Activity Number: Concepts & Contexts

• Activity number is a context-dependent term defined differently across disciplines, representing counts, measured signals, or activation rates depending on the research domain.
• It can denote the number of activity classes in recognition datasets, per-epoch actigraphy values in sleep studies, or even sunspot numbers as proxies for solar activity.
• In combinatorics and temporal network models, the term represents active element counts or node activation rates, playing a key role in generating functions and epidemic threshold analysis.

Searching arXiv for the cited papers to ground the article. I’ll look up the relevant arXiv records by id. Searching arXiv for (Sener et al., 2022, Marquez-Carpintero et al., 4 Mar 2025, Alruban et al., 2022, Xu et al., 2024, McIntosh et al., 2020, Flórez et al., 2022, Almeida et al., 2023, Perra et al., 2012, Ozaki et al., 2022, Zhou et al., 2023, Reiners et al., 2014), and (Fang et al., 2018). Activity number is a context-dependent technical expression rather than a single standardized variable. In current arXiv usage, it can denote the cardinality of activity classes in an activity-recognition dataset, the number of object instances around which procedural activities are organized, an actigraphy-derived per-epoch activity value, the sunspot number as a proxy for solar magnetic activity, the cardinality of an active-element set in matroidal and tree-based combinatorics, or an activity potential or activity rate governing temporal network formation. In stellar magnetism, closely related usage often replaces a discrete “number” with an activity indicator such as $L_{\rm X}/L_{\rm bol}$ or chromospheric excess emission. The term therefore requires domain-specific disambiguation before any quantitative interpretation is possible (Sener et al., 2022, Marquez-Carpintero et al., 4 Mar 2025, McIntosh et al., 2020, Flórez et al., 2022, Perra et al., 2012, Reiners et al., 2014).

1. Dataset-centered meanings in activity recognition

In activity-recognition research, “activity number” most often denotes the size of the prediction space, but the counted entity varies by dataset. In some cases it is the number of activity categories; in others it is the number of task instances or objects that induce procedural variation.

Dataset	Meaning of the number	Value
Assembly101	Unique take-apart toy vehicles	101
CADDI	In-class activities	19
Smartphone HAR	Total recorded activities; separate merged-result setting	6 total; 4 in one reported result
ARIC	Classroom activity categories	32

Assembly101 is an explicit example of nontrivial disambiguation. Its defining number, 101, refers to “101 unique toys” or “101 ‘take-apart’ toy vehicles,” not to the action vocabulary. The dataset contains 4321 videos, 362 unique sequences, more than 100K coarse and 1M fine-grained action segments, and 18M 3D hand poses. It also reports 1380 fine-grained action labels, formed from 24 verbs and 90 objects, and 202 coarse actions composed of 11 verbs and 61 objects. Accordingly, when “activity number” is queried for Assembly101, the correct value is 101 unique toy vehicle instances rather than 1380 action classes or 202 coarse action categories (Sener et al., 2022).

CADDI uses the term in the more standard class-count sense. The dataset contains 19 in-class activities collected from 12 participants in classroom scenarios, using smartwatch IMU data and synchronized stereo images. These 19 activities are partitioned into 8 continuous activities and 11 instantaneous activities, and the benchmark analysis notes that instantaneous actions are easier to classify than continuous ones. Here the activity number defines the full recognition label space for classroom behavior detection (Marquez-Carpintero et al., 4 Mar 2025).

The smartphone-sensor human activity recognition study distinguishes between the total number of activities recorded and the number of categories used in a particular reported result. Six activities were recorded—normal walk, fast walk, walk with a bag, downstairs walking, upstairs walking, and sitting—yet the headline 98% accuracy refers to a four-class setting in which walking variants are merged into broader categories such as walking, walking upstairs, walking downstairs, and sitting. The same paper also evaluates five-class and six-class configurations. This demonstrates that an “activity number” can depend on whether one is referring to raw data collection or to a specific evaluation protocol (Alruban et al., 2022).

ARIC returns to the direct class-count meaning. It contains 36,453 surveillance images covering 32 classroom activities, with three modalities—image, audio, and text—and viewpoints from front, middle, and rear classroom positions. The 32 categories are indexed from 0 through 31 in Table 1. The dataset also formalizes continual learning settings 8 + 6 × 4 and 12 + 5 × 4, and few-shot continual learning settings 16 + 4 × 4 and 20 + 4 × 3, but these settings are derived from, rather than replacements for, the base activity number of 32 classes (Xu et al., 2024).

2. Sensor-derived activity values and functional activity indicators

A different usage appears when “activity” denotes a measured signal rather than a class label. In sleep analysis and accelerometry, the relevant quantity is often an activity value or binary active/inactive indicator attached to each time window.

