Device and method
The apparatus and method enhance life event estimation by processing time-series and non-time-series data to predict various events, addressing the limitations of existing models in predicting infrequent events like travel or pet purchase.
Patent Information
- Application Number
- PCT/JP2025/017392
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-05-13
- Publication Date
- 2026-11-19
AI Technical Summary
Existing life event estimation models struggle to accurately predict relatively infrequent events such as travel or pet purchase, in addition to milestone events like marriage and childbirth.
An apparatus and method that utilize a life event estimation model to process first and second input data, which include time-series element data with assigned embedding vectors and position information, to calculate attention levels and estimate the occurrence of various events by combining these data with non-time-series data through a learning model.
Accurately estimates the occurrence of a wide range of events, including both milestone and infrequent events, by improving data handling and interpolation techniques to enhance prediction accuracy.
Smart Images

Figure JP2025017392_19112026_PF_FP_ABST
Abstract
Description
Apparatus and Method
[0001] The present disclosure relates to an apparatus and method for estimating occurrence of an event.
[0002] The estimation device described in Patent Document 1 uses a life event estimation model, takes a feature quantity generated based on input information as an input, and outputs an estimated value related to the occurrence of a predetermined life event. The feature quantity generation unit included in the estimation device generates, based on the input information, a numerical value such as contract information included in the input information, an aggregated value obtained by aggregating the numerical values, or information indicating an attribute at one time point in a time series or in a predetermined period, as a feature quantity. The feature quantity generation unit generates, for example, numerical values, aggregated values (such as average value, maximum value, and minimum value) and attribute values in the most recent predetermined period as the feature quantity.
[0003] Japanese Unexamined Patent Application Publication No. 2019-220100
[0004] In the life event estimation model as described above, the probability of occurrence of a life event, which is a milestone event in the entire life of a user, is output as an estimated value. On the other hand, it is desired to estimate the occurrence of not only life events but also relatively infrequent events such as travel or pet purchase.
[0005] An object of the present disclosure is to provide an apparatus and method capable of estimating the occurrence of various events.
[0006] An apparatus according to the present disclosure includes: a first input unit to which first input data including a plurality of first element data arranged in time series is input; a processing unit that calculates estimation input data based on a plurality of first attention levels obtained by quantifying mutual attention levels between the plurality of first element data; and an event estimation unit that outputs an estimation result of a user's event based on the estimation input data.
[0007] According to the present disclosure, it is possible to estimate the occurrence of various events.
[0008] Figure 1 is a block diagram showing the functional configuration of the estimation device of this embodiment. Figure 2(a) is a diagram showing an example of first input data. Figure 2(b) is a diagram showing an example of second input data. Figures 3(a) and 3(b) are diagrams illustrating an example of data adjustment. Figure 4 is a diagram showing an example of multiple first attention values. Figures 5(a), 5(b), and 5(c) are diagrams showing an example of complementing multiple first attention values. Figures 6(a), 6(b), and 6(c) are diagrams showing an example of complementing multiple second attention values. Figures 7(a) and 7(b) are diagrams showing an example of calculating estimation input data. Figure 8 is a diagram showing an example of first input data. Figure 9 is a diagram showing an example of event estimation. Figure 10 is a flowchart showing an example of the estimation method of this embodiment. Figure 11 is a diagram showing an example of the hardware configuration of the estimation device according to one embodiment of the present disclosure.
[0009] Embodiments of this disclosure will be described in detail below with reference to the drawings. In each drawing, the same or corresponding parts are denoted by the same reference numerals, and redundant descriptions are omitted.
[0010] [Configuration of Estimation Device] Figure 1 is a block diagram showing the functional configuration of the estimation device of this embodiment. The estimation device 10 comprises a first input unit 1, a second input unit 2, a first position embedding unit 3, a second position embedding unit 4, a processing unit 5, a third input unit 6, a coupling unit 7, and an event estimation unit 8. The estimation device 10 acquires multiple element data contained in each of multiple types of input data in time series or non-time series, and estimates events related to the user using the degree of attention between elements for input data acquired in time series. The degree of attention is a numerical representation of the relationship between preceding and succeeding element data in multiple element data arranged in time series. In other words, the degree of attention indicates how much the user's past actions influence their current actions. According to the estimation device 10, it is possible to evaluate how much the user's past important actions influence future events as input data for the estimation model. Therefore, it is possible to accurately estimate the occurrence of a wide range of events. Here, events include life events, which are milestone events in the user's lifetime, and relatively infrequent events that indicate changes in the user's life. Life events include, for example, marriage and childbirth. Relatively infrequent events include moving, traveling, and keeping pets.
[0011] The first input unit 1 receives the first input data D11. The second input unit 2 receives the second input data D21. Figure 2(a) shows an example of the first input data. Figure 2(b) shows an example of the second input data. The estimation device 10 communicates with an external data management server via a network, for example. The estimation device 10 obtains the first input data D11 and the second input data D21 from the data management server, for example. The data management server is a server for storing data necessary for estimating events in the estimation device 10.
[0012] The first input data D11 is defined for each user, for example. The first input data D11 includes multiple first element data arranged in chronological order. For example, the multiple first element data may be a history of multiple stores where the user made purchases using a specific payment method within a predetermined period. Each of the multiple stores is a first element data. As shown in Figure 2(a), the first input data D11 indicates that user U1 made a purchase at convenience store A and then made a purchase at a station kiosk. The predetermined period may be one day, one week, or one month. The specific payment application may include, for example, electronic payment or mobile payment using a mobile device such as a smartphone.
[0013] The second input data D21, like the first input data D11, is defined for each user, for example. The second input data D21 includes multiple second element data arranged in chronological order. The second element data is different from the multiple first element data. For example, the multiple second element data may be a history of multiple locations (stores and facilities, etc.) visited by the user within a predetermined period. Each of the multiple locations is a second element data. The predetermined period may be one day, one week, or one month. The predetermined period during which the second input data D21 is acquired may be the same as or different from the predetermined period during which the first input data D11 is acquired.
[0014] The first input unit 1 assigns an embedding vector to each of the multiple first element data. The second input unit 2 assigns an embedding vector to each of the multiple second element data. In other words, the first input unit 1 and the second input unit 2 function as embedding units that assign embedding vectors. This allows multiple first element data and multiple second element data, which are discrete pieces of information, to be processed as easy-to-handle vector data. The first input unit 1 outputs first input data D11, which includes multiple first element data to which embedding vectors have been assigned, to the first position embedding unit 3. The second input unit 2 outputs second input data D21, which includes multiple second element data to which embedding vectors have been assigned, to the second position embedding unit 4.
