Campaign: Coordinated Actions & Outcomes

• Campaign is a coordinated process with a defined objective, target population, and specific actions that yield measurable outcomes.
• Research models employ optimization, diffusion, and strategic interventions to influence voting behavior, social opinions, and market dynamics.
• Campaign strategies are assessed using domain-specific metrics such as cost-efficiency, outreach, engagement rates, and detection performance.

Across the research literature, “campaign” denotes a coordinated effort organized around a target state: making a preferred candidate win, maximizing social-network outreach, steering equilibrium opinion, predicting or optimizing the effect of overlapping marketing actions, disseminating public-health content through a funnel from awareness to conversion, coordinating observations during a rare astronomical event, or benchmarking systems under common rules in a shared-task setting (Faliszewski et al., 2020, Kotnis et al., 2016, Hegselmann et al., 2014, Chu et al., 2022, Yan et al., 27 Aug 2025, Leadbeater, 2011, Aepli et al., 2023). The term therefore spans intervention, observation, and evaluation. What remains stable is the combination of a specified objective, a defined population or dataset, an action repertoire, and explicit success criteria such as winner determination, outreach, cost-efficiency, completeness contrast, or macro-averaged classification performance (Elkind et al., 2010, Kotnis et al., 2016, Wahhaj et al., 2013).

1. Semantic range and basic structure

In computational social choice and opinion dynamics, a campaign is an intervention process over preferences or opinions. The action set may consist of shift bribery, degree-dependent incentivization, activation of activists, or the placement of a strategic opinion in each period. The objective is usually formalized as a constrained optimization problem: minimize the cost of making a preferred candidate win, maximize the eventual sum of expressed opinions, or maximize the number of agents inside a conviction interval at a terminal time (Elkind et al., 2010, Faliszewski et al., 2020, Gionis et al., 2013, Böttcher et al., 2018, Hegselmann et al., 2014).

In marketing and advertising research, the same term refers to a coordinated set of business-facing or public-facing actions delivered across users, events, keywords, platforms, and time periods. A campaign may be modeled as an individual-level prediction problem over observational industrial data with multiple intertwined events, as a semi- and fully-automated pay-per-click workflow for keyword extraction, ad generation, monitoring, and bid optimization, or as a national digitally-enabled health promotion effort built around an explicit funnel of awareness, consideration, and conversion (Chu et al., 2022, Thomaidou et al., 2012, Yan et al., 27 Aug 2025).

In astronomy and benchmark design, “campaign” has a different operational meaning. It may denote an organized observing program, such as the International Epsilon Aurigae Campaign 2009–2011, the Z CamPaign, or the Gemini NICI Planet-Finding Campaign, where the central tasks are coordination, data collection, and pipeline validation rather than persuasion. It may also denote a structured evaluation exercise, as in the VarDial Evaluation Campaign 2023, where organizers release data and rules, participants submit systems, and standardized metrics determine comparative performance (Leadbeater, 2011, Simonsen, 2011, Wahhaj et al., 2013, Aepli et al., 2023).

This range suggests that campaign is unified less by domain than by structure. Across the cited work, campaigns are typically finite-horizon, target-oriented processes with explicit observables, and their study depends on formal models that connect local actions to aggregate outcomes.

2. Campaigns as intervention in elections, opinion dynamics, and social networks

A major research tradition treats campaigning as a computational intervention on electoral or social-diffusion dynamics. In “campaign management” under voting rules, an external party seeks to help a preferred candidate $p$ by paying to move $p$ upward in voters’ rankings. For an election $E=(C,V)$, shift bribery is represented by a shift-action $t=(t_1,\dots,t_n)$, with total cost $\Pi(t)=\sum_{i=1}^n \pi^i(t_i)$. The central optimization problem is to find a successful shift-action of minimum cost. For scoring rules, there is a polynomial-time $2$-approximation even with weighted voters; for Copeland$^\alpha$ there is a polynomial-time $m$-approximation, and for maximin a polynomial-time $O(\log m)$-approximation (Elkind et al., 2010).

