DRES: Dynamic Representation and Ensemble Selection

• DRES is an instance-adaptive text classification framework that dynamically selects the best representation and ensemble for each query.
• It measures instance hardness using kDN across multiple textual embeddings (e.g., BERT, Word2Vec) to guide local competence estimation.
• Empirical results on LIAR, COVID, and GM datasets show that DRES significantly outperforms static baselines and fixed-representation models.

Dynamic Representation and Ensemble Selection (DRES) denotes an instance-adaptive classification framework in which the system first chooses the representation space in which a query appears easiest to classify and then chooses, within that representation, the classifier subset judged most competent for the same query. In the arXiv literature cited here, the acronym is used explicitly for a text-only fake news detection method that combines test-time instance hardness estimation across multiple textual embeddings with dynamic ensemble selection over representation-specific classifier pools (Farhangian et al., 21 Sep 2025). Methodologically, DRES sits at the intersection of dynamic selection, local competence estimation, meta-learning, and region-of-competence design, but differs from earlier work by making representation choice itself a query-dependent decision rather than a fixed preprocessing step.

1. Definition and scope

DRES, as explicitly formulated in the recent fake news detection literature, is a two-stage adaptive procedure for supervised text classification. Given a news statement, tweet, or article $\mathbf{x}$, the method first embeds it in multiple textual feature spaces, estimates the classification difficulty of that same instance in each space, and selects the representation with minimum estimated hardness. It then restricts attention to the classifier pool trained on that chosen representation and applies a dynamic ensemble selection (DES) method to choose the most competent local experts for final prediction (Farhangian et al., 21 Sep 2025).

This formulation is narrower than the broader dynamic selection literature. Earlier dynamic selection research typically assumed a fixed feature representation and concentrated on choosing one classifier or one ensemble per query from a predefined pool. The standard distinction, codified in DESlib, is between dynamic classifier selection (DCS), which selects one classifier, and dynamic ensemble selection (DES), which selects a subset of classifiers; both are usually organized around the region of competence, the source of competence information, and the selection rule (Cruz et al., 2018). DRES preserves the DES stage but adds a preceding representation-selection stage.

The immediate application in which the acronym appears explicitly is text-only fake news detection. That setting is especially compatible with DRES because different textual encoders induce markedly different neighborhood structures, and the same article can therefore be locally easy in one representation and hard in another (Farhangian et al., 21 Sep 2025).

2. Conceptual lineage in dynamic selection

The conceptual substrate of DRES comes from the dynamic selection literature, especially work showing that local competence estimation is more informative than fixed global combination when classifier expertise is heterogeneous. A central result is that dynamic selection often outperforms $K$-NN even when both use the same neighborhood, because $K$-NN uses the neighborhood to infer the local class majority whereas dynamic selection uses it to infer which classifier is locally competent. This difference is most consequential for high-hardness samples near decision borders, and the same literature uses $k$Disagreeing Neighbors ($kDN$) as a local hardness measure (Cruz et al., 2018).

A second strand is meta-learning for competence estimation. META-DES recasts dynamic ensemble selection as a binary meta-classification problem in which each query-classifier pair is represented by multiple competence descriptors, including local feature-space behavior, posterior confidence, decision-space behavior, and classifier confidence relative to its decision boundary. The selector then predicts whether each classifier is competent enough to participate in the ensemble (Cruz et al., 2018). This is directly relevant to DRES because it shows that competence estimation can itself be treated as a learned representation problem.

A third strand concerns the quality of the local evidence used for selection. FIRE-DES++ improves DES by filtering the dynamic selection dataset with Edited Nearest Neighbors and replacing standard $K$-NN by K-Nearest Neighbors Equality, thereby making Dynamic Frienemy Pruning operate on a cleaner and more class-balanced region of competence (Cruz et al., 2018). KNORA-B and KNORA-BI alter how the neighborhood is reduced so that the reduced region preserves either all classes or at least the minority class (Oliveira et al., 2018). The local overlap reduction procedure of Souza and coauthors removes the locally hardest samples rather than the farthest samples when searching for local oracles in imbalanced settings (Souza et al., 2022). FH-DES replaces instance-based neighborhoods entirely by fuzzy hyperboxes, thereby treating the local competence region as a compressed geometric representation rather than a raw nearest-neighbor set (Davtalab et al., 2022). DRES differs from all of these in that the representation space itself becomes query-adaptive before the ensemble is chosen.

A fourth related direction is meta-selection over dynamic selectors. PS-DES generates candidate ensembles from several DES methods and then post-selects the ensemble with highest estimated potential for the current query (Cordeiro et al., 2023). That mechanism is close in spirit to DRES, but the object being switched dynamically is the DES output, not the input representation.

3. Formal architecture

The DRES framework defines a set of representation functions

$\phi = \{\phi_1, \phi_2, \ldots, \phi_n\}.$

Applying representation $\phi_j$ to the training set $\mathcal{T}$ yields

$$\mathbf{R}_j = \phi_j(\mathcal{T}), \qquad \mathbf{R}_j \in \mathbb{R}^{|\mathcal{T}| \times d_j},$$

where $K$0 is the dimensionality of representation $K$1 (Farhangian et al., 21 Sep 2025).

