Boosting the Learning for Ranking Patterns

Nassim Belmecheri; Noureddine Aribi; Nadjib Lazaar; Yahia Lebbah; Samir Loudni

cs.AIOct 12, 2026

Boosting the Learning for Ranking Patterns

Authors: Nassim Belmecheri, Noureddine Aribi, Nadjib Lazaar, Yahia Lebbah, Samir Loudni

Organizations: Lab. LITIO, Universit´e Oran1, 31000, Oran, Algeria. · LIRMM, Universit´e de Montpellier, CNRS, 34000, Montpellier, France. · TASC (LS2N-CNRS), IMT Atlantique, 44307, Nantes, France.

Abstract

Pattern mining is a valuable tool for exploratory data analysis, but identifying relevant patterns for a specific user is challenging. Various interestingness mea- sures have been developed to evaluate patterns, but they may not efficiently estimate user-specific functions. Learning user-specific functions by ranking patterns has been proposed, but this requires significant time and training samples. In this paper, we present a solution that formulates the problem of learning pattern ranking functions as a multicriteria decision making problem. Our approach uses Analytic Hierarchy Process (AHP) to elicit weights for different interestingness measures based on user preference. We also propose an active learning mode with a sensitivity-based heuristic to minimize user ranking queries while still providing high-quality results. Experiments show that our approach significantly reduces running time and returns precise pattern ranking while being robust to user error, compared to state-of-the-art approaches.

Figures & tables

$user$	$S$	$M_{1}$ – $\mathit{rank}$		$M_{2}$ – $\mathit{rank}$		$M_{3}$ – $\mathit{rank}$		$M_{4}$ – $\mathit{rank}$		$M_{5}$ – $\mathit{rank}$		$g_{w}$ – $\mathit{rank}$
1	$P_{7}$	0.95	1	0.48	8	0.79	2	0.30	9	0.80	1	0.72	1
2	$P_{3}$	0.75	3	0.72	1	0.78	3	0.70	2	0.61	2	0.68	2
3	$P_{6}$	0.80	2	0.49	7	0.50	9	0.65	4	0.60	3	0.61	3
4	$P_{1}$	0.47	9	0.47	9	0.76	5	0.56	6	0.59	4	0.54	4
5	$P_{8}$	0.56	6	0.65	4	0.63	8	0.69	3	0.40	5	0.53	5
6	$P_{10}$	0.57	5	0.50	6	0.80	1	0.4	8	0.02	10	0.34	8

Table 1: Running example with

10

patterns and

5

measures.

Groups		Measures
1		Yules Q , Yules Y, Odds Ratio
2		Cosine , Jaccard
3		Laplace , Support
4		$\phi$ coefficient , CollectiveStrength, piatetskyShapiro’s
5		Goodman Kruskal’s , Gini Index
6		Interest factor ,added value, Klosgen K

Table 2: Independent Subjective Measure Groups

Dataset	$\lvert\mathcal{T}\rvert$	$\lvert\mathcal{I}\rvert$	Density(%)	Type of Data	${\tt\#Rules}$
$\tt Hepatitis$	137	68	50.00	Disease	0.5 M
$\tt Connect$	67,557	129	33.33	Game steps	1 M
$\tt Mushroom$	8,124	119	18.75	Species of mushrooms	1.5 M
$\tt T40$	100,000	1,000	4.20	Synthetic dataset	2 M
$\tt Retail$	88,162	16,470	0.06	Retail market basket data	2,5 M
$\tt T40$ : $\tt T40I10D100K$

Table 3: Dataset Characteristics.

Datasets	measures
Datasets	(1)	(2)	(3)	(4)	(5)	(6)	(7)	${\tt VBM}$
Rand-Emu
$\tt Hepatitis$	0.78	0.92	0.29	0.97	0.79	0.85	0.45	0.97
$\tt Connect$	0.68	0.62	0.95	0.70	0.60	0.45	0.57	0.95
$\tt Mushroom$	0.84	0.97	0.28	0.83	0.55	0.73	0.36	0.97
$\tt T40$	0.71	0.99	0	0.76	0.99	0.99	0	0.99

Table 4: Correlation results with user ranking for using the emulators.

Rand-Emu
	${\tt RankingSVM.0}$				${\tt AHPRank.0}$
	$\rho$	${\tt R@{10\%}}$	${\tt R@{1\%}}$	t(s)	$\rho$	${\tt R@{10\%}}$	${\tt R@{1\%}}$	t(s)
$\tt Hepatitis$	0.94	0.74	0.77	851	0.93	0.79	0.71	9
$\tt Connect$	-	-	-	TO	0.95	0.72	0.63	26
$\tt Mushroom$	-	-	-	TO	0.89	0.69	0.66	32
$\tt T40$	-	-	-	TO	0.99	0.93	0.58	63

Table 5: 5-fold cross-validation results (passive learning).