A more general formalization appears in campaigning on society graphs. There, voters are not modeled individually but as types or clusters. If there are $\tau$ types, the society is a vector $p$0, and the corresponding society graph is $p$1. In the basic version, an edge links two types whose preference orders differ by one adjacent swap. Diffusion is deterministic: a vertex $p$2 is assimilated into a neighbor $p$3 whenever

$p$4

Campaigning is modeled as unit-cost shift bribery applied before diffusion. The decision problem $p$5-BSG asks whether at most $p$6 unit-cost actions can make the preferred candidate $p$7 an $p$8-winner after diffusion. Synchronous diffusion converges in at most $p$9 steps, whereas asynchronous diffusion is order-dependent; deciding whether some asynchronous order makes $E=(C,V)$0 a Plurality winner is NP-hard, and Borda-BSG is NP-hard in both synchronous and asynchronous settings. At the same time, synchronous $E=(C,V)$1-BSG is fixed-parameter tractable with respect to the number $E=(C,V)$2 of candidates for every ILP-expressible voting rule $E=(C,V)$3, and the asynchronous optimistic and pessimistic variants inherit fixed-parameter tractability via ILP and Lenstra’s algorithm (Faliszewski et al., 2020).

A related line of work studies campaigning as degree-based incentivization in social networks. In the percolation-based model of cost-effective campaigning, the network is an undirected, uncorrelated random graph generated by the configuration model, and selected nodes become more effective spreaders, with transmissibility $E=(C,V)$4. If $E=(C,V)$5 is the probability of incentivizing a degree-$E=(C,V)$6 node, then the key scalar is

$E=(C,V)$7

the probability that a randomly traversed link reaches an incentivized node. When $E=(C,V)$8, outbreak size is strictly monotone in $E=(C,V)$9, which allows both “minimum cost for target outreach” and “maximum outreach for fixed budget” to be reduced to linear programs. A closely related threshold-based model of incentivized campaigning uses reliability-theory arguments to prove uniqueness and monotonicity of the fixed point $t=(t_1,\dots,t_n)$0, then solves the dual optimization problems in $t=(t_1,\dots,t_n)$1 time by sorting degree classes and greedily filling capacities (Kotnis et al., 2016, Kotnis et al., 2016).

Other models make the campaigner’s influence explicit at the level of individual opinions. In the CAMPAIGN problem over a Friedkin–Johnsen/DeGroot-style equilibrium model, each node has an internal opinion $t=(t_1,\dots,t_n)$2 and an expressed opinion $t=(t_1,\dots,t_n)$3; the campaign chooses a set $t=(t_1,\dots,t_n)$4 of size $t=(t_1,\dots,t_n)$5 and clamps those nodes to value $t=(t_1,\dots,t_n)$6, with the objective of maximizing the total equilibrium opinion $t=(t_1,\dots,t_n)$7. The problem is NP-hard, but the objective is monotone and submodular, so the standard greedy algorithm yields a $t=(t_1,\dots,t_n)$8 approximation. In the “campaign problem” of optimal opinion control, one strategic agent chooses a control opinion $t=(t_1,\dots,t_n)$9 at each stage $\Pi(t)=\sum_{i=1}^n \pi^i(t_i)$0 to maximize the number of normal agents inside a target interval $\Pi(t)=\sum_{i=1}^n \pi^i(t_i)$1 at time $\Pi(t)=\sum_{i=1}^n \pi^i(t_i)$2. Under bounded-confidence dynamics, the model is discontinuous and computationally difficult even for very small instances, which motivates mixed-integer linear programming, upper- and lower-bound formulations using a safety margin $\Pi(t)=\sum_{i=1}^n \pi^i(t_i)$3, and heuristics such as the strongest-guy method, genetic algorithms, and model predictive control (Gionis et al., 2013, Hegselmann et al., 2014).