In the sleep-wake classification model based on photoplethysmography and activity signals, “activity” means the actigraphy-derived activity value supplied in the MESA Sleep dataset by Actiware-Sleep version 5.59. The labels are aligned to 30 s windows, and the final feature vector uses exactly three features per epoch: HR, HRV, and ACT. No additional activity statistics are introduced; ACT is the sole activity-related feature dimension. The proposed XGBoost model, built on HR + HRV + ACT, achieves Accuracy $= 77.62 \pm 0.56\%$, Sensitivity $= 91.14 \pm 1.15\%$, Specificity $= 53.66 \pm 1.11\%$, F1 $= 83.88 \pm 0.56\%$, and Kappa $= 0.480 \pm 0.008$. In this setting, the relevant “activity number” is not a category count but a per-window actigraphy value (Almeida et al., 2023).

The NHANES GM-FPCA study moves further from any scalar count interpretation. For subject $i$, day $j$, and minute $k$, activity is encoded as the binary indicator

$Z_{ijk} \in \{0,1\},$

derived from thresholding MIMS activity units:

$= 77.62 \pm 0.56\%$0

The cleaned accelerometry sample contains $= 77.62 \pm 0.56\%$1 participants, with 4,445 retained for the mortality analysis, up to seven days per person, and $= 77.62 \pm 0.56\%$2 minutes per day. The data are binary, multilevel, and functional simultaneously. GM-FPCA represents the latent predictor as

$= 77.62 \pm 0.56\%$3

thereby turning minute-level activity indicators into subject-level and day-level principal-component scores. The first three level-1 scores are significantly associated with mortality, with hazard ratios 0.85, 1.14, and 0.89, respectively. A plausible implication is that in statistical functional-data settings, “activity number” is best understood as a compressed representation of an activity process rather than as a discrete class cardinality (Zhou et al., 2023).

3. Solar-activity numbers and sunspot-number formulations

In solar physics, “activity number” refers to the sunspot number used as an observable proxy for magnetic activity. The quantity is not a class count but a physically interpretable time series.

In the Hale-cycle terminator analysis, the observable is the monthly mean total sunspot number from SILSO version 2.0, and the cycle maximum is treated as the sunspot-cycle amplitude. The paper defines terminator events as low-latitude magnetic-band cancellations at the equator, identified from the analytic signal

$= 77.62 \pm 0.56\%$4

where $= 77.62 \pm 0.56\%$5 is the sunspot-number time series and $= 77.62 \pm 0.56\%$6 its Hilbert transform. After smoothing the sunspot number and subtracting a 40-year rlowess trend, terminators are defined by zero crossings of the analytic phase. The interval between consecutive terminators, $= 77.62 \pm 0.56\%$7, measures the temporal spacing of magnetic-cycle terminations and hence the degree of overlap between Hale-cycle bands. The central empirical result is that the amplitude of the next cycle is strongly anti-correlated with this spacing:

$= 77.62 \pm 0.56\%$8

with Pearson correlation coefficient $= 77.62 \pm 0.56\%$9, significant at the 99.999% level. Using a predicted SC24 terminator at $= 91.14 \pm 1.15\%$0, the paper forecasts Sunspot Cycle 25 to have amplitude about $= 91.14 \pm 1.15\%$1, with a 68% confidence interval of 204–254 and a 95% interval of 153–305. In this usage, the activity number is the sunspot number itself, and its dynamics are interpreted through Hale-cycle overlap rather than minimum-to-minimum cycle length (McIntosh et al., 2020).

The multimessenger reconstruction of solar activity extends the same basic idea over much longer timescales. There, the reconstructed activity number is a decadal sunspot-number time series on the modern SILSO Version 2 scale, spanning 6755 BC to 2015 AD through a combination of cosmogenic-isotope reconstruction and more recent solar and geomagnetic proxies. The reconstruction uses $= 91.14 \pm 1.15\%$2Be and $= 91.14 \pm 1.15\%$3C archives for the deep past, and then extends to the present using Observed Sunspot Number, Group Number, the geomagnetic diurnal variation range $= 91.14 \pm 1.15\%$4, and the Interdiurnal Variation index IDV. The paper argues that the scaling from the Wu et al. reconstruction to SN V2 should be about $= 91.14 \pm 1.15\%$5, rather than 1.667. It reports agreement between the multimessenger and cosmogenic reconstructions with correlation $= 91.14 \pm 1.15\%$6, finds no secular uptick in activity over the last three hundred years, and identifies a sharp 87.6-year spectral peak but no significant Hallstatt-period power near 2300 years. Here again, “activity number” means sunspot number on a specified calibration scale (1810.11952).

