Benchmarking Attribute Selection Techniques for Discrete Class Data Mining

Mark A. HallGeoffrey Holmes

article2003TKDE1,314 citations

Evaluates six major attribute selection methods across multiple benchmark and high-dimensional datasets to determine how ranking-based feature selection impacts the classification performance and efficiency of C4.5 and naive Bayes.

Listen

Modern data mining applications frequently suffer from poor predictive accuracy and heavy computational overhead due to datasets contaminated with irrelevant, redundant, and noisy variables. Identifying a compact subset of highly predictive attributes is critical for building efficient, interpretable models, yet practitioners lack comprehensive benchmarks to guide the selection among numerous competing techniques.

The article systematically evaluates and compares six major attribute selection methods for supervised classification on discrete class data. The primary objective is to demonstrate how these ranking-based dimensionality reduction techniques affect predictive accuracy, model complexity, and computational runtime across different types of learning algorithms.

The authors conducted rigorous experimental evaluations using sixteen standard benchmark classification datasets alongside three large-scale datasets from the UCI repository containing up to 1,557 features. Six attribute selection techniques were implemented: Information Gain, ReliefF, Principal Component Analysis, Correlation-based Feature Selection (CFS), a Consistency-based evaluator, and a Wrapper method. These techniques produced ranked lists of variables that were cross-validated to identify optimal subsets for two fundamentally distinct learning schemes: a probabilistic model (naive Bayes) and a decision tree inducer (C4.5).

The benchmark revealed four key findings. First, no single attribute selection method was best across all situations, as performance strongly depended on the underlying inductive bias of the learning algorithm. Second, the Wrapper approach achieved the highest classification accuracy for naive Bayes (scoring 34 wins and 4 losses), but it was computationally prohibitive on large datasets, taking an estimated 140 days on the largest benchmark. Third, for decision trees (C4.5), ReliefF and the Consistency method delivered the top predictive performance because they effectively capture attribute interactions that decision trees exploit during early node splitting. Fourth, CFS consistently eliminated the most unnecessary data, retaining only 42% to 48% of the original attributes on average, while producing the smallest, most interpretable decision trees and executing significantly faster than wrapper or instance-based techniques. Conversely, unsupervised Principal Component Analysis consistently degraded predictive accuracy across models.

These findings indicate that reducing data dimensionality by roughly 50% generally maintains or improves classification performance while producing simpler, more explainable models. For decision-makers, choosing the right attribute selection technique directly affects operational efficiency, project timelines, and model interpretability. Implementing computationally heavy wrapper models introduces severe timeline risks on large datasets without guaranteeing better results than faster heuristic filters.

Organizations should align their feature selection method with their modeling approach and resource constraints. When computational budget allows, Wrapper methods remain the top choice for naive Bayes, whereas ReliefF or Consistency methods are recommended for tree-based models where feature interactions matter. When processing speed, scalability, and compact model size are essential, teams should deploy CFS as a fast, balanced default. Highly complex pipelines should avoid unsupervised Principal Component Analysis for supervised discrete classification.

Confidence in these findings is high for standard tabular classification tasks given the extensive cross-validation methodology. However, caution is warranted when extrapolating these results to non-discrete regression problems, extremely high-dimensional sparse datasets, or architectures beyond naive Bayes and decision trees, which were outside the scope of this evaluation.

association rules data mining techniques engineering discrete class data mining attribute selection techniques feature selection

Hall et al (2003).pdf

Paper: Correlation-based Feature Selection for Discrete and Numeric Class Machine Learning, Mark A. Hall (1999). Introduces Correlation-based Feature Selection (CFS) and ranking-based attribute evaluation metrics that serve as direct baselines for the attribute selection benchmark.
Paper: Irrelevant Features and the Subset Selection Problem, George H. John et al. (1994). Formalizes the distinction between filter and wrapper approaches for subset selection and evaluates their performance on standard decision tree and naive Bayes classifiers.
Paper: Toward Optimal Feature Selection, Daphne Koller et al. (1996). Provides the foundational information-theoretic framework for feature utility estimation and redundant attribute elimination evaluated in classification benchmarks.
Paper: Feature Selection: Evaluation, Application, and Small Sample Performance, Anil K. Jain et al. (1997). Establishes a systematic benchmark methodology for comparing variable selection techniques across synthetic and empirical datasets.
Paper: A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection, Ron Kohavi (1995). Defines the cross-validation and accuracy estimation protocols used to evaluate the utility of attribute selection with C4.5 and Naive Bayes.
Paper: Supervised and Unsupervised Discretization of Continuous Features, James Dougherty et al. (1995). Explores preprocessing and discretization strategies for continuous features in discrete-class learning environments using C4.5 and Naive Bayes.
Paper: On the Optimality of the Simple Bayesian Classifier under Zero-One Loss, Pedro Domingos et al. (1997). Analyzes the behavior and optimality conditions of the Naive Bayes classifier under varying degrees of attribute dependence.