A further political-campaign model emphasizes activists, political clout, and budget exhaustion. Nodes are divided into activists $\Pi(t)=\sum_{i=1}^n \pi^i(t_i)$4, persuadable individuals $\Pi(t)=\sum_{i=1}^n \pi^i(t_i)$5, and empty nodes $\Pi(t)=\sum_{i=1}^n \pi^i(t_i)$6. Activists are mobile; they may move to empty nodes and try to persuade neighboring persuadables subject to a threshold condition such as $\Pi(t)=\sum_{i=1}^n \pi^i(t_i)$7. The mean-field system has a unique stationary solution. Under equal clout and equal costs, the side with more activists wins; under unequal clout or costs, higher convincing probability $\Pi(t)=\sum_{i=1}^n \pi^i(t_i)$8 or lower cost $\Pi(t)=\sum_{i=1}^n \pi^i(t_i)$9 can overturn an activist or budget disadvantage. The paper uses the 2016 U.S. presidential election as an illustration of the claim that political clout can overcome a substantial budget disadvantage or a lower number of activists (Böttcher et al., 2018).

Taken together, these models make campaign design a branch of constrained control and optimization. The operative variables differ—rank shifts, diffusion schedules, degree-class incentives, activist deployment, or control opinions—but the common task is to choose a small set of interventions that alter a collective end state under explicit computational, budgetary, or dynamical constraints.

3. Marketing, advertising, and digitally-enabled campaigns

In marketing research, campaigns are often evaluated under observational rather than interventional data. HapNet treats marketing campaign effect prediction as an individual-level regression problem in which each user may be exposed to several campaigns simultaneously, the overlap is highly unbalanced, and the number of possible combinations is $2$0 for $2$1 campaigns. The paper explicitly distinguishes this from classical causal effect estimation: the task is to predict the factual outcome under the actual campaign assignment, even though the assignment is biased. HapNet represents the problem as a hierarchical parse-tree-like structure,

$2$2

implemented by four capsule stages: Disentangled Feature Capsule, Single Event Capsule, Event Cluster Capsule, and Outcome Capsule. Its key mechanism for variable-size event combinations is property-aware additive attention (PAAA), and its loss combines mean squared error with feature reconstruction, $2$3 with $2$4. On Alipay datasets, HapNet reports lower MAPE than all listed baselines: on Promotion, $2$5; on Demotion, $2$6. The ablation results show that removing PAAA or reconstruction loss degrades performance, especially as the number of events grows (Chu et al., 2022).

A different strand addresses campaign creation and optimization in pay-per-click advertising. The AD-MAD framework integrates campaign generation, monitoring, and optimization. Its GrammAds component extracts and expands keywords from landing-page content, while Adomaton handles deployment, statistics collection, and budget-constrained optimization. Keyword weighting is based on HTML tag importance, n-gram construction, search-snippet expansion through Google JSON/Atom Custom Search API and Apache Lucene, and ad-text creation through summarization under character constraints. The optimization problem is cast as a Multiple-Choice Knapsack Problem over keyword-bid options, with a genetic algorithm used to maximize profit or traffic under budget constraints. The prototype is evaluated on real Google AdWords campaigns and is reported to outperform systematically competing manually maintained campaigns, with better average CPC and better ad positions over the comparison period (Thomaidou et al., 2012).

The 2025 Singapore digitally-enabled health promotion campaign shows campaign design in a public-health setting. It was a 3-month, nation-wide campaign conducted from February to April 2025 across TikTok, Facebook, Instagram, and YouTube. Content included narrative videos, infographics, watchlist-style posts, Q&A materials, testimonials, and direct links to mindline.sg. The campaign used an adapted funnel,

$2$7

with narrative videos as awareness content, “Watchlist” and “Ask Mili Anything” materials for consideration, and testimonials with calls-to-action for conversion. The flagship creative was the trilogy “Inspector Quek: Dark Place Casefiles,” with episodes on Anxiety, Unsaid Misery, and Internal Torment. The campaign generated 3.49 million total impressions and reached 1.39 million unique residents, equivalent to 33.3% of Singapore’s resident population of 4.18 million. It accumulated over 630,000 video views and 18,768 engagements, with reported CPM of \$2$829.33, and <a href="https://www.emergentmind.com/topics/cyber-physical-awareness-cpa" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">CPA</a> of \$6.06. Stronger engagement was observed among young adults aged 25–34 and, more broadly, the 18–44 age range; female users had lower CPC than male users; TV screens produced the highest view counts, whereas mobile phones generated the greatest click-through activity (Yan et al., 27 Aug 2025).

