This paper presents the participation of the GPLSI team in the Quantum CLEF (QClef) Lab at CLEF 2025, focusing on Task 2 - Instance Selection and Task 3 - Clustering. The QClef Lab explores the applicability of quantum and quantum-inspired techniques to core AI tasks, emphasizing optimization eficiency and data reduction. In Task 2, we propose three multi-paradigm approaches for selecting representative training instances for sentiment classification, leveraging sentiment-aware pairing, local set-based criteria, and classical heuristics. In Task 3, we introduce a single quantum-inspired clustering framework that integrates four distinct pivot selection strategies for document grouping in embedding space. Our methods achieved competitive performance across both tasks. In particular, our LocalSets method achieved the highest efectiveness in Task 2 while substantially reducing the training set, and our FPS-Medoids approach delivered the best results for Task 3 in terms of nDCG@10. Overall, our ifndings support the potential of annealing-based techniques to deliver efective trade-ofs between performance and computational eficiency in realistic machine learning pipelines.
The Quantum CLEF (QClef) Lab at CLEF 2025 [
QClef 2025 is organized into three primary tasks: Task 1 - Feature Selection, Task 2 - Instance Selection, and Task 3 - Clustering. Each task presents a diferent machine learning challenge, ofering opportunities to apply quantum approaches in data preprocessing, optimization, and representation learning.
Our team, GPLSI, participated in Task 2 and Task 3. The main hypothesis guiding our work is that quantum-inspired optimization methods can lead to more representative and compact data subsets, thereby improving downstream learning and analysis tasks such as classification and clustering. In particular, we explore the utility of Quadratic Unconstrained Binary Optimization (QUBO) formulations and hybrid clustering strategies to address instance selection and structure discovery in sentiment analysis datasets.
Our specific objectives are: • To develop and compare three instance selection approaches for Task 2. • To design a quantum-inspired clustering method for Task 3.
• To evaluate all methods using the oficial datasets and protocols provided by QClef 2025.
The remainder of this report is structured as follows. Sections 2 and 3 detail our methodology for instance selection and clustering, respectively. Section 4 presents the experimental setup and discusses results and observations. Finally, Section 5 concludes with a summary and future directions.
This section outlines the methodology employed for Task 2, which focuses on instance selection (IS) aimed at improving the eficiency and performance of downstream sentiment analysis models. We explore three distinct approaches, each designed to identify representative and informative subsets of data. Section 2.1 introduces a pairing strategy based on sentiment polarity and semantic similarity to capture nuanced contrasts. Section 2.2 details a technique that leverages local neighborhood bounded by the local set-concept for instance selection. Finally, Section 2.3 presents a hybrid approach combining heuristic pre-filtering with statistical optimization. Together, these methodologies ofer complementary perspectives on selecting high-quality training instances.
This approach focuses on selecting a balanced and representative subset of sentiment-labeled instances by leveraging clustering and quantum-inspired optimization. The goal is to construct a reduced dataset that maintains semantic diversity and class balance. The method consists of three main stages: (I) balancing through clustering, (II) pair construction and selection, and (III) subset optimization via QUBO.
To ensure a balanced set of instances from both sentiment classes, we first independently apply
KMedoids clustering [
Let + and − be the sets of positive and negative instances, respectively. We apply K-Medoids to each set to obtain:
ℳ+ = KMedoids(+, ), ℳ− = KMedoids(− , ) where is the number of desired medoids per class. The value of controls the number of selected instances from each class and ensures that |ℳ+| = |ℳ− |, enforcing class balance.
After balancing the dataset through medoid extraction (Section 2.1.1), we construct cross-class pairs by combining positive and negative medoids in a greedy manner. Each positive medoid is paired with its closest unpaired negative medoid based on cosine similarity in the feature space. This process ensures that each negative medoid is used at most once, and the resulting sequence of pairs reflects a decreasing order of similarity —from the most semantically aligned pairs to less similar ones.
Formally, let: • ℳ+ = {1+, . . . , + } be the set of positive medoids, • ℳ− = {−1 , . . . , − } be the set of negative medoids, • sim(+, − ) denote the cosine similarity between any two medoids.
We initialize an empty set = ∅. For each + ∈ ℳ+, we select: − = arg
max − ∈ℳ− ∖ sim(+, − ) where is the set of already paired negative medoids. We then add the pair (+, − ) to and update ← ∪ {− }.
This greedy construction results in semantically aligned positive-negative pairs, each with an associated similarity score.
Pair Selection Strategies To reduce the number of candidate pairs while maintaining semantic and similarity diversity, we apply an agglomerative clustering approach that integrates two complementary views of pair dissimilarity: (1) the diference in intra-pair similarity scores and (2) the semantic distance between mean document vectors of the pairs.