Paper: Toward integrating feature selection algorithms for classification and clustering, Huan Liu et al. (2005). Synthesizes and extends benchmark evaluations of feature selection by proposing an integrated categorization framework across classification and clustering tasks.
Paper: Efficient Feature Selection via Analysis of Relevance and Redundancy, Lei Yu et al. (2004). Extends attribute evaluation techniques by explicitly decoupling relevance ranking from redundancy analysis to scale feature selection for ultra-high-dimensional datasets.
Paper: Laplacian Score for Feature Selection, Xiaofei He et al. (2005). Generalizes attribute ranking and filter methods from discrete-class supervised mining to graph-based unsupervised feature selection.
Paper: Auto-WEKA: combined selection and hyperparameter optimization of classification algorithms, Chris J. Thornton et al. (2012). Builds on discrete-class benchmark evaluations by automating the combined selection of feature selection methods and classification algorithm hyperparameters.

Cover for Benchmarking Attribute Selection Techniques for Discrete Class Data Mining

Abstract

Data engineering is generally considered to be a central issue in the development of data mining applications. The success of many learning schemes, in their attempts to construct models of data, hinges on the reliable identification of a small set of highly predictive attributes. The inclusion of irrelevant, redundant and noisy attributes in the model building process phase can result in poor predictive performance and increased computation.

Attribute selection generally involves a combination of search and attribute utility estimation plus evaluation with respect to specific learning schemes. This leads to a large number of possible permutations and has led to a situation where very few benchmark studies have been conducted.

This paper presents a benchmark comparison of several attribute selection methods for supervised classification. All the methods produce an attribute ranking, a useful devise for isolating the individual merit of an attribute. Attribute selection is achieved by cross-validating the attribute rankings with respect to a classification learner to find the best attributes. Results are reported for a selection of standard data sets and two diverse learning schemes C4.5 and naive Bayes.

Topics

association rules data mining techniques engineering genetic selection naive bayes supervised classification knowledge discovery in database classification models attribute data estimation method

Keywords

discrete class data mining attribute selection techniques feature selection attribute subset selection multi-class data sets benchmark data sets decision tree attribute ranking

1 Introduction
2 Attribute Selection Techniques
2.1 Information Gain Attribute Ranking
2.2 Relief
2.3 Principal Components
2.4 CFS
2.5 Consistency-based Subset Evaluation
2.6 Wrapper Subset Evaluation
3 Experimental Methodology
3.1 Weka Experiment Editor
4 Results on Sixteen Benchmark Data Sets
5 Results on Large Data Sets
6 Conclusions
References

Knowls

Knowl 1 — Modified Forward Selection for Attribute Ranking
algorithm
The modified forward selection search procedure adapts heuristic subset evaluators to generate a complete ranked list of individual features by forcing greedy forward hill-climbing search to continue until the entire feature set is included.
Input: Full feature set $F = \{f_1, f_2, \dots, f_d\}$ , subset evaluation function $M(S)$
Output: Ranked list of attributes $R$
$S \leftarrow \emptyset$
$R \leftarrow []$
$U \leftarrow F$
while $U \neq \emptyset$ do
     $best\_f \leftarrow \operatorname{argmax}_{f \in U} M(S \cup \{f\})$
     $S \leftarrow S \cup \{best\_f\}$
     $U \leftarrow U \setminus \{best\_f\}$
     $R \leftarrow \operatorname{append}(R, best\_f)$
end while
return $R$
Unlike standard forward hill-climbing search—which terminates as soon as no single feature addition strictly increases the subset evaluation score $M(S)$ —the modified algorithm forces the search to continue until all attributes are exhausted. In each iteration $i$ , the remaining feature that yields the maximum subset score when combined with the currently selected $i-1$ features is identified, added to the subset, and recorded at rank $i$ in the output list. This converts subset evaluators such as Correlation-based Feature Selection (CFS), Consistency-based evaluation (CNS), and Wrapper evaluation (WRP) into attribute rankers.
Knowl 2 — Optimal Subset Determination via Cross-Validated Feature Ranking Evaluation
experimental setup
To identify the best attribute subset from an ordered attribute ranking $R = (f_{(1)}, f_{(2)}, \dots, f_{(d)})$ generated on training data, a forward cross-validation procedure is executed strictly on the training partition.

