Alex Open Research Wiki

HUMI

3 min read

HUMI

Summary

HUMI is the official repository name for the nonparametric mixed-type mutual-information estimator introduced in Estimating Mutual Information between Time Series and Temporal Event Sequences Across Diverse Analysis Tasks. It measures dependence between numeric time-series observations and categorical events without binning the numeric values or imposing an order on event IDs.

The paper itself does not expand HUMI as an acronym; it usually says “the proposed estimator.”

Method Contract

  • Input: temporally aligned categorical events and numeric observations.
  • Numeric representation: values observed once enter a continuous nearest-neighbor entropy estimate; all occurrences of repeated values enter a discrete empirical-probability estimate.
  • Optional event representation: hierarchical clustering merges event types with similar conditional numeric-response distributions.
  • Output: clipped, normalized mixed-type mutual information.
  • Demonstrated uses: lagged-dependence scans, global/local repetition, discrete covariate ranking for forecasting, and continuous feature ranking for classification.
flowchart LR
  E["event labels"] --> H["HUMI"]
  S["numeric observations"] --> H
  H --> L["lagged dependence"]
  H --> R["repetition / seasonality"]
  H --> C["covariate ranking"]
  H --> F["feature ranking"]

Official Artifacts

No repository license, tagged release, package metadata, or focused test suite was verified at the audited commit.

Role In The Wiki

HUMI is a pre-model analysis and context-selection operator. It can identify which categorical events or covariates are statistically informative about a numeric target before those variables are supplied to a forecaster or latent-state model. It is especially relevant when sensor quantization produces repeated values and when multiple event names induce similar response distributions.

HUMI is not a forecasting model, representation learner, causal-discovery method, temporal point process, or action-conditioned world model. Time-delayed MI is a dependence diagnostic and MUST NOT be interpreted as causal intervention evidence by itself.

Relation To Foundation TSFM Agenda

Use the slot-level evidence table on the HUMI source page. At the entity level, HUMI contributes an event-to-numeric context-selection interface and a benchmark warning about quantization and repeated values. It does not maintain latent state or model event-conditioned transitions.

Implementation Caveats

  • The paper says event distributions use Wasserstein distance, while the pinned implementation uses energy distance.
  • The paper describes a fixed clustering distance threshold of 0.9; the code derives a threshold from the 90th percentile of pairwise distances.
  • The paper equation and pseudocode include mixture-weight terms, while the pinned implementation uses a different entropy decomposition without those terms.
  • Numeric range normalization has no explicit constant-series guard.
  • Negative MI estimates are clipped rather than accompanied by uncertainty or bias diagnostics.
  • The unique/repeated split changes with precision, rounding, and sample size.

Graph View

Backlinks