[0015] The first input unit 1 assigns a unique index to each of the multiple first element data. For example, the first input unit 1 assigns "Index IN1" to "Convenience Store A" and "Index IN2" to "Station Kiosk" in Figure 2(a). The first input unit 1 may assign the same index to the same first element data. Then, the first input unit 1 assigns an embedded vector corresponding to each index. Taking the first input data D11 of user U1 in Figure 2(a) as an example, for example, the index is represented as [1, 2, 3, 1...] and the embedded vector is represented as [E1, E2, E3, E1...]. The second input unit 2 may, similar to the first input unit 1, assign a unique index to each of the multiple second element data and then assign an embedded vector corresponding to each index.
[0016] As shown in Figure 1, the first position embedding unit 3 is located between the first input unit 1 and the processing unit 5. The first input data D11 is input to the first position embedding unit 3 from the first input unit 1. At this time, embedding vectors are assigned to multiple first element data. The second position embedding unit 4 is located between the second input unit 2 and the processing unit 5. The second input data D21 is input to the second position embedding unit 4. At this time, embedding vectors are assigned to multiple second element data. The first position embedding unit 3 assigns position information to each of the multiple first element data. The second position embedding unit 4 assigns position information to each of the multiple second element data. As a result, the multiple first element data and the multiple second element data can be processed as data that includes time-series sequence information. As will be described in detail later, the first transformer 53 and the second transformer 54 included in the processing unit 5 need to assign position information to each element data in order to calculate the degree of attention between elements. The first position embedding unit 3 outputs first input data D11, which includes a plurality of first element data to which position information is assigned, to the processing unit 5. The second position embedding unit 4 outputs second input data D21, which includes a plurality of second element data to which position information is assigned, to the processing unit 5.
[0017] The first position embedding unit 3 encodes position information for multiple first element data using a fixed mathematical formula. The first position embedding unit 3 uses mathematical formulas such as equations (1) and (2), for example: PE_{(pos, 2i)} = sin(pos / 10000^{2i / d}) ... (1) PE_{(pos, 2i+1)} = cos(pos / 10000^{2i / d}) ... (2) Here, pos is the time-series position of each first element data, i is the dimension of the embedding vector (index of each first element data), and d is the number of first element data.
[0018] For example, if the embedding vectors of multiple first element data are represented as [E1, E2, E3, E1...], the first position embedding unit assigns position information such as [P1, P2, P3, P4...] to each of the multiple first element data. The first position embedding unit 3 may add up each embedding vector and each position information. In this case, the final embedding vector of the multiple first element data will be, for example, [E1+P1, E2+P2, E3+P3, E1+P4...]. The second position embedding unit 4 may, similar to the first position embedding unit 3, assign position information to each of the multiple second element data by adding up the embedding vectors and position information of the multiple second element data.
[0019] The processing unit 5 calculates estimated input data based on a plurality of first attention values, which are numerical representations of the degree of attention each of the plurality of first element data is paying to each other. Similarly, the processing unit 5 calculates estimated input data based on a plurality of second attention values, which are numerical representations of the degree of attention each of the plurality of second element data is paying to each other. As shown in Figure 1, the processing unit 5 includes a first data adjustment unit 51, a second data adjustment unit 52, a first transformer 53, a second transformer 54, a first interpolation unit 55, a second interpolation unit 56, and a calculation unit 57.
[0020] The first data adjustment unit 51 receives the first input data D11 from the first position embedding unit 3. At this time, embedding vectors and position information are assigned to multiple first element data. The first data adjustment unit 51 generates first adjustment data D12 that matches the number of multiple first element data to a target number. For example, if the number of multiple first element data at a predetermined time point does not match the target number, the first data adjustment unit 51 adjusts the number of multiple first element data included in the first input data D11. A predetermined time point is a specific period, which may be daily or weekly. The first data adjustment unit 51 outputs the adjusted first input data D11 as first adjustment data D12 to the first transformer 53. The second data adjustment unit 52 receives the second input data D21 from the second position embedding unit 4. At this time, embedding vectors and position information are assigned to multiple second element data. The second data adjustment unit 52 generates second adjustment data D22 that matches the number of multiple second element data to a target number. For example, if the number of multiple second element data at a predetermined time point does not match the target number, the second data adjustment unit 52 adjusts the number of multiple second element data included in the second input data D21. The second data adjustment unit 52 outputs the adjusted second input data D21 as second adjustment data D22 to the second transformer 54.
[0021] The first data adjustment unit 51 adds at least one first dummy element to the first input data D11 in a number corresponding to the difference between the number of first element data and the target number, when the number of multiple first element data at a predetermined time point is less than the target number. Figures 3(a) and 3(b) are diagrams illustrating an example of data adjustment. In Figures 3(a) and 3(b), each predetermined time point is shown as a day. In Figure 3(a), multiple first element data 110 included in the first input data D11 are arranged in chronological order. In the example in Figure 3(a), it is shown that multiple first element data 110 were acquired over several days from day 1 to day N. As shown in Figure 3(a), if the number of first element data 110 present on multiple days is not uniform, there may be cases where the number of elements included in the first input data D11 is less than the target number X. For example, with respect to a target number X, the number of multiple first element data 110 on the first day is A, and the number of multiple first element data 110 on the second day is B. The numbers A and B are less than the target number X. Note that the target number X, the number A, and the number B are integers greater than or equal to 1.
[0022] Therefore, as shown in Figure 3(b), the first data adjustment unit 51 adds first dummy elements 112 to the first input data D11 to fill in the gaps in the first element data 110, thereby generating the first adjusted data D12. For example, the first data adjustment unit 51 adds a number of first dummy elements 112 corresponding to the difference between the target number X and the number A to a plurality of first element data 110 on the first day. For example, the first data adjustment unit 51 adds a number of first dummy elements 112 corresponding to the difference between the target number X and the number B to a plurality of first element data 110 on the second day. As a result, the number of first element data 110 included in the first adjusted data D12 becomes uniform in terms of the target number X for each of the days. Note that the first dummy elements 112 may be set arbitrarily. For example, the embedding vector of the first dummy element 112 is the embedding vector E1, and in this case, the first dummy element 112 may be "Convenience Store A", which is the first element data corresponding to the embedding vector E1.
[0023] The first data adjustment unit 51 may generate the first adjusted data D12 by deleting at least one first element data 110 from the first input data D11, in a number corresponding to the difference between the number of first element data 110 and the target number X, if the daily number of multiple first element data 110 is greater than the target number X. In this case, the first data adjustment unit 51 may delete older data from the multiple first element data 110 arranged in time series for each of the days.
[0024] The second data adjustment unit 52 adjusts the second input data D21 in the same manner as the first data adjustment unit 51. For example, if the daily number of multiple second element data 210 (see, for example, Figure 6) is less than the target number X, the second data adjustment unit 52 adds at least one second dummy element 212 to the second input data D21 in a number corresponding to the difference between the number of multiple second element data 210 and the target number X. The second dummy element 212 may be set arbitrarily. Alternatively, for example, if the daily number of multiple second element data 210 is greater than the target number X, the second data adjustment unit 52 deletes at least one second element data 210 from the second input data D21 in a number corresponding to the difference between the number of multiple second element data 210 and the target number X. As a result, the number of first element data 110s per day included in the first adjustment data D12 and the number of second element data 210s per day included in the second adjustment data D22 may both be the same. In the following explanation, we will use as an example the case where the number of multiple first element data 110s and the number of multiple second element data 210s per day are each less than the target number X, and the first adjustment data D12 includes at least one first dummy element 112 and the second adjustment data D22 includes at least one second dummy element 212.