For each candidate subset size $k \in \{1, 2, \dots, d\}$ , a feature subset $S_k = \{f_{(1)}, \dots, f_{(k)}\}$ containing the top $k$ ranked features is formed. A 10-fold cross-validation is then performed on the training split using the target induction algorithm (e.g., Naive Bayes or C4.5) trained only on $S_k$ . The subset size $k^*$ that yields the highest cross-validated classification accuracy across the training folds is chosen as optimal:

$k^* = \operatorname{argmax}_{k \in \{1, \dots, d\}} \operatorname{Accuracy}_{\text{CV}}(S_k)$

The final classifier is subsequently trained on the complete training split using only the top $k^*$ features and evaluated on the unseen test split. This wrapper evaluation isolates feature ranking from final evaluation and prevents information leakage between training and testing sets.
Knowl 3 — Correlation-Based Feature Selection Metric and Search
model/method
Correlation-Based Feature Selection (CFS) evaluates an attribute subset $S$ containing $k$ features based on the principle that an optimal feature subset contains features highly correlated with the target class yet mutually uncorrelated with one another. The heuristic subset merit $Merit_S$ is computed as:

$Merit_S = \frac{k \, \overline{r_{cf}}}{\sqrt{k + k(k - 1) \, \overline{r_{ff}}}}$

where $k$ is the number of features in $S$ , $\overline{r_{cf}}$ is the average correlation between the features in $S$ and the discrete class variable $C$ , and $\overline{r_{ff}}$ is the average pairwise correlation across all feature pairs within $S$ .

For discrete nominal variables, correlation between any two variables $X$ and $Y$ is measured using symmetrical uncertainty ( $SU$ ):

$SU(X, Y) = 2.0 \times \left[ \frac{H(X) + H(Y) - H(X, Y)}{H(X) + H(Y)} \right]$

where $H(X) = -\sum_{x \in X} p(x) \log_2 p(x)$ is Shannon entropy and $H(X, Y)$ is the joint entropy of $X$ and $Y$ . Continuous numeric attributes are discretized prior to correlation computation using the entropy-based Minimum Description Length (MDL) discretization method of Fayyad and Irani.
Knowl 4 — Consistency-Based Attribute Subset Evaluation Metric
model/method
The consistency-based subset evaluator assesses a candidate attribute subset $s$ by measuring the degree to which the attribute value combinations of $s$ partition the training data into pure, single-class groups.

For a dataset with $N$ instances and an attribute subset $s$ , instances are grouped by their distinct combinations of attribute values on $s$ . Let $J$ denote the total number of distinct attribute value combinations present in the dataset. For the $i$ -th combination, let $|D_i|$ denote the total number of instances exhibiting that attribute combination, and let $|M_i|$ denote the number of instances belonging to the majority class within that combination. The subset consistency score $Consistency_s$ is given by:

