Aggregation Methods: Principles & Applications

• Aggregation methods are formal algorithms that fuse multiple data sources into one representative outcome while complying with algebraic properties and optimization criteria.
• They are applied in statistics, machine learning, and distributed systems to improve efficiency, scalability, and decision-making through techniques like linear and nonlinear pooling.
• Advanced designs incorporate robust, consensus, and element-wise aggregation to handle heterogeneity, adversarial data, and application-specific constraints.

Aggregation methods constitute a fundamental class of mathematical and algorithmic procedures for combining multiple sources of information, parameter estimates, models, scores, or data summaries into a single representative object. These procedures are pervasive across statistical inference, machine learning, distributed computing, expert elicitation, symbolic data analysis, principal component analysis, multi-criteria decision-making (MCDM), numerical solution of PDEs, graph and neural network models, and networked systems. The technical design of an aggregation method addresses both the structure of the objects being combined and the objectives or constraints of the application domain, such as statistical efficiency, robustness, scalability, privacy, interpretability, or decision-theoretic optimality.

1. Definitions and Foundational Principles

An aggregation method is a formal rule or algorithm for mapping a collection $\mathcal{X}=(x_1,\dots,x_n)$ of objects—these may be scalar values, probability distributions, models, matrices, graphs, trees, or other structured data—into a single object $A(\mathcal{X})$ that summarizes or fuses the collective information. The design of aggregation rules is usually informed by algebraic properties (e.g., commutativity, associativity, idempotence), optimization perspectives (e.g., minimization of loss or divergence), and domain-specific requirements.

In statistical inference, aggregation often operates on probability distributions or parameter estimates, as exemplified by linear or non-linear pooling in prior elicitation (Williams et al., 2020). In distributed and federated learning, aggregation acts on model parameters or gradients to update a global model (Mächler et al., 2023, Mächler et al., 2024, Hu et al., 2024, Deng et al., 2022, Song et al., 2024). In MCDM, aggregation is the core mechanism for collapsing an alternatives-criteria matrix into scores (Wang et al., 8 Sep 2025). For exactly mergeable summaries, aggregation corresponds to homomorphic operations on summary statistics, preserving algebraic structure (Batagelj, 2023). In numerical methods for PDEs, aggregation underlies algebraic multigrid coarsening and multilevel techniques (Gandham et al., 2014).

2. Taxonomy of Aggregation Methods

Aggregation methods can be classified by the mathematical objects involved, the algebraic structure of aggregation, and the optimization criteria. Major classes include:

• Linear Pooling: Arithmetic average (possibly weighted) of probability densities, model parameters, or scores. For priors from multiple experts: $f_{agg}(\theta) = \sum_{i} w_i f_i(\theta)$, with $w_i$ normalized weights (Williams et al., 2020).
• Nonlinear Pooling: e.g., logarithmic pooling, decision-theoretic and behavioral consensus mechanisms like SHELF (Williams et al., 2020).
• Convex Mixtures: Weighted sums where weights may depend on data size, performance, or dynamic criteria (Mächler et al., 2023, Mächler et al., 2024, Li et al., 26 Feb 2025).
• Truth Discovery/Consensus Methods: Aggregation of structured objects, such as trees or graphs, via metric minimization and inference of source reliabilities. CPTAM aggregates constituency parse trees by minimizing weighted Robinson–Foulds distances, inferring parser weights from interparser congruence (Kulkarni et al., 2022).
• Robust Aggregation: Use of robust M-estimators or Huber-type functions to mitigate the influence of contaminated or adversarial local contributions (Li et al., 26 Feb 2025).
• Combinatorial and Feedback Aggregation: Enumeration of spanning trees or reconstructions for consistent fusion of incomplete or partial comparative judgments, with explicit feedback and agreement indices (Tsyganok et al., 2017).
• Element-wise Aggregation: At the finest granularity, as in EWWA-FL, weights are assigned per parameter element using adaptive moment statistics and normalized via elementwise Softmax across clients (Hu et al., 2024).
• Optimization-based Aggregation in Algorithms: e.g., iterative aggregation in PCA using coarse-level models to accelerate power iteration or in aggregation algebraic multigrid (AMG) (Bulgakov, 2016, Gandham et al., 2014).
• Decision/Score Aggregation: Additive, multiplicative, hybrid, and rank-based aggregators in MCDM (e.g., SAW, MEW, AHP, ANP, COPRAS, MOORA, WASPAS, FUCA) (Wang et al., 8 Sep 2025).