[0025] As shown in Figure 1, the first transformer 53 receives first adjustment data D12 from the first data adjustment unit 51. Based on the first adjustment data D12, the first transformer 53 sets multiple first attention levels. The first transformer 53 outputs first attention level data D13, which includes multiple first attention levels, to the first interpolation unit 55. The second transformer 54 receives second adjustment data D22 from the second data adjustment unit 52. Based on the second adjustment data D22, the second transformer 54 sets multiple second attention levels. The second transformer 54 outputs second attention level data D23, which includes multiple second attention levels, to the second interpolation unit 56.
[0026] The first transformer 53 and the second transformer 54 are types of models that use an attention mechanism in natural language processing. The first transformer 53 can quantify the relationships between the first element data 110 of the input first adjustment data D12 as weights (scores). The first attention score is, for example, a score indicating the degree to which the first element data 110 are related to each other. The second transformer 54 can quantify the relationships between each of the second element data 210 of the input second adjustment data D22 as weights (scores). The second attention score is, for example, a score indicating the degree to which the second element data 210 are related to each other.
[0027] Figure 4 shows an example of multiple first attention scores. In the example in Figure 4, the multiple first attention scores 120 are decimal numbers with one decimal place between 0.1 and 1.0. In this case, the multiple first attention scores 120 may be closer to 1.0 as the relationship between the first element data 110 is greater (higher attention score), and closer to 0.1 as the relationship between the first element data 110 is smaller (lower attention score). Examples of numerical values for the multiple first attention scores 120 and the multiple second attention scores are not limited to these; for example, they may be integers between 1 and 10, or decimal numbers with two or more decimal places.
[0028] The first transformer 53, for example, arranges a plurality of first element data 110 included in the first adjustment data D12 in the first direction DR1, which is the row direction, and also in the second direction DR2, which is the column direction, to comprehensively set the first attention level 120 among the first element data 110. That is, the first attention level data D13 may be in matrix format. The number of columns and rows of the first attention level data D13 is the value obtained by multiplying the target number X by the number of days N in the first adjustment data D12.
[0029] Since positional information is assigned to each first element data 110 by the first position embedding unit 3, the first transformer 53 may set the first attention level 120 based on the chronological order of the multiple first element data 110. For example, if the embedding vectors in the multiple first element data 110 are [E1+P1, E2+P2, E3+P3, E1+P4, E2+P5, ...], and the embedding vector E2 appears frequently after the embedding vector E1, then the first attention level 120 of the element data "station kiosk," which corresponds to the embedding vector E2, may be higher than the attention levels of the other elements, from the perspective of the element data "convenience store A," which corresponds to the embedding vector E1. The first attention level 120 may differ between adjacent first element data 110 depending on the chronological order. For example, the first Transformer 53 may set the first attention level of "station kiosk" as seen from "convenience store A" to be higher than the first attention level of "convenience store A" as seen from "station kiosk".
[0030] The multiple first attention levels 120 include at least one first normal attention level 121 and at least one first dummy attention level 122. The at least one first normal attention level 121 is the attention level between first normal elements 111, which are element data among the multiple first element data 110 that are not first dummy elements 112. The at least one first dummy attention level 122 is the attention level to at least one first dummy element 112. For example, the first dummy attention level 122 is the attention level between first dummy elements 112, and the attention level between the first normal element 111 and the first dummy element 112.
[0031] The second transformer 54, like the first transformer 53, may set multiple second attention levels 220 (see, for example, Figure 6) based on the time-series order of multiple second element data 210. The second attention level data D23 may be in matrix format. In this case, the number of columns and rows of the second attention level data D23 is the value obtained by multiplying the target number X by the number of days of the second adjustment data D22. Furthermore, the multiple second attention levels 220 include at least one second normal attention level 221 and at least one second dummy attention level 222.
[0032] As shown in Figure 1, the first interpolation unit 55 receives the first attention level data D13 from the first transformer 53. The second interpolation unit 56 receives the second attention level data D23 from the second transformer 54. The first dummy attention level 122 included in the first attention level data D13 and the second dummy attention level 222 included in the second attention level data D23 are unlikely to be valid attention levels. This is because the first dummy element 112 and the second dummy element 212 are not actually observed valid element data. Therefore, the first interpolation unit 55 interpolates at least one first dummy attention level 122 based on the first attention level data D13. The first interpolation unit 55 outputs the first interpolation data D14, which includes the interpolated first dummy attention level 122, to the calculation unit 57. Similarly, the second interpolation unit 56 interpolates at least one second dummy attention level 222 based on the second attention level data D23. The second interpolation unit 56 outputs the second interpolation data D24, which includes the interpolated second dummy attention score 222, to the calculation unit 57.
[0033] Figure 5 shows an example of the interpolation of multiple first attention levels. Figure 5(a) shows multiple first attention levels 120 calculated by the first transformer 53. In Figure 5(a), the first attention level data D13 shown in Figure 4 is omitted. The multiple first attention levels 120 include at least one first normal attention level 121, which is the attention level between normal element data, and at least one first dummy attention level 122, which is the attention level for at least one first dummy element 112. In the example in Figure 5(a), there are two first normal attention levels 121 and four first dummy attention levels 122.
[0034] The first interpolation unit 55 replaces at least one first dummy focus 122 with at least one missing value. Figure 5(b) is a figure obtained by replacing the first dummy focus 122 in Figure 5(a) with the missing value 100. The missing value is, for example, 0.0. Next, the first interpolation unit 55 interpolates the at least one missing value 100. Non-negative Matrix Factorization (NMF) or bilinear interpolation may be used as the interpolation method. NMF is a method of non-negative matrix factorization. For example, the first interpolation unit 55 may decompose the matrix data in Figure 5(b), in which the first dummy focus 122 has been replaced with the missing value 100, into a non-negative basis matrix and a non-negative coefficient matrix. The first interpolation unit 55 may then interpolate the missing value 100 by updating the basis matrix and coefficient matrix to approximate the matrix data in Figure 5(b). On the other hand, if bilinear interpolation is used, the first interpolation unit 55 may perform linear interpolation using the first degree of interest 120 around the missing value 100.
[0035] Next, the first interpolation unit 55 interpolates at least one missing value 100 with at least one first interpolated value 123, as shown in Figure 5(c). As an example, the first interpolation unit 55 calculates the average value of the data obtained by applying the interpolation method described above to the matrix data in Figure 5(b) and the first attention level data D13 in Figure 5(a), and replaces at least one missing value 100 with at least one first interpolated value 123. As a result, the first interpolation unit 55 generates the first interpolated data D14. The multiple first attention levels 120 included in the first interpolated data D14 include at least one first normal attention level 121 and at least one first dummy attention level 122, with at least one first interpolated value 123 replacing them.