4. Activity number in matroids, NBC sets, and rooted-tree correspondences

In combinatorics, activity number has a precise structural meaning. It is the cardinality of an active-element set associated with a basis, and it governs generating functions such as the Tutte polynomial and its relatives.

For a matroid $= 91.14 \pm 1.15\%$7 on a linearly ordered ground set $= 91.14 \pm 1.15\%$8, a basis element $= 91.14 \pm 1.15\%$9 is interiorly active if it is the smallest element of the fundamental cocircuit $= 53.66 \pm 1.11\%$0, while an element $= 53.66 \pm 1.11\%$1 is exteriorly active if it is the smallest element of the fundamental circuit $= 53.66 \pm 1.11\%$2. The paper denotes the interior and exterior active sets by $= 53.66 \pm 1.11\%$3 and $= 53.66 \pm 1.11\%$4, with cardinalities $= 53.66 \pm 1.11\%$5 and $= 53.66 \pm 1.11\%$6. These determine the Tutte expansion

$= 53.66 \pm 1.11\%$7

For NBC bases, exterior activity is 0, so only interior activity remains. The activity polynomial is

$= 53.66 \pm 1.11\%$8

where $= 53.66 \pm 1.11\%$9 counts NBC bases of activity $= 83.88 \pm 0.56\%$0, and $= 83.88 \pm 0.56\%$1 is the number of bounded regions. The paper explicitly states that when it says “activity number,” it means the interior activity number. In the generalized covering-system framework $= 83.88 \pm 0.56\%$2, an activity is a function $= 83.88 \pm 0.56\%$3 satisfying interval-partition properties, and the activity number is $= 83.88 \pm 0.56\%$4. The corresponding activity polynomial becomes

$= 83.88 \pm 0.56\%$5

Within this framework, $= 83.88 \pm 0.56\%$6, $= 83.88 \pm 0.56\%$7, and $= 83.88 \pm 0.56\%$8 count respectively bases of activity 0, all bases, and all independent sets (Flórez et al., 2022).

The same paper translates the combinatorial notion into a tree statistic. For colored decreasing or non-increasing rooted trees on vertex set $= 83.88 \pm 0.56\%$9, the activity $= 0.480 \pm 0.008$0 is the set of edges having $= 0.480 \pm 0.008$1 as parent and colored with 1. The activity number of such a tree is therefore the number of edges of the form $= 0.480 \pm 0.008$2 colored 1. The case of activity $= 0.480 \pm 0.008$3, where $= 0.480 \pm 0.008$4, is singled out because it matches bounded regions in the associated gainic hyperplane arrangement. For $= 0.480 \pm 0.008$5, the number of NBC bases of activity $= 0.480 \pm 0.008$6 equals the number of $= 0.480 \pm 0.008$7-colored trees having $= 0.480 \pm 0.008$8 edges of the form $= 0.480 \pm 0.008$9 colored 1; in particular, the number of bounded regions equals the number of such trees with no edge of that form. The paper gives concrete examples for the Linial arrangement:

$i$0

and

$i$1

The constant term 4 is the bounded-region count for the activity-0 case (Flórez et al., 2022).

5. Activity as a dynamical driver in temporal networks and epidemic risk models

In time-varying network theory, the term does not denote a label count at all. It instead describes a node-level propensity to initiate interactions.

The activity-driven network model defines the activity potential of node $i$2 as the number of interactions performed by that node in a given time window divided by the total number of interactions made by all nodes in the same window:

$i$3

The associated activity distribution $i$4 characterizes interaction dynamics at the population level. The model then defines the activity rate

$i$5

which is the probability per unit time that node $i$6 creates new contacts. At each discrete step, one starts with $i$7 disconnected nodes; node $i$8 becomes active with probability $i$9; if active, it creates $j$0 links to randomly chosen vertices; all edges are deleted at the next step. The expected number of active nodes per unit time is $j$1, the total number of edges per unit time is $j$2, and the mean instantaneous degree is $j$3. The integrated-network degree satisfies

$j$4

which yields the sparse-limit asymptotic relation

$j$5

In SIS dynamics, the epidemic threshold depends on the first two moments of the activity distribution:

$j$6

Thus, activity heterogeneity, not preferential attachment, generates hubs and shapes spreading thresholds (Perra et al., 2012).

