Scoring relevancy of features based on combinatorial analysis of Lasso with application to lymphoma diagnosis
 Habil Zare^{1, 2}Email author,
 Gholamreza Haffari^{3},
 Arvind Gupta^{2} and
 Ryan R Brinkman^{4}
https://doi.org/10.1186/1471216414S1S14
© Zare et al.; licensee BioMed Central Ltd. 2013
Published: 21 January 2013
Abstract
One challenge in applying bioinformatic tools to clinical or biological data is high number of features that might be provided to the learning algorithm without any prior knowledge on which ones should be used. In such applications, the number of features can drastically exceed the number of training instances which is often limited by the number of available samples for the study. The Lasso is one of many regularization methods that have been developed to prevent overfitting and improve prediction performance in highdimensional settings. In this paper, we propose a novel algorithm for feature selection based on the Lasso and our hypothesis is that defining a scoring scheme that measures the "quality" of each feature can provide a more robust feature selection method. Our approach is to generate several samples from the training data by bootstrapping, determine the best relevanceordering of the features for each sample, and finally combine these relevanceorderings to select highly relevant features. In addition to the theoretical analysis of our feature scoring scheme, we provided empirical evaluations on six real datasets from different fields to confirm the superiority of our method in exploratory data analysis and prediction performance. For example, we applied FeaLect, our feature scoring algorithm, to a lymphoma dataset, and according to a human expert, our method led to selecting more meaningful features than those commonly used in the clinics. This case study built a basis for discovering interesting new criteria for lymphoma diagnosis. Furthermore, to facilitate the use of our algorithm in other applications, the source code that implements our algorithm was released as FeaLect, a documented R package in CRAN.
Keywords
Introduction
To build a robust classifier, the number of training instances is usually required to be more than the number of features. In many real life applications such as bioinformatics, natural language processing, and computer vision, a high number of features might be provided to the learning algorithm without any prior knowledge about which ones should be used. Therefore, the number of features can drastically exceed the number of training instances and the model is subject to overfit the training data. Many regularization methods have been developed to prevent overfitting and to improve the generalization error bound of the predictor in this learning situation.
Most notably, Lasso [1] is an ℓ_{1}regularization technique for linear regression which has attracted much attention in machine learning and statistics. The same approach is useful in classification because any binary classification problem can be reduced to a regression problem by treating the class labels as real numbers, and consider the sign of the model prediction as the class label. The features selected by the Lasso depends on the regularization parameter, and the set of solutions for all values of this free parameter is provided by regularization path [2]. Although efficient algorithms exist for recovering the whole regularization path for the Lasso [3], finding a subset of highly relevant features which leads to a robust predictor is a prominent research question.
In this paper, we propose a novel algorithm for feature selection based on the Lasso and our hypothesis is that defining a scoring scheme that measures the "quality" of each feature can provide a more robust feature selection method. Our approach is to generate several samples from the training data by bootstrapping, determine the best relevanceordering of the features for each sample, and finally combine these relevanceorderings to select highly relevant features. In addition to the theoretical analysis of our feature scoring scheme, we provided empirical evaluations using a reallife lymphoma dataset as well as several UCI datasets, which confirms the superiority of our method in exploratory data analysis and prediction performance.
Background and previous work
where the response random variable Y ∈ ℝ is dependent on a ddimensional covariate X ∈ ℝ^{ d }, and the training data $D={\left\{\left({\mathbf{x}}_{i},{y}_{i}\right)\right\}}_{i=1}^{n}$ is independently and identically sampled from a fixed joint distribution P_{ XY }. It is well known that the ℓ_{1}regularization term shrinks many components of the solution to zero, and thus performs feature selection [4]. There has been also some variants, such as elastic nets [5], to select highlycorrelated predictive features. The number of selected features in eqn (1) is controlled by the regularization parameter λ.