[0036] Figure 6 shows an example of the interpolation of multiple second attention levels. As shown in Figure 6(a), the multiple second attention levels 220 include at least one second normal attention level 221, which is the attention level between normal element data, and at least one second dummy attention level 222, which is the attention level for at least one second dummy element 212. As shown in Figure 6(b), the second interpolation unit 56 replaces at least one second dummy attention level 222 with at least one missing value 100. The missing value 100 that replaces the second dummy attention level 222 may be the same as the missing value 100 that replaces the first dummy attention level 122. The second interpolation unit 56 then interpolates at least one missing value 100. Subsequently, as shown in Figure 6(c), the second interpolation unit 56 interpolates at least one missing value 100 with at least one second interpolated value 223 to generate the second interpolated data D24. The multiple second attention values 220 included in the second supplemental data D24 include at least one second normal attention value 221 and at least one second supplemental value 223 in place of at least one second dummy attention value 222.
[0037] As shown in Figure 1, the calculation unit 57 receives first supplementary data D14 from the first supplementary unit 55 and second supplementary data D24 from the second supplementary unit 56. Based on the multiple first attention values 120 contained in the first supplementary data D14 and the multiple second attention values 220 contained in the second supplementary data D24, the calculation unit 57 calculates estimated input data D51. The calculation unit 57 outputs the calculated estimated input data D51 to the subsequent coupling unit 7.
[0038] Figures 7(a) and 7(b) show an example of the calculation of estimated input data. The calculation unit 57 sums up multiple first attention values 120 and multiple second attention values 220 as shown in Figure 7(a) to calculate multiple sum values 510 as shown in Figure 7(b). The estimated input data D51 is data that includes multiple sum values 510. The number of columns and rows in the first supplementary data D14 is the same as the first attention value data D13, obtained by multiplying the target number X by the number of days N in the first adjustment data D12. The number of columns and rows in the second supplementary data D24 is the same as the second attention value data D23, obtained by multiplying the target number X by the number of days in the second adjustment data D22. The calculation unit 57 sums up the first attention values 120 and second attention values 220 at the same row and column positions.
[0039] Here, if the number of days in the second adjustment data D22 is different from the number of days N, the number of columns and rows in the second supplementary data D24 may differ from the number of columns and rows in the first supplementary data D14. Even if the number of days in the second adjustment data D22 is M (where M is an integer smaller than N), since the number of daily occurrences of multiple first element data 110 and the number of daily occurrences of multiple second element data 210 are all the same at the target number X, all the second attention values 220 up to the Mth row and Mth column in the second supplementary data D24 are added together with the first attention values 120.
[0040] The calculation unit 57 sums the multiple first attention values 120 and the multiple second attention values 220. As a result, the first attention values with larger values among the multiple first attention values 120 and the second attention values with larger values among the multiple second attention values 220 are summed up and can be calculated in subsequent stages with a higher degree of confidence. In the examples of Figures 7(a) and 7(b), the first attention value 120 of the second row and first column of the first supplementary data D14, "0.89", is relatively high compared to the other first attention value 120 values, and the second attention value 220 of the second row and first column of the second supplementary data D24, "0.8", is relatively high compared to the other second attention value 220 values. Therefore, the sum of the second row and first column of the estimated input data D51, 510, is relatively high compared to the other sum of 510 values. This increases confidence that the first attention level of 120 in the second row and first column, and the second attention level of 220 in the second row and first column, are important.
[0041] On the other hand, if one of the multiple first attention values 120 and multiple second attention values 220 has a small value, it can be complemented by the other attention value. In the example in Figure 7, the first attention value 120 of the first row and first column of the first complementary data D14, "0.1", is relatively low compared to the other first attention values 120, but the second attention value 220 of the first row and first column of the second complementary data D24, "0.8", is relatively high compared to the other second attention values 220. Therefore, the sum of the first row and first column of the estimated input data D51, 510, can be close to the median value among the multiple sums 510.
[0042] As shown in Figure 1, the third input unit 6 receives third input data D31, which includes multiple non-time-series third element data. The estimation device 10 obtains the third input data D31 from, for example, a data management server. Figure 8 shows an example of the third input data D31. For example, the multiple third element data are the number of times a user has visited a location within a predetermined period. The number of visits for each location constitutes the third element data. In this case, the data management server may, for example, aggregate the number of visits to the locations visited by the user and output the aggregated result as the third input data D31 to the estimation device 10. The third input data D31 is data without order, and in the example in Figure 8, the multiple third element data may be arranged randomly.
[0043] The third input unit 6 includes a learning model that has been trained to take the third input data D31 as input and output a third output value D32. The learning model may be generated without machine learning, or it may be generated by either supervised or unsupervised machine learning. The third output value D32 is, for example, a non-negative integer. The third input unit 6 outputs the third output value D32 to the coupling unit 7.
[0044] As shown in FIG. 1, estimated input data D51 is input to the combining unit 7 from the arithmetic unit 57, and a third output value D32 is input to the combining unit 7 from the third input unit 6. The combining unit 7 combines the estimated input data D51 and the third output value D32. For example, the combining unit 7 may calculate a plurality of combined values by multiplying each sum value 510 of the estimated input data D51 by the third output value D32. The combining unit 7 outputs combined data D81 including the plurality of combined values to the event estimation unit 8.
[0045] The combined data D81 calculated based on the estimated input data D51 is input to the event estimation unit 8. The event estimation unit 8 includes an event estimation model trained to receive the combined data D81 and output an estimation result D82 of a user's event. The event estimation model may be generated without using machine learning, or may be generated by either supervised or unsupervised machine learning. For example, when the event estimation model is generated by supervised machine learning, the event estimation model is generated by applying techniques such as regression, classification, and clustering. More specifically, the event estimation model may be generated by techniques such as decision trees, linear regression, logistic regression, support vector machines, random forests, neural networks, and gradient boosting.
[0046] FIG. 9 is a diagram illustrating an example of event estimation. The combined data D81 includes a plurality of combined values 810. The plurality of combined values 810 are values obtained by multiplying an integer k, which is the third output value D32 output from the third input unit 6, by each sum value 510 of the estimated input data D51. The event estimation unit 8 outputs an estimation result D82 of the user's event. As an example, the event estimation unit 8 estimates a probability "0.2" that the event "marriage" occurs, a probability "0.8" that the event "birth of a child" occurs, and a probability "0.1" that the event "housing purchase" occurs. The event estimation unit 8 outputs the estimation result D82 to, for example, a proposing unit external to the estimation device 10. The proposing unit may generate information related to a product or service to be proposed to the user based on the association between the estimation result D82 and a predetermined product or service.
