Device for processing time series data having irregular time interval and operating method thereof
View Patent ↗Disclosed is a time-series data processing device that includes a preprocessor, a learner, and a predictor. The preprocessor generates time-series interval data based on a time interval of time-series data, generates feature interval data based on a time interval of each of features of the time-series data, and preprocesses the time-series data. The learner generates a weight group of a prediction model for generating a prediction result based on the time-series interval data, the feature interval data, and the preprocessed time-series data. The predictor generates a time-series weight, which depends on a feature weight of each of the features and a time flow of the time-series data, based on the time-series interval data, the feature interval data, and the preprocessed time-series data and generates a prediction result based on the feature weight and the time-series weight.
1. A time-series data processing device comprising:
a preprocessor configured to generate time-series interval data based on a time interval of time-series data, to generate feature interval data based on a time interval of each of features of the time-series data, and to preprocess the time-series data; and
a learner configured to generate a weight group of a prediction model for generating a prediction result based on the time-series interval data, the feature interval data, and the preprocessed time-series data,
wherein the preprocessor is further configured to:
preprocess the time-series data by putting an interpolation value into a missing value of the time-series data, and
generate masking data for identifying the missing value,
wherein the learner is further configured to generate the weight group further based on the masking data,
wherein the time-series data includes the features corresponding to each of a plurality of times, and
wherein the preprocessor is configured to:
generate the time-series interval data corresponding to a first time based on a difference between the first time and a second time preceding the first time among the plurality of times, and
generate the feature interval data corresponding to a first target feature of the first time based on a time interval between the first target feature corresponding to the first time and a second target feature corresponding to the second time preceding the first time.
2. The time-series data processing device of claim 1 , wherein, when the second target feature corresponding to the second time is a missing value and a third target feature corresponding to a third time preceding the second time is present, the preprocessor is configured to generate the feature interval data corresponding to the first target feature of the first time based on a difference between the first time and the third time.
3. The time-series data processing device of claim 2 , wherein the preprocessor is configured to generate the feature interval data corresponding to the second target feature of the second time based on a difference between the second time and the third time.
4. The time-series data processing device of claim 2 , wherein each of the time-series interval data and the feature interval data corresponding to an initial time among a plurality of times of the time-series data has an initial value.
5. A time-series data processing device comprising:
a preprocessor configured to generate time-series interval data based on a time interval of time-series data, to generate feature interval data based on a time interval of each of features of the time-series data, and to preprocess the time-series data; and
a learner configured to generate a weight group of a prediction model for generating a prediction result based on the time-series interval data, the feature interval data, and the preprocessed time-series data,
wherein the preprocessor is further configured to:
preprocess the time-series data by putting an interpolation value into a missing value of the time-series data, and
generate masking data for identifying the missing value,
wherein the learner is further configured to generate the weight group further based on the masking data, and
wherein the learner includes:
a feature learner configured to:
calculate a feature weight of each of the features based on a first parameter group of the weight group, the feature interval data, and the preprocessed time-series data, and
generate a first learning result of the preprocessed time-series data based on the calculated feature weight;
a time-series learner configured to:
calculate a time-series weight of each of times of the time-series data based on a second parameter group of the weight group, the time-series interval data, and the first learning result, and
generate a second learning result of the preprocessed time-series data based on the time-series weight; and
a weight controller configured to adjust the first parameter group or the second parameter group based on the first learning result or the second learning result.
6. The time-series data processing device of claim 5 , wherein the feature learner includes:
a feature irregularity processor configured to generate encoding data corresponding to each of the features by encoding the preprocessed time-series data and the feature interval data;
a feature weight calculator configured to calculate the feature weight by assigning the first parameter group to the encoding data;
a feature weight applicator configured to generate a feature application result by applying the encoding data or an intermediate result of the feature weight calculator to the calculated feature weight; and
a missing value processor configured to generate the first learning result by processing the feature application result based on a missing value of the time-series data.
7. The time-series data processing device of claim 5 , wherein the feature learner is configured to:
generate merged data corresponding to each of the features by classifying the preprocessed time-series data and the feature interval data based on the features,
generate encoding data by encoding the merged data, and
calculate the feature weight by assigning the first parameter group to the encoding data.
8. The time-series data processing device of claim 5 , wherein the feature learner is configured to:
generate merged data by merging the preprocessed time-series data and the feature interval data,
generate encoding data corresponding to each of the features by encoding the merged data, and
calculate the feature weight by assigning the first parameter group to the encoding data.
