Multivariate Time Series Clustering via Vector Auto Regression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing time series clustering approaches are limited in scalability and utility due to treating each time series individually, which restricts their application in clustering millions of Multivariate Time Series (MTS) and fails to satisfy stationarity criteria for short time series data, such as mobile subscriber usage trends.
Innovation Solution
A system and method for multivariate time series clustering using Vector Auto Regression (VAR) models that learn parameters for multiple time series instances, allowing for simultaneous estimation and allocation of MTS to clusters based on minimal prediction error, with iterative refinement until convergence, enabling effective clustering of short, non-stationary time series data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing time series clustering approaches treat each time series individually with separate models, then measurement precision of individual series is maintained, but scalability deteriorates and cannot handle clustering of millions of MTS
Solution Approach 1:
The patent combines multiple individual time series models into a single unified VAR model that processes multiple time series simultaneously. This merging approach maintains clustering accuracy by capturing inter-series relationships while achieving scalability through shared parameter estimation across millions of MTS instances.
Solution Approach 2:
The VAR model serves as a universal framework that can handle diverse multivariate time series data from different sources and domains. The model's ability to simultaneously estimate parameters for multiple series makes it multi-functional, addressing both precision requirements for individual series and scalability needs for large-scale clustering.
2Adaptability or versatility
If traditional clustering methods are used on short time series data, then individual series analysis is possible, but generalization deteriorates due to sparsity and stationarity criteria cannot be satisfied
Solution Approach 1:
By combining information from multiple short time series into a unified VAR model, the patent overcomes the sparsity problem inherent in individual short series. The aggregated data from multiple series provides sufficient information for reliable parameter estimation while maintaining the ability to handle short series lengths that would otherwise fail stationarity criteria.
3Adaptability or versatility
If separate models are created for each time series, then model specificity is maintained, but device complexity increases and cannot handle millions of MTS
Solution Approach 1:
The VAR model functions as a universal modeling framework that maintains specificity for individual time series through its ability to capture unique inter-series relationships while avoiding the complexity of separate models. The single model structure handles millions of MTS efficiently through shared parameter estimation and unified computation.
Data Source
AI summary
Disclosed herein are methods and systems for providing multivariate time series clustering for customer segmentation. The system comprises of a model management unit that devices a customer segmentation procedure based on temporal variations of user preferences, using MTS clustering, and utilize the discovered clusters to learn association rules specific to each clusters, and improves campaign targeting. The order of the VAR model is fixed based on the nature of the data and length of the time series.