A common practice is to find the best value for λ by crossvalidation to maximize the prediction accuracy. Having found the best value for the regularization parameter, the features are selected based on the nonzero components of the global and unique minimizer of the training objective in equation (1). However, recent research on the consistency of the Lasso [4, 6–10] shows that a fixed value of λ for all n will not result in a consistent estimate for the parameter vector [7]. Now, the question is what would be a proper value for λ as a function of n with a theoretical basis?
Various decaying schemes of the regularization parameter were studied [4, 7, 8, 11] and it is shown that under specific settings, Lasso selects the relevant features with probability one and the irrelevant features with a positive probability less than one, provided that the number of training instances tends to infinity. To do a better feature selection, note that each run of the crossvalidation gives the value of the regularization parameter λ and the corresponding selectedfeatures. If several samples were available from the underlying data distribution, irrelevant features could be removed by simply intersecting the set of selected features for each sample. The idea in [7] is to provide such datasets by resampling with replacement from the given training dataset using the bootstrap method [12]. This approach leads to Bolasso algorithm for feature selection that is theoretically motivated by the proposition 1.
where m > 1 is the number of bootstrap samples, and all A_{ i } s are positive constants.
Now, if we send m to infinity slower than ${e}^{{A}_{2}n}$, then with probability tending to one Bolasso will select J, exactly the relevant features. The proposition 1 guarantees the performance of Bolasso only asymptoticly, i.e. when n → ∞. However, in real applications where the number of training samples is often limited, the probability of selecting relevant features can be significantly less than 1. One of the main goals of our proposed framework in this paper is to address this problem by scoring the features.
Previous studies have shown that there is room for improving Bolasso [7, 8]. For example, while on synthetic data it outperforms similar methods such as ridge regression, Lasso, and bagging of Lasso estimates [13], Bolasso is sometimes too strict on real data because it requires the relevant features to be selected in all bootstrap runs. BolassoS, a soft version of Bolasso, performs better in practice because it relaxes this condition and selects a feature if it is chosen in at least a userdefined fraction of the bootstrap replicates (a threshold of 90% is considered to be enough). BolassoS is more flexible and thus, more appropriate for the practical models that are not extremely sparse [8].
Our contributions
In this paper, we develop FeaLect algorithm that is softer than Bolasso in the following three directions:

For each bootstrap sample, Bolasso considers only one model that minimizes the training objective $\mathcal{L}$ in eqn (1), whereas we include information provided by the whole regularization path,

Instead of making a binary decision of inclusion or exclusion, we compute a score value for each feature that can help the user to select the more relevant ones,

While BolassoS relies on a threshold, our theoretical study of the behaviour of irrelevant features leads to an analytical criterion for feature selection without using any predefined parameter.
We compared the performance of Bolasso, FeaLect, and Lars algorithms for feature selection on six real datasets in a systematic manner. The source code that implements our algorithm was released as FeaLect, a documented R package in CRAN.
Feature scoring and mathematical analysis
In this section, we describe our novel algorithm that scores the features based on their performance on samples obtained by bootstapping. Afterwards, we present the mathematical analysis of our algorithm which builds the theoretical basis for its proposed automatic thresholding in feature selection.
The FeaLect algorithm
Algorithm 1 Feature Scoring
1: for t = 1 to m do
2: Sample (without replacement) a random subset B ⊂ D with size γD
3: Run Lars on B to obtain ${F}_{1}^{B}$, ...,${F}_{d}^{B}$
4: Compute ${S}_{1}^{B}$, ...,${S}_{d}^{B}$ using eqn (2)
5: for k ∈ {1, ..., d} do
6: Update the feature scores for all feature $f:S(f\leftarrow S\left(f\right)+{S}_{k}^{B}\left(f\right)/m$
7: end for
8: end for
9: Fit a 3segment spline (g_{1}(.), g_{2}(.), g_{3}(.)) on logscale feature score curve (see the text for more information)
10: return features corresponding to g_{3} as informative features
Before describing the rest of the algorithm, let us have a look at the feature scores for our lymphoma classification problem (The task and data set is described in details in the Experiment section). Figure 2 depict the total score of features in logscale, where features are sorted according to their increasing total scores. The feature score curve is almost linear in the middle and bending at both ends. We hypothesize that features with a "very high score" in the top nonlinear and bending part of the curve are good candidates for informative features. Furthermore, the linear middlepart of the curve consists of features that are responsible for the model to get overfitted and therefore we call them irrelevant features. In the next section, a formal definition will be provided to clarify this intuitive idea and we show how this insight can be very helpful in identifying informative features.
