KDD 2026

HUMI

Mutual Information between Time Series and Temporal Event Sequence

A direct way to quantify the interaction between a time series and a temporal event sequence, without learning a neural network model.

Input A continuous time series and a temporal event sequence, such as sales with promotions or traffic volume with holidays.
Output A normalized mutual information score in [0, 1] measuring how much the event sequence explains the signal.

Introduction.

Temporal analysis often requires measuring dependence between two heterogeneous data types: a continuous time series, such as sales or traffic volume, and a temporal event sequence, such as holidays, promotions, or system states. Existing measures do not directly fit this setting: Pearson correlation assumes two continuous signals, while pointwise mutual information is typically defined for discrete events.

We introduce HUMI (Mutual Information between Time Series and Temporal Event Sequence), a method for directly estimating mutual information between a time series and a temporal event sequence. HUMI produces a single dependence score analogous to Pearson correlation and supports downstream tasks including lag and causality analysis, seasonality measurement, repeated-pattern discovery, covariate selection, and feature selection.

Existing approaches, however, do not reach this combination of accuracy and generality: they usually run into one of two gaps:

1 Limited aligned training pairs

Neural estimators need many matched time series and event sequence pairs, but these pairs are expensive to label by hand and rarely exist as ready to use data.

2 Unstable on real data

Existing non-cross-modal measurement approaches often rely on ad hoc data transformations, and existing MI estimators struggle with rounded values, repeated measurements and redundant event states.

Neither gap is addressed directly, so existing methods cannot produce a reliable estimate, which raises a basic question: can we trust the estimated interaction score?

This kind of interaction estimate shows up across many research problems and settings, so an unreliable score causes trouble throughout downstream analysis. A few examples:

01 Causality analysis

Time-delayed mutual information (TDMI) can measure lagged dependence, but existing methods are not built to directly measure cross-modal causality strength between a time series and an event sequence.

02 Seasonality analysis

Classical methods can detect seasonality, but noisy or disrupted data can distort the result. Even when they find seasonality, they may not tell how much noise or deviation from a clean seasonal pattern is present.

03 Repeated pattern analysis

Seasonality is a global repeated pattern at a fixed interval across the whole series. A local pattern that holds only under certain conditions, such as a specific month, could be measured by manually splitting the series first, but no existing tool measures it directly.

04 Covariate selection

Forecasting often needs discrete covariates such as promotions, holidays or calendar attributes, ranked by how strongly each one relates to a continuous target like sales. Existing methods can already do this, but only after cutting the continuous target into bins first, and that binning step can distort the ranking.

These tasks may look different, but they share the same missing piece. HUMI addresses it by combining three key ideas:

01 Theoretical modeling

It provides a theoretical estimation foundation for mutual information between discrete and continuous variables, without forcing one side to become the other.

02 Time series duality

It models a recorded time series as both continuous variation and repeated value mass caused by finite measurement precision.

03 Redundancy clustering

It groups similar event states together, so repeated or overlapping event information does not overweight the final score.

Together, these components provide a more reliable measure of heterogeneous temporal dependence.

Background and related work.

Most existing dependence measures assume homogeneous inputs. A common workaround is to transform one data type into another, for example by discretizing continuous values,2 assigning numerical codes to event labels,3 or decomposing event states into binary variables. Such transformations can distort the result: binning may create or hide dependence, numerical codes may impose artificial ordering, and binary decomposition may lose the structure of the original event sequence.

Mixed discrete-continuous mutual-information estimators5,3 avoid some of these transformations, but they remain sensitive to repeated values caused by finite measurement precision and to redundant or highly correlated event states. Neural estimators offer greater flexibility,6 but require large collections of aligned time-series and event-sequence pairs and may produce dataset-dependent, difficult-to-interpret scores.

These limitations directly affect downstream tasks. Lag analysis requires reliable dependence estimates across temporal shifts,8 seasonality and repeated-pattern analysis require distinguishing strong patterns from noisy or conditional ones,9, 10, 11 and covariate selection requires consistently ranking discrete temporal factors.12,13 HUMI estimates dependence directly from a single observed pair, without binning or learned cross-modal representations, while explicitly accounting for repeated measurements and event redundancy.

1. Church and Hanks, Word Association Norms, Mutual Information, and Lexicography, 1990.

2. Liu et al., Discretization: An Enabling Technique, 2002.

3. Ross, Mutual Information between Discrete and Continuous Data Sets, 2014.

4. Luo et al., Correlating Events with Time Series for Incident Diagnosis, KDD 2014.

5. Kraskov et al., Estimating Mutual Information, 2004.

6. Belghazi et al., Mutual Information Neural Estimation, 2018.

7. Radford et al., Learning Transferable Visual Models From Natural Language Supervision, 2021.

8. Fraser and Swinney, Independent Coordinates for Strange Attractors from Mutual Information, 1986.

9. Box et al., Time Series Analysis: Forecasting and Control, 2015.

10. Cleveland et al., STL: A Seasonal-Trend Decomposition, 1990.

11. Brillinger, Time Series: Data Analysis and Theory, 2001.

12. Guyon and Elisseeff, An Introduction to Variable and Feature Selection, 2003.

13. Belmonte et al., Hierarchical Shrinkage in Time-Varying Parameter Models, 2014.

14. Dougherty et al., Supervised and Unsupervised Discretization of Continuous Features, 1995.

Approach.