Let each pair = (+, − ) ∈ consist of a positive and negative medoid, along with their cosine similarity score . We define a combined distance measure between any two pairs and as follows: 1. Similarity Score Distance. The first component quantifies the absolute diference in intra-pair similarity:
sim(, ) = | − | 2. Semantic Mean Vector Distance Each pair is represented by the mean of its two constituent document vectors:
1 = 2 (+ + − )
The semantic distance between two pairs is then computed using cosine distance: 3. Combined Distance Metric The final distance between pairs is given by the average of the two components: vec(, ) = 1 − cos( , )
Using this combined pairwise distance matrix, we apply Agglomerative Clustering [
min ∈cluster ∈cluster ∑︁ (, )
This yields a final set ′ ⊆ of representative and diverse pairs that form the candidate pool for the subsequent QUBO-based optimization (Section 2.1.3). By integrating both intra-pair and interpair semantics, this selection process ensures that the reduced dataset maintains coverage over the underlying structure of the sentiment space.
Once a reduced set of candidate cross-class pairs ′ is obtained (Section 2.1.2), we aim to select a
representative subset that is both informative and non-redundant. While auxiliary procedures such
as K-Medoids clustering for balancing, greedy algorithms for pair construction, and agglomerative
clustering for pair selection operate within polynomial time complexity, the overall problem of instance
selection often involves solving a combinatorial core that is inherently hard. In our case, this core is
formulated as a Quadratic Unconstrained Binary Optimization (QUBO) problem [
Relevance Score The relevance (or importance) of each pair is quantified based on how far its similarity score deviates from the average of the minimum and maximum similarity scores among all selected pairs. This discourages the selection of semantically central pairs (i.e., neither too similar nor too dissimilar):
Let denote the cosine similarity score of pair ∈ ′, and let: = min∈′ + max∈′
2 relevance = | − | Then, the relevance score for is defined as: This assigns higher importance to pairs that diverge from the central similarity mass, promoting diversity in semantic strength.
Redundancy Score To penalize semantic overlap among selected pairs, we define a redundancy score that measures similarity between every pair of pairs. For two candidate pairs and , redundancy is computed as the sum of cosine similarities across three views: 1. Between their positive document vectors (+). 2. Between their negative document vectors (− ). 3. Between their mean vectors ( ).
The total redundancy between and is:
redundancy, = cos(+, +) + cos(− , − ) + cos( , )
This encourages diversity by penalizing the inclusion of multiple pairs that represent similar semantic regions in the embedding space.
QUBO Objective Finally, we define a QUBO objective function that combines relevance and redundancy:
min x∈{0,1} =1 ∑︁ − relevance · + ∑︁ ∑︁ redundancy, · + penalty
=1 =1 where ∈ {0, 1} indicates whether pair is selected, is the total number of candidate pairs, the ifrst term rewards the inclusion of relevant pairs, and the second term penalizes redundant selections. To restrict the number of selected pairs to ′, we add a soft constraint: (∑︀ − ′)2. This term is incorporated into the QUBO as an additional quadratic penalty.
All submissions—labeled as SentimentPairs—are based on variations of key parameters used in the instance selection and QUBO optimization pipeline. The two main configurations difer in how output pairs are compiled—either from the final QUBO solution only ( just-final) or from QUBO-selected pairs enriched with related instances (pair-related). In the latter case, related instances refer to pairs that belong to the same cluster as the QUBO-selected pairs, based on the clustering step used to reduce the candidate set from to (Section 2.1.2). Each submission uses a fixed number of annealing reads (2000), enforces a balance limit of 90% of the minority class size, sets the number of candidate cross-class pairs = 150, and restricts the QUBO-selected final subset to ′ = 0.5 × = 75 pairs. Table 1 summarizes the explored configurations.
The main idea of this method is to exploit the concept of local set [
Our approach relies on a categorization of instances based on their contribution to classification accuracy, where each category can be quantified using a measure based on local sets.
• Noise Instances Noise instances disagree in classification with their neighborhoods, thereby obscuring the relationship between the features of an instance and its label [11]. Datasets like Yelp Reviews [12] and Vader NYT [13], where labels originate from subjective human ratings, are prone to such noise. Within the Local Set-Based Smoother (LSSm) framework [9], noise instances are identified with a measure of harmfulness, tied to how many instances identify them as their nearest enemy. • Redundant/Useless Instances By definition, these instances do not significantly influence classification. Redundancy arises when excessive same-class neighbors exist. LSSm characterizes redundancy with the usefulness measure, based on the number of local sets an instance belongs to. • Border instances Instances near class borders can be identified within the LSBo framework as instances having the lowest Local Set Cardinality (LSC) among the members of their local sets. • Central/typical instances Typical instances, also called prototypes, are instances central enough to represent a local region. They are determined within the scope of the Local Set-based Centroids selector (LSCo) [9], which selects a set of centroids of clusters obtained with LS-clustering.
In the instance selection literature, methods classified as edition methods are typically directed at removing noise, while condensation methods aim at removing redundancy. Methods combining both, such as our method, are called hybrid [14]. Generally, the selection criterion also fluctuates between focusing on border instances or central instances, as both contribute positively to classification accuracy. Finding an optimal criterion is subject to the dataset characteristics [10]. Our method is a combination of both main criteria, as we retain both a set of border instances and a set of central instances.