The final step of our feature selection algorithm is to fit a 3segment spline model to the feature score curve: the first quadratic lowerpart captures the low score features, the linear middlepart captures irrelevant features, and the last quadratic upperpart captures highscore informative features. As discussed below, the middle linearpart provides an analytic threshold for the score of relevant features: The features with score above this threshold are reported as informative features which can be used for training the final predictor and/or explanatory data analysis.
The analysis
The aim of this analysis is to provide a mathematical explanation for the linearity of the middle part of the scoring function (Figure 2), and also a justification for why the features corresponding to this part can be excluded. We first present a probabilistic interpretation of the feature scores. and then we provide a precise definition of an irrelevant feature. Our definition formalizes the fact that such a feature is selected by the Lasso if and only if a particular fixed finite subset U of instances is included in the random training set, whereas a relevant feature should be selected for almost any general U. We estimate the probability that a random sample B ⊂ D contains U as n grows to infinity. Finally, we show that asymptotically, the log of the scores for irrelevant features is linear in U. This explains the linearity of the middle part of the feature score curve in Figure 2.
The following definition formalizes the idea that irrelevant features depend only on a specific subset of the whole data set.
In words, f_{ i } is selected in ${F}_{k}^{B}$ if and only if B contains U. Next, we derive the probability of including a specific set U in a randomly generated sample.
where r is the number of samples in U and γ is the fraction of samples chosen for a random set B.
The first line of the above proof relies on the assumption that the members of the random set B are chosen without replacement, and the claim derives from the fact that γ is a fixed constant. □
The following theorem concludes our argument for the exponential behavior of total score of irrelevant features. It relates the probability of selecting a feature f_{ i } irrelevant on U to the probability of including U in the sample.
The last equation was proved in lemma 4, and the one before that from definition 3. □
Although we presented the above arguments for the Lasso, it also should work for any other feature selection algorithm which exhibits linearity in its feature score curve. That is, features corresponding to the linear part of the scoring curve are indeed the irrelevant features for that algorithm, and therefor, the features on nonlinear upperpart should be considered as informative ones. Obviously the features on the nonlinear lowerpart are not interesting for the any prediction task because their scores are even less than the irrelevant features. We speculate that these features do not present a linear behavior because not only they are not relevant to the outcome, but also they are not associated with any particular set U, meaning they are not even included in an overfitted model. A followup study may investigate this hypothesis further.
Experiment with real data
We applied FeaLect on several datasets to test the performance of our feature selection algorithm in real life conditions.
Lymphoma
Lymphoma is a cancer that begins in the lymphatic cells of the immune system, and is presented as a solid tumor of lymphoid cells [14]. Just as cancer represents many different diseases, lymphoma represents many different cancers of lymphocytes [15]. We applied our algorithm for automatic diagnosis of lymphoma types based on flow cytometry (FCM) data [16]. Usually 1530 markers are used for each patient, where each marker distinguishes a particular cell type based on its protein content. We analyzed flow cytometry data of 85 lymphoma patients who had been diagnosed at BC Cancer Agency, Vancouver, Canada between 20042007. The patients were grouped into four toplevel disease subgroups and the goal was to build a classifier that could diagnosis 20 test patients based on their FCM data. For each group, we trained a classifier to distinguish that group versus the others. These four classifiers were then combined to provide the toplevel diagnosis.