[0047] [Estimation Method Using Estimation Device] FIG. 10 is a diagram illustrating an example of a user event estimation method using the estimation device 10 described above. The estimation method includes steps ST1 to ST9, and steps ST1 to ST9 can be performed in sequence. The estimation method is started, for example, when first input data D11, second input data D21, and third input data D31 are input to the estimation device 10 from an external data management server.
[0048] In step ST1, the first input data D11 is input to the first input unit 1, and the second input data D21 is input to the second input unit 2. The first input data D11 includes a plurality of first element data 110 arranged in time series. The second input data D21 includes a plurality of second element data 210 arranged in time series. Further, in step ST1, the third input data D31 including a plurality of non-time-series third element data is input to the third input unit 6. The third input unit 6 includes a learned model that has been trained to receive the third input data D31 and output a third output value D32.
[0049] Subsequently, in step ST2, the first input unit 1 assigns an embedding vector to each of the plurality of first element data 110, and the second input unit 2 assigns an embedding vector to each of the plurality of second element data 210. Thereby, the plurality of first element data 110 and the plurality of second element data 210, which are discrete information, can be processed as easy-to-handle vector data. Subsequently, in step ST3, the first position embedding unit 3 assigns position information to each of the plurality of first element data 110, and the second position embedding unit 4 assigns position information to each of the plurality of second element data 210. Thereby, the plurality of first element data 110 and the plurality of second element data 210 can be processed as data including time-series order information.
[0050] In step ST4, the first data adjustment unit 51 adjusts the number of first element data 110 included in the first input data D11 if the number of multiple first element data does not match the target number X. Similarly, the second data adjustment unit 52 adjusts the number of multiple second element data included in the second input data D21 if the number of multiple second element data 210 does not match the target number X. For example, if the daily number of multiple first element data 110 is less than the target number X, the first data adjustment unit 51 adds at least one first dummy element 112 to the first input data D11 in a number corresponding to the difference between the daily number of multiple first element data 110 and the target number X. Similarly, for example, if the daily number of multiple second element data 210 is less than the target number X, the second data adjustment unit 52 adds at least one second dummy element 212 to the second input data D21 in a number corresponding to the difference between the daily number of multiple second element data 210 and the target number X.
[0051] Next, in step ST5, the first transformer 53 sets a plurality of first attention levels 120 based on the first adjustment data D12, and the second transformer 54 sets a plurality of second attention levels 220 based on the second adjustment data D22. The first transformer 53 can quantify the relationships between the first element data 110 of the input first adjustment data D12 as weights (scores). The second transformer 54 can quantify the relationships between each of the second element data 210 of the input second adjustment data D22 as weights (scores). The first attention level data D13, which includes a plurality of first attention levels 120, and the second attention level data D23, which includes a plurality of second attention levels 220, may be in matrix format.
[0052] Next, in step ST6, the first interpolation unit 55 interpolates at least one first dummy attention value 122 based on the first attention value data D13 to generate the first interpolated data D14. Similarly, the second interpolation unit 56 interpolates at least one second dummy attention value 222 based on the second attention value data D23 to generate the second interpolated data D24. The first interpolation unit 55, for example, replaces at least one first dummy attention value 122 with at least one missing value 100 and interpolates at least one missing value 100 into at least one first interpolated value 123. The second interpolation unit 56, for example, replaces at least one second dummy attention value 222 with at least one missing value 100 and interpolates at least one missing value 100 into at least one second interpolated value 223.
[0053] Next, in step ST7, the calculation unit 57 calculates the estimated input data D51 based on the multiple first attention values 120 included in the first supplementary data D14 and the multiple second attention values 220 included in the second supplementary data D24. Then, in step ST8, the merging unit 7 combines the estimated input data D51 and the third output value D32 to output the combined data D81. The merging unit 7 may, for example, calculate multiple combined values 810 by multiplying each sum of the estimated input data D51's values 510 by the third output value D32. Finally, in step ST9, the event estimation unit 8 receives the combined data D81 as input and outputs the estimated result D82 of the user's event.
[0054] [Function and Effects] In the estimation device 10 and the method described above, the multiple first attention values 120 are numerical representations of the relationship between preceding and succeeding element data in a series of multiple first element data 110. In other words, the multiple first attention values 120 indicate how much a user's past actions influence their current actions. According to the estimation device 10, by using estimation input data as input data for the estimation model, it is possible to evaluate how much a user's past important actions influence future events. Therefore, it is possible to accurately estimate the occurrence of a wide range of events.
[0055] The estimation device 10 further includes a second input unit 2 to which second input data D21, which includes a plurality of second element data 210 different from a plurality of first element data 110, is input. The processing unit 5 includes a first transformer 53 for setting a plurality of first attention levels 120, a second transformer 54 for setting a plurality of second attention levels 220 which are numerical representations of the mutual attention levels among the plurality of second element data 210, and a calculation unit 57 for calculating estimation input data D51 based on the plurality of first attention levels 120 and the plurality of second attention levels 220. In this case, by calculating estimation input data D51 based on the second input data D21 in addition to the first input data D11, the accuracy of estimation input data D51 is improved, and the occurrence of a wide range of events can be estimated more accurately.
[0056] The calculation unit 57 calculates the estimated input data D51 by summing the multiple first attention values 120 and the multiple second attention values 220. In this case, the first attention values 120 with larger values among the multiple first attention values 120 and the second attention values 220 with larger values among the multiple second attention values 220 are summed up, allowing for calculations in subsequent stages with a higher degree of confidence. In addition, attention values with smaller values among the multiple first attention values 120 and the multiple second attention values 220 can be compensated for by the other attention value. This makes it possible to further improve the accuracy of the estimated input data D51.
[0057] The processing unit 5 further includes a first data adjustment unit 51 that generates first adjustment data D12 in which the number of multiple first element data 110 at predetermined time points is adjusted to a target number X, and a second data adjustment unit 52 that generates second adjustment data D22 in which the number of multiple second element data 210 at predetermined time points is adjusted to a target number X. The first data adjustment unit 51 outputs the first adjustment data D12 to the first transformer 53, and the second data adjustment unit 52 outputs the second adjustment data D22 to the second transformer 54. In this case, by matching the number of data in the first input data D11 and the number of data in the second input data D21 to the target number X, calculations using the first input data D11 and the second input data D21 in subsequent stages can be made easier.
[0058] The first data adjustment unit 51 adjusts the number of first element data 110 to the target number X using at least one first dummy element 112 if the number of first element data 110 at a predetermined time point is less than the target number X, and the second data adjustment unit 52 adjusts the number of second element data 210 to the target number X using at least one second dummy element 212 if the number of second element data 210 is less than the target number X. The processing unit 5 further includes a first interpolation unit 55 that interpolates at least one first dummy attention level 122, which is the attention level for at least one first dummy element 112 among a plurality of first attention levels 120, and a second interpolation unit 56 that interpolates at least one second dummy attention level 222, which is the attention level for at least one second dummy element 212 among a plurality of second attention levels 220. The calculation unit 57 calculates estimated input data D51 based on the interpolated plurality of first attention levels 120 and the interpolated plurality of second attention levels 220. In this case, the accuracy of the plurality of first attention levels 120 and the plurality of second attention levels 220 can be improved by interpolating the low accuracy of the first dummy attention levels 122 and the second dummy attention levels 222.
