Fixed Aggregation Features Overview

• Fixed Aggregation Features are non-trainable mechanisms that convert variable-sized or structured data into fixed-dimensional representations through preset, permutation-invariant reducers.
• They are applied in diverse areas such as graph node classification, speaker verification, and business analytics to simplify complex data into actionable features.
• Empirical studies show that FAF-based methods often rival trainable models, offering competitive performance with enhanced interpretability and efficiency.

Fixed Aggregation Features (FAFs) are a class of aggregation mechanisms and representations in machine learning and data analysis where the aggregation operations—reducers such as mean, max, sum, or variants thereof—are fixed and non-trainable, in contrast to parameterized or attention-based aggregation. FAFs appear in diverse application domains, including business analytics for default measure-grouping selection, deep architectures for utterance-level speaker verification in audio, and as a tabular baseline approach in modern node and graph representation learning. Their central characteristic is the transformation of variable-sized or structured data into fixed-dimensional representations using prescribed (non-learned) aggregation mappings, enabling subsequent standard learning or analytical tasks.

1. Core Definitions and Mechanisms

FAFs are instantiated by applying a predetermined set of permutation-invariant aggregation operators to some set or sequence of feature-vectors. Depending on context, this is performed over:

• columns in a tabular business analytics dataset for measure aggregation (Chinaei et al., 2015),
• variable-length sequences in audio (e.g., frame-level features aggregated to fixed-size utterance-level representations) (Xie et al., 2019),
• neighborhoods in a graph (k-hop aggregations of node features) (Rubio-Madrigal et al., 27 Jan 2026),
• non-local feature aggregations using canonical data structures (e.g., loop, sequence, tensor-based structures) (Rahmani et al., 2021).

The shared property is that, for each entity (node, utterance, column), the final representation is an explicit, fixed-dimensional feature vector, constructed solely by applying fixed (often multi-hop or multi-scale) aggregation operations, optionally followed by non-adaptive postprocessing (e.g., normalization, dimensionality reduction).

2. FAFs in Graph and Network Representation Learning

Fixed Aggregation Features provide a training-free transformation of graph-structured data into tabular form, suitable for input into any standard classifier or regressor (Rubio-Madrigal et al., 27 Jan 2026). The canonical construction is as follows:

Let $G = (V, E)$ be an undirected graph with node features $x_v \in \mathbb{R}^d$. The k-hop closed neighborhood of node $v$ is $N_0(v) = \{v\}$, $N_k(v) = \cup_{u \in N_{k-1}(v)} N(u)$. For a fixed set of reducers $R$ (e.g., mean, sum, max, variance), define recursively: $h^{(0)}_r(v) = x_v$

$$h^{(k)}_r(v) = r( \{ h^{(k-1)}_r(u) : u \in N(v) \} )$$

for $k=1,\,\dots,\,K$. Construct the FAF vector for node $v$ as the concatenation: $$\mathrm{FAF}(v) = [\, h^{(0)}(v)\, \|\, h^{(1)}_r(v)\, \|\,\cdots\, \|\, h^{(K)}_r(v)\,]_{r \in R}$$ This yields a feature vector of dimension $d \cdot (1 + |R| \cdot K)$ per node.

Such representations allow node-classification or regression tasks to be reframed as standard tabular learning. Empirically, multilayer perceptrons (MLPs) trained on FAFs (using $R=\{mean, sum, max, min\}$, $K$ up to typical GNN depth) rival or outperform GNNs and graph transformers on 12 out of 14 benchmark datasets, with the most prominent gains where local information is predictive (Rubio-Madrigal et al., 27 Jan 2026). Mean alone is frequently sufficient except for rare benchmarks requiring more expressive aggregation.

Theoretically, a fixed aggregator is injective on 1-hop multisets for orthogonal basis features (one-hot or orthonormal embeddings). Kolmogorov–Arnold representation theorems imply that, with discontinuous injective mappings (e.g., using Cantor set encodings), all continuous multiset functions can in principle be represented, though in practice continuous reducers (mean, sum) trade injectivity for numerical stability.

3. FAFs in Business Analytics and Case-Based Aggregation

In business analytics, FAFs are used as features for case-based reasoning systems that recommend default aggregation operators (sum, mean, last-period) on measure columns based on the data's column semantics, statistical trends, and associations (Chinaei et al., 2015). Here, three explicit FAFs are defined:

1. Semantic column annotations: Each measure/category column is tagged with semantic types (e.g., monetary, metric, temporal). The intersection of these sets is used as a feature for similarity in case-based retrieval.
2. Association type: Cardinality between category and measure columns (one-to-one, one-to-many, etc.) informs if aggregation is even meaningful.
3. Coefficient of Variation (CoV) trends: Within each category, compute the mean and standard deviation across records for the measure, yielding CoV = σ/μ. The average CoV is a strong separator: CoV ≈ 1 indicates sum, ≈0.5 suggests average, ≈0 last-period.

These FAFs are used as similarity features within a CBR framework. Combined, they yield 86% accuracy in predicting correct default aggregation over labeled test cases (Chinaei et al., 2015). Extracting only the CoV feature achieves 63% accuracy alone.