Data preparation and feature extraction
The blood sample of each patient was divided into 7 portions, and each portion is examined in a different tube by the cytometer. Each tube gives 5 dimensional data of 20,00070,000 blood cells. In the first analysis step, we used a spectral clustering approach to cluster the cells in each tube into cell populations. It was not possible to directly apply classical spectral clustering [17–20] to the lymphoma data because it involved computing eigenvectors of a big nbyn matrix where n ranges from 20,000 to 70,000. Instead, we have made use of SamSPECTRAL that is our enhanced spectral clustering method capable of analyzing large amount of data in a reasonable amount of time; it has also a good memory footprints [21].
 1.
Set all points to be unregistered and assume the parameter h is adjusted appropriately.
 2.
Pick a random unregistered point p (the representative of a community) and find all unregistered data points within distance h from p.
 3.
Put all of these points in a set called community p, and label them as registered.
 4.
Repeat the above two steps until no unregistered points are left.
After the above steps, the similarity between the communities is defined by summing up similarities between their members, and the resulting similarity matrix is passed to a classical spectral clustering algorithm. Because this matrix is much smaller than the original similarity matrix (3000by3000 instead of 20,000by20,000 in our experiments), its eigenvectors can be efficiently computed in reasonable time.
Each cluster computed by SamSPECTRAL was regarded as a "cell population" that could potentially have information about the lymphoma type. Without imposing any a priori knowledge on the importance of any population, we considered their sizes and their means in all dimensions as features. In total, 276 features were obtained and ignoring those with very low variance, 224 were kept for feature selection and classification.
Feature selection and classification
Since the number of features was considerably larger than the number of training samples (p = 224, n = 85), a careful feature selection scheme was needed. To reduce the computation time required, we imposed a predefined upper bound 60 on the number of features based on a priori knowledge from the biology. We initially applied ℓ_{1} regularization technique, and it was not by its own enough to prevent overfitting. Reducing the regularization parameter did not improve the results as we observed that some of the features that were known to be biologically and clinically interesting were ignored. We also applied Bolasso [7] to select relevant features. For most bootstrap samples of our data, the global error defined by equation (1) was minimized when only a few (less than 4) features were selected. Because the intersection of selected features from several (more than 10) samples was empty, Bolasso could not result in appropriate feature selection.
Next, we applied our feature selection algorithm. In our experiment, we set $\gamma =\frac{3}{4}$ to be the fraction of instances used in each iteration for training. Figures 2(a) and 2(b) depict the resulting feature scores for follicular lymphoma type after m = 1000 random runs. The logscale plot consisted of a linear part that confirmed our hypothesis experimentally. Similarly, the plots for other types of lymphoma had also linear parts. Furthermore, we reran the experiments with m = 5000 random samples and the results did not vary significantly indicating a fast convergence rate for the feature scores.
To select the informative features, we fitted a 3segment spline model to each curve. The features corresponding to the middle linear segment were considered as irrelevant ones, and ignored for the rest of analysis. Features with score higher than score of these irrelevant features were selected as informative features. We observed that unlike the pure Lasso, all features that were known to be biologically and clinically interesting were selected by our approach. Prediction accuracy was improved confirming the efficiency of our feature selection method. We used our selected features to build a linear classifier that had precision, recall and Fmeasure 98%, 94% and 96%, respectively while the best result we obtained with the pure Lasso was 93%, 82% and 87%, respectively.
For further evaluation in a data explarotary setting, we interrogated the selected features together with our clinical collaborators for novel biomarker discovery. A task which would be challenging otherwise, due to large number of features and huge amount of clinical work required to evaluate each individual feature. We narrowed down our attention to those features which were relevant to lymphoma types based on our feature selection algorithm, but were not previously reported to be biologically relevant. This approach resulted in the discovery of interesting new criteria for lymphoma diagnosis that have clinical applications in practice [22].
Additional real datasets
Comparsion of area under the ROC curve between FeaLect, lars, and Bolasso on six different datasets.