Our method aims to enhance the LSBo, which already achieves a good tradeof between accuracy and reduction, supported with empirical evidence [9], by selecting an additional set of instances through a high complexity step executed with quantum annealing.
To maximize accuracy with this additional set, rather than prioritizing centrality, our strategy will be to select the less-redundant set among the non-border instances. Our heuristic to detect redundancy is based on the distribution of same-class instances in local regions delimited by borders. More concretely, a pair of same-class instances (or clusters, as will be explained hereafter) will be awarded or penalized for selection by comparing their inner distance to the sum of the distances to their relative nearest border.
The complete algorithm pseudocode can be found at Algorithm 1.
In the following sections, the algorithm steps 3, 4 and 5 will be explained in detail. A more complete presentation of the algorithms LSSm and LSBo of steps 1 and 2 can be found in the paper [9]. Section 2.2.3 details the method used for clustering, Section 2.2.4 presents our heuristic for preselecting clusters and Section 2.2.5 explains how we obtain the weights of the QUBO matrix determining the final selection.
Our strategy to reduce the problem size is to associate each decision variable to a cluster containing same-class local regions of instances. To that end, non-border, non-noisy instances are clustered using a modified version of the local set-based clustering [ 10]. The original algorithm selects local sets with a high LSC, resulting in clusters containing wide same-class regions. We modify the clustering algorithm by processing small local sets first, obtaining numerous clusters containing smaller patches of instances of the same region. This will allow us to select the less redundant patch of each region.
The algorithm pseudocode can be found at Algorithm 2.
Keeping only the clusters closest to borders Removing redundant clusters with quantum annealing (QA)
Adding selected cluster members Algorithm 2 Modified Local Set-based Clustering Require: Set of instances Require: Set of noise instances and border instances
Set of clusters ← ∅ ← ∖ ( ∪ ) Compute the s (Local Sets) for each instance in Sort instances in ascending order of (Local Set Cardinality) added ← ∅ for each ∈ do if ∈/ added then added ← added ∪ {} Create cluster with medoid and members () ∖ added added ← added ∪ ()
Add to end if end for return
Filter noise and border instances Smaller local sets are processed first Instances already added to a cluster Due to the hardware constraints, the number of problem decision variables is restricted to 150, meaning that the quantum annealer can handle at most 150 candidate clusters at once. To meet this constraint, we adopt a heuristic that preselects the 150 clusters whose centroids are closest to their nearest same-class border instance. The process is as follows: 1. For each cluster , compute its centroid as the mean of its member instances. = 1 ∑︁ || ∈ * = ∈m, in = ‖ − ‖2 2. Let be the set of border instances obtained using the LSBo method, and ∈ {0, 1} the binary class of each cluster. For each cluster , find the distance * to the nearest same-class border instance ∈ .
3. Select the 150 clusters with the smallest distances * to form the preselection set.
The time complexities of the diferent algorithms described so far —LSBo, modified LS-clustering and algorithm for computing nearest borders— are ( 2), ( ) and () respectively, where T represents the set of initial instances, C represents the total number of clusters, B the number of border instances and d the feature dimensionality. Therefore, running these algorithms before the step that will be executed in the quantum annealer does not compromise eficiency.
The problem of selecting the less-redundant set of clusters based on their location around border instances is formulated using a QUBO model. Specifically, a pair of same-class clusters is rewarded if the distance between their centroids is greater than the sum of their respective distances to the nearest border instance, as such cases are likely to correspond to clusters belonging to distinct local regions separated by borders. The QUBO formulation is described below:
Let each preselected cluster be associated with: • Centroid , • Nearest border distance * , • Class label ∈ {0, 1}.
We define the inter-cluster distance and the separation margin as:
inter_dist = ‖ − ‖, = inter_dist − (* + * + ) where ≥ 0 is a tolerance parameter that increases the required separation between same-class clusters to avoid being considered redundant. Larger values of reflect stricter redundancy criteria. For both datasets (Vader NYT and Yelp Reviews), we use a value of 0.2 for the parameter .
We define the QUBO matrix ∈ R× as: ⎧⎪self_bias ⎪⎪⎪ 1 ⎪⎪⎪ · inter_dist + [, ] = ⎨ 1 ⎪⎪− · inter_dist + ⎪ ⎪ ⎪ ⎪ ⎪⎩0 if = if < 0 and = if ≥ 0 and = if ̸= with the following parameters (we use the same parameter values for both datasets): • = 1.0: penalty scaling factor for redundant same-class cluster pairs, • = 1.0: reward scaling factor for well-separated same-class cluster pairs, • self_bias = -1.0: negative diagonal term that promotes sparsity in the selection. • = 10− 6: small constant added to avoid division by zero.
