Causal analysis with time series data

Inventors

Divakaran, Ajay • Yao, Yi • Kruk, Julia • Hostetler, Jesse • Huang, Jihua

Assignees

SRI International Inc

Abstract

In general, various aspects of the techniques are directed to causal analysis using large scale time series data. A computing system may convert large scale time series data to first time period records and second time period records according to a multi-scale time resolution. The computing system may implement a hierarchical machine learning model to generate embeddings that capture temporal characteristics of features of the large scale time series data. The computing system may generate a graph data structure indicating cause and effect correlations between features of the large scale time series data based on temporal dynamics captured in the cause and second time period records and/or the embeddings.

Core Innovation

The invention relates to an automated causal discovery system for large-scale time series that converts continuous time series data into multi-resolution discrete time period records by aggregating feature values for data points within time windows. The system generates a first time period record and a second time period record, where each record includes entries indicating combined time period feature values associated with features.

Based on the first time period record and the second time period record, the system generates a graph data structure indicating cause and effect correlations between features for purposes of indicating causal analysis of the time series data. Causal discovery is performed using temporal dynamics, including statistical and pruning causal discovery on the time-period records using independence tests, and/or using hierarchical transformer models that produce field embeddings and sequence embeddings capturing temporal trends between time periods, followed by graph construction that is pruned or validated using attribution.

The invention further includes counterfactual analysis to assess defect rate using counter-factual hypotheses, and outputs causal graphs to support causal questions. Example computing and user interface output is described where graphs support causal analysis of time series data in applications including healthcare productivity and throughput.

Claims Coverage

The independent claims cover two main causal-analysis pipelines: (i) time period record generation from feature values followed by causal graph generation, and (ii) an embedding-based pipeline that generates embeddings and sequence embeddings for temporal trends before producing a causal graph. Across the independent claims, the core coverage includes generating first and second time period records (or first and second embeddings), deriving a representation of temporal dynamics, generating a graph data structure indicating cause and effect correlations, and outputting an indication including the graph for causal analysis.

Overall, the independent claim set covers producing causal graphs from time series by either aggregating feature values into first and second time period records and generating a graph indicating cause-and-effect correlations, or by generating embeddings for feature values in first and second time periods, deriving a sequence embedding for temporal trends, and then generating a graph indicating cause-and-effect correlations. The claim coverage also explicitly includes outputting the graph as part of an indication for causal analysis and, in one independent claim, outputting data for a graphical user interface to include the graph data structure.

Stated Advantages

Documented Applications