HUMI starts by turning the temporal question into a dependence question. For each aligned pair, X is the event state, such as weekday, holiday, promotion or medication status, and Y is the corresponding signal value, such as traffic volume, sales, temperature or heart rate. In simple terms, we ask: how much does knowing the event state reduce uncertainty about the signal value?

The core object is the paper's mutual information formulation for one discrete variable and one continuous variable. The estimator computes this quantity through entropy terms: the event probabilities come from empirical frequencies, while the signal distribution is estimated nonparametrically. Instead of predicting the next signal value, HUMI measures dependence directly, which keeps it closer to a measurement operator than to a trained classifier, embedding model or forecasting model.

I(X, Y) = Σx∈Xy∈Y p(x,y) log( p(x,y) / (p(x)p(y)) ) dy

MI is high when knowing X makes Y less uncertain.

Two modeling details make this practical for real temporal data. First, a recorded time series is not always a clean continuous curve. Take temperature as an example: the real temperature changes continuously, but a sensor may only report one decimal place, so many timestamps can share the exact same value, such as 20.1, 20.1 and 20.2. Classical nearest-neighbor entropy estimators are designed for continuous samples, where exact duplicates should almost never happen. When duplicates appear, the nearest-neighbor distance can become zero, and the estimate becomes unstable. HUMI handles this with continuous-discrete duality: values that appear once are estimated as continuous samples, while repeated values are treated as repeated recorded states. This keeps the numeric meaning of the signal without pretending that duplicated readings are ordinary continuous samples.

Second, event sequences often contain labels that are different symbols but lead to similar signal behavior. For example, two promotion types in a sales dataset, or two calendar states in a traffic dataset, may produce similar distributions of aligned signal values. Treating every raw event label as fully separate can fragment the data and make the estimate unstable. HUMI optionally collects the aligned signal values for each event type, compares those empirical conditional distributions, and uses hierarchical clustering to replace redundant labels with latent event groups before estimating MI.

As a result, the estimator keeps the two modalities close to their native forms: the signal remains numeric, the event sequence remains categorical, and the final output is a normalized MI score for ranking event sequences, discovering lags and selecting covariates.

Core idea: estimate HUMI on native temporal data Align event states with signal values, model real recording artifacts, then return a normalized MI score. 01 Pair native inputs X: event state Y: aligned signal value 02 Model data artifacts DUAL SIGNAL Repeated values are modeled as discrete mass, not noise. EVENT GROUPING Merge redundant labels. 03 Estimate dependence I(X, Y) normalized HUMI score rank events, lags and covariates

What makes the estimator robust:

A Continuous-discrete duality

Unique values use continuous entropy; repeated values become discrete mass.

B Latent event grouping

Event labels with similar signal distributions can share one latent label.

Pseudocode

def temporal_mi(events, values): aligned = align_on_time(events, values) groups = collect_values_by_event(aligned) events2 = optional_cluster_events(groups) parts = split_repeated_values(events2, aligned.values) score = estimate_mixed_mi(parts) return normalize(score)
The actual implementation uses entropy estimates, repeated-value partitioning and optional event clustering; this captures the practical shape.

Downstream tasks.

The experiments ask whether one score can work as a practical dependence operator whenever a continuous signal must be compared with discrete temporal context.

We start from seasonality because it is the classic version of this problem. Traditional methods are good at finding a period, but they usually treat seasonality as a global signal pattern and can be hard to interpret when noise, missing values or holidays weaken the cycle. HUMI improves this by turning seasonality into a dependence score between the time series and calendar context: it can still identify the period, but it also says how strong the seasonal effect is. From there, the same idea leads to a more general temporal repeated pattern task, where repetition can depend on local context rather than a single global cycle.

The remaining tasks test the same idea in other settings: cross-modal TDMI measures delayed dependence between events and signals, covariate selection ranks discrete forecasting covariates for a continuous target, and feature selection shows that the estimator can also work outside temporal data when the input is simply continuous values paired with discrete labels.

01

Seasonality with strength

Classical tools such as ACF, seasonal decomposition and Fourier analysis are good at searching for a fixed period. But they mostly answer a yes or no style question: is there a dominant period? Real traffic is messier. Holidays, missing values and irregular disruptions can blur the pattern, while decomposition can still report a strong weekly cycle even when the observed signal is clearly partly broken.

HUMI treats seasonality as dependence between traffic volume and calendar context. This keeps the familiar weekly seasonality task, but adds a more useful score: it detects the 7 day pattern and also reflects how strongly that pattern appears. In the paper, June-July receives a higher MI score than December-January, matching the visual fact that winter traffic is more disrupted by holidays.

Dataset

Minneapolis daily traffic volume. The event sequence is day-of-week labels. June-July has clearer weekly seasonality; December-January is disrupted by holidays.