To enforce class balance in the selected set of clusters, we add a penalty to the QUBO formulation: ← + · ww⊤ where = 1 − 2 where regulates the strength of the penalty term. We use a value of = 5.0 for both datasets. The sum of the terms 1 − 2 for all clusters will be minimized when there is the same number of clusters of each class in the final solution. As the optimal solution is the solution that minimizes the expression , adding this term with a high enough allows us to promote class balance. After obtaining the final cluster selection, all the selected cluster member instances are added to the submission set containing the border instances.
We propose PREST, Pre-Selection and Statistical Tuning, for QUBO Annealing a three-step pipeline that begins with classical heuristics and ends with the exact same optimization task running on both CPUs and quantum hardware. First, we build candidate panels greedily and wrap them into a single QUBO [15]. Because the objective and hyper-parameters stay identical, the classical-versus-quantum comparison is strictly like-for-like. The workflow comprises three stages: 1. Stage I – Representation & Pre-filtering : Sentence-BERT embeddings are enriched with scalar features. Nine classical heuristics generate candidate subsets for a grid of target sizes . A metric bundle {silhouette, Davies–Bouldin, balance, diversity, coverage} [16] ranks the outputs. 2. Stage II – Combinatorial Optimization: each of the four best heuristics selected in Stage I is cast as a QUBO whose coeficients encode diversity, class-frequency fairness, emotional contrast , and sentiment confidence . A SA sweep over and QUBO weights keeps the two top-scoring panels per heuristic. 3. Stage III – Quantum-Ready Aggregation: To comply with the ≤ 150-variable limit of current quantum annealers, the retained subsets are clustered into super-nodes. We solve two QUBO variants—unconstrained and cardinality-constrained—with both SA and QA in a D-Wave QPU, yielding the final instance sets.
Augmented sentence embeddings Each sentence is embedded with the all-mpnet-base-v2 Sentence-BERT model [17], producing a 768-dimensional vector. To capture additional information we concatenate two task-specific scalars: emotional contrast () defined as the gap between the two highest emotion logits, and sentiment confidence (), given by the maximum soft-max probability produced by the sentiment classifier.
Subset generators We generate nine deterministic subsets for each target size ∈ {200, 400, . . . , 2000}, leveraging a range of complementary heuristics grounded in core-set theory, graph analysis, and uncertainty modeling. Classical core-set methods—Max–Min, K-Means Centroids, and Farthest-First—are commonly used for feature selection in Natural Language Processing (NLP) [18]. To exploit structural cues, we include Community selection via modularity maximization [19] and Closeness Centrality [20], two graph-based sampling heuristics. The Density-Weighted heuristic prioritizes low-density regions in the embedding space, enhancing representational diversity by sampling from semantically sparse areas [21]. Emotional Conflict , a custom heuristic introduced in this work, identifies semantically similar sentence pairs with opposing afective signals—highlighting emotionally ambiguous cases relevant to afective representation [ 22]. Diversity Sampling maximizes coverage by iteratively picking points farthest from the current subset, thereby enhancing generalization and reducing feature redundancy [23]. Finally, Uncertainty-Based selection focuses on items near sentiment boundaries, following Bayesian active-learning and margin-sampling principles [24]. Metric-Driven Filtering Each subset is evaluated with five criteria [ 16] —silhouette (), Davies–Bouldin (), class balance (ℬ), diversity (), and coverage (). For every dataset we gather the raw scores of all nine heuristics and all target sizes , compute the global mean and standard deviation of each metric , and obtain -scores = ( − )/ . The normalized scores are ifnally averaged into Ψ = 15 (︀ − + ℬ + + )︀ , and the four highest-ranked heuristics progress to Stage II.
After the Stage I filtering we retain Max–Min, K-Means Centroids, Emotional Conflict and Uncertainty-Based . For each pool = {1, . . . , } and target size ∈ {1000, 1500, 2000, 2500} (a reduced grid chosen for runtime reasons) we flip a binary switch x ∈ {0, 1} ( = 1 selects ). The QUBO combines four signals: • Diversity , amplified when the two sentences come from diferent sentiment classes and inverted when they coincide. • Class-frequency fairness: diagonal bias /, where is the number of instances in class and is the total number of instances, that favours minority classes. • Emotional contrast emo on the diagonal; • Sentiment confidence sent︀( 1 − | − 0.5|)︀ on the diagonal (prefer items whose softmax probability is close to 0.5).
The QUBO matrix is ⎧ + emo + sent︀( 1 − | − 0.5|)︀ , = , ⎪⎪⎪⎨
Hyper-parameter sweep We explore emo ∈ {0.5, 1.0, 2.0, 2.5} and sent ∈ {0.25, 0.5, 1.0}, crossed with the four values, i.e. 4 × 3 × 4 = 48 QUBOs per heuristic.
Optimization Each pair (, ) defines a QUBO instance, solved with Neal’s SA sampler (100 reads, default settings). Panels are re-scored with the same metric bundle and the two best configurations per heuristic are retained for Stage III.
