Descriptor Aggregation

Descriptor Aggregation is a process in behavioral signal processing that combines descriptors to enhance pattern recognition and system behavior analysis.

Descriptor Aggregation is the explicit construction of a higher-level characterization derived from a declared collection of eligible Behavioral Signal Descriptor instances within a specified aggregation domain. This construction applies a defined aggregation operator, weighting or denominator rule, and yields output with particular semantics. Aggregation serves to summarize properties such as location (central tendency), dispersion, extremes, quantiles, prevalence, rates, temporal trends, distributions, or other statistical or structural characteristics of descriptor instances. It does so by discarding or preserving various amounts of structural information from the original descriptors.

It is critical to establish that Descriptor Aggregation is distinct from multiple related but different operations: it is not the enlargement of Descriptor Support by pooling raw evidence before descriptor calculation; it is not recomputing a descriptor directly on pooled raw signals; it is not merely concatenating multiple descriptors end-to-end; it is not normalizing descriptor values; it is not fusing signals from multiple sources; it is not feature selection; nor is it representation construction such as vector or matrix assembly without an explicit aggregation mapping. Each of these operations has different scientific and computational implications, and conflating them undermines interpretability and reproducibility.

Meaning and Boundaries of Descriptor Aggregation

Descriptor Aggregation is formally defined as a declared mapping from a collection of compatible descriptor instances to a new scalar, vector, distributional, structured, or otherwise explicitly defined characterization of that collection. The aggregated result is not just a restatement of any one contributing descriptor's meaning; rather, it possesses its own scientific semantics and provenance because it describes properties of the descriptor instances themselves or their grouping.

Several terms often appear in literature related to combining descriptor values but differ in scope and meaning:

Aggregation: A declared, explicit mapping combining descriptor instances into a new characterization with defined semantics.
Reduction: Often synonymously used with aggregation but may imply dimensionality reduction or loss of complexity.
Pooling: Generally refers to combining raw data or feature maps, sometimes without explicit semantics.
Functional: A function applied to a set of values; aggregation is a form of functional but with explicit domain and semantics.
Summary Statistic: A scalar or vector summary measure; a subtype of aggregation result.
Concatenation: Linking descriptors sequentially without combining or summarizing values.
Normalization: Rescaling or transforming descriptor values to a common standard, not combining.
Signal Fusion: Integrating multiple signals or modalities, often before descriptor extraction.
Descriptor Recalculation on Pooled Evidence: Computing a descriptor anew on combined raw data, not combining descriptors post hoc.
Representation Construction: Forming structured sets or arrays of descriptors without necessarily aggregating.

Although these terms overlap in some software or applied contexts, a defensible Descriptor Aggregation must explicitly state which objects are combined, what mapping or operator is applied, which structural aspects are retained or discarded, and what property the resulting object characterizes.

Operation	Input Objects	Output Meaning	Critical Non-Equivalence
Descriptor Aggregation	Compatible descriptor instances	New characterization summarizing descriptor set	Explicit mapping; output semantics differ from any single input; not raw data recomputation or concatenation
Descriptor Recalculation on Pooled Evidence	Raw evidence pooled across supports	Descriptor recomputed on pooled data	Input is raw data, not descriptors; can differ from aggregation of descriptors
Concatenation	Descriptor instances	Vector or array combining instances end-to-end	No summarization or mapping; no new semantics beyond collection
Normalization	Descriptor values	Rescaled or transformed descriptor values	Alters scale but does not combine or summarize
Signal Fusion	Multiple raw or preprocessed signals	Integrated signal or combined modality data	Operates before descriptor extraction; different processing stage
Feature Selection	Descriptor or feature vectors	Subset of features chosen	Selection, not aggregation or summarization
Representation Construction	Descriptor values or aggregates	Structured representation (e.g., vector, matrix)	Assembly without explicit aggregation mapping

An aggregation result itself constitutes a descriptor if it has a stable definition with declared input descriptor semantics, a specified aggregation domain, operator, parameters, and interpretable output property. Such a higher-level descriptor might represent, for example, mean event duration, median local speed, variability of segment energy, or distribution of interaction delays. Importantly, a stored summary is not scientifically interpretable merely because software produced a numerical column; it requires explicit semantic definition and provenance.

