Hierarchical Time Series Clustering for Model Fitting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for time series forecasting are inefficient when dealing with large volumes of data, as they either focus on individual model fitting, which is time-consuming, or use a single generalized model that loses individual time series characteristics, reducing accuracy.
Innovation Solution
A processor-implemented method that clusters time series into hierarchical clusters using Dynamic Time Warping (DTW) distance measure, identifying a best fit model for each time series and iteratively determining models for successive cluster heights based on Error Tolerance and Error Difference thresholds, optimizing model fitting time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If individual model fitting is performed for each time series, then forecast accuracy is improved, but model fitting time increases significantly
Solution Approach 1:
The patent segments the time series data into hierarchical clusters based on similarity, allowing different modeling strategies to be applied at different levels of the hierarchy. This segmentation enables the system to treat similar time series as a group while still preserving individual characteristics when needed, thus balancing accuracy and efficiency.
Solution Approach 2:
The patent applies partial individual model fitting only to representative time series from each cluster rather than to every single time series. By fitting models to cluster representatives and then applying them to similar time series within the cluster, the system achieves sufficient accuracy without the excessive time cost of individual fitting for all data points.
2Loss of time
If a single generalized model is used for all time series in hierarchical clusters, then model fitting time is reduced, but forecast accuracy deteriorates due to loss of individual characteristics
Solution Approach 1:
The patent implements local quality by allowing different time series within the same hierarchical cluster to have different model assignments based on their individual characteristics. While the cluster level provides a generalized model for efficiency, individual time series can deviate from the cluster model when their specific characteristics warrant a different approach, thus maintaining local accuracy while benefiting from global clustering.
Solution Approach 2:
The patent introduces dynamics by making the model assignment flexible rather than static. Time series can be assigned to cluster models dynamically based on similarity metrics, and individual time series can opt out of cluster models when their characteristics significantly differ. This dynamic approach allows the system to adapt between generalized and individualized modeling based on actual data characteristics.
3Productivity
If time series are clustered using existing methods, then processing efficiency is improved, but individual time series characteristics are generalized away, reducing accuracy
Solution Approach 1:
The patent creates a nested hierarchical structure where time series are organized into clusters at multiple levels. Individual time series maintain their unique identity at the base level while being nested within progressively more generalized cluster groups. This nesting allows the system to efficiently process data at higher hierarchical levels while preserving the ability to access and treat individual time series characteristics when needed at lower levels.
Data Source
AI summary
Method and system for model fitting to hierarchical time series clusters is described. A plurality of time series to be analyzed are clustered as hierarchical time series clusters using Dynamic Time Warping (DTW) as the optimal distance measure to create time series hierarchical clusters. The method disclosed recognizes least dissimilarity time series in the hierarchical time series clusters, and a best fit model is identified for the time series. The same model is continued up the hierarchy along the branch of the hierarchical time series clusters till the model identified satisfies Error Tolerance (ET) and Error Difference (ED) criteria. The method reduces the model fitting time or model building time by more than 50%.