The table below gives representative examples:

Application	Aggregation Type	Reference
Prior elicitation	Linear pool, Classical method, SHELF	(Williams et al., 2020)
Tree-structured data	Truth discovery, min weighted RF	(Kulkarni et al., 2022)
Distributed statistics	Robust Huber aggregation	(Li et al., 26 Feb 2025)
Federated learning	PID-inspired weighted mixtures, element-wise Softmax	(Mächler et al., 2023, Mächler et al., 2024, Hu et al., 2024)
MCDM	Additive/multiplicative/scoring/rank-based	(Wang et al., 8 Sep 2025)
Symbolic time series	Sorting-based group aggregation	(Chen et al., 2022)

3. Design Criteria and Weighting Schemes

Different aggregation scenarios require specialized weighting schemes:

• Uniform Weights: All sources contribute equally (Equal-Weight or EW in prior pooling, baseline FedAvg).
• Performance-derived Weights: In federated learning, weights may be proportional to data size, recent loss decrease (“derivative”), cumulative loss improvement (“integral”; PID), or dynamic topological/model-similarity metrics (Mächler et al., 2023, Mächler et al., 2024, Deng et al., 2022, Wang et al., 8 Sep 2025).
• Calibration and Informativeness: Classical Method (Cooke’s) uses seed questions to derive calibration and informativeness for expert priors (Williams et al., 2020).
• Reliability/Truth Discovery: CPTAM infers parser weights from parser–consensus tree distances (Kulkarni et al., 2022).
• Robustness and Outlier Control: Huber-type aggregation discounts outlier local estimators, with tuning for robustness/efficiency tradeoff (Li et al., 26 Feb 2025).
• Competence/Information Content: Combinatorial feedback uses multi-level competence and scale information for multi-expert aggregation (Tsyganok et al., 2017).

In MCDM, weight vectors for criteria are typically obtained from subjective rankings (as in AHP/ANP) or set by decision-makers (Wang et al., 8 Sep 2025).

4. Evaluation Metrics and Theoretical Guarantees

The appropriateness of an aggregation method is application-dependent, and is often evaluated through:

• Proper Scoring Rules: For prior aggregation, log score, Brier score, and quadratic score assess calibration and informativeness (Williams et al., 2020).
• Generalization and Accuracy: In federated or distributed learning, metrics include test accuracy, convergence, and robustness to heterogeneity or adversarial clients (Mächler et al., 2023, Mächler et al., 2024, Li et al., 26 Feb 2025, Hu et al., 2024).
• Convergence Rates: Theoretical results may establish minimax risk rates, oracle inequalities, or convergence in statistical distance measures (Li et al., 26 Feb 2025, Liu, 2014, Bulgakov, 2016, Carrillo et al., 2018).
• Consensus and Agreement: Double-entropy inter-rater indices for aggregation of subjective pairwise comparisons (Tsyganok et al., 2017).
• Decision Consistency: In MCDM, sensitivity to normalization, compensation effects, and rank-reversal phenomena are scrutinized (Wang et al., 8 Sep 2025).

5. Computational and Algorithmic Aspects

Efficient computation is critical in large-scale, distributed, or complex-structured aggregation problems:

• Alternating Minimization/EM: For truth-discovery structures (CPTAM), block coordinate descent estimates both aggregate and reliabilities (Kulkarni et al., 2022).
• Stochastic Sampling: When the combinatorial space (as in model aggregation or spanning tree enumeration) is large, approximation using MCMC or heuristic reduction is employed (Liu, 2014, Tsyganok et al., 2017).
• Data Fusion via Parallelism: Exactly mergeable summaries enable single-pass, streaming aggregation with efficient parallel or tree-reduce patterns (Batagelj, 2023).
• Adaptive Partitioning: Aggregation in multi-level algorithms (e.g., algebraic multigrid) relies on graph-based partitioning or aggregation of nodes (Gandham et al., 2014).
• Element-wise Computation: EWWA-FL performs per-parameter updates and Softmax normalization, incurring $O(dC)$ overhead per round for $d$ parameters and $C$ clients (Hu et al., 2024).
• Sorting-based Symbolic Aggregation: For time series, one-pass, norm-based grouping enables $O(n \log n)$ efficiency and adaptive symbol selection (Chen et al., 2022).
• Privacy and Security: In federated contexts, parameter-level aggregation with differential privacy noise addition is used to resist membership inference attacks (Song et al., 2024).

6. Theoretical Properties and Robustness

Aggregation methods’ theoretical attributes depend on the structure and assumptions:

• Optimality: Under appropriate loss functions or scoring rules, certain aggregation schemes attain minimax or oracle-optimal risk (e.g., exponential weighting mixtures, robust $M$-estimators) (Liu, 2014, Li et al., 26 Feb 2025).
• Robustness: Huber aggregators protect against a small fraction of contaminated estimates, maintaining statistical efficiency (Li et al., 26 Feb 2025).
• Consistency: Properly designed aggregation (e.g., under truth-discovery paradigms or Cooke’s method) reliably identifies credible sources or experts in the absence of ground truth (Kulkarni et al., 2022, Williams et al., 2020).
• Exact Mergeability: Summaries satisfying associativity and commutativity (e.g., sums, counts, top-$k$) allow exact, lossless merge operations, supporting distributed and streaming computation (Batagelj, 2023).
• Convergence Rates: Second-order numerical aggregation methods for PDEs retain formal convergence guarantees and handle blow-up regimes with minimal loss of accuracy (Carrillo et al., 2018).
• Decision-theoretic Guarantees: In MCDM, aggregation schemes can be tailored to enforce additivity, monotonicity, or non-compensatory behavior as required by the problem (Wang et al., 8 Sep 2025).

