data of FIG. 1 described above, the GPT2 model of FIG. 2 can proceed with training while deleting the first 1 token and adding no token to the end.

Prior Art Documents

Patent Documents

[0012]

Patent Document 1

Non-Patent Documents

[0013]

Non-Patent Document 1

Non-Patent Document 2

Non-Patent Document 3

[0014] According to the above-mentioned conventional technology, LSTM (conceptually RNN (Recurrent Neural Network)) is used for supervised learning, but the longer the sequence data, the more likely it is that early data will be forgotten, making it difficult to understand the entire sequence data. In addition, since calculations are performed sequentially and recursively, parallelization using a GPU (Graphics Processing Unit) becomes difficult. Therefore, the longer the sequence data, the slower the processing speed becomes. As such, since LSTM has a limit to the sequence length, only location information on a daily basis, for example, can be used when predicting a movement trajectory.

[0015] The aforementioned GPT2 can also be applied to predict the user's movement trajectory. In that case, the user's movement unit corresponds to the text of GPT2, and the user's position corresponds to the token of GPT2. Even the position can be applied by considering it as equivalent to a word string.

[0016] However, user movement trajectories have periodicity. Human social activity has periodic units such as a week or a day. If GPT2 is applied as is without considering such periodicity, a problem occurs in that it continues to make predictions ignoring human social activity, such as predicting daytime behavior from late-night behavior.

[0017] In addition, in the case of a user's movement trajectory, the sequence length cannot be specified in advance. Of course, the sequence length can be fixed as a known value, but it is not realistic to apply this to predicting diverse and real-time user positions.

[0018] That is, if the GPT2 data structure is applied as is to predict a user's movement trajectory and sequence data as shown in FIG. 1 is constructed, multiple time series positions are packed consecutively. If periodicity is ignored in this way, and there is no correlation between late-night movement trajectories and daytime movement trajectories, the daytime movement trajectories will be predicted from the late-night movement trajectories. Note that in the case of language generation tasks in the prior art, a large amount of long sentences are often trained, and periodicity is not an issue.

[0019] Therefore, an object of the present invention is to provide a prediction device, program, and method for predicting a user's movement trajectory using a sequence prediction model, taking into account temporal periodicity based on human social activity. [Means for solving the problem]

[0020] According to the present invention, a prediction device for predicting the next token by inputting a token string of a predetermined length, the device comprising: Data for multiple periods has periodicity as a time series, A maximum number of tokens including the periodic period data including a plurality of period data for one period and the subsequent period data following the periodic period data is set as a predetermined length of the sequence data; a padding means for filling a portion of each of the sequence data that does not reach a predetermined length with padding; a prediction control means for inputting a token sequence of a predetermined length including cyclic period data to the sequence prediction model for each padded sequence data, and predicting a token sequence of a predetermined length in which the tokens of the subsequent period data following the cyclic period data are incremented by one; The present invention is characterized by having the following.

[0021] As another embodiment of the prediction device of the present invention, The predictive control means is A token sequence of a predetermined length including cyclic period data is input, and a token sequence of a predetermined length is predicted by incrementing the first token of the subsequent period data following the cyclic period data by 1, or A token sequence including a periodic period data and one or more tokens of subsequent period data is input, and a token sequence of a predetermined length is predicted by incrementing the tokens following the periodic period data and subsequent period data by one. It is also preferable.

[0022] As another embodiment of the prediction device of the present invention, In the training phase, the sequence prediction model inputs a token sequence of a predetermined length while incrementing the tokens of the subsequent period data following the cycle period data by one, and repeats training by outputting the tokens of the subsequent period data following the cycle period data while incrementing them by one, and recognizes the end of one sequence data by padding. It is also preferable.

[0023] As another embodiment of the prediction device of the present invention, The period data consists of multiple tokens for a unit time. The periodic period data is made up of period data having a periodicity consisting of a plurality of unit times, The subsequent period data is made up of a plurality of tokens for a unit time following the periodic period data. It is also preferable.

[0024] As another embodiment of the prediction device of the present invention, For each period data, a starting special token is further included at the beginning, or an ending special token is further included at the end. It is also preferable.

[0025] As another embodiment of the prediction device of the present invention, The period data consists of multiple tokens for one day. The cycle period data consist of one week's worth of period data. Subsequent period data consists of multiple daily tokens following a one-week period data. It is also preferable.

[0026] As another embodiment of the prediction device of the present invention, The sequence data including multiple period data is divided by user, The tokens included in the period data are location information arranged in chronological order. It is also preferable.