[0059] The first interpolation unit 55 replaces at least one first dummy attention score 122 with at least one missing value 100 and interpolates at least one missing value 100 into at least one first interpolated value 123, and the second interpolation unit 56 replaces at least one second dummy attention score 222 with at least one missing value 100 and interpolates at least one missing value 100 into at least one second interpolated value 223. In this case, the interpolation process can be carried out simply by having the first interpolation unit 55 and the second interpolation unit 56 replace the dummy attention scores with a common missing value 100.
[0060] The first input unit 1 assigns an embedding vector to each of the multiple first element data 110, and the second input unit 2 assigns an embedding vector to each of the multiple second element data 210. In this case, the multiple first element data 110 and the multiple second element data 210, which are discrete pieces of information, can be processed as easily manageable vector data.
[0061] The estimation device 10 further comprises a first position embedding unit 3 positioned between the first input unit 1 and the processing unit 5, and a second position embedding unit 4 positioned between the second input unit 2 and the processing unit 5. The first position embedding unit 3 assigns position information to each of the plurality of first element data 110, and the second position embedding unit 4 assigns position information to each of the plurality of second element data 210. In this case, the plurality of discrete first element data 110 and the plurality of second element data 210 can be processed as data that includes time-series sequence information.
[0062] The estimation device 10 further includes a third input unit 6 that includes a learning model trained to take third input data D31, which includes a plurality of non-time-series third-element data, as input and output a third output value D32, and a coupling unit 7 that combines the third output value D32 and the estimation input data D51. The event estimation unit 8 includes an event estimation model trained to take data formed by combining the third output value D32 and the estimation input data D51 as input and output an estimation result. In this case, by calculating the input data of the event estimation model based on the third input data D31, the accuracy of the input data of the event estimation model is improved, and the occurrence of a wide range of events can be estimated more accurately. [Modification]
[0063] The present invention is not limited to the embodiments described above. The estimation device 10 does not need to include a third input unit 6 and a coupling unit 7. In this case, the estimation input data D51 output from the calculation unit 57 may be input to the event estimation unit 8. Furthermore, the first data adjustment unit 51 does not need to adjust the first input data D11 if the daily counts of all of the multiple first element data 110 match the target X. In this case, the first adjustment data D12 may be the same data as the first input data D11. Similarly, the second data adjustment unit 52 does not need to adjust the second input data D21 if the daily counts of all of the multiple second element data 210 match the target X. In this case, the second adjustment data D22 may be the same data as the second input data D21. Furthermore, the first interpolation unit 55 does not need to do anything to the first attention data D13 if the multiple first attention levels 120 do not include the first dummy attention level 122. In this case, the first supplementary data D14 may be the same data as the first attention level data D13. Similarly, if the second supplementary unit 56 does not include the second dummy attention level 222 among the multiple second attention levels 220, it may do nothing to the second attention level data D23. In this case, the second supplementary data D24 may be the same data as the second attention level data D23.
[0064] For example, in step ST7 of the estimation method in Figure 10, if the number of columns and rows of the second supplementary data D24 differs from the number of columns and rows of the first supplementary data D14, the calculation unit 57 may process the data to make the number of columns and rows of the first supplementary data D14 and the second supplementary data D24 match, and then sum the first supplementary data D14 and the second supplementary data D24. If the number of columns and rows of the second supplementary data D24 is less, the calculation unit 57 may expand the number of columns and rows of the second supplementary data D24. The calculation unit 57 may, for example, use bilinear interpolation to expand the number of columns and rows of the second supplementary data D24.
[0065] The estimation device of the present disclosure has the following configuration: [1] A device comprising: a first input unit into which first input data including a plurality of first element data arranged in a time series is input; a processing unit that calculates estimation input data based on a plurality of first attention values which are numerical representations of the degree of attention each of the plurality of first element data is; and an event estimation unit that outputs an estimation result of a user's event based on the estimation input data. [2] The device according to [1], further comprising: a second input unit into which second input data including a plurality of second element data different from the plurality of first element data is input, wherein the processing unit comprises: a first transformer for setting the plurality of first attention values; a second transformer for setting a plurality of second attention values which are numerical representations of the degree of attention each of the plurality of second element data is; and a calculation unit for calculating the estimation input data based on the plurality of first attention values and the plurality of second attention values. [3] The device according to [2], wherein the calculation unit calculates the estimation input data by summing the plurality of first attention values and the plurality of second attention values. [4] The apparatus according to [2] or [3], wherein the processing unit further includes a first data adjustment unit that generates first adjustment data that adjusts the number of the plurality of first element data at predetermined time points to a target number, and a second data adjustment unit that generates second adjustment data that adjusts the number of the plurality of second element data at predetermined time points to a target number, wherein the first data adjustment unit outputs the first adjustment data to the first transformer, and the second data adjustment unit outputs the second adjustment data to the second transformer.[5] The apparatus according to [4], wherein the first data adjustment unit adjusts the number of the plurality of first element data to the target number using at least one first dummy element when the number of the plurality of first element data at each predetermined time is less than the target number, the second data adjustment unit adjusts the number of the plurality of second element data to the target number using at least one second dummy element when the number of the plurality of second element data at each predetermined time is less than the target number, the processing unit further includes a first complementation unit that complements at least one first dummy attention level which is the attention level for at least one first dummy element among the plurality of first attention levels, and a second complementation unit that complements at least one second dummy attention level which is the attention level for at least one second dummy element among the plurality of second attention levels, and the calculation unit calculates the estimated input data based on the complemented plurality of first attention levels and the complemented plurality of second attention levels. [6] The apparatus according to [5], wherein the first interpolation unit replaces the at least one first dummy attention value with at least one missing value and interpolates the at least one missing value into at least one first interpolated value, and the second interpolation unit replaces the at least one second dummy attention value with at least one missing value and interpolates the at least one missing value into at least one second interpolated value. [7] The apparatus according to any one of [2] to [6], wherein the first input unit assigns an embedding vector to each of the plurality of first element data, and the second input unit assigns an embedding vector to each of the plurality of second element data. [8] The apparatus according to any one of [2] to [7], further comprising a first position embedding unit disposed between the first input unit and the processing unit, and a second position embedding unit disposed between the second input unit and the processing unit, wherein the first position embedding unit assigns position information to each of the plurality of first element data, and the second position embedding unit assigns position information to each of the plurality of second element data.[9] The apparatus according to any one of [1] to [8], further comprising: a third input unit including a learning model that has been trained to take a third input data including a plurality of non-time-series third element data and output a third output value; and a coupling unit that combines the third output value and the estimated input data, wherein the event estimation unit includes an event estimation model that has been trained to take data formed by combining the third output value and the estimated input data and output the estimation result.