Only the subsets generated by the K-Means Centroids and Emotional Conflict heuristics outperformed all others in Stage II across all datasets, so we focus on those two for the quantum run. Both methods generate datasets that exceed the variable budget of current QPUs, so each is compressed into = 150 super-nodes via -means and reformulate the problem as a second QUBO optimized with both classical Simulated Annealing (SA) and Quantum Annealing (QA). A super-node = ⟨z¯, ^, ¯, ¯⟩ stores centroid, majority class, mean emotional-contrast and mean sentiment-confidence.
To keep the search flexible, we run the QUBO in two versions: one gently caps the solution at roughly half of the available super-nodes (cardinality-constrained), while the other drops that limit altogether and lets the annealer decide for itself how many super-nodes are worth keeping.
QUBO without cardinality constraint Let = |{}| and x ∈ {0, 1} . For super-nodes , define the class indicator () = 1 if ^ = ^ and 0 otherwise, and the Euclidean distance = ‖z¯ − z¯ ‖2. Following our implementation build_qubo_qclef_no_cardinality(), the unconstrained QUBO reads ⎧ ^ + emo ¯ + sen t︀( 1 − | ¯ − 0.5|)︀ , = , ⎪⎪⎪⎨ Here ^ is the frequency of class ^ in the panel; emo and sent ≥ 0 weight emotional contrast and sentiment confidence, respectively.
QUBO with cardinality constraint [25] To steer the solution towards = ⌈/2⌉ selected supernodes we build a Binary Quadratic Model (BQM) from unc and add a combination penalty generated with dimod.generators.combinations. Its strength is set automatically to the maximum energy spread of the unconstrained BQM: card = unc + (︀ − ∑︁ ℓ)︀ 2, ℓ
= bqm.maximum_energy_delta().
Optimization and evaluation Both QUBOs are solved independently with SA (same schedule as Stage II) and QA (2 000 reads, default settings). The binary solutions are projected back to their original sentences and rescored with the five-metric bundle.
This section outlines the methodology employed for Task 3, which focuses on centroid selection aimed at accelerating document retrieval over the ANTIQUE sentence-embedding corpus. We investigate four complementary clustering pipelines, each designed to generate a small set of representative centroids that preserve cluster quality and maximize downstream retrieval efectiveness.
Our hybrid, QA-driven k-medoids pipeline unfolds in three steps: 1. Pivot selection: we start from the 6 513 sentence embeddings provided with the Antique corpus; this full set of ≈ 6500 vectors is filtered down to = 150 well-spread candidate pivots. 2. Annealing optimization : cast a QUBO on those pivots to identify the final medoids. 3. Global assignment: assign every document to its nearest medoids, producing the final clusters. function Following the standard QUBO framework [15] and Bauckhage’s quadratic reduction [26], we embed the pivots as rows of X ∈ R × (ℓ2-normalized). Let = 1 − ∈ {0, 1}, with = 1 if pivot is selected. The k-medoids objective becomes the pseudo-Boolean cos(x, x ) and introduce binary ︁( ∑︁ − ︁) 2
+ < 1 ∑︁ , which expands to the QUBO ∑︁ + 2, = ≤ ⎧ ⎨ − ⎩ − 2 , 2 + ∑︀ =1 − 2, = , ̸= , where we fix = 1/, = 1/ , and = 2, balancing dispersion, linear penalties, and enforcing ∑︀ = . The parameter acts as a Lagrange multiplier, which is crucial in balancing the minimization of the objective function against the constraint that the number of selected medoids is equal to . This same parameter set is kept across all four methods (A–D) for both the exact- reduction and the final exact- optimization. These values are chosen according to the guidelines provided by Bauckhage et al. [26] to ensure a balanced contribution of the diferent terms in the objective function. The resulting BinaryQuadraticModel is submitted to a D-Wave annealer (2 000 reads); the returned vector y* lfags the selected medoids.
This Bauckhage QUBO ofers several practical advantages: by specifying only its upper-triangular entries, the matrix is symmetric ( = ), simplifying minor embedding on quantum hardware; the diagonal shift − 2 cancels the linear-term bias from the squared cardinality penalty, avoiding skew toward too many or too few pivots; the constant ofset 0 maintains equivalence with the original objective yet can be dropped after solving without afecting cluster assignments; and, since = 0 and most of-diagonal are small, is naturally sparse, reducing the solver’s memory footprint.
With the QUBO core fixed, we isolate the efect of classical preprocessing by testing four distinct pivot-selection strategies that each deliver = 150 candidates. quantum stage.
Method A: qIIMAS strategy (SubMedoids).