[0027] As another embodiment of the prediction device of the present invention, The padding means, when a missing portion occurs in the period data in the periodic period data, fills the missing portion with padding so that the average number of tokens of the period data having the same periodicity as the missing portion becomes equal to the average number of tokens of the period data having the same periodicity as the missing portion. It is also preferable.

[0028] According to the present invention, a program for inputting a token string of a predetermined length and causing a computer to function to predict the next token includes the steps of: Data for multiple periods has periodicity as a time series, A maximum number of tokens including the periodic period data including a plurality of period data for one period and the subsequent period data following the periodic period data is set as a predetermined length of the sequence data; a padding means for filling a portion of each of the sequence data that does not reach a predetermined length with padding; a prediction control means for inputting a token sequence of a predetermined length including cyclic period data to the sequence prediction model for each padded sequence data, and predicting a token sequence of a predetermined length in which the tokens of the subsequent period data following the cyclic period data are incremented by one; The present invention is characterized in that it causes a computer to function.

[0029] According to the present invention, a prediction method of a prediction device for inputting a token string of a predetermined length and predicting the next token, the sequence data being made up of a plurality of period data, each period data being made up of a plurality of tokens, comprising: Data for multiple periods has periodicity as a time series, The prediction device is Data for multiple periods has periodicity as a time series, A maximum number of tokens including the periodic period data including a plurality of period data for one period and the subsequent period data following the periodic period data is set as a predetermined length of the sequence data; A first step of filling in any part of sequence data that does not reach a predetermined length with padding; a second step of inputting a token sequence of a predetermined length including the periodic period data for each padded periodic data into the sequence prediction model, and predicting a token sequence of a predetermined length in which the tokens of the subsequent period data following the periodic period data are incremented by 1; The present invention is characterized by carrying out the following steps. Effect of the Invention

[0030] According to the prediction device, program, and method of the present invention, it is possible to predict a user's movement trajectory using a sequence prediction model, taking into account the time periodicity based on human social activity. [Brief description of the drawings]

[0031] [Figure 1] FIG. 1 is an explanatory diagram showing sequence data of the GPT2 model in the prior art. [Diagram 2] FIG. 1 is an explanatory diagram showing the process of training a GPT2 model in the prior art. [Diagram 3] FIG. 2 is a functional configuration diagram of a prediction device according to the present invention. [Figure 4] FIG. 2 is an explanatory diagram illustrating sequence data in the present invention. [Diagram 5] FIG. 1 is a first explanatory diagram showing a process during training of a prediction device according to the present invention. [Figure 6] FIG. 2 is a second explanatory diagram showing the processing during training of the prediction device in the present invention. [Figure 7] FIG. 3 is a third explanatory diagram showing the processing during training of the prediction device in the present invention. [Figure 8] FIG. 13 is an explanatory diagram showing the inclusion of a starting or ending for each period data in sequence data. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0032] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0033] FIG. 3 is a functional configuration diagram of a prediction device according to the present invention.

[0034] 3, the prediction device 1 predicts sequence data and includes a sequence prediction model 10, a padding unit 11, and a prediction control unit 12. These functional components are realized by executing a program that causes a computer installed in the prediction device to function. The process flow of these functional components can also be understood as a method for predicting sequence data.

[0035] FIG. 4 is an explanatory diagram showing sequence data in the present invention.

[0036] The sequence data includes multiple "period data" and is classified for each user. The period data may be, for example, a day, or of course, an hour, a week, a month, or a year. In other words, it is a time unit that has periodicity in human social activities.

[0037] Each period data is composed of multiple "tokens" arranged in chronological order. In the present invention, a token is an information unit including a position or padding. Multiple tokens in one period data may represent a "position" of a user in chronological order. Multiple period data represent a periodic movement trajectory of a user.

[0038] 4, the period data is in units of one day. Specifically, for example, one week plus one day is expressed as follows: <Period data> <Token sequence (position in time series)> January 17, 2022 · Monday: 0,1,2,3 January 18, 2022 · Tuesday: 0,1,2 January 19, 2022 · Wednesday: 0,1,2,3,4,5 January 20, 2022 · Thursday: 0,1,2,3 January 21, 2022 · Friday: 0,1,2,3,4 January 22, 2022 · Saturday: 0,1,2,3,4,5 January 23, 2022 · Sunday: 0,1,2,3 January 24, 2022 · Monday: 0,1,2,3 The token sequence represents the user's movement trajectory. The location may be acquired by the user's smartphone or may be collected by a carrier's base station or access point.

