US-12626097-B2 - Ensemble time series model for forecasting

US12626097B2US 12626097 B2US12626097 B2US 12626097B2US-12626097-B2

Abstract

An ensemble time series prediction system that makes predictions based on observed data. The disclosed ensemble time series prediction system may leverage different types of datasets and information from different resources for making predictions. The disclosed ensemble time series prediction system may extract time dependent features from autoregressive time dependent data, embedding features from sparse datasets, continuous features from continuous dataset, and time lagged features from data that include time-lag information. The disclosed ensemble time series prediction system may then consolidate the features extracted from the different types of datasets and generate a set of consolidated input features for training a neural network, which may include a recurrent neural unit that finds sequential pattern for the sequence of input features and a regression unit that performs regression and predictions. The ensemble time series prediction system may output a set of outputs that include predicted values and associated confidence intervals.

Inventors

Nibhrat Lohia
Peyman Yousefian
Sayantan Mitra
Rajiv Gumpina

Assignees

HUMANA INC.

Dates

Publication Date: 20260512
Application Date: 20220427

Claims (14)

1 . An ensemble model stored on a non-transitory computer readable storage medium, the model associated with a set of parameters, and configured to receive a set of features, wherein the model is manufactured by a process comprising: obtaining a training dataset, wherein the training dataset is generated by concatenating two or more different datasets comprising historical observed values and time lagged features, wherein the time lagged features are extracted from a dataset including data with time lag information; and forming, from the training dataset, a set of consolidated input feature vectors by concatenating features extracted from the two or more different datasets, wherein the set of consolidated input feature vectors comprises a concatenation of (i) the historical observed values (ii) continuous features (iii) fixed-length embeddings generated from high-dimension sparse datasets, and (iv) fixed-length embeddings of time-lag sequences; for the ensemble model associated with the set of parameters, repeatedly iterating the steps of: performing a forward pass by providing the set of consolidated input feature vectors as inputs to a neural network of the ensemble model to generate predicted outputs; obtaining an error term from a loss function associated with the ensemble model; backpropagating the error term to update the set of parameters associated with the ensemble model; stopping the backpropagation after the error term satisfies a predetermined criterion; and storing the set of parameters on the computer readable storage medium as a set of trained parameters of the ensemble model and creating, using the stored set of parameters, a trained ensemble model configured to generate prediction results responsive to inputted consolidated feature vectors.
2 . The ensemble model of claim 1 , wherein the time lagged features are generated using a Long Short Term Memory (LSTM) neural network.
3 . The ensemble model of claim 1 , wherein the feature embeddings are of a first dimensionality and are generated based on input data that is of a second dimensionality, wherein the first dimensionality is smaller than the second dimensionality.
4 . The ensemble model of claim 2 , wherein the time dependent features are user parameters that change over time.
5 . The ensemble model of claim 1 , wherein the time-lagged data represents records storing information provided by a user.
6 . The ensemble model of claim 1 , wherein the ensemble model outputs predictions, wherein each prediction is associated with a confidence interval.
7 . The ensemble model of claim 1 , further comprising a recurrent unit that repeatedly passes outputs from one or more hidden layers of the recurrent unit back to the one or more hidden layers.
8 . The ensemble model of claim 1 , wherein the two or more different datasets are chosen from the following: an observed values dataset, an autoregressive time dependent dataset, a continuous dataset, a high-dimension dataset, and a time lagged dataset.
9 . The ensemble model of claim 8 , wherein the ensemble model generates a respective set of features for each of the autoregressive time dependent dataset, the continuous dataset, the sparse dataset, and the time lagged dataset.
10 . The ensemble model of claim 1 , wherein the historical observed values comprise observed medication adherence rate values aggregated by time period, and wherein the time-lagged features are extracted from claims data.
11 . The ensemble model of claim 10 , wherein the trained ensemble model is configured to generate a prediction of a medication adherence rate for a future time period.
12 . The ensemble model of claim 1 , wherein the ensemble model is configured to generate risk profile or risk prediction results for an entity, and wherein generating the training dataset comprises forming consolidated input feature vectors by concatenating features representing historical observed values associated with the risk profile and time-lag features derived from records that include time-lag information, including claims records.
13 . The ensemble model of claim 1 , wherein the historical observed values comprise generic dispense rate (GDR) values aggregated by time period for a population, and wherein the time-lagged features are extracted from claims data.
14 . The ensemble model of claim 13 , wherein the trained ensemble model is configured to generate a forecast of GDR for a future time period.