Following the technique introduced by team qIIMAS in last year’s competition [27], each document is assigned a binary variable ∈ {0, 1} indicating whether it remains a pivot. An “exact- ” QUBO with cost matrix ′ = / forces the cardinality constraint ∑︀ = while minimizing pairwise similarity within the chosen subset. Solving this model on the same annealing backend used downstream yields a maximally diverse pivot set that captures even the rarest semantic regions of the corpus. Any surplus indices are deterministically trimmed, whereas deficits are filled with previously unused, randomly sampled documents to ensure || = before the Method B: Farthest-Point Sampling (FPS). FPS adopts a simple yet powerful greedy rule: starting from one random seed, it maintains each document’s distance to its nearest selected pivot (measured in cosine distance) and repeatedly adds the document with the largest recorded distance until exactly QUBO receives a well-spread candidate pool for the final k-medoids step. pivots have been chosen [28]. Each iteration is a single pass over the corpus, yielding a cost of ( ) time and ( ) memory. The resulting pivots provide broad geometric coverage, so the subsequent Method C: MiniBatch-KMeans.
Running MiniBatch-KMeans [29] with clusters, we pick as pivot the document closest to each centroid in cosine distance. These representatives reflect data density, furnishing the annealer with high-quality starting points at a per-epoch cost of ( ). Method D: CLARA–CLARANS hybrid. We combine two classic medoid heuristics in a single, twostep routine. CLARA first draws several disjoint subsamples (each roughly 20 % of the corpus), applies PAM to every subsample, and retains the medoids that minimise the within-subset distance. Building on this provisional set, CLARANS performs up to twenty random medoid–non-medoid swaps, accepting a swap only when it reduces the overall clustering cost [30]. The combination lets CLARA explore diverse regions of the embedding space, while CLARANS fine-tunes the most promising configuration, yielding a pivot set that achieves both broad coverage and strong internal cohesion.
For every pivot selector—qIIMAS strategy (A), FPS (B), MiniBatch-KMeans (C), and CLARA–CLARANS (D)—and each target size ∈ {10, 25, 50}, we run the six-step pipeline below: 1. Initial set-up. Start with the full corpus of ℓ2-normalized sentence embeddings and fix . 2. Coarse compression. Apply the chosen method to obtain ≪ coarse clusters; their centroids define three candidate pivot budgets ∈ {100, 125, 150}. Validation retains =150 for A and =100 for B–D (seed = 42). 3. Fast QUBO sweep (100 SA reads). For every ⟨, method, ⟩ build the exact- QUBO, solve it once with 100 SA reads, and log validation Davies–Bouldin and nDCG@10. 4. Pivot-budget selection. Choose the that minimizes Davies–Bouldin (nDCG@10 breaks ties); this reproduces the preferences in step 2. 5. Full Annealing run (2 000 reads). Rebuild the exact- QUBO with the selected and solve it with 2 000 reads on the designated annealer (SA or QA) to obtain the medoids. 6. Global assignment and metrics. Attach each of the embeddings to its nearest medoid, then compute final Davies–Bouldin and mean nDCG@10.
This section presents the experimental setup and results for evaluating our proposed approaches across two main tasks: Task 2 (Instance Selection) and Task 3 (Clustering). We detail the evaluation scenarios, including datasets, submission configurations, and whether Simulated Annealing (SA) or Quantum Annealing (QA) was used. The results are reported as computed by the challenge organizers. Finally, we provide a discussion analyzing the trade-ofs between performance and eficiency achieved by our methods compared to baselines and competing submissions.
The evaluation was divided into two main scenarios corresponding to the oficial tasks defined in the challenge. The selected tasks were: (i) Task 2: Instance Selection and (ii) Task 3: Clustering. Each task involved the use of SA and QA approaches.
Task 2 focuses on selecting the most representative subsets of training data for fine-tuning a sentiment classification model (Llama3.1), aiming to reduce computational cost while preserving efectiveness. The evaluation considered both the macro 1 score and the reduction rate. We submitted several configurations using both SA and QA methods on the Vader NYT and Yelp datasets. Table 2 summarizes our submitted runs for Task 2, indicating which approaches were applied to each dataset using SA and QA.
Task 3 involves clustering sentence embeddings to support eficient document retrieval and exploration. Quality was evaluated both intrinsically with the Davies–Bouldin Index and extrinsically with nDCG@10 on test queries. We submitted clustering results on the large and small variants of the ANTIQUE dataset using both SA and QA. Table 3 summarizes our submitted runs for Task 3, indicating which approaches were applied to each dataset using SA and QA.
Tables 4 and 5 present a summary of the results for Task 2, using the Vader and the Yelp datasets respectively. They show the performance of our approaches—SentimentPairs (SP) and LocalSet for both datasets, and Pre-Selection and Statistical Tuning submitted in the variants emoconflictCard and SentimentKmeansCard only for the Yelp dataset—compared to the baseline and the top-performing submission in terms of average macro 1 score. Results are averaged over five cross-validation folds, where available. The Evaluation Time column combines fine-tuning and prediction time (in seconds), while the Annealing Time refers to the total execution time for QA approaches, including programming, sampling, and post-processing. Submissions marked with an asterisk (* ) did not provide results for all ifve folds due to issues in the challenge’s execution infrastructure, which were beyond the control of both participants and organizers. The SentimentPairs (SP) variant was executed with reads=2000 and limit=True.