Aggregation Domain, Eligibility, and Compatibility

The aggregation domain is the declared collection or grouping scope of descriptor instances eligible to contribute to one aggregated characterization. Domains can be scientifically defined sets such as temporal intervals, windows, events, segments, episodes, trials, sessions, participants, entities, relations, modalities, spatial regions, or other meaningful groupings. The grouping key (the attribute or identifier partitioning instances) and the membership rule (the criteria for inclusion) must be explicit.

It is critical to distinguish Descriptor Support from the aggregation domain or aggregation support. Each contributing descriptor instance has its own evidence support — the raw data segment or context over which it was originally characterized. Aggregation operates over a collection of descriptor instances or their declared groups. The support of the aggregated characterization may relate to the union, intersection, envelope, event set, participant set, or another declared scope, but this relation must be explicitly declared rather than inferred from the stored bounds of contributing instances.

Semantic compatibility is a prerequisite for aggregation. Descriptor instances should normally share:

Compatible descriptor definitions,
Common input-property meaning,
Units or measurement scale,
Parameterization,
Preprocessing state,
Support semantics,
Entity interpretation.

An explicit mapping must justify any combination of instances lacking these compatibilities. Equal column names, numeric ranges, or storage types alone do not establish commensurability.

Aggregation eligibility states distinguish among:

Valid instances: fully defined, reliable outputs,
Invalid or undefined outputs: failed or erroneous computations,
Missing instances: not computed due to unavailable evidence,
Unavailable evidence: raw data or context missing,
Out-of-domain values: values outside valid parameter ranges,
Censored values: limited or truncated measurements,
Quality-excluded instances: filtered for quality reasons,
Provisional instances: preliminary or tentative computations,
Structurally inapplicable cases: instances not meaningful in the particular aggregation context.

The aggregation rule must clearly state which states contribute, which are excluded, which remain explicit in the output, and whether exclusion affects the denominator or weighting.

A single descriptor instance can belong to multiple scientifically meaningful aggregation groups, for example overlapping time windows, nested summaries, or multiple relational contexts. Aggregation must preserve group identity and repeated membership explicitly to avoid mistaking reused instances as independent evidence.

Element	Question It Answers	Risk If Unspecified
Input Descriptor Identity	Which descriptor instances are combined?	Mixing incompatible descriptors; loss of provenance
Aggregation Domain	What is the scope or grouping of instances?	Ambiguous or invalid aggregation population
Grouping Key	How are instances partitioned into groups?	Misassigned or overlapping groups; misinterpretation
Eligibility Rule	Which instances contribute?	Inclusion of invalid or inappropriate data; biased aggregates
Operator	How are instances combined?	Meaningless or invalid aggregation results
Weights	How is unequal contribution handled?	Incorrect influence of instances; biased estimands
Denominator	What normalizes counts or averages?	Invalid rates or proportions; misinterpreted summaries
Order Treatment	Is ordering preserved or ignored?	Loss of temporal or relational meaning; incorrect trend summaries
Missing/Invalid Policy	How are missing or invalid instances treated?	Hidden bias; invalid denominators; inconsistent aggregates
Output Semantics	What property does the aggregate characterize?	Misinterpretation; invalid scientific conclusions

Aggregation Operators and Functionals

Location and Central-Tendency Aggregation includes arithmetic mean, weighted mean, median, trimmed mean, geometric mean, and harmonic mean, used only when their scale assumptions are scientifically appropriate. For example, arithmetic mean is meaningful for interval or ratio scales but not for nominal categorical identifiers; averaging ordinal codes imposes unsupported metric assumptions and is generally invalid. The choice of operator must match the descriptor’s scale and meaning.

Robust Aggregation reduces sensitivity to outliers or contaminated descriptor instances using medians, trimmed or winsorized means, M-estimators, or other robust functionals. Robustness does not automatically distinguish genuine extremes from artifacts; aggressive robustness can suppress scientifically important rare events.

Dispersion and Variability Aggregation measures heterogeneity among descriptor instances using variance, standard deviation, interquartile range, median absolute deviation, range, or coefficient-like relative spread when scale semantics permit. It is important to distinguish variability among instances from uncertainty in a single descriptor instance or variability in the underlying raw signal.

Extrema, Quantile, and Threshold-Oriented Aggregation includes minimum, maximum, quantiles, exceedance fractions, tail summaries, and threshold counts. These preserve rare or boundary behaviors that central tendency hides but are sensitive to sample size, support coverage, noise, censoring, and threshold choice. The scientific meaning of thresholds and tail regions must be explicit.