Description

FIELD OF INVENTION This invention relates generally to making predictions using machine learning models, and more particularly to making predictions using an ensemble time series prediction model. BACKGROUND Several applications generate time series data, for example, sensors, online systems, processes executing on computing systems, and so on. These applications often make predictions based on the time series data. A machine learning based predictive model is a commonly used technique for predicting outcomes based on such data. The predictive model may use historical values associated with an entity and predict a possibility of an event for the entity (e.g., predicting a probability that a person is associated with a particular event). Predictive models require a certain amount of training data for training parameters associated with the models. However, in some scenarios, the quantity of available training data is limited, which leads to sub-optimal performance of the predictions. SUMMARY Systems and methods are disclosed herein for an ensemble time series prediction system for making predictions based on observed data. The disclosed ensemble time series prediction system may leverage different types of datasets and information from different resources for making predictions. The disclosed ensemble time series prediction system may extract time dependent features from autoregressive time dependent data, embedding features from sparse datasets, continuous features from continuous dataset, and time lagged features from data that include time-lag information. The disclosed ensemble time series prediction system may then consolidate the features extracted from the different types of datasets and generate a set of consolidated input features for training a neural network, which may include a recurrent neural unit that finds sequential pattern for the sequence of input features and a regression unit that performs regression and predictions. The ensemble time series prediction system may output a set of outputs that include predicted values and associated confidence intervals. The ensemble time series prediction system may leverage multiple categories of data and construct an ensemble model with sub-models processing the multiple categories of data. For example, the ensemble time series prediction system may include four sub-models for processing and extracting time dependent features, embedding features, continuous features, and time-lagged features. The sub-models may perform different functionalities such as feature embeddings, feature extraction, etc. The sub-models may also include additional neural networks for processing datasets with specific characteristics. For example, the sub-model may further include a neural network such as an LSTM (Long short-term memory) for processing sequential data. The sub-models may output feature vectors for each input dataset and then a concatenation module may concatenate the outputs from the sub-models and generate a consolidated feature vector that include features from all input datasets. The consolidated feature vectors are used as input for a recurrent neural network of the ensemble time series prediction system. The recurrent neural network may perform forward pass and backpropagation and generate regression predictions. In one embodiment, the predictions may be associated with confidence intervals, that indicate a range and a likelihood that the predicted values fall within the range. The disclosed ensemble time series prediction system improves prediction performance when limited amount of data is available. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is an exemplary system environment including an ensemble time series prediction system, according to one embodiment. FIG. 2 illustrates an exemplary embodiment of modules in an ensemble time series prediction system, according to one embodiment. FIG. 3 illustrates an exemplary architecture of an ensemble time series prediction system, according to one embodiment. FIG. 4 illustrates an exemplary prediction process using an ensemble time series prediction system, according to one embodiment. The figures depict various embodiments of the present invention for purposes of illustration only. One skilled in the art will readily recognize from the following discussion that alternative embodiments of the structures and methods illustrated herein may be employed without departing from the principles of the invention described herein. DETAILED DESCRIPTION System Overview The Figures (FIGS.) and the following description relate to preferred embodiments by way of illustration only. It should be noted that from the following discussion, alternative embodiments of the structures and methods disclosed herein will be readily recognized as viable alternatives that may be employed without departing from the principles of what is disclosed. Reference will now be made in detail to several embodiments, examples of which are illustrated in t