$Consistency_s = 1 - \frac{\sum_{i=1}^{J} (|D_i| - |M_i|)}{N}$

A subset achieves maximum consistency ( $Consistency_s = 1.0$ ) if and only if every unique combination of attribute values maps exclusively to instances of one class. Continuous attributes are converted to discrete intervals using Fayyad and Irani's MDL discretizer prior to consistency evaluation.
Knowl 5 — ReliefF Attribute Ranking for Multi-Class Noisy Domains
algorithm
ReliefF is an instance-based attribute ranking algorithm that estimates feature importance by measuring how effectively feature values distinguish between nearest neighbor instances from the same versus different classes.
Input: Training set $D$ , sampling iterations $m = 250$ , neighbor count $k = 10$ , feature set $A$
Output: Weight vector $W$ for all attributes in $A$
for each attribute $a \in A$ do
     $W[a] \leftarrow 0.0$
end for
for $i = 1$ to $m$ do
    Randomly select an instance $R \in D$
    Find $k$ nearest instances $H_j$ to $R$ from class $(R)$ (nearest hits)
    for each class $C \neq \operatorname{class}(R)$ do
        Find $k$ nearest instances $M_j(C)$ to $R$ from class $C$ (nearest misses)
    end for
    for each attribute $a \in A$ do
         $W[a] \leftarrow W[a] - \sum_{j=1}^k \frac{\operatorname{diff}(a, R, H_j)}{m \cdot k} + \sum_{C \neq \operatorname{class}(R)} \left[ \frac{P(C)}{1 - P(\operatorname{class}(R))} \sum_{j=1}^k \frac{\operatorname{diff}(a, R, M_j(C))}{m \cdot k} \right]$
    end for
end for
return $W$
The difference function $\operatorname{diff}(a, I_1, I_2)$ between instances $I_1$ and $I_2$ on attribute $a$ is defined as:

$\operatorname{diff}(a, I_1, I_2) = \begin{cases} 0 & \text{if } I_1[a] = I_2[a] \text{ (for discrete } a\text{)} \\ 1 & \text{if } I_1[a] \neq I_2[a] \text{ (for discrete } a\text{)} \\ \frac{|I_1[a] - I_2[a]|}{\max(a) - \min(a)} & \text{for continuous } a \end{cases}$

where $P(C)$ is the prior probability of class $C$ estimated from the training data.

Knowl 6 — Classification Accuracy Benchmark for Naive Bayes

data/table

Six attribute selection methods—Information Gain (IG), ReliefF (RLF), Consistency-based selection (CNS), Principal Components Analysis (PC), Correlation-based Feature Selection (CFS), and Wrapper selection (WRP)—were evaluated with Naive Bayes against the unselected baseline (NB) across 15 standard UCI datasets using 10 runs of 10-fold cross-validation.

Data Set	NB	IG	RLF	CNS	PC	CFS	WRP
zoo	95.04	94.34	93.37	93.85	93.86	93.94	94.34
heart-c	83.83	82.54	82.12	82.28	81.85	82.64	82.68
ionosphere	82.60	88.78	89.52	89.95	90.72	89.75	91.28
soybean	92.90	92.43	92.56	92.81	90.93	92.46	92.64
glass2	62.33	67.42	63.83	68.31	66.74	71.08	75.06
vote	90.19	95.63	95.33	95.82	92.32	95.63	95.93
heart-stat	84.37	85.11	86.00	83.48	82.07	85.07	85.00
lymph	83.24	82.63	81.47	82.55	79.67	82.35	84.11
labor	93.93	89.17	90.97	92.00	89.77	89.20	85.77
diabetes	75.73	76.24	75.95	75.64	74.42	76.19	76.12
breast-c	73.12	72.84	70.99	71.79	73.54	73.01	72.28
credit-g	74.98	74.36	74.49	74.06	73.30	74.33	74.35
segment	80.10	87.17	86.97	85.98	90.03	89.03	89.57
horse colic	78.28	83.20	82.58	82.77	78.56	83.01	82.61
anneal	86.51	87.06	89.17	89.71	90.65	87.16	92.91

Pairwise statistical significance was tested at the 1% level using a two-sided paired $t$ -test across all pairs of schemes. Ranking methods by net wins (Wins minus Losses) demonstrates that Wrapper selection and CFS achieve the highest performance when paired with Naive Bayes:

Scheme	Wins $-$ Losses	Wins	Losses
WRP	30	34	4
CFS	7	21	14
CNS	2	21	19
IG	-2	17	19
RLF	-3	19	22
NB	-7	28	35
PC	-27	17	44

Knowl 7 — Classification Accuracy Benchmark for C4.5 Decision Trees

data/table

Six attribute selection methods—Information Gain (IG), Correlation-based Feature Selection (CFS), Consistency-based selection (CNS), ReliefF (RLF), Wrapper (WRP), and Principal Components Analysis (PC)—were benchmarked with C4.5 decision trees (release 8) against the unselected C4.5 baseline across 15 standard UCI datasets using 10 runs of 10-fold cross-validation.