[0039] The multiple period data are divided into "period period data" including multiple period data for one period, and "subsequent period data" following the period data. "Rolling Period Data" means data for the one-week period beginning on Monday, January 17, 2022 and ending on Sunday, January 23, 2022. "Subsequent Period Data" means the period data following the Cycle Period Data for Monday, January 24, 2022.

[0040] According to Figure 4, data for the period from Monday, January 17, 2022 to Monday, January 24, 2022 is linked. Here, the maximum number of tokens including "cyclic period data" including multiple period data for one period and the following "subsequent period data" is set as the predetermined length of the sequence data. In other words, the number of tokens that can be included in the cyclic period data and the subsequent period data is set as the predetermined length of the sequence data. This unifies the number of tokens in all the sequence data, and prevents a plurality of tokens constituting the cyclic period data and the subsequent period data from being divided into different sequence data.

[0041] According to the present invention, the "period data" consists of a number of tokens for a unit time. The "cyclical period data" is made up of period data having a periodicity consisting of a plurality of unit times. The "subsequent period data" consists of multiple tokens for a unit time following the cycle period data. According to Figure 4, the following settings are made depending on the application: "Period data" consists of multiple tokens for one day. "Cycle period data" consists of one week's worth of data. The "subsequent period data" consists of multiple tokens on a daily basis following the one-week period data. In the above example, the cyclical period data is described as one week from Monday to Sunday, but of course, it is not limited to this. For example, if the subsequent period data is Tuesday, the cyclical period data is one week from Tuesday to Monday.

[0042] [Padding section 11] The padding unit 11 fills the rear of each piece of sequential data with padding until the length reaches a predetermined length (for example, 40). The padding indicates the end of the sequential data.

[0043] According to FIG. 4, for example, sequence data with a predetermined length of 40 includes periodic period data (period data from Monday, January 17, 2022 to Sunday, January 23, 2022) and subsequent period data (period data from Monday, January 24, 2022). In addition, padding is added to the last four tokens that do not reach the predetermined length. In this way, the length of the sequence data is unified.

[0044] [Series Prediction Model 10] The sequence prediction model 10 may be, for example, a GPT2 model. According to the present invention, an existing language generation model is applied to predict a user's movement trajectory. By using a Transformer (Self Attention) based sequence prediction model, long sequence data can be handled. The sequence prediction model 10 is trained to input a token sequence of a predetermined length n=40 and output a token sequence that is the correct answer, and also recognizes the end of one piece of sequence data by padding.

[0045] The sequence prediction model 10 sets a loss function that reduces the error between the token sequence of the output data as prediction data and the token sequence of the correct answer data. The loss function may be, for example, a cross entropy error. This allows training to be performed so that the token sequence of the output data is trained to reduce the error between the token sequence of the correct answer data and the output data. In addition, when padding is input to a token of sequence data, the location of that token is trained so that the loss function is not calculated and no error backpropagation (feedback) is performed.

[0046] The predetermined length n of the sequence data depends on the sequence prediction model 10. For this reason, the combined length of the cycle period data and the subsequent period data must be equal to or less than the predetermined length n.

[0047] [Predictive Control Unit 12] The prediction control unit 12 inputs a token string of a predetermined length including “cyclic period data” to the sequence prediction model 10 for each piece of padded sequence data, and predicts a token string of a predetermined length in which the tokens of the “succeeding period data” following the cyclic period data are incremented by 1. Here, two predictions are made as follows. (1) A token string of a predetermined length including "cyclic period data" is input, and a token string of a predetermined length in which the "first token of the subsequent period data" following the cyclic period data is incremented by 1 is predicted. (2) A token string including "cyclical period data" and "one or more tokens of subsequent period data" is input, and a token string of a predetermined length in which the "cyclical period data" and the "token following the subsequent period data" are incremented by one is predicted. That is, the prediction control unit 12 repeatedly trains the sequence prediction model 10 in a training stage by inputting a token string of a predetermined length while incrementing the tokens of the subsequent period data following the cyclic period data by one, and outputting the tokens of the subsequent period data following the cyclic period data while incrementing them by one, and recognizes the end of one piece of sequence data by padding.

[0048] FIG. 5 is a first explanatory diagram showing the processing during training of the prediction device in the present invention.

[0049] According to FIG. 5(a), for example, sequence prediction model 10 is trained as follows. Input data token sequence: 0, 1, 2, 3, 0, , 1, 2, 3, 0, P, P, P, P, P, P, P Token sequence of the correct data: 0, 1, 2, 3, 0, , 1, 2, 3, 0, 1, P, P, P, P, P, P Output data token sequence:*,*,*,*,*,·····,*,*,*,*,*,*,*,*,*,*,*,*,* In this case, it is trained to predict one token, "1", in the following period of data.