4. FAFs in Sequence and Audio Processing

In utterance-level speaker recognition, FAFs denote the fixed-dimensional utterance representations obtained by aggregating frame-level features through a non-adaptive NetVLAD or GhostVLAD pooling layer (Xie et al., 2019). The canonical model processes a log-magnitude spectrogram using a thin-ResNet-34; outputs are $T' = T/32$ frame descriptors $x_t \in \mathbb{R}^{512}$. NetVLAD acquires soft-assignment weights $a_{tk}$ using per-cluster linear assignments, then forms residual vectors relative to learnable cluster centers: $v_k = \sum_{t=1}^{T'} a_{tk} (x_t - c_k)$ Stacked over $k$, L2-normalized, and reduced via a fully connected layer, this produces a fixed 512-dimensional FAF for the utterance. The GhostVLAD extension introduces ghost clusters to absorb noisy/irrelevant frames, reducing their impact on the core representation.

Empirical ablations confirm that NetVLAD/GhostVLAD-based FAFs yield significantly better verification error rates (EER) than temporal average pooling. GhostVLAD with 8 clusters and 2 ghosts plus softmax loss achieves 3.22% EER, compared to 10.48% for average-pooling (Xie et al., 2019).

5. Advanced Variants: Latent Fixed Data Structures (LFDS)

FAFs are further extended to non-local feature aggregation on graphs via Latent Fixed Data Structures (LFDS) (Rahmani et al., 2021). Here, node features are distributed onto a fixed structure (e.g., 3D tensor, sequence, loop, data-driven graph), over which a small CNN or GNN operates to aggregate the feature information globally. The assignment is performed by softmax over learned prototype slots, inflating or distributing features over the LFDS. The aggregated representation is invariant to graph size and node ordering.

CNN or GNN applied to the LFDS yields a dense fixed-dimensional embedding, which is then used with a standard feed-forward classifier. This approach remains linear in graph size. Experimental results show that multiple LFDS variants (e.g., loop, array, tensor) outperform previously proposed pooling strategies and compete with trainable hierarchical methods, especially on harder graph-level tasks (Rahmani et al., 2021).

6. Empirical Performance and Theoretical Insights

The effectiveness of FAFs across contexts is well-established:

• In graph node classification, MLPs on FAFs with mean, sum, max, min, and $K$ up to 4 often rival or exceed GNNs; only deep, nonlocal benchmarks (e.g., Minesweeper, Roman Empire) favor GNNs (Rubio-Madrigal et al., 27 Jan 2026).
• In business analytics, FA-based CBR reaches 86% default aggregation accuracy, with CoV trending alone at 63% (Chinaei et al., 2015).
• In utterance-level speaker recognition, NetVLAD/GhostVLAD FAFs consistently outperform average pooling across all tested protocol variations (Xie et al., 2019).
• Non-local graph pooling via LFDS-based FAFs outperforms max-pooling and hierarchical approaches while remaining efficient and scalable (Rahmani et al., 2021).

A plausible implication is that, for many real-world datasets where local or low-order statistics suffice, FAFs offer a powerful and interpretable alternative to fully trainable aggregation.

7. Interpretability, Efficiency, and Directions for Further Research

FAFs offer strong interpretability: their construction is transparent, with each feature traceable to an explicit aggregation operation, facilitating domain insights and explanation (e.g., SHAP, Rashomon set analyses). FAF pipelines separate the (fixed) aggregation from the subsequent classifier, leveraging mature tabular learning methodologies for downstream tasks.

Empirical results suggest tuning the choice and number of reducers $R$ and hop-depth $K$ is beneficial but subject to diminishing returns beyond 2–4 hops in most datasets (Rubio-Madrigal et al., 27 Jan 2026). Recommendations in the literature include always benchmarking new graph learning methods against strong FAF-based tabular baselines, and developing datasets where simple FAFs cannot saturate accuracy—particularly those requiring complex, nonlocal, or permutation-sensitive aggregations.

Ongoing research explores meta-learning or attention mechanisms to dynamically select aggregators/hops, hybrid approaches combining FAFs with structural graph features, and partially-learned or data-adaptive aggregation operators that balance expressivity and learnability (Rubio-Madrigal et al., 27 Jan 2026). Extensions to the CBR framework suggest including time-series features, domain-specific metadata, or intent-based features as new FAFs for richer default aggregation behavior (Chinaei et al., 2015).

Summary Table: FAF Instantiations Across Domains

Application Domain	FAF Construction	Notable Feature/Result
Graph/node classification	Multi-hop aggregators (mean, sum, max, var) over neighborhood	MLP on FAFs rivals/outperforms GNNs (Rubio-Madrigal et al., 27 Jan 2026)
Business analytics (CBR)	Semantic annotation, association type, CoV trend as descriptive FAFs	CBR with 3 FAFs: 86% accuracy (Chinaei et al., 2015)
Audio/utterance aggregation	NetVLAD/GhostVLAD (soft cluster-based residual aggregation)	>3× EER improvement over average pooling (Xie et al., 2019)
Global non-local graph pooling	Soft-assignment to LFDS (loop, sequence, tensor); CNN/GNN on LFDS	Outperforms max-pool and hierarchical pool (Rahmani et al., 2021)