Counts, Proportions, Occupancy, and Rates aggregate event-like, categorical, binary, or thresholded descriptors. Counts are raw event numbers; proportions normalize counts by eligible instances; rates normalize by valid time, observation opportunity, events, participants, distance, or other exposure measures. A count should not be interpreted as a rate when exposure differs.

Order-Aware Aggregation captures temporal or relational ordering via slope, first-to-last change, monotonic trend, transition summaries, run structures, temporal quantiles, or other order-dependent characterizations. Unlike permutation-invariant summaries like mean or median, these require explicit ordering coordinates and spacing assumptions.

Distribution-Preserving or Distribution-Oriented Aggregation outputs histograms, empirical distributions, quantile vectors, kernel or parametric summaries, mixture models, or other distributional forms that preserve heterogeneity beyond scalar summaries. Normalization, binning, bandwidth, parameterization, support, and sample-size assumptions affect interpretation. Distinguish empirical descriptor distribution from a probabilistic model of behavioral truth.

Aggregation Family	Property Preserved	Information Commonly Lost	Major Assumption or Risk
Location	Central tendency	Dispersion, rare events	Scale appropriateness; averaging nominal/ordinal invalid
Robust Location	Central tendency less sensitive to outliers	Some extremes or rare events	Aggressive robustness can hide meaningful extremes
Dispersion	Variability or spread	Order, exact distribution shape	Scale assumptions; conflation of uncertainty and variability
Extrema/Quantiles	Boundary or tail behavior	Central tendency, sample size effects	Threshold meaning; sensitivity to censoring/noise
Counts/Proportions/Rates	Event frequency or prevalence	Timing, order, exact distribution	Correct denominator choice; exposure definition
Order-Aware Trend	Temporal or relational ordering	Permutation-invariant properties	Explicit order and spacing assumptions
Distributional Summary	Full or partial distribution	Individual value identities	Binning, normalization, sample size, model assumptions

Weighting, Denominators, and Unequal Contribution

A representative weighted aggregation is rendered as:

a_{g} = \frac{\sum_{i \in E_{g}} w_{i} d_{i}}{\sum_{i \in E_{g}} w_{i}}

Here,

(g) is the aggregation group,
(E_g) is the set of eligible descriptor instances in group (g),
(i) indexes contributing descriptor instances,
(d_i) is a commensurate scalar descriptor value,
(w_i) is its declared nonnegative aggregation weight,
(a_g) is the resulting weighted aggregate.

This formula applies only when weighted averaging is scientifically meaningful for the descriptor scale. Normalized weights do not imply evidential validity or justify aggregation without declared semantics.

Weight semantics vary by context and should represent a single declared scientific or statistical role, for example:

Equal-instance weighting reflects typical instance,
Duration weighting reflects typical unit time,
Exposure weighting reflects opportunity or risk,
Event or participant weighting reflects sampling scheme,
Quality, uncertainty, confidence, or reliability weights reflect measurement properties.

Weights combining multiple meanings implicitly are discouraged. Confidence or quality weights require explicit justification rather than assumption as superior to equal weighting.

When aggregating over unequal descriptor supports, weighting affects interpretation. An unweighted mean of one value per behavioral segment weights each segment equally, regardless of duration. Duration-weighting gives longer segments greater influence. Event-weighting, participant-weighting, or exposure-weighting define still different estimands. Aggregation must state whether the target quantity is typical instance, typical unit time, typical event, typical participant, or another population.

Overlapping source supports pose a risk of double-counting. Descriptor instances from heavily overlapping windows, nested intervals, repeated events, or reused evidence can receive repeated numerical weight though they share source information. Aggregation can legitimately summarize a dense contour but the result should not be interpreted as if every contributing instance provides independent evidence.

Denominator identity is essential for rates, proportions, averages, and coverage-aware aggregates. Denominators can be:

Number of eligible instances,
Number of valid instances,
Total nominal duration,
Valid duration,
Total exposure,
Number of opportunities,
Total distance,
Or other defined quantities.

Excluding invalid or missing instances can change numerator and denominator; this should be explicit and not hidden behind generic labels like “mean” or “rate.”