[0050] According to FIG. 5, it appears that training is being performed to predict only token "1" of the subsequent period data, but inside the sequence prediction model 10, when only token 0 in the cyclic period data is input, it learns to be able to predict the next token 1. Also, when tokens 0 and 1 in the cyclic period data are input, it learns to be able to predict the next token 2. By repeating this, eventually, it will also learn about the cyclic period data, and when a known token is input in the subsequent period data, it will be able to predict the next token.

[0051] According to Fig. 5(b), following Fig. 5(a), it is trained as follows: Input data token sequence: 0, 1, 2, 3, 0, , 1, 2, 3, 0, , 1, P, P, P, P, P, P Token sequence of the correct data: 0, 1, 2, 3, 0, , 1, 2, 3, 0, 1, 2, P, P, P, P, P Output data token sequence:*,*,*,*,*,·····,*,*,*,*,*,*,*,*,*,*,*,*,* In this case, it is trained to predict one token, "2", in the subsequent period data.

[0052] FIG. 6 is a second explanatory diagram showing the processing during training of the prediction device in the present invention.

[0053] According to Fig. 6(a), followed by Fig. 5(b), it is trained as follows: Input data token sequence: 0, 1, 2, 3, 0, , 1, 2, 3, 0, , 1, 2, P, P, P, P, P Token sequence of the correct answer data: 0, 1, 2, 3, 0, , 1, 2, 3, 0, 1, 2, 3, P, P, P, P Output data token sequence:*,*,*,*,*,·····,*,*,*,*,*,*,*,*,*,*,*,*,* In this case, it is trained to predict one token, "3", in the subsequent period data.

[0054] According to Fig. 6(b), following Fig. 6(a), it is trained as follows: Input data token sequence: 0, 1, 2, 3, 0, , 1, 2, 3, 0, 1, 2, 3, P, P, P, P Token sequence of the correct answer data: 0, 1, 2, 3, 0, , 1, 2, 3, 0, 1, 2, 3, P, P, P, P Output data token sequence:*,*,*,*,*,·····,*,*,*,*,*,*,*,*,*,*,*,*,* In this case, it is trained to predict one token "P" in the subsequent period data, thereby recognizing the end of the series data.

[0055] As shown in Figures 5 and 6 described above, according to the present invention, by training a long-term movement trajectory for each user, it is possible to predict the "location" of the next movement destination that is characteristic of the user by inputting the most recent cycle period data and the subsequent period data that has passed.

[0056] Here, according to the present invention, it becomes possible to predict subsequent period data in a time series for each token. For example, if the sequence data has one week of rolling period data and padding embedded in it, then the first token (position) of the day of the subsequent period data is predicted. For example, in the case of sequence data in which the first token (position) of a day of a one-week period data and subsequent period data are embedded with padding, the second token (position) of the day of the subsequent period data is predicted. In other words, by inputting the tokens (positions that have moved) that have already passed in one day of the subsequent period data to be predicted, the next token (position that will move next) can be predicted. In this way, the next token can be predicted by inputting the past one week of cycle period data and the subsequent period data up to the elapsed token.

[0057] FIG. 7 is a third explanatory diagram showing the processing during training of the prediction device in the present invention.

[0058] According to Fig. 7, unlike Fig. 4, there is a missing period data in the cycle period data. In such a case, there is a concern that the token of the subsequent period data may be erroneously predicted. In such a case, according to the present invention, the padding unit 11 fills the missing portion with padding so that the average number of tokens becomes equal to the average number of tokens of period data having the same periodicity as that of the period data.

[0059] As shown in Figure 7, there is a gap in the period data for Thursday in the cyclical period data. Because the cyclical period data has periodicity, a large amount of Thursday period data is input even in the training stage. For this reason, padding is added to the average number of tokens in the period data for Thursday. This ensures that the length of the period data for Thursday is as long as possible. As described above, when padding is input to a token of sequence data, the location of the token is trained not to calculate the loss function and not to perform backpropagation (feedback). This makes it possible to avoid affecting the prediction of subsequent period data.

[0060] FIG. 8 is an explanatory diagram showing the inclusion of a starting or ending for each period data in sequence data.

[0061] According to FIG. 8(a), each period data further includes a starting special token at the beginning. According to FIG. 8(b), each period data further includes an ending special token at the end. As a result, even if the period data is consecutive, the period data is classified by each period data. For example, it is possible to recognize by the special token that the last token of the previous period data and the first token of the next period data are not consecutively generated. In addition, the location of the special token is trained without calculating the loss function and without backpropagating the error. Therefore, it is possible to prevent the special token from affecting the prediction of the data for the subsequent period.