Figure 1: Learning accuracy comparison between

{\tt RankingSVM.0}

and

{\tt AHPRank.0}

learning Chi-Emu on different training data size.

Figure 2: CPU time comparison between

{\tt RankingSVM.0}

and

{\tt AHPRank.0}

learning using Chi-Emu emulator on different training data sizes averaged overall datasets.

Figure 3: Qualitative comparison of Random vs. Sensitivity-Based Query Generator using the Spearman correlation

\rho

and the recall at

10\%

and

1\%

of each iteration.

Approaches	Metrics	$p$ - $value$	$z$ - $score$	${\tt CI}$
${\tt RankingSVM.1}$	$\rho$	$.00226$	$-2.8373$	$99\%$
	${\tt R@{10\%}}$	$.12302$	$-1.1573$	$70\%$
	${\tt R@{1\%}}$	$.06301$	$-1.5306$	$90\%$
${\tt AHPRank.1}$	$\rho$	$.00034$	$-3.3973$	$99.9\%$
	${\tt R@{10\%}}$	$.00205$	$-2.8746$	$99\%$
	${\tt R@{1\%}}$	$.00139$	$-2.9866$	$99\%$

Table 6: Wilcoxon Signed-Rank Test (RG vs. SBG).

10 Queries
		(a)	(b)	(c)	(d)	(e)
(1)	$\rho$	0.73	0.04	0.69	0.99	0.97
	${\tt R@{10\%}}$	0.61	0.28	0.38	0.92	0.91
	${\tt R@{1\%}}$	0.44	0.49	0.36	0.45	0.67
	${\tt Time}\,$	0	0	0	0	0
(2)	$\rho$	0.78	0.10	0.93	0.99	0.99

Table 7: Qualitative evaluation of

{\tt RankingSVM.1}

{\tt AHPRank.1}

Figure 4: Learning under mistakes (

20

queries).

Figure 5: Learning under swaps (

20

queries)

Learning to Rank

Learning to rank develops models that order items by relevance or utility for search, recommendations, and other decisions.

See all 267 papers

Active Learners as Efficient PRP Rerankers
May 14, 2026Jeremías Figueiredo Paschmann, Juan Kaplan, Francisco Nattero +3Pairwise Preference Learning Pairwise Comparison
ELIAS Lab, Departamento de Ingeniería, Universidad de San Andrés
Alternative Strategies for Active Learning with Bayesian Bradley Terry Model
Oct 21, 2026Saman Jayasinghe, Sean Walton, Alma Rahat +2Pairwise Ranking Active Preference Learning
Swansea University, United Kingdom · Loughborough University, United Kingdom · Bath Spa University, United Kingdom +1
Metric-agnostic Learning-to-Rank via Boosting and Rank Approximation
Apr 16, 2026Camilo Gomez, Pengyang Wang, Yanjie FuLearning to Rank
Dept. of Statistics & Data Science University of Central Florida Orlando, USA · Department of CIS, SKL-IOTSC University of Macau Macao, China · School of Computing and AI Arizona State University Tempe, USA

May 14, 2026J/K move · Enter open · S save

Active Learners as Efficient PRP Rerankers

Jeremías Figueiredo Paschmann, Juan Kaplan, Francisco Nattero, Santiago Barron+2

ELIAS Lab, Departamento de Ingeniería, Universidad de San Andrés

Pairwise Preference Learning Pairwise Comparison Pairwise Preference Evaluation Learning to Rank Active Learning LLM Reranking+3

Open paper AskPDF arXiv

Pairwise Ranking Prompting (PRP) elicits pairwise preference judgments from an LLM, which are then aggregated into a ranking, usually via classical sorting algorithms. However, judgments are noisy, order-sensitive, and sometimes intransitive, so sorting assumptions do not match the setting. Because sorting aims to recover a full permutation, truncating it to meet a call budget does not produce a dependable top-K. We thus reframe PRP reranking as active learning from noisy pairwise comparisons and show that active rankers are drop-in replacements that improve NDCG@10 per call in the call-constrained regime. Our noise-robust framework also introduces a randomized-direction oracle that uses a single LLM call per pair. This approach converts systematic position bias into zero-mean noise, enabling unbiased aggregate ranking without the cost of bidirectional calls.

Abstract

Learning to Rank

Explore similar work

Active Learners as Efficient PRP Rerankers

Alternative Strategies for Active Learning with Bayesian Bradley Terry Model

Metric-agnostic Learning-to-Rank via Boosting and Rank Approximation