Post hoc analysis of campaign effectiveness is also a major theme. The study of the 2016 Russian Facebook ads campaign analyzes 3,516 ads released by the U.S. House Intelligence Committee, with 2,600 ads having at least one impression and 2,539 ads both having at least one impression and being paid in RUB. Effectiveness is defined at the individual-ad level by CTR, at the campaign level by CPM and CPC, and at the party level by CTR, CPM, and total cost. Using topic metadata, FastText-based similarity, overlap coefficients, and Louvain community detection, the authors move from 9 initial clusters to 21 final campaigns, including Police Brutality, Black Lives Matter, Immigration, Anti-immigration, Conservative, Veterans, 2nd Amendment, Religious, and Anti-Islam. More effective ads have lower compounded and positive sentiment, greater use of past-tense verbs and adverbs, and more personalized language. The duration and thematic pairing of campaigns such as Anti-Islam/Islam and Anti-Immigration/Immigration are interpreted as suggesting a desire to sow division rather than sway the election (Dutt et al., 2018).

These studies establish several domain-specific features of campaign research in marketing and media systems. Campaigns are embedded in platform analytics, audience segmentation, content design, and budget allocation; they are often optimized against measurable proxies such as MAPE, clicks, CTR, CPC, CPA, or conversions; and they may be evaluated either prospectively, as in automated optimization, or retrospectively, as in ad-effectiveness analysis. At the same time, some of the most important results explicitly caution that predictive success on observational logs is not equivalent to causal identification or behavioral change.

4. Scientific observing campaigns

In astronomy, campaign commonly denotes an organized, often long-duration observing effort designed to exploit a rare event or settle a disputed classification. The International Epsilon Aurigae Campaign 2009–2011 was a coordinated pro-am effort organized around the eclipse of $2$9 Aurigae, a naked-eye magnitude 3 star in an eclipsing binary with a 27.1-year period and an approximately 2-year flat-bottomed eclipse of about 0.8 mag in $^\alpha$0. The campaign was organized primarily by Bob Stencel, Jeff Hopkins, and Robin Leadbeater, and it distinguished itself from the broader AAVSO Citizen Sky project by focusing on experienced amateurs contributing research-grade photometry and spectroscopy. By May 2010 the spectroscopic team included 12 amateur observers from five countries, using mostly LHIRES III at $^\alpha$1 and two eShel systems at $^\alpha$2; more than 330 spectra were available through the campaign website. The K I 7699 Å line was the principal diagnostic because it is absent outside eclipse apart from a small constant interstellar component. The line appeared more than two months before the photometric brightness had started to drop, widened, shifted slightly redward, and by May 2010 was moving blueward. The growth in equivalent width occurred in steps, interpreted as possible density variations such as ring-like structure, spiral arms, or arcs. Measurements of the maximum radial velocity at the red edge of the K I line were consistent with a rotating Keplerian disk and with the low-mass model of Hoard, Howell, and Stencel (Leadbeater, 2011).

The Z CamPaign is another observing campaign centered on contested classification. Launched on 25 September 2009 by the AAVSO Cataclysmic Variable Section, it targeted Z Camelopardalis-type dwarf novae, or UGZ systems, because catalogues disagreed on which few dozen systems truly belong to the class. The campaign’s central observational criterion is the presence of standstills, defined as prolonged states about one magnitude below outburst maximum. Its goal is to build detailed light curves covering the entire range of variability for candidate stars, beginning with $^\alpha$3-band observations and extending later to other bandpasses. Early results confirmed ten systems as UGZ stars, including RX And, Z Cam, AT Cnc, SY Cnc, AH Her, and HX Peg, while identifying several likely imposters such as TW Tri, KT Per, BI Ori, CN Ori, SV CMi, and AB Dra. The campaign also highlighted unusual repeated patterns in IW And and V513 Cas consisting of outburst, standstill, another outburst, then rapid fade to quiescence, repeating cyclically (Simonsen, 2011).