9. The time-series data processing device of claim 5 , wherein the time-series learner includes:
a time-series irregularity processor configured to encode the time-series interval data;
a time-series weight calculator configured to calculate the time-series weight by assigning the second parameter group to the encoded time-series interval data and the first learning result; and
a time-series weight applicator configured to generate the second learning result by applying the first learning result or an intermediate result of the time-series weight calculator to the calculated time-series weight.
10. A time-series data processing device comprising:
a preprocessor configured to generate time-series interval data based on a time interval of time-series data, to generate feature interval data based on a time interval of each of features of the time-series data, and to preprocess the time-series data; and
a predictor configured to generate a time-series weight, which depends on a feature weight of each of the features and a time flow of the time-series data, based on the time-series interval data, the feature interval data, and the preprocessed time-series data and to generate a prediction result based on the feature weight and the time-series weight,
wherein the predictor includes:
a feature predictor configured to generate a first result of the preprocessed time-series data based on the feature weight;
a time-series predictor configured to generate a second result of the preprocessed time-series data based on the time-series weight; and
a result generator configured to calculate the prediction result corresponding to a target time based on the second result.
11. The time-series data processing device of claim 10 , wherein the preprocessor is further configured to:
preprocess the time-series data by putting an interpolation value into a missing value of the time-series data, and
generate masking data for identifying the missing value, and
wherein the feature predictor includes:
a feature irregularity processor configured to generate encoding data by encoding the feature interval data and the preprocessed time-series data;
a feature weight calculator configured to generate the feature weight by applying a prediction model to the encoding data;
a feature weight applicator configured to generate a feature application result by applying the encoding data or an intermediate result of the prediction model to the feature weight; and
a missing value processor configured to generate the first result by applying the masking data to the feature application result.
12. The time-series data processing device of claim 10 , wherein the time-series predictor includes:
a time-series irregularity processor configured to encode the time-series interval data;
a time-series weight calculator configured to generate the time-series weight by applying a prediction model to the encoded time-series interval data and the first result; and
a time-series weight applicator configured to generate the second result by applying the first result or an intermediate result of the prediction model to the time-series weight.
13. The time-series data processing device of claim 10 , wherein, when a value for a first target feature is present at a first time among a plurality of times of the time-series data and a second target feature is a missing value at a second time preceding the first time, the time-series interval data corresponding to the first time is different from the feature interval data corresponding to the first target feature of the first time.
14. A time-series data processing device, comprising:
a preprocessor configured to generate time-series interval data based on a time interval of time-series data, to generate feature interval data based on a time interval of each of features of the time-series data, and to preprocess the time-series data; and
a predictor configured to generate a time-series weight, which depends on a feature weight of each of the features and a time flow of the time-series data, based on the time-series interval data, the feature interval data, and the preprocessed time-series data and to generate a prediction result based on the feature weight and the time-series weight,
wherein the predictor includes:
a feature analyzer configured to generate the feature weight based on the feature interval data and the preprocessed time-series data;
a time-series analyzer configured to generate the time-series weight based on the time-series interval data and the preprocessed time-series data; and
an integrated weight applicator configured to generate the prediction result by applying the preprocessed time-series data to the feature weight and the time-series weight, and
wherein, when a value for a first target feature is present at a first time among a plurality of times of the time-series data and a second target feature is a missing value at a second time preceding the first time, the time-series interval data corresponding to the first time is different from the feature interval data corresponding to the first target feature of the first time.
15. A method of calculating, based on time-series data including a missing value, a prediction result corresponding to a target time, the method comprising:
generating interpolation data by putting an interpolation value into the missing value of the time-series data;
generating time-series interval data based on a time interval of the time-series data including generating the time-series interval data corresponding to a first time based on a difference between the first time and a second time preceding the first time among a plurality of times;
generating feature interval data based on a time interval of each of features of the time-series data including generating the feature interval data corresponding to a first target feature of the first time based on a time interval between the first target feature corresponding to the first time and a second target feature corresponding to the second time preceding the first time;
generating masking data based on the missing value;
generating a feature weight of each of the features based on the interpolation data and the feature interval data;
generating a first result based on the feature weight and the masking data;
generating a time-series weight, which depends on a time flow of the time-series data, based on the first result and the time-series interval data;
generating a second result based on the time-series weight, and
calculating the prediction result corresponding to the target time based on the second result.
16. The method of claim 15 , further comprising:
adjusting a weight group for generating the feature weight or the time-series weight based on the second result.