7. Application Domains and Empirical Results

Aggregation methods pervade numerous domains:

• Expert Bayesian Prior Elicitation: Linear and behavioral pooling, with SHELF methods outperforming classical and equal-weight approaches on proper scoring rules in clinical-trial prior elicitation (Williams et al., 2020).
• Federated Learning: PID-inspired, topology-graph, element-wise, or performance-weighted aggregation establish best-practice for unstable, heterogeneous data settings, surpassing FedAvg in segmentation, classification, and privacy robustness (Mächler et al., 2023, Mächler et al., 2024, Deng et al., 2022, Hu et al., 2024, Song et al., 2024).
• Distributed Inference: Robust Huber aggregation with spatial-median variance estimation enables reliable inference even in the presence of arbitrarily corrupted nodes or heavy-tailed distributions (Li et al., 26 Feb 2025).
• Parse Tree Aggregation: CPTAM yields improved F1 and structure metrics over all baselines across natural language and bioinformatics corpora (Kulkarni et al., 2022).
• Symbolic Data Reduction: Sorting-based fABBA produces faster and more accurate time-series compression than k-means-based approaches, outperforming SAX and 1d-SAX (Chen et al., 2022).
• Multicriteria Decision-making: Side-by-side comparisons of additive (SAW), multiplicative (MEW), hybrid (WASPAS), rank-based (FUCA), and network-based (ANP) aggregation yield different tradeoffs in compensation, sensitivity, and interpretability (Wang et al., 8 Sep 2025).
• PCA and Numerical Linear Algebra: Two-level aggregation accelerates power iteration and subspace computation in massive document-term matrices (Bulgakov, 2016).
• Algebraic Multigrid Solvers: GPU-accelerated aggregation-based AMG achieves superior setup/solve times on large sparse linear systems (Gandham et al., 2014).
• Social Choice in Databases: Quota, distance, and merge-based aggregation in multi-source database integration can be designed to preserve, or fail to preserve, various classes of integrity constraints and query-answer commutation (Belardinelli et al., 2019).
• IoT and Networked Systems: Adaptive, learning-automata-driven aggregation in distributed sensor networks optimizes traffic and resource consumption (Homaei et al., 2019).

Empirical evaluations consistently demonstrate that appropriately designed aggregation methods outperform naïve (equal-weight, data-size-only) baselines, and can robustly tolerate data and model heterogeneity, missing data, adversarial nodes, and feedback-driven adaptation.

---

• (Williams et al., 2020): A Comparison of Prior Elicitation Aggregation using the Classical Method and SHELF
• (Kulkarni et al., 2022): CPTAM: Constituency Parse Tree Aggregation Method
• (Mächler et al., 2023): FedPIDAvg: A PID controller inspired aggregation method for Federated Learning
• (Mächler et al., 2024): FedPID: An Aggregation Method for Federated Learning
• (Hu et al., 2024): An Element-Wise Weights Aggregation Method for Federated Learning
• (Deng et al., 2022): FedGraph: an Aggregation Method from Graph Perspective
• (Li et al., 26 Feb 2025): On Robust Aggregation for Distributed Data
• (Batagelj, 2023): Exactly mergeable summaries
• (Tsyganok et al., 2017): A Method with Feedback for Aggregation of Group Incomplete Pair-Wise Comparisons
• (Wang et al., 8 Sep 2025): Chapter 8 Multi-Criteria Decision-Making: Aggregation-Type Methods
• (Bulgakov, 2016): Iterative Aggregation Method for Solving Principal Component Analysis Problems
• (Gandham et al., 2014): A GPU Accelerated Aggregation Algebraic Multigrid Method
• (Carrillo et al., 2018): A second-order numerical method for the aggregation equations
• (Belardinelli et al., 2019): Social Choice Methods for Database Aggregation
• (Huang et al., 2020): Anchor Attention for Hybrid Crowd Forecasts Aggregation
• (Liu, 2014): An Aggregation Method for Sparse Logistic Regression
• (Homaei et al., 2019): Enhanced distributed data aggregation method in the internet of things
• (Chen et al., 2022): An efficient aggregation method for the symbolic representation of temporal data

These works define the state of the art and theoretical foundations for aggregation methods in modern computational and scientific domains.