The Gemini NICI Planet-Finding Campaign uses the term in still another technical sense: a large, dedicated direct-imaging survey. Using the Gemini South 8.1-m telescope and NICI, an adaptive-optics coronagraphic imager with dual cameras, the campaign observed over 200 young nearby stars from December 2008 to April 2012. The paper analyzing the early 2009 subset presents the companion-detection pipeline for 45 stars and reports median contrasts of 12.6 magnitudes at $^\alpha$4 and 14.4 magnitudes at $^\alpha$5. Its principal methodological contribution is a 95% completeness contrast curve based on artificial companions injected into the raw data and recovered while accounting for the false-positive rate. The adopted detection criteria require both exceeding the false-positive threshold derived from a source-free reduction and attaining a detection strength $^\alpha$6. Using this metric, the standard NICI pipeline outperformed nominally lower-noise LOCI reductions in terms of actual recovery completeness (Wahhaj et al., 2013).

These cases show that astronomical campaigns are not merely observation schedules. They are organizational frameworks that align instrumentation, temporal coverage, data sharing, and pipeline methodology around a rare or contentious target. Amateur participation, classification disputes, and model competition are integral rather than incidental features of this usage.

5. Evaluation campaigns and benchmark governance

In natural language processing, an evaluation campaign is a structured benchmarking exercise with fixed tasks, shared data, common rules, and standardized metrics. The VarDial Evaluation Campaign 2023 was the shared-task component of the tenth Workshop on NLP for Similar Languages, Varieties and Dialects, co-located with EACL 2023. It ran on a tight schedule, with the call for participation in early January, training data release on 23 January, and submission deadline on 27 February. Three tasks were offered for the first time in that edition: SID4LR, DSL-TL, and DSL-S (Aepli et al., 2023).

SID4LR addressed slot and intent detection for low-resource language varieties without standard orthography, using Bernese Swiss German, South Tyrolean German, and Neapolitan. The evaluation used accuracy for intent detection and span F1 for slot detection. UBC employed several multilingual Transformer models and augmentation strategies, while Notre Dame used a zero-shot approach with character-level noise. For intent detection, Notre Dame obtained the best results on all three languages, with 0.9420 on DE-ST, 0.8860 on GSW, and 0.8900 on NAP. For slot detection, UBC exceeded the baseline on DE-ST and GSW but not on NAP, and Swiss German was the hardest variety in both subtasks (Aepli et al., 2023).

DSL-TL reformulated discriminating between similar languages by introducing human-annotated “true labels,” allowing each short text to be assigned to variety A, variety B, or a common both/neither class when no language-specific evidence is present. The dataset contains 12,900 instances in English, Portuguese, and Spanish, with annotation via Amazon Mechanical Turk restricted to annotators from relevant countries. On the three-way closed track, the best submitted F1 was 0.5318 by UnibucNLP, but the mBERT and XLM-R baselines achieved 0.5400 and 0.5360 respectively. On the binary closed track, ssl obtained F1 0.7604, yet the adaptive Naive Bayes baseline reached 0.7990. The open tracks benefited substantially from external data, with VaidyaKane reaching 0.5854 in Track 1 and 0.8561 in Track 2 (Aepli et al., 2023).

DSL-S extended the campaign into speech-based similar-language identification using Mozilla Common Voice v12 across nine languages. No teams submitted systems, so the reported results consist only of baselines. A linear SVM on x-vectors performed poorly, with Macro F1 0.0876. Transformer-based baselines were much stronger: direct XLS-R classification reached 0.5856, while an ASR-plus-Naive-Bayes pipeline reached 0.7031, outperforming direct speech classification by more than 10 F1 points (Aepli et al., 2023).