[0062] As described above in detail, the prediction device, program, and method of the present invention make it possible to predict a user's movement trajectory using a sequence prediction model, taking into account temporal periodicity based on human social activity.

[0063] Furthermore, this will "promote the use of big data to predict user locations," which will contribute to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs), which is to "build resilient infrastructure, promote sustainable industrialization and foster innovation."

[0064] Regarding the various embodiments of the present invention described above, various changes, modifications, and omissions within the scope of the technical idea and perspective 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 in any way. The present invention is limited only by the scope of the claims and their equivalents. [Explanation of symbols]

[0065] 1 Prediction device 10 Series Prediction Model 11 Padding section 12 Predictive control unit

Claims

1. A prediction device for predicting the next token by inputting a token sequence of a predetermined length, the sequence data being made up of a plurality of period data, each period data being made up of a plurality of tokens, comprising: Data for multiple periods has periodicity as a time series, A maximum number of tokens including the periodic period data including a plurality of period data for one period and the subsequent period data following the periodic period data is set as a predetermined length of the sequence data; a padding means for filling a portion of each of the sequence data that does not reach a predetermined length with padding; a prediction control means for inputting a token sequence of a predetermined length including cyclic period data to the sequence prediction model for each padded sequence data, and predicting a token sequence of a predetermined length obtained by incrementing the token of the subsequent period data following the cyclic period data by one; A prediction device comprising:

2. The predictive control means is A token string of a predetermined length including cyclic period data is input, and a token string of a predetermined length is predicted by incrementing the first token of the subsequent period data following the cyclic period data by 1, or A token string including one or more tokens of periodic period data and subsequent period data is input, and a token string of a predetermined length is predicted by incrementing the tokens following the periodic period data and subsequent period data by one. The prediction device according to claim 1 .

3. In the training phase, the sequence prediction model inputs a token string of a predetermined length while incrementing the token of the subsequent period data following the cycle period data by one, and repeats training so as to output the token of the subsequent period data following the cycle period data while incrementing it by one, and recognizes the end of one sequence data by padding. The prediction device according to claim 2 .

4. The period data consists of multiple tokens for a unit time. The periodic period data is made up of periodic data having a periodicity consisting of a plurality of unit times, The subsequent period data is made up of a plurality of tokens for a unit time following the periodic period data. The prediction device according to any one of claims 1 to 3.

5. For each period data, a starting special token is further included at the beginning, or an ending special token is further included at the end. The prediction device according to claim 4 .

6. The period data consists of multiple tokens for one day. The cycle period data consists of one week's worth of period data, Subsequent period data consists of multiple tokens on a daily basis following one week of period data.

6. The prediction device according to claim 4 or 5.

7. The sequence data including multiple period data is divided by user, The tokens included in the period data are location information arranged in chronological order. The prediction device according to any one of claims 1 to 6.

8. The padding means, when a missing portion occurs in the period data in the periodic period data, fills the missing portion with padding so that the average number of tokens of the period data having the same periodicity as the missing portion becomes equal to the average number of tokens of the period data having the same periodicity as the missing portion. The prediction device according to any one of claims 1 to 7.

9. A program for inputting a token string of a predetermined length and causing a computer to function to predict the next token, the program comprising: Data for multiple periods has periodicity as a time series, A maximum number of tokens including the periodic period data including a plurality of period data for one period and the subsequent period data following the periodic period data is set as a predetermined length of the sequence data; a padding means for filling a portion of each of the sequence data that does not reach a predetermined length with padding; a prediction control means for inputting a token sequence of a predetermined length including cyclic period data to the sequence prediction model for each padded sequence data, and predicting a token sequence of a predetermined length in which the tokens of the subsequent period data following the cyclic period data are incremented by one; A program for causing a computer to function.

10. A prediction method for a prediction device which receives input of a token sequence of a predetermined length and predicts the next token, the prediction method comprising: Data for multiple periods has periodicity as a time series, The prediction device is Data for multiple periods has periodicity as a time series, A maximum number of tokens including the periodic period data including a plurality of period data for one period and the subsequent period data following the periodic period data is set as a predetermined length of the sequence data; a first step of filling a portion of each sequence data that does not reach a predetermined length with padding; a second step of inputting a token sequence of a predetermined length including the periodic period data for each padded periodic data into the sequence prediction model, and predicting a token sequence of a predetermined length in which the tokens of the subsequent period data following the periodic period data are incremented by one; A prediction method comprising the steps of:

Images