Table 6 summarises the Task 3 outcomes on the ANTIQUE corpus obtained using only SA. For each target panel size ∈ {10, 25, 50} we report the nDCG@10 and the Davies-Bouldin computed on the entire dataset. The Annealing time column aggregates programming, sampling, and post-processing of the SA search, executed with the default neal schedule (2000 reads, defaults settings).
DBI
Anneal time ( s)
Type
This section discusses the main findings of our participation in the QuantumCLEF Lab, organized by task. We analyze the results obtained in the instance selection task (Section 4.3.1) and the clustering task (Section 4.3.2), highlighting the efectiveness of our proposed methods, their trade-ofs, and their relative performance compared to other participating systems and the baseline.
As shown in Table 4, our approaches—SentimentPairs (SP) and LocalSet—demonstrate efective trade-ofs between data reduction and performance. Compared to the baseline, which uses the full dataset and achieves the highest macro 1, our best-performing approach enables a 50% reduction in data, translating into nearly half the fine-tuning and prediction time, at the cost of a moderate decrease in macro 1. When compared to the top competing system in terms of macro 1, our approach achieves only 2.6% lower macro 1 on average while using half as much data, efectively doubling the reduction rate (25% → 50%).
Among our methods, the LocalSet approach yields the highest macro 1 (63.3) with an average data reduction of 50%, striking a strong balance between eficiency and accuracy. Our SentimentPairs variants that include pair-related documents trade only 1.1% of macro 1 for up to 70% data reduction. In contrast, the strictest SentimentPairs configurations, which retain only the most relevant documents, reach reduction rates of 83–95%, though at a more pronounced cost in performance (macro 1 drops from 62% to 47%). These results highlight the flexibility of our methods in tailoring the reduction-performance trade-of for diferent application needs.
The results in Table 5 also reveal a good balance between efectivity and reduction for our methods. Our best-performing method on this dataset achieves a 50% reduction of the dataset without afecting the macro 1 score with respect to the baseline. As compared to the top-performing submission in terms of macro 1, our approach supposes only a 0.1% decrease of this metric, and doubles the dataset reduction.
Our methods present varied and versatile solutions to the instance selection bi-objective problem using the Yelp dataset. Our LocalSet approach achieves the highest macro 1 score (99.4) while halving the Yelp dataset size. Our SentimentPairs variant including pair-related documents achieves a higher reduction (62.7%), resulting in a decrease of the 59.4% of the fine-tuning and prediction time by compromising only a 0.2% of the 1 score with respect to the baseline. The maximum evaluation time reduction (91.6%) is achieved by our SentimentPairs variant including just-final documents, at the expense of decreasing the 1 metric only an 8.6%. Finally, our two variants of the Pre-Selection and Statistical Tuning method are at a middle ground between our methods achieving high performance (LocalSet and SP pair-related) and high reduction (SP just-final ). These variants have been executed using both SA and QA. The emoconflictCard achieves slightly better results, with an average macro 1 score only 0.7% behind the baseline and an average reduction of 71.5% of the dataset, resulting in an average reduction of 67.8% of the fine-tuning and prediction time.
Overall, our results in Task 2 show that it is possible to substantially reduce the size of sentimentlabeled datasets while maintaining competitive classification performance. Among our methods, the LocalSet approach stands out for achieving the highest efectiveness across both datasets, consistently delivering top macro 1 scores while halving the training data. This makes it particularly attractive for scenarios requiring balanced trade-ofs between performance and computational eficiency. The diversity of our approaches—ranging from aggressive reduction strategies to more conservative selections—also highlights the adaptability of quantum-inspired instance selection techniques to various application needs and constraints.
As shown in Table 6, the SA runs submitted by our team outperform the oficial baseline for every tested cluster size and in every evaluation metric. Compared to other submissions, ours match or outperform them in all scenarios except when = 10, where the diference is negligible. For = 10, our 3-FPS-Medoids configuration achieves an nDCG@10 of 0.578, a relative gain of 5 % over the baseline (0.551) and virtually the same efectiveness as the best performing submission regarding nDCG@10 (0.580). Additionally, all our variants reduce the Davies–Bouldin from 7.99 to at most 6.43; this improvement comes at a computational cost of approximately 15 s (15 M s) of annealing time versus the 0.008 second runtime of other participant submissions. By increasing the cluster size to = 25, our 3-FPS-Medoids approach remains in the lead with nDCG@10 = 0.547 (+4 % over baseline), and our remaining methods collectively reduce the DBI from 6.12 to 5.56, although their runtime increases to 20–40 s (20–40 M s). The diference increases for = 50: our 3-FPS-Medoids method reaches 0.559 nDCG@10, corresponding to a gain of 20 % over the baseline (0.466), and achieves the lowest DBI (4.45). In contrast, the submission with the best DBI has a runtime of less than 0.3 s, but its nDCG@10 drops dramatically to 0.006.