The evaluation-campaign model highlights a distinct function of campaign as institutionalized comparison. Its outcome is not persuasion, conversion, or observation, but a public ranking of methods under a common task design. At the same time, the VarDial results show that benchmark campaigns also expose negative results: no DSL-S submissions, difficult low-resource varieties, and strong baselines that outperform participant systems on some tracks.

6. Objectives, metrics, limitations, and contested interpretations

Campaign research is strongly shaped by formal objective functions and performance metrics. Electoral campaign management minimizes $^\alpha$7 subject to making $^\alpha$8 a winner, or minimizes bribery cost $^\alpha$9 in society-graph ILPs (Elkind et al., 2010, Faliszewski et al., 2020). Incentivized and cost-effective campaigning maximize outreach or minimize expected cost under degree-dependent policies $m$0, subject to budget caps and fixed-point or percolation constraints (Kotnis et al., 2016, Kotnis et al., 2016). Opinion-control models maximize the number of agents inside $m$1 at a terminal horizon $m$2 (Hegselmann et al., 2014). Marketing and advertising campaigns are commonly evaluated through MAPE, clicks, conversions, CTR, CPM, CPC, CPA, and profit (Chu et al., 2022, Thomaidou et al., 2012, Yan et al., 27 Aug 2025, Dutt et al., 2018). Astronomical campaigns rely on diagnostic observables such as equivalent width, radial velocity evolution, standstill detection, and 95% completeness contrast (Leadbeater, 2011, Simonsen, 2011, Wahhaj et al., 2013). Evaluation campaigns standardize accuracy, span F1, precision, recall, and macro F1 (Aepli et al., 2023).

Several recurring limitations are explicit in the literature. In campaigning on society graphs, asynchronous diffusion is order-dependent, and even deciding whether some order yields a favorable Plurality winner is NP-hard; the full BSG problem is NP-hard in general despite fixed-parameter tractability under parameterization by the number of candidates (Faliszewski et al., 2020). In optimal opinion control, bounded-confidence dynamics is discontinuous, numerically fragile under floating-point arithmetic, and computationally difficult even for small benchmark instances (Hegselmann et al., 2014). In marketing campaign prediction, HapNet does not estimate counterfactual effects under unassigned campaigns; it predicts factual outcomes under biased observational assignments (Chu et al., 2022). In the Singapore health-promotion study, the evaluation uses anonymized observational analytics, has no control group, and mostly measures proximal outcomes such as impressions, views, clicks, and traffic rather than direct behavior change (Yan et al., 27 Aug 2025).

Controversy and ambiguity are also intrinsic to several campaign domains. The Z CamPaign exists because catalogues disagree on which systems are genuinely UGZ, and the campaign’s purpose is partly to separate true members from imposters (Simonsen, 2011). The $m$3 Aurigae campaign operated amid competition between a long-standing high-mass picture and a revised lower-mass interpretation supported by UV/visual/IR analysis and CHARA imaging (Leadbeater, 2011). DSL-TL was introduced because the assumption that every short text has a single unambiguous variety label is too simplistic; the both/neither class is a direct response to this problem (Aepli et al., 2023). The Russian Facebook ads analysis further shows that campaign “effectiveness” may refer to engagement efficiency rather than persuasion in any narrow electoral sense, and the authors explicitly interpret the thematic structure as suggesting polarization and social division (Dutt et al., 2018).

Ethical concern is most explicit in opinion-control work. The bounded-confidence campaign problem is presented as mathematically neutral but politically charged basic research, with the authors noting that the same tools can support propaganda or disinformation (Hegselmann et al., 2014). A plausible implication is that campaign research, especially when framed as optimization over human beliefs, must be read simultaneously as a technical literature on control and as a literature on the governance of influence.

In this broader sense, campaign is a general research construct for organized action under constraints. Whether the task is changing votes, steering opinions, maximizing outreach, optimizing ad spend, validating a detection pipeline, or benchmarking language technologies, campaigns are studied as systems in which local actions, intermediate models, and global outcomes are explicitly linked and made measurable.