Data Set	C4.5	IG	CFS	CNS	RLF	WRP	PC
zoo	92.26	91.65	91.06	93.65	92.95	90.45	91.49
heart-stat	78.67	84.52	85.33	84.11	82.00	82.11	82.22
ionosphere	89.74	89.40	91.09	91.05	91.43	91.80	88.80
diabetes	73.74	73.92	73.67	73.71	73.58	73.50	71.51
vote	96.46	95.84	95.65	95.98	95.79	95.74	92.07
credit-g	71.18	72.72	72.99	72.20	71.63	72.23	69.34
soybean	92.48	92.40	91.14	92.43	92.43	92.19	83.75
heart-c	76.64	78.95	79.11	80.23	80.40	77.00	82.65
glass2	77.97	78.35	78.53	77.05	79.53	76.53	66.41
labor	80.20	80.60	81.00	79.73	79.53	78.33	88.60
lymph	75.50	73.09	73.41	75.43	76.83	76.63	74.60
breast-c	73.87	73.75	73.70	72.24	72.77	73.43	70.62
segment	96.90	96.81	96.94	96.87	96.89	96.92	93.95
anneal	98.58	98.72	98.47	98.65	98.73	98.66	96.26
horse colic	85.44	84.18	83.94	84.00	84.90	84.14	78.18

Pairwise comparison (two-sided paired $t$ -test at the 1% level) demonstrates that ReliefF and Consistency-based selection achieve the highest net accuracy rankings when paired with C4.5:

Scheme	Wins $-$ Losses	Wins	Losses
RLF	15	22	7
CNS	12	20	8
C4.5	7	23	16
CFS	5	21	16
WRP	5	17	12
IG	2	15	13
PC	-46	12	58

Knowl 8 — Inducer-Specific Biases in Attribute Selection Effectiveness
empirical result
The relative effectiveness of attribute selection techniques is strongly dependent on the learning algorithm's inductive bias:
1. Naive Bayes: Naive Bayes relies on the conditional independence assumption among attributes given the class. As a result, Wrapper selection (which tailors the subset directly to Naive Bayes) and Correlation-based Feature Selection (CFS, which explicitly penalizes redundant, inter-correlated features) achieve the best performance, yielding $+30$ and $+7$ net wins respectively over other schemes across benchmark datasets.
2. C4.5 Decision Trees: C4.5 constructs univariate hierarchical splits and can exploit multi-feature interactions if relevant interactive features are selected together. Methods capable of detecting attribute interactions—specifically ReliefF ( $+15$ net wins) and Consistency-based selection ( $+12$ net wins)—outperform Wrapper selection ( $+5$ net wins) and the unselected C4.5 baseline ( $+7$ net wins).
3. Wrapper Limitations with C4.5: The Wrapper underperforms ReliefF and Consistency on C4.5 for two primary reasons:
  - Greedy forward search: Forward selection evaluates attributes incrementally and can fail to identify strongly interacting attribute groups during early iterations.
  - Sample size reduction in internal CV: The Wrapper's internal 5-fold cross-validation on training data reduces the number of instances available to construct trees during evaluation, leading to noisy and unreliable subset evaluations for sample-sensitive decision tree learners.

Knowl 9 — Decision Tree Complexity Reduction via Attribute Selection

data/table

Applying attribute selection prior to decision tree induction with C4.5 significantly reduces final tree complexity (measured by total node count) across 15 standard UCI datasets.

Data Set	C4.5	IG	CFS	CNS	RLF	WRP	PC
zoo	15.64	13.22	13.74	13.44	13.04	13.98	13.02
heart-stat	34.84	12.12	11.98	13.52	13.66	14.92	4.82
ionosphere	26.58	21.84	16.64	17.14	17.22	13.90	20.04
diabetes	41.54	14.62	15.92	16.54	16.74	17.06	30.52
vote	10.64	9.44	8.64	9.92	9.00	9.72	20.44
credit-g	125.05	57.34	60.39	61.82	68.52	63.48	10.98
soybean	92.27	86.50	88.29	92.25	91.21	90.75	88.84
heart-c	42.34	19.72	19.45	22.48	23.17	24.20	8.16
glass2	23.78	14.88	16.28	16.26	17.12	16.22	11.18
labor	6.96	6.22	6.10	6.18	5.48	6.13	5.88
lymph	27.41	14.71	14.35	12.26	14.56	14.43	18.18
breast-c	12.38	10.47	10.26	15.09	11.80	8.42	7.72
segment	81.86	80.82	80.26	79.44	80.96	79.50	119.00
anneal	49.75	48.45	50.06	46.83	46.73	48.63	38.94
horse colic	8.57	21.18	25.75	8.81	20.64	20.90	6.42