Baselines

ACF, seasonal decomposition plus ACF and Fourier transform. These methods mainly detect a candidate period.

Result

HUMI finds the same 7-day pattern in both periods, but the score drops from 0.6501 to 0.5664 when the pattern becomes weaker.

Minneapolis traffic volume seasonality comparison
DataMethodMeasureSeasonality
Jun-JulACF0.4651No
Jun-JulSD+ACF0.98177 days
Jun-JulFT0.137-8 days
Jun-JulOurs0.65017 days
Dec-JanACF0.2181No
Dec-JanSD+ACF0.91577 days
Dec-JanFT0.0811-12 days
Dec-JanOurs0.56647 days

ACF and FT can miss or drift under noise. SD+ACF detects weekly seasonality but stays close to 1 even when holidays disrupt the pattern; HUMI gives a lower score for the disrupted period.

02

General repeated patterns

After seasonality, HUMI extends repeated pattern analysis beyond one global fixed period. A local repeated pattern may appear only under specific contextual conditions, rather than across the whole series as a single periodic structure. This is why HUMI uses temperature as the next experiment: day/night labels provide a coarse temporal context, while combining month with day/night gives a finer context. The finer context narrows the possible temperature range, so the repeated pattern becomes stronger and easier to quantify.

This is where normalized MI is useful. If the event context explains the signal better, the score should increase in a stable and interpretable way. In the temperature experiment, richer contexts move from DN to TwoMon to DNTwoMon, and HUMI increases monotonically from 0.6124 to 0.8932 to 0.9596.

Dataset

Minneapolis 2023 air temperature sampled at 12-hour intervals. Event sequences describe day/night, two-month blocks and their joint context.

Baselines

Ross gives negative values, violating MI non-negativity. Mixture stays positive, but its scale is unbounded, making it hard to tell whether different contextual constructions are strong enough.

Result

As the context becomes richer, HUMI increases from 0.6124 to 0.8932 to 0.9596, giving a bounded score for local repeated patterns.

Minneapolis 2023 air temperature repeated pattern comparison
MethodDNTwoMonDNTwoMon
Ross-1.8418-0.7864-0.2332
Mixture0.37491.28071.5544
Ours0.61240.89320.9596

The score increases as the temporal context becomes more informative, while Ross gives invalid negative values and Mixture is unbounded.

03

Cross-modal TDMI

TDMI measures delayed dependence, which is useful for causality analysis. The usual setting compares two continuous time series. HUMI changes the input pair: it can compare an event type at time t with a continuous signal value at time t + tau, without turning either side into the other.

Dataset

Synthetic event sequence with three event types. Each event controls a different Gaussian distribution for the signal at a true delay of tau = 5.

Baselines

Series2Seq and Seq2Series force the data into one type; Ross and Mixture estimate discrete-continuous MI but do not fully handle repeated rounded values.

Result

HUMI recovers the delay peak at tau = 5 and closely matches the ground truth, showing that continuous-discrete duality improves stable TDMI estimation.

LagGround TruthOurs
00.03760.0353
51.09841.0988
90.03840.0355

The estimator recovers the true delayed dependence peak without converting event labels into numbers or discretizing the signal.

04

Covariate selection

Forecasting often needs discrete covariates such as promotions, holidays or calendar attributes. The missing tool is a direct way to rank those event covariates for a continuous target series before training a forecasting model.

Dataset

Rossmann and M5 sales forecasting. The target is a continuous sales time series; candidate covariates are discrete event or calendar attributes.

Baselines

Ross and Mixture are used to rank covariates by MI. Forecasting models then test whether the selected covariates actually help downstream prediction.

Result

Ours-Cluster gives the best NDCG in nearly all settings, showing that grouping redundant discrete states improves covariate ranking.

ModelDatasetOursOurs-Cluster
CatBoostRossmann0.940.95
CatBoostM50.810.86
DeepARRossmann0.900.92
DeepARM50.760.82
Chronos-2Rossmann0.920.95
Chronos-2M50.780.83
05

Feature selection

HUMI was motivated by temporal data, but the estimator itself does not depend on the order of instances. It only needs paired samples from a continuous variable and a discrete variable. This means the same formulation can be used for tabular feature selection: a continuous feature becomes the continuous variable, and the class label becomes the discrete variable.

Dataset

Mixed-type tabular classification datasets. Each continuous feature is paired with a discrete class label and ranked by mixed MI.

Baselines

Ross and Mixture are compared as MI-based feature selectors. Selected features are evaluated through RF, SVM and logistic regression.

Result

HUMI wins against Ross and Mixture in most settings, and clustering usually improves the ranking by reducing redundant feature-label information.

ComparisonWinTieLoseInterpretation
Ours vs Ross891417Ours wins more often.
Ours vs Mixture931512Strongest pairwise result.
With cluster vs Without cluster772023Clustering usually helps.
2 data typescontinuous signals plus timestamped events
0 traininga measurement method, not another prediction model
5 testslag, seasonality, repeated patterns and feature selection
[0, 1]normalized scores make event types easier to compare