Order, Grouping, and Nested Aggregation

Aggregation methods can be distinguished as permutation-invariant or order-sensitive. Means, medians, many quantiles, and ordinary histograms ignore the order of contributing descriptor instances. In contrast, slopes, transition summaries, sequence functionals, event-relative changes, and trajectory-shape summaries depend on ordering. Aggregation should preserve order when temporal, spatial, symbolic, or relational arrangement is integral to the property characterized.

Aggregation order affects results, especially where operations are non-associative. Applying one aggregation inside groups and then aggregating group summaries can differ from aggregating all instances directly. This is particularly true for medians, quantiles, variances, nonlinear transformations, robust estimators, thresholds, and unequal group sizes. Even arithmetic means require appropriate group-size or exposure weights for a mean of group means to equal the overall mean.

Nested grouping occurs across behavioral scales, e.g., local windows within events, events within episodes, episodes within sessions, or repeated observations within participants. Each stage may weight lower-level instances, groups, or exposure differently. Nested aggregation thus defines different population-weighting schemes and should not be treated as a computational optimization without scientific acknowledgment.

It is essential to distinguish aggregating descriptor instances from recomputing a descriptor on pooled underlying evidence. Equivalence occurs only for particular descriptor/operator combinations and support structures. For example, a correctly sample-weighted mean of disjoint local means can equal a global mean, but a median of medians generally does not equal a global median; averaging local standard deviations does not yield pooled standard deviation. Equivalence must be demonstrated explicitly, not assumed.

Multiscale aggregation involves descriptor instances computed at several support scales. These can be retained as scale-indexed values, summarized within each scale, or combined by an explicitly defined cross-scale functional. Combining scales can obscure which behavioral scale produced a pattern; hence scale identity should remain visible unless invariance or cross-scale integration is the intended descriptor property.

Quantity	When Results Can Coincide	Why They Can Differ	Interpretation Requirement
Mean	When weighted by support size or exposure	Unequal group sizes or weights	Appropriate weighting for global mean
Median	Rarely; median of medians ≠ global median	Nonlinear, order-sensitive nature	Avoid naive median-of-medians without validation
Variance / Standard Deviation	Only if pooled with correct formula	Variance is not additive; depends on group means	Use pooled variance formulas, not average of std devs
Event Rate	When denominators sum and events are disjoint	Overlapping events or exposures	Explicit denominator and overlap treatment
Quantiles	Only with identical samples or distributions	Sample size, grouping, and ordering differences	Interpret group quantiles carefully
Threshold Exceedance	When thresholds, supports, and exposure align	Different thresholds or censoring	Explicit threshold and denominator specification

Structured, Multivariate, and Relational Aggregation

Component-wise aggregation of vector-valued descriptors applies means, medians, quantiles, or other functionals separately to each component. This is valid only when components have stable, aligned semantics, units, coordinate frames, and ordering across instances. However, component-wise aggregation can destroy covariance, geometry, constraints, or directional relationships present in the original vector and should not be described as preserving the full structured descriptor.

Aggregation of categorical, set-valued, ordinal, symbolic, and other non-scalar descriptors produces outputs such as category proportions, empirical distributions, modes, ordered summaries, set unions or intersections (with justified semantics), transition summaries, or structure-aware counts. Arithmetic operators should not be applied merely due to numeric encoding of categories or symbols.

Relational and multisignal aggregation involves descriptors defined for entity pairs, channel pairs, modalities, or event correspondences. Aggregation across such relations is valid only when pair identity, directionality, correspondence, and support semantics remain explicit. Averaging several pairwise relations produces a group-level characterization but does not imply every individual pair shares that value.

It is necessary to distinguish aggregation from representation assembly. Collections of aggregated values can form descriptor sets or become coordinates in vectors, matrices, tensors, graphs, or other representations. However, assembly alone is not aggregation unless a declared mapping combines values into a new characterization. Dimensionality reduction or learned pooling changes representations under model-specific assumptions and should not be treated as ordinary descriptor aggregation without stable descriptor semantics.

Missingness, Uncertainty, and Dependence in Aggregation

Descriptor instances may be missing, invalid, censored, undefined, or structurally unavailable, representing distinct aggregation states. Strategies include:

Listwise exclusion (dropping instances),
Available-case aggregation (using all valid instances),
Imputation (filling missing values),
Explicit missing-category treatment,
Unresolved aggregate status.