Pairwise comparison across all pairs of schemes ranks the methods by net tree size reduction (Wins minus Losses):

Scheme	Wins $-$ Losses	Wins	Losses
PC	21	47	26
CFS	15	30	15
IG	13	29	16
RLF	7	25	18
CNS	6	26	20
WRP	0	22	22
C4.5	-62	6	68

Although Principal Components yields the smallest trees, it causes severe degradation in predictive accuracy. CFS is the most effective selector for reducing tree size while maintaining high classification accuracy, producing significantly smaller trees than baseline C4.5 on 11 datasets.

Knowl 10 — Dimensionality Reduction and Computational Speed Trade-Offs
empirical result
Systematic evaluation of attribute selectors across benchmark datasets highlights distinct trade-offs between feature reduction rate, computational runtime, and scaling behavior:
1. Degree of Feature Reduction: Correlation-based Feature Selection (CFS) consistently selects the most compact attribute subsets, retaining on average $\approx 48\%$ of attributes for Naive Bayes and $\approx 42\%$ for C4.5 (achieving $+24$ net wins in feature reduction rankings for both learners). The Wrapper retains $\approx 50\%$ for Naive Bayes and $\approx 42\%$ for C4.5. ReliefF retains slightly larger feature subsets ( $\approx 52\%$ for C4.5, ranking lowest in feature reduction among filters), which enables it to maintain superior classification accuracy on interaction-heavy datasets.
2. Performance on Large Datasets: On large datasets (arrhythmia with 227 attributes, anonymous with 293 attributes, internet-ads with 1557 attributes), CFS reduces the feature space down to $1\%\text{--}22\%$ of original attributes (e.g., selecting 6 of 227 features on arrhythmia and 11 of 293 on anonymous for Naive Bayes) while matching or improving upon baseline accuracy.
3. Computational Efficiency: CFS and Information Gain are the fastest selectors overall ( $+50$ and $+49$ net wins in execution time for Naive Bayes). Consistency-based selection is the fastest method for C4.5 ( $+34$ net wins) because identifying high-quality attribute rankings early accelerates the internal cross-validation loop by producing smaller, faster-to-build decision trees. The Wrapper is by far the slowest scheme ( $-66$ net wins for both learners) and becomes computationally intractable on high-dimensional domains (e.g., requiring an estimated 140 days of CPU time to complete forward selection on internet-ads).

Coverage note — None was omitted; all key methods, equations, algorithms, experimental designs, benchmark data tables, and empirical analyses from the paper are fully represented.