For each representation $K$2, DRES performs two training-time operations. First, it computes an instance-hardness matrix

$K$3

whose entry $K$4 is the $K$5 hardness of training instance $K$6 under representation $K$7: $K$8 Second, it trains a representation-specific classifier pool

$K$9

Across all views, this gives

$K$0

for a total of $K$1 classifiers (Farhangian et al., 21 Sep 2025).

At inference time, a query $K$2 is embedded in all views,

$K$3

Because the true label of $K$4 is unknown, DRES does not compute $K$5 directly. Instead, it estimates test-time hardness in each representation by averaging the precomputed hardness values of the query’s $K$6-nearest training neighbors in that same representation: $K$7 The selected representation is then

$K$8

If several representations tie, DRES chooses the one with the lowest average hardness over the training set (Farhangian et al., 21 Sep 2025).

After representation selection, DRES restricts the search to the classifier pool $K$9 trained on the chosen representation and applies a DES method to obtain a subset

$k$0

The final prediction is produced by majority voting over $k$1. A faithful reconstruction is

$k$2

This yields a strict two-stage architecture: dynamic representation selection followed by dynamic ensemble selection (Farhangian et al., 21 Sep 2025).

4. Representation spaces and hardness modeling

The representation stage is the defining addition of DRES. The appendix of the DRES paper lists the following embedding families and dimensions: Word2Vec, GloVe, and FastText at 300 dimensions; ELMo at 1024; BERT, DistilBERT, ALBERT, RoBERTa, BART, ELECTRA, and XLNet at 768; Falcon and LLaMA3 at 2048; and LLaMA and Mistral at 4096 (Farhangian et al., 21 Sep 2025). The paper also notes an internal inconsistency: the main experimental section reports $k$3 representations, while the appendix table enumerates 15. TF and TF-IDF appear in the static-baseline appendix, but the core DRES representation table emphasizes pretrained embedding models rather than listing TF or TF-IDF as part of the main DRES view set (Farhangian et al., 21 Sep 2025).

The hardness mechanism is based entirely on local label disagreement. In DRES, hardness is not a classifier-specific margin or posterior-entropy score; it is a representation-specific estimate of local class overlap. This choice aligns DRES with the instance-hardness literature in dynamic selection, where $k$4 is used as a proxy for boundary proximity and local ambiguity (Cruz et al., 2018). DRES extends that idea by computing hardness across multiple representation spaces and then using those hardness profiles as a representation-selection criterion (Farhangian et al., 21 Sep 2025).

The reported empirical justification is that hardness varies strongly across views. On the GM dataset, over 50% of instances have a max-minus-min hardness gap above 0.5 across representations, and 25% exceed 0.7; the same analysis also reports that LIAR exhibits consistently higher hardness than the easier binary datasets (Farhangian et al., 21 Sep 2025). This means that a fixed representation can be locally mismatched even when another available representation induces a much cleaner neighborhood for the same instance.

A plausible implication is that DRES treats the representation layer as a query-conditioned neighborhood generator. The chosen representation is not merely an embedding with better global quality; it is the one whose local geometry appears least ambiguous for the current instance. That interpretation is consistent with the paper’s minimum-hardness rule, but the explicit selection criterion remains Eq. (3) rather than a more general learned routing objective (Farhangian et al., 21 Sep 2025).

5. Ensemble selection after representation choice

Within the selected representation, DRES re-enters the established DES setting. The classifier pool contains ten learner families: Support Vector Machines, Logistic Regression, $k$5-Nearest Neighbors, Naive Bayes, Multi-layer Perceptron, Random Forest, AdaBoost, XGBoost, Convolutional Neural Networks, and Bidirectional LSTMs. A separate pool is trained for each representation, so the full system contains all representation-classifier combinations (Farhangian et al., 21 Sep 2025).

The DES stage is deliberately modular. The DRES experiments use KNORA-E, DES-P, and META-DES, adapted from DESlib to support multiple text representations (Farhangian et al., 21 Sep 2025). This choice places DRES in continuity with the wider DES ecosystem exposed by DESlib, where dynamic selection methods are implemented as scikit-learn-style modules and organized into DCS, DES, and static baselines (Cruz et al., 2018). It also means that the representation-selection stage is orthogonal to the specific competence model used in the ensemble-selection stage.

The local region used by the DES methods is neighborhood-based. The DRES paper reports $k$6 both for hardness estimation and for defining the region of competence used by dynamic selection (Farhangian et al., 21 Sep 2025). A faithful reconstruction of the region-of-competence notation is

$k$7

where the neighborhood is taken in the selected representation space. Competence estimation itself is then delegated to the chosen DES algorithm. For KNORA-E, this means selecting classifiers that correctly classify all samples in the current local region; for DES-P, selecting classifiers whose local performance exceeds random classification; and for META-DES, using a meta-classifier to predict whether each classifier is competent (Farhangian et al., 21 Sep 2025).