[10] A method in the apparatus, comprising: a step of taking a first input data including a plurality of first element data arranged in time series as input; a step of calculating estimated input data based on a plurality of first attention values that quantify the degree of attention each of the plurality of first element data is paying to each other; and a step of outputting an estimation result of the user's event based on the estimated input data.
[0066] The block diagrams used in the description of the above embodiments show functional units. These functional blocks (components) are realized by any combination of at least one of hardware and software. Furthermore, the method of realizing each functional block is not particularly limited. That is, each functional block may be realized using one device that is physically or logically coupled, or it may be realized using two or more physically or logically separated devices that are directly or indirectly connected (for example, using wired or wireless connections). A functional block may also be realized by combining software with the one or more of the above devices.
[0067] Functions include, but are not limited to, judgment, decision, determination, calculation, calculation, processing, derivation, investigation, exploration, confirmation, reception, transmission, output, access, resolution, selection, selection, establishment, comparison, assumption, expectation, assumption, broadcasting, notifying, communicating, forwarding, configuring, reconfiguring, allocating (mapping), and assigning. For example, a functional block (configuration part) that enables transmission is called a transmitting unit or transmitter. In all cases, as mentioned above, the method of implementation is not particularly limited.
[0068] For example, the estimation device 10 in one embodiment of the present disclosure may function as a computer that processes the wireless communication method of the present disclosure. Figure 11 is a diagram showing an example of the hardware configuration of the estimation device 10 according to one embodiment of the present disclosure. The estimation device 10 described above may be physically configured as a computer device including a processor 1001, memory 1002, storage 1003, communication device 1004, input device 1005, output device 1006, bus 1007, etc.
[0069] In the following explanation, the term "device" can be replaced with "circuit," "device," "unit," etc. The hardware configuration of the estimated device 10 may include one or more of the devices shown in the figure, or it may be configured to omit some of the devices.
[0070] Each function in the estimation device 10 is realized by loading predetermined software (programs) onto hardware such as the processor 1001 and memory 1002, which causes the processor 1001 to perform calculations, control communication by the communication device 1004, and control at least one of data reading and writing in the memory 1002 and storage 1003.
[0071] The processor 1001 controls the entire computer, for example, by running an operating system. The processor 1001 may be composed of a central processing unit (CPU) that includes interfaces with peripheral devices, control devices, arithmetic units, registers, etc. For example, each function in the estimation device 10 may be implemented by the processor 1001. The processor 1001 may also be configured to include a GPU (Graphics Processing Unit). For example, the rendering unit 15 shown in Figure 1 may be implemented by the processor 1001.
[0072] Furthermore, the processor 1001 reads programs (program code), software modules, data, etc., from at least one of the storage 1003 and the communication device 1004 into the memory 1002 and executes various processes accordingly. The program used is one that causes the computer to execute at least a part of the operations described in the above embodiment. For example, each function in the estimation device 10 may be implemented by a control program stored in the memory 1002 and operated on the processor 1001, and other functional blocks may be implemented similarly. The above-described processes have been explained as being executed by one processor 1001, but they may be executed simultaneously or sequentially by two or more processors 1001. The processor 1001 may be implemented by one or more chips. The program may also be transmitted from the network via a telecommunications line.
[0073] The memory 1002 is a computer-readable recording medium and may consist of at least one of the following: ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), RAM (Random Access Memory), etc. The memory 1002 may also be called a register, cache, main memory, etc. The memory 1002 can store executable programs (program code), software modules, etc., for carrying out a wireless communication method according to one embodiment of the present disclosure.
[0074] The storage 1003 is a computer-readable recording medium and may consist of at least one of the following: an optical disc such as a CD-ROM (Compact Disc ROM), a hard disk drive, a flexible disk, a magneto-optical disk (e.g., a compact disk, a digital multipurpose disk, a Blu-ray® disc), a smart card, flash memory (e.g., a card, a stick, a key drive), a floppy® disk, a magnetic strip, etc. The storage 1003 may also be called an auxiliary storage device. The storage medium provided by the estimation device 10 may be, for example, a database, server, or other suitable medium including at least one of the memory 1002 and the storage 1003.
[0075] The communication device 1004 is hardware (transceiver / receiver device) for communicating between computers via at least one of a wired network and a wireless network, and is also referred to as a network device, network controller, network card, communication module, etc.
[0076] The input device 1005 is an input device that accepts input from an external source (e.g., a keyboard, mouse, microphone, switch, button, sensor, etc.). The output device 1006 is an output device that outputs to an external source (e.g., a display, speaker, LED lamp, etc.). The input device 1005 and the output device 1006 may be configured as an integrated unit (e.g., a touch panel).
[0077] Furthermore, each device, such as the processor 1001 and memory 1002, is connected by a bus 1007 for communicating information. The bus 1007 may be configured using a single bus, or different buses may be configured for each device.
[0078] Furthermore, the estimation device 10 may be configured to include hardware such as a microprocessor, a digital signal processor (DSP), an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), and an FPGA (Field Programmable Gate Array), and some or all of each functional block may be realized by such hardware. For example, the processor 1001 may be implemented using at least one of these hardware components.
[0079] The processing procedures, sequences, flowcharts, etc., of each aspect / embodiment described in this disclosure may be reordered, provided they do not contradict each other. For example, the methods described in this disclosure present various step elements using exemplary order and are not limited to the specific order presented.
[0080] Input and output information may be stored in a specific location (e.g., memory) or managed using a management table. Input and output information may be overwritten, updated, or appended to. Output information may be deleted. Input information may be transmitted to other devices.
[0081] The determination may be made by a value represented by one bit (0 or 1), by a boolean value (true or false), or by a numerical comparison (for example, a comparison with a predetermined value).
[0082] Each aspect / embodiment described in this disclosure may be used individually, in combination, or switched between as needed during implementation. Furthermore, notification of specific information (e.g., notification that "X is") is not limited to explicit notification, but may also be implicit (e.g., by not providing such notification).
[0083] Although the present disclosure has been described in detail above, it will be clear to those skilled in the art that the present disclosure is not limited to the embodiments described herein. The present disclosure can be implemented in modified and altered forms without departing from the intent and scope of the present disclosure as defined by the claims. Therefore, the descriptions in the present disclosure are illustrative and not intended to be restrictive in any way.
[0084] Software should be broadly interpreted to mean instructions, instruction sets, code, code segments, program code, programs, subprograms, software modules, applications, software applications, software packages, routines, subroutines, objects, executable files, execution threads, procedures, functions, and so on, whether they are called software, firmware, middleware, microcode, hardware description languages, or by any other name.
[0085] Furthermore, software, instructions, information, etc., may be transmitted and received via a transmission medium. For example, if software is transmitted from a website, server, or other remote source using at least one of wired technology (such as coaxial cable, fiber optic cable, twisted pair, or digital subscriber line (DSL)) and wireless technology (such as infrared or microwave), then at least one of these wired and wireless technologies is included in the definition of a transmission medium.