Overall, each one of our variant ofers a superior efectiveness—cohesion profile relative to the baseline, and the SA framework scales robustly with list length, unlike the faster but markedly less stable that the other submissions approaches. Among our methods, 3-FPS-Medoids provides the best trade-of between retrieval quality and cluster compactness; MBK-Medoids consistently yields the tightest clusters; CLARA–CLARANS occupies a middle ground, delivering efectiveness and cohesion above the baseline with the lowest runtime of our SA runs, making it attractive when computational budget is limited; and SubMedoids-QUBO maintains solid efectiveness at the expense of the highest computational cost.
It is important to note that, while the comparison with other submissions provides useful insights, it may not be entirely conclusive due to a structural diference in how centroids were represented. Three of our four submissions—SubMedoids-QUBO, CLARA–CLARANS, and MBK-Medoids—generate centroids with lower dimensionality than the original embedding. This dimensionality reduction, although beneficial for improving cluster compactness and interpretability, may have introduced a disadvantage under the evaluation protocol, particularly afecting metrics like nDCG@10 that are sensitive to representation format. As the evaluation setup did not anticipate lower-dimensional centroids, the relative efectiveness of these methods should be interpreted with care.
This work presented our approaches and findings for the instance selection and clustering tasks defined in the Quantum CLEF Lab. Our methods achieved competitive results in both scenarios, obtaining top-ranked performance in multiple evaluations. In Task 2, our LocalSets-based strategy achieved the highest efectiveness while maintaining a substantial reduction in training data. In Task 3, our FPS-Medoids approach yielded the best clustering results in terms of nDCG@10. Overall, our results demonstrate efective trade-ofs between performance and reduction, showing that it is possible to significantly reduce the amount of training data or clustering complexity without incurring major performance losses.
Key contributions of this work include the following.
• We present three multi-paradigm approaches to the instance selection problem, each leveraging complementary principles: (i) sentiment-aware document pairing, (ii) local similarity-based criteria, and (iii) classical heuristic methods for data reduction. • We propose a unified clustering framework specifically designed for quantum annealing, which integrates four distinct pivot selection strategies to optimize centroid-based grouping in highdimensional embedding spaces.
Future Work. Future work will focus on addressing current limitations and exploring extensions of our proposed methods. Due to infrastructure constraints during the challenge, quantum annealing could not be fully leveraged across all configurations; future experiments will aim to complete the evaluation of QA-based approaches. In Task 2, our SentimentPairs and LocalSets methods relied solely on precomputed embeddings, omitting potentially informative textual features such as readability or syntactic structure. Preliminary results from other configurations suggest that integrating custom embeddings or textual signals may enhance performance. Additionally, solving Task 2 involved tackling several clustering-related subproblems. Some of the techniques developed, such as LocalSets, could be adapted for Task 3. However, this would require automatically labeling the unlabeled documents in Task 3 to enable the application of supervised or semi-supervised strategies.
Funding: The research work conducted is part of the R&D projects “CORTEX: Conscious Text Generation” (PID2021-123956OB-I00), funded by MCIN/ AEI/10.13039/501100011033/ and by “ERDF A way of making Europe”; “CLEAR.TEXT:Enhancing the modernization public sector organizations by deploying Natural Language Processing to make their digital content CLEARER to those with cognitive disabilities” (TED2021-130707B-I00), funded by MCIN/AEI/10.13039/501100011033 and “European Union NextGenerationEU/PRTR”; “VIVES: “Pla de Tecnologies de la Llengua per al valencià” project (2022/TL22/00215334) as subproject of “ILENIA” project (grant number 2022/TL22/00215337) from the Ministerio de Transformación Digital y por el Plan de Recuperación, Transformación y Resiliencia – Financiado por la Unión Europea – NextGenerationEU; and the project “NL4DISMIS: Natural Language Technologies for dealing with dis- and misinformation with grant reference (CIPROM/2021/021)" funded by the Generalitat Valenciana and “GEO.IA PLATAFORMA DE GEOINTELIGENCIA ARTIFICIAL PARA RESOLVER PROBLEMAS A LOS CIUDADANOS Y FACILITAR LA TOMA DE DECISIONES ESTRATÉGICAS EN LAS ADMINISTRACIONES PÚBLICAS” (INNEST/2023/344) funded by Agencia Valenciana de la Innovación under the “Programa Comunitat Valenciana Fondo Europeo de Desarrollo Regional (FEDER) 2021-2027”. Additionally, this work is funded by the Ministerio para la Transformación Digital y de la Función Pública and Plan de Recuperación, Transformación y Resiliencia - Funded by EU – NextGenerationEU within the framework of the project Desarrollo Modelos ALIA.
During the preparation of this work, the author(s) used ChatGPT, Grammarly in order to: Grammar and spelling check. After using these tools/services, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.