Dataset  Total samples  # of features  20 training samples  40 training samples  Reference  

Bolasso  lars  FeaLect  Bolasso  lars  FeaLect  
Lymphoma  258  505  0.62  0.81  0.84  0.67  0.87  0.88  current 
Colon  62  2000  0.50  0.57  0.65  0.47  0.64  0.75  [23] 
Arcene  100  10000  0.51  0.59  0.64  0.50  0.66  0.72  [26] (UCI) 
SECOM  208  590  0.51  0.57  0.61  0.52  0.61  0.64  [25] (UCI) 
Connectionist  208  60  0.63  0.76  0.78  0.67  0.78  0.79  [27] (UCI) 
ISOLET  479  617  0.90  0.99  1.00  0.91  1.00  1.00  [28] (UCI) 
Conclusion
We have presented FeaLect, a novel feature selection algorithm, based on Lasso (Figure 1). The idea of FeaLect is to combine the selected feature sets to score the features according to their relevancy and prediction power. An advantage of FeaLect compared to many other feature selection methods is to provide a ranking for features relevance, which can be leveraged in better prediction models and/or exploratory data analysis. We reported a cancer classification problem (lymphoma diagnosis) for which distinguishing the most relevant features is of great interest from the biological and clinical point of view. FeaLect has led to the discovery of novel biomarkers for this disease to help clinicians in lymphoma subtype diagnosis [22]. The logscale score curve can be studied in more detail and explaining its behavior in the nonlinear parts is potentially a source of insight. Shedding more light on the Lasso performance by studying feature scores is a possible future direction of this study.
Furthermore, we provided empirical and quantitative evaluations on five other realworld datasets (from different fields) to confirm the superiority of our method, in prediction performance, compared to the baselines.
Declarations
The research and publication costs for this article were funded by NIH grants 1R01EB008400 and 1R01EB005034, the Michael Smith Foundation for Health Research, the National Science and Engineering Research Council and the MITACS Network of Centres of Excellence.
This article has been published as part of BMC Genomics Volume 14 Supplement 1, 2013: Selected articles from the Eleventh Asia Pacific Bioinformatics Conference (APBC 2013): Genomics. The full contents of the supplement are available online at http://www.biomedcentral.com/bmcgenomics/supplements/14/S1.
Declarations
Acknowledgements
The authors would like to thank Andrew Weng and Randy Gascoyne for providing data and valuable clinical insight, and Nima Aghaeepour for his scientific comments.
Authors’ Affiliations
References
 Tibshirani R: Regression Shrinkage and Selection via the Lasso. Journal of the Royal Statistical Society (Series B). 1996, 58: 267288.Google Scholar
 Hastie T, Rosset S, Tibshirani R, Zhu J: The entire regularization path for the support vector machine. Journal of Machine Learning Research. 2005, 5 (2): 1391Google Scholar
 Efron B, Hastie T, Johnstone L, Tibshirani R: Least angle regression. Annals of Statistics. 2004, 32: 407499. 10.1214/009053604000000067.View ArticleGoogle Scholar
 Zhao P, Yu B: On Model Selection Consistency of Lasso. J Mach Learn Res. 2006, 7: 25412563.Google Scholar
 Zou H, Hastie T: Regularization and variable selection via the Elastic Net. Journal of the Royal Statistical Society, Series B. 2005, 67: 301320. 10.1111/j.14679868.2005.00503.x.View ArticleGoogle Scholar
 Wainwright MJ: Sharp thresholds for highdimensional and noisy sparsity recovery using l1constrained quadratic programming (Lasso). IEEE Trans Inf Theor. 2009, 55 (5):Google Scholar
 Bach FR: Bolasso: model consistent Lasso estimation through the bootstrap. ICML '08: Proceedings of the 25th international conference on Machine learning. 2008Google Scholar