A distinct but related epidemiological usage appears in GPS-based COVID-19 modeling, where activity is categorical and linked to different infection-rate coefficients. The model defines four daily activity categories—home, work, move, and stay-out—on a 15-minute, 1 km-grid representation of the Tokyo metropolitan area. The effective reproduction number is written as

$j$7

where

$j$8

The fitted coefficients are

$j$9

per 15 min. The paper states that stay-out is more than 28 times larger than other activities, and specifically 28 and 55 times larger than home and move/work, respectively. It further finds that high-risk zones are concentrated in downtown Tokyo, especially around Tokyo station and the Yamanote line, and that suppressing infection rate to 10% in the top 5, 10, or 20 highest-risk squares reduces total $k$0 by 17%, 25%, and 36% in the first period considered. This suggests that “activity number” in epidemic mobility models can refer either to the number of activity states or to activity-specific transmission coefficients attached to those states (Ozaki et al., 2022).

6. Activity indicators in stellar magnetism

In stellar astrophysics, “activity number” is rarely a formal variable name. Instead, activity is quantified by observables that trace magnetic heating, especially X-ray and chromospheric emission.

The generalized investigation of the rotation–activity relation analyzes 821 stars with masses below $k$1 and treats the main activity indicator as the normalized X-ray luminosity $k$2, with total X-ray luminosity $k$3 also examined directly. In the non-saturated regime, the best generalized scaling is

$k$4

equivalently

$k$5

The best-fit exponents are $k$6 and $k$7, with essentially the same scatter for the rounded values $k$8 and $k$9; the minimum scatter is about 0.346 dex. The Rossby-number formulation has larger scatter, about 0.371 dex. In the saturated regime, activity levels off near

$Z_{ijk} \in \{0,1\},$0

with critical period

$Z_{ijk} \in \{0,1\},$1

In this literature, activity is an emission indicator, and the debate concerns which control parameter—rotation period or Rossby number—best organizes it (Reiners et al., 2014).

The LAMOST open-cluster study uses chromospheric activity diagnostics for more than 700 late-type stars in the Pleiades, M34, Praesepe, and Hyades. Its quantitative measures are excess equivalent widths and excess fractional luminosities:

$Z_{ijk} \in \{0,1\},$2

$Z_{ijk} \in \{0,1\},$3

with analogous definitions for H$Z_{ijk} \in \{0,1\},$4 and Ca II K. The study identifies two activity sequences paralleling the rotational C and I sequences, finds a saturated regime typically at $Z_{ijk} \in \{0,1\},$5, and reports approximate saturation levels

$Z_{ijk} \in \{0,1\},$6

In the unsaturated regime, the eye-fit trends are roughly

$Z_{ijk} \in \{0,1\},$7

The paper also reports that fully convective slow rotators follow a similar rotation–chromospheric-activity relation, and that chromospheric diagnostics correlate with coronal emission and, more weakly, with spot coverage and photometric variability. A plausible implication is that, in stellar work, “activity number” is best interpreted as a shorthand for a calibrated activity indicator rather than for a literal count (Fang et al., 2018).

7. Disambiguation principles

Across these literatures, the first interpretive task is to identify what exactly is being counted, measured, or summarized. In activity-recognition datasets, the quantity may be the number of objects driving procedural variation, the number of categories, or the number of evaluation labels after class merging; Assembly101, CADDI, the smartphone HAR study, and ARIC illustrate all three possibilities (Sener et al., 2022, Marquez-Carpintero et al., 4 Mar 2025, Alruban et al., 2022, Xu et al., 2024).

In sensor and statistical modeling, the relevant quantity may be a time-window activity value or an entire activity process reduced to latent scores; the ACT feature in sleep-wake classification and the binary indicator process $Z_{ijk} \in \{0,1\},$8 in GM-FPCA are representative cases (Almeida et al., 2023, Zhou et al., 2023).

In physics and mathematics, the phrase becomes even more domain-specific. Solar-activity number refers to a sunspot-number proxy on a defined calibration scale, often with cycle amplitude treated as a key derived statistic (McIntosh et al., 2020, 1810.11952). In combinatorics, activity number is the cardinality of an active-element set, especially interior activity for NBC bases and its tree analogues (Flórez et al., 2022). In temporal networks, the analogous quantity is an activity potential or activity rate governing contact initiation (Perra et al., 2012). In stellar astrophysics, activity is quantified through emission indicators such as $Z_{ijk} \in \{0,1\},$9 or $= 77.62 \pm 0.56\%$00 rather than through a formal “number” variable (Reiners et al., 2014, Fang et al., 2018).

The common misconception is to assume that “activity number” always means the number of activity classes. The record across these papers shows otherwise. Depending on context, it may denote a label cardinality, a task-instance count, a per-epoch actigraphy value, a binary functional indicator, a sunspot-based activity index, a combinatorial cardinality, or a node-level activation propensity. Any rigorous use of the term therefore depends on preserving the local definition supplied by the relevant field.