This stage is where DRES differs from static representation ensembling. Group C in the paper aggregates all classifiers and all representations through stacked generalization, using logistic regression as the meta-classifier. DRES does not stack all views. It first commits to one representation for the current instance and only then performs dynamic local expert selection inside that representation (Farhangian et al., 21 Sep 2025). The distinction is substantial: representation selection is instance-wise and precedes ensemble construction, whereas static stacking aggregates all views globally.

6. Empirical findings and reported limitations

The DRES evaluation uses macro $k$8 as the primary metric and compares against static ensemble baselines and prior text-only fake news detectors. The datasets are LIAR, COVID, and GM, although the GM statistics are internally inconsistent: the table reports 11,000 instances, the prose reports 14.1K, and the class counts shown sum to 6310 (Farhangian et al., 21 Sep 2025). LIAR is a six-class benchmark with 12,836 political statements; COVID has 10,700 fact-checked tweets with two classes (Farhangian et al., 21 Sep 2025).

The best static-versus-dynamic results reported for macro $k$9 are as follows (Farhangian et al., 21 Sep 2025):

Dataset	Best static baseline	Best DRES result
LIAR	0.260	0.385
COVID	0.950	0.973
GM	0.951	0.986

On LIAR, the strongest DRES variant is DRES(DES-P) at 0.385, while on COVID and GM the strongest reported results are DRES(KNORA-E) at 0.973 and 0.986 respectively (Farhangian et al., 21 Sep 2025). The paper reports repeated-measures ANOVA with Tukey post-hoc testing and $kDN$0 for the main comparisons.

The ablation study is especially informative because it isolates the two constituent stages. Relative to Group C, the full static stack over all representations and classifiers, dynamic ensemble selection alone improves LIAR/COVID/GM from 0.243/0.941/0.950 to 0.299/0.948/0.963. Representation selection alone improves the same scores to 0.345/0.955/0.971. Full DRES with KNORA-E reaches 0.371/0.961/0.982 (Farhangian et al., 21 Sep 2025). The reported pattern is therefore additive rather than redundant: dynamic representation selection contributes more than DES alone, but the DES stage still provides additional gains after the representation has been chosen.

The oracle analysis shows a substantial remaining gap. Oracle representation only yields 0.832 on LIAR, 0.996 on COVID, and 0.999 on GM, while oracle representation plus oracle classifier reaches 0.995, 0.999, and 1.000 respectively (Farhangian et al., 21 Sep 2025). This indicates that DRES reduces, but does not remove, the mismatch between the selected representation-ensemble pair and the per-instance optimum.

The limitations reported are also explicit. Evaluation is restricted to English datasets. The method currently selects only one representation per instance, even though several views may be similarly favorable. Hardness estimation relies only on $kDN$1. Training incurs substantial upfront cost because hardness must be computed across multiple representations and all representation-specific classifier pools must be trained. Some representations are selected only infrequently, suggesting redundancy and possible pruning. The paper also notes possible training-data contamination concerns for newer LLM encoders such as LLaMA3 and Mistral, although it argues that this does not alter the central finding that hardness varies across representations and that dynamic selection improves performance (Farhangian et al., 21 Sep 2025).

7. Relation to adjacent formulations

DRES is distinct from several neighboring formulations that also use dynamic selection language. AutoDESS performs dataset-level AutoML search over preprocessing, classifier subsets, and ensemble strategies, including DES methods, but its strategy choice is global for the dataset rather than instance-wise for each query (Zhao et al., 2022). PS-DES performs post-selection among candidate ensembles produced by several DES techniques for the same query, but it does not introduce an explicit representation-selection stage (Cordeiro et al., 2023). Stream-oriented frameworks such as scikit-dyn2sel and DESDD adapt dynamic selection to evolving data streams and concept drift, yet their emphasis is on online validation sets, evolving pools, and drift response rather than on choosing among alternative representations for each instance (Cavalheiro et al., 2020, Albuquerque et al., 2019).

A further distinction concerns the selector signal. In adversarial robustness, dynamic ensemble selection based on Dirichlet uncertainty ranks sub-models by predictive uncertainty and selects the least uncertain expert for each input (Qin et al., 2023). That approach is close to DRES in its per-instance gating logic, but its dynamic variable is model uncertainty rather than cross-representation hardness. Conversely, classical DES methods such as META-DES provide rich competence modeling within a fixed representation, but do not decide which representation should define the region of competence in the first place (Cruz et al., 2018).

The most precise characterization is therefore narrow rather than generic. DRES is not simply any dynamic ensemble method with multiple views. It is an instance-adaptive architecture in which representation choice is itself a local decision variable, estimated through representation-specific hardness, and ensemble choice is subsequently performed inside the chosen representation. In the current literature, that combination is what distinguishes DRES from earlier DES systems and from global strategy-selection frameworks (Farhangian et al., 21 Sep 2025).