 Bach F: Modelconsistent sparse estimation through the bootstrap. Tech rep, HAL00354771. 2009Google Scholar
 Meinshausen N, Yu B: Lassotype recovery of sparse representations for highdimensional data. Annals of Statistics. 2009, 37: 246270. 10.1214/07AOS582.View ArticleGoogle Scholar
 Lounici K et al: Supnorm convergence rate and sign concentration property of Lasso and Dantzig estimators. Electronic Journal of statistics. 2008, 2: 90102. 10.1214/08EJS177.View ArticleGoogle Scholar
 Meinshausen N, Buhlmann P: Consistent neighborhood selection for sparse highdimensional graphs with the Lasso. Statist Surv. 2004, 2: 6193.Google Scholar
 Efron B, Tibshirani RJ: An Introduction to the Bootstrap (Chapman & Hall/CRC Monographs on Statistics & Applied Probability). 1998, Chapman and Hall/CRCGoogle Scholar
 Breiman L: Bagging predictors. Machine learning. 1996, 24 (2): 123140.Google Scholar
 Matasar MJ, Zelenetz AD: Overview of Lymphoma Diagnosis and Management. Radiologic Clinics of North America. 2008, 46 (2): 175198. 10.1016/j.rcl.2008.03.005.View ArticlePubMedGoogle Scholar
 Hiddemann W, Longo D, Coiffier B, Fisher R, Cabanillas F, Cavalli F, Nadler L, De Vita V, Lister T, Armitage J: Lymphoma classificationthe gap between biology and clinical management is closing. Blood. 1996Google Scholar
 Hawley TS, Hawley RG: Flow Cytometry Protocols, Methods in Molecular Biology. 2005, Humana Press, 2Google Scholar
 von Luxburg U: A tutorial on spectral clustering. Statistics and Computing. 2007, 17 (4): 395416. 10.1007/s112220079033z.View ArticleGoogle Scholar
 von Luxburg U, Belkin M, Bousquet O: Consistency of Spectral Clustering. Annals of Statistics. 2008, 36 (2): 555586. 10.1214/009053607000000640.View ArticleGoogle Scholar
 Ng AY, Jordan MI, Weiss Y: On Spectral Clustering: Analysis and an algorithm. Advances in Neural Information Processing Systems 14. 2001Google Scholar
 Azran A, Ghahramani Z: Spectral Methods for Automatic Multiscale Data Clustering. Computer Vision and Pattern Recognition, 2006 IEEE Computer Society Conference on, Volume 1. 2006, 190197.Google Scholar
 Zare H, Shooshtari P, Gupta A, Brinkman R: Data reduction for spectral clustering to analyze high throughput flow cytometry data. BMC Bioinformatics. 2010, 11: 40310.1186/1471210511403.PubMed CentralView ArticlePubMedGoogle Scholar
 Zare H, Bashashati A, Kridel R, Aghaeepour N, Haffari G, Connors J, Gascoyne R, Gupta A, Brinkman R, Weng A: Automated analysis of multidimensional flow cytometry data improves diagnostic accuracy between mantle cell lymphoma and small lymphocytic lymphoma. American Journal of Clinical Pathology. 2012, 137: 7585. 10.1309/AJCPMMLQ67YOMGEW.PubMed CentralView ArticlePubMedGoogle Scholar
 Alon U, Barkai N, Notterman D, Gish K, Ybarra S, Mack D, Levine A: Broad patterns of gene expression revealed by clustering analysis of tumor and normal colon tissues probed by oligonucleotide arrays. Proc Natl Acad Sci USA. 1999, 96 (12): 67456750. 10.1073/pnas.96.12.6745.PubMed CentralView ArticlePubMedGoogle Scholar
 Stingo F, Vannucci M: Variable selection for discriminant analysis with Markov random field priors for the analysis of microarray data. Bioinformatics. 2011, 27 (4): 49510.1093/bioinformatics/btq690.PubMed CentralView ArticlePubMedGoogle Scholar
 Frank A, Asuncion A: UCI Machine Learning Repository. 2010Google Scholar
 Guyon I, Hur AB, Gunn S, Dror G: Result analysis of the NIPS 2003 feature selection challenge. Advances in Neural Information Processing Systems 17. 2004, MIT Press, 545552.Google Scholar
 Gorman R, Sejnowski T: Analysis of hidden units in a layered network trained to classify sonar targets. Neural networks. 1988, 1: 7589.View ArticleGoogle Scholar
 Fanty MA, Cole R: Spoken Letter Recognition. NIPS. 1990, 220Google Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.