References

1.H. Almuallim and T. G. Dietterich. Learning with many irrelevant features. In Proceedings of the Ninth National Conference on Artificial Intelligence, pages 547–552. AAAI Press, 1991.
2.C. Blake, E. Keogh, and C. J. Merz. UCI Repository of Machine Learning Data Bases. University of California, Department of Information and Computer Science, Irvine, CA, 1998. [http://www.ics.uci.edu/~mlearn/MLRepository.html].
3.Avrim Blum and Pat Langley. Selection of relevant features and examples in machine learning. Artificial Intelligence, 97(1-2):245–271, 1997.
4.M. Dash and H. Liu. Feature selection for classification. Intelligent Data Analysis, 1(3), 1997.
5.S. Dumais, J. Platt, D. Heckerman, and M. Sahami. Inductive learning algorithms and representations for text categorization. In Proceedings of the International Conference on Information and Knowledge Management, pages 148–155, 1998.
6.U. M. Fayyad and K. B. Irani. Multi-interval discretisation of continuous-valued attributes. In Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence, pages 1022–1027. Morgan Kaufmann, 1993.
7.M. A. Hall. Correlation-based feature selection for machine learning. PhD thesis, Department of Computer Science, University of Waikato, Hamilton, New Zealand, 1998.
8.Mark Hall. Correlation-based feature selection for discrete and numeric class machine learning. In Proc. of the 17th International Conference on Machine Learning (ICML2000), 2000.
9.K. Kira and L. Rendell. A practical approach to feature selection. In Proceedings of the Ninth International Conference on Machine Learning, pages 249–256. Morgan Kaufmann, 1992.
10.Ron Kohavi and George H. John. Wrappers for feature subset selection. Artificial Intelligence, 97:273–324, 1997.
11.I. Kononenko. Estimating attributes: Analysis and extensions of relief. In Proceedings of the Seventh European Conference on Machine Learning, pages 171–182. Springer-Verlag, 1994.
12.P. Langley, W. Iba, and K. Thompson. An analysis of Bayesian classifiers. In Proc. of the Tenth National Conference on Artificial Intelligence, pages 223–228, San Jose, CA, 1992. AAAI Press. [Langley92.ps.gz, from http://www.isle.org/~langley/papers/bayes.aaai92.ps].
13.H. Liu and R. Setiono. A probabilistic approach to feature selection: A filter solution. In Proceedings of the 13th International Conference on Machine Learning, pages 319–327. Morgan Kaufmann, 1996.
14.J. R. Quinlan. C4.5: Programs for Machine Learning. Morgan Kaufmann, San Mateo, CA., 1993.
15.M. Sikonja and I. Kononenko. An adaptation of relief for attribute estimation in regression. In Proceedings of the Fourteenth International Conference (ICML’97), pages 296–304. Morgan Kaufmann, 1997.
16.Yiming Yang and Jan O. Pedersen. A comparative study on feature selection in text categorization. In International Conference on Machine Learning, pages 412–420, 1997.

Citation

MLA

Hall, M. A., and G. Holmes. “Benchmarking Attribute Selection Techniques for Discrete Class Data Mining”. IEEE Transactions on Knowledge and Data Engineering, vol. 15, no. 6, 2003, pp. 1437–47, https://doi.org/10.1109/TKDE.2003.1245283.

APA

Hall, M. A., & Holmes, G. (2003). Benchmarking attribute selection techniques for discrete class data mining. IEEE Transactions on Knowledge and Data Engineering, 15(6), 1437–1447. https://doi.org/10.1109/TKDE.2003.1245283

Chicago

Hall, M. A., and G. Holmes. 2003. “Benchmarking Attribute Selection Techniques for Discrete Class Data Mining”. IEEE Transactions on Knowledge and Data Engineering 15 (6): 1437–47. https://doi.org/10.1109/TKDE.2003.1245283.

Harvard

Hall, M.A. and Holmes, G. (2003) “Benchmarking attribute selection techniques for discrete class data mining”, IEEE Transactions on Knowledge and Data Engineering, 15(6), pp. 1437–1447. Available at: https://doi.org/10.1109/TKDE.2003.1245283.

Vancouver

1. Hall MA, Holmes G (2003) Benchmarking attribute selection techniques for discrete class data mining. IEEE Transactions on Knowledge and Data Engineering 15:1437–1447

BibTeX

@article{Hall_2003, title={Benchmarking attribute selection techniques for discrete class data mining}, volume={15}, ISSN={1041-4347}, url={http://dx.doi.org/10.1109/TKDE.2003.1245283}, DOI={10.1109/tkde.2003.1245283}, number={6}, journal={IEEE Transactions on Knowledge and Data Engineering}, publisher={Institute of Electrical and Electronics Engineers (IEEE)}, author={Hall, M.A. and Holmes, G.}, year={2003}, month=Nov, pages={1437–1447} }

Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

Benchmarking Attribute Selection Techniques for Discrete Class Data Mining

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Modified Forward Selection for Attribute Ranking

Knowl 2 — Optimal Subset Determination via Cross-Validated Feature Ranking Evaluation

Knowl 3 — Correlation-Based Feature Selection Metric and Search

Knowl 4 — Consistency-Based Attribute Subset Evaluation Metric

Knowl 5 — ReliefF Attribute Ranking for Multi-Class Noisy Domains

Knowl 6 — Classification Accuracy Benchmark for Naive Bayes

Knowl 7 — Classification Accuracy Benchmark for C4.5 Decision Trees

Knowl 8 — Inducer-Specific Biases in Attribute Selection Effectiveness

Knowl 9 — Decision Tree Complexity Reduction via Attribute Selection

Knowl 10 — Dimensionality Reduction and Computational Speed Trade-Offs

References

Citation

BibTeX

Access the Paper