These strategies answer different scientific questions and affect the effective aggregation domain. Undefined or missing instances should never be replaced with zero unless zero has the correct scientific meaning.

Descriptor uncertainty arises from finite support, noise, uncertain boundaries, parameter estimation, model-assisted descriptors, or prior processing. Aggregation may reduce random variation, preserve systematic uncertainty, or conceal heterogeneity. Distinguish variability among instances from uncertainty of the aggregate; do not infer narrow aggregate uncertainty solely from many dependent instances.

Dependence among contributing instances arises due to overlapping supports, repeated participants, nested segments, shared events, common preprocessing, correlated channels, or repeated derivation. Dependence does not prohibit aggregation but changes uncertainty, weighting, and inference. It forbids treating the raw number of descriptor instances as the number of independent observations.

Worked Example: Local Movement-Speed Descriptors

Consider local movement-speed descriptor instances computed over behavioral segments with unequal duration. Suppose we have:

Five segments with speeds: 1.2, 1.5, 1.8, invalid (missing), and 1.4 m/s,
Durations (seconds): 10, 5, 15, 8, and 12,
The fourth instance is invalid,
The first and second segments partially overlap in time.

Define the aggregation domain as all segments within a session. Eligibility excludes the invalid instance; grouping key is the session ID.

Aggregation options:

Equal-segment mean: Average speed treating each valid segment equally, ignoring duration and overlap.
- Eligible instances: 1.2, 1.5, 1.8, 1.4
- Mean = (1.2 + 1.5 + 1.8 + 1.4)/4 = 1.475 m/s
- Interpretation: Typical segment speed, regardless of segment length.
Duration-weighted mean: Weight speeds by segment duration.
- Weights: 10, 5, 15, 12 seconds for valid segments
- Weighted mean = (101.2 + 51.5 + 151.8 + 121.4) / (10 + 5 + 15 + 12)
  = (12 + 7.5 + 27 + 16.8) / 42 = 63.3 / 42 ≈ 1.507 m/s
- Interpretation: Typical speed weighted by time spent moving at that speed.
Median: Median speed ignoring weights.
- Speeds sorted: 1.2, 1.4, 1.5, 1.8
- Median = (1.4 + 1.5)/2 = 1.45 m/s
- Interpretation: Central tendency robust to outliers.
Recompute mean from pooled valid raw samples: Combine all raw movement samples underlying valid segments (including overlapping portions) and compute mean speed directly.
- Overlapping supports may double-count some samples.
- Result can differ from weighted mean unless corrected for overlap.
- Interpretation: Overall average speed across all valid time points.

Each aggregation answers a different scientific question. The equal-segment mean treats segments equally, the duration-weighted mean accounts for time spent, the median is robust to outliers, and recomputation from pooled evidence represents a global average but may be biased by overlapping supports.

Aggregation Adequacy, Sensitivity, and Provenance

Aggregation adequacy and sensitivity require verifying:

Input descriptor instances are semantically compatible,
Eligibility and grouping rules are correctly applied,
Operator assumptions match the descriptor scale and meaning,
Weights and denominators have declared meanings,
Order treatment is appropriate,
Missingness and dependence are handled explicitly,
Output is reproducible.

Sensitivity analysis compares plausible operators, weights, groupings, denominators, missing-data policies, overlap handling, robust vs. ordinary summaries, and nested-aggregation orders to determine if aggregated values, rankings, distributions, or conclusions materially change. Sensitivity to defensible alternatives should remain visible, not hidden by selecting only the most favorable aggregate.

Descriptor Aggregation provenance comprises all information needed to reproduce and interpret an aggregated characterization, including:

Aggregation-definition identity and version,
Contributing descriptor definitions and versions,
Aggregation domain and grouping keys,
Eligible and excluded instance identities,
Original support relations,
Ordering coordinate,
Operator and parameters,
Weighting semantics,
Numerator and denominator rules,
Missing or invalid-state policy,
Robust or trimming rules,
Distributional settings,
Nested aggregation sequence,
Dependence or repeated-membership information,
Uncertainty treatment,
Implementation and version,
Sensitivity results,
Resulting aggregate identity.

A defensible Descriptor Aggregation states what descriptor instances were summarized, why they were compatible, how contributions were defined, which structure was lost or preserved, and what higher-level property the aggregate legitimately characterizes.