[0086] The terms “system” and “network” as used in this disclosure are interchangeable.
[0087] Furthermore, the information, parameters, etc., described in this disclosure may be expressed using absolute values, relative values from a predetermined value, or corresponding other information.
[0088] As used in this disclosure, the terms “determining” and “determining” may encompass a wide variety of actions. “Determining” may include, for example, judging, calculating, computing, processing, deriving, investigating, looking up, searching, or inquiring (e.g., searching in a table, database, or other data structure), or ascertaining. “Determining” may also include receiving (e.g., receiving information), transmitting (e.g., sending information), inputting, outputting, or accessing (e.g., accessing data in memory). Furthermore, "judgment" and "decision" can include considering something as having been "judged" or "decided" after resolving, selecting, choosing, establishing, comparing, etc. In other words, "judgment" and "decision" can include considering something as having been "judged" or "decided" after some action. Also, "judgment (decision)" can be reinterpreted as "assuming," "expecting," or "considering."
[0089] The terms “connected,” “coupled,” or any variation thereof, mean any direct or indirect connection or coupling between two or more elements, and may include the presence of one or more intermediate elements between two elements that are “connected” or “coupled” with each other. The coupling or connection between elements may be physical, logical, or a combination thereof. For example, “connection” may be reinterpreted as “access.” As used in this disclosure, two elements may be considered to be “connected” or “coupled” with each other using at least one of one or more wires, cables, and printed electrical connections, and, in some non-limiting and non-exclusive examples, electromagnetic energy having wavelengths in the radio frequency domain, microwave domain, and optical (both visible and invisible) domain.
[0090] In this disclosure, the phrase "based on" does not mean "based solely on" unless otherwise specified. In other words, the phrase "based on" means both "based solely on" and "based at least on."
[0091] Any reference to elements using the designations “first,” “second,” etc., as used in this disclosure does not generally limit the quantity or order of those elements. These designations may be used in this disclosure as a convenient way to distinguish between two or more elements. Accordingly, references to the first and second elements do not imply that only two elements may be employed, or that the first element must precede the second element in any way.
[0092] Where the terms “include,” “including,” and variations thereof are used in this disclosure, these terms are intended to be inclusive, as is the term “comprising.” Furthermore, the term “or” as used in this disclosure is not intended to mean exclusive OR.
[0093] In this disclosure, if articles are added through translation, such as a, an, and the in English, this disclosure may include the fact that the noun following these articles is plural.
[0094] 1...First input unit, 2...Second input unit, 3...First position embedding unit, 4...Second position embedding unit, 5...Processing unit, 6...Third input unit, 7...Combination unit, 8...Event estimation unit, 10...Estimation device (device), 51...First data adjustment unit, 52...Second data adjustment unit, 53...First transformer, 54...Second transformer, 55...First completion unit, 56...Second completion unit, 57...Calculation unit, 100...Missing value, 110...First element data, 111...First normal element, 112...First dummy element, 120...First attention level, 121...First Normal attention score, 122...First dummy attention score, 123...First imputed value, 210...Second element data, 211...Second normal element, 212...Second dummy element, 220...Second attention score, 221...Second normal attention score, 222...Second dummy attention score, 223...Second imputed value, 510...Summary value, 810...Combined value, D11...First input data, D12...First adjustment data, D21...Second input data, D22...Second adjustment data, D31...Third input data, D51...Estimated input data, D81...Combined data, D82...Estimated result.
Claims
1. A device comprising: a first input unit that receives first input data including a plurality of first element data arranged in a time series; a processing unit that calculates estimated input data based on a plurality of first attention values that quantify the degree of attention each of the plurality of first element data is paying to each other; and an event estimation unit that outputs an estimated result of a user event based on the estimated input data.
2. The apparatus according to claim 1, further comprising a second input unit into which second input data including a plurality of second element data different from the plurality of first element data is input, wherein the processing unit includes a first transformer for setting the plurality of first attention levels, a second transformer for setting a plurality of second attention levels that quantify the mutual attention levels among the plurality of second element data, and a calculation unit for calculating the estimated input data based on the plurality of first attention levels and the plurality of second attention levels.
3. The apparatus according to claim 2, wherein the calculation unit calculates the estimated input data by summing the plurality of first attention values and the plurality of second attention values.
4. The apparatus according to claim 2, wherein the processing unit further includes a first data adjustment unit that generates first adjustment data that adjusts the number of the plurality of first element data at predetermined time points to a target number, and a second data adjustment unit that generates second adjustment data that adjusts the number of the plurality of second element data at predetermined time points to a target number, the first data adjustment unit outputs the first adjustment data to the first transformer, and the second data adjustment unit outputs the second adjustment data to the second transformer.
5. The apparatus according to claim 4, wherein the first data adjustment unit adjusts the number of the plurality of first element data to the target number using at least one first dummy element when the number of the plurality of first element data at each predetermined time point is less than the target number, the second data adjustment unit adjusts the number of the plurality of second element data to the target number using at least one second dummy element when the number of the plurality of second element data at each predetermined time point is less than the target number, the processing unit further includes a first interpolation unit that interpolates at least one first dummy attention level which is the attention level for at least one first dummy element among the plurality of first attention levels, and a second interpolation unit that interpolates at least one second dummy attention level which is the attention level for at least one second dummy element among the plurality of second attention levels, and the calculation unit calculates the estimated input data based on the interpolated plurality of first attention levels and the interpolated plurality of second attention levels.
6. The apparatus according to claim 5, wherein the first interpolation unit replaces the at least one first dummy attention value with at least one missing value and interpolates the at least one missing value into at least one first interpolated value, and the second interpolation unit replaces the at least one second dummy attention value with at least one missing value and interpolates the at least one missing value into at least one second interpolated value.
7. The apparatus according to claim 2, wherein the first input unit assigns an embedding vector to each of the plurality of first element data, and the second input unit assigns an embedding vector to each of the plurality of second element data.
8. The apparatus according to claim 2, further comprising: a first position embedding unit disposed between the first input unit and the processing unit; and a second position embedding unit disposed between the second input unit and the processing unit, wherein the first position embedding unit assigns position information to each of the plurality of first element data; and the second position embedding unit assigns position information to each of the plurality of second element data.
9. The apparatus according to claim 1, further comprising: a third input unit including a learning model trained to take third input data including a plurality of non-time-series third-element data and output a third output value; and a coupling unit for coupling the third output value and the estimated input data, wherein the event estimation unit includes an event estimation model trained to take data obtained by coupling the third output value and the estimated input data and output the estimation result.
10. A method in a device comprising: receiving first input data including a plurality of first element data arranged in a time series; calculating estimated input data based on a plurality of first attention values that quantify the degree of attention each of the plurality of first element data is paying to each other; and outputting an estimated result of a user event based on the estimated input data.
