Wouter Duivesteijn

Nearest Neighbour Classification with Monotonicity Constraints

In many application areas of machine learning, prior knowledge concerning the monotonicity of relations between the response variable and predictor variables is readily available. Monotonicity may also be an important model requirement with a view toward explaining and justifying decisions, such as acceptance/rejection decisions. We propose a modified nearest neighbour algorithm for the construction of monotone classifiers from data. We start by making the training data monotone with as few label changes as possible. The relabeled data set can be viewed as a monotone classifier that has the lowest possible error-rate on the training data. The relabeled data is subsequently used as the training sample by a modified nearest neighbour algorithm. This modified nearest neighbour rule produces predictions that are guaranteed to satisfy the monotonicity constraints. Hence, it is much more likely to be accepted by the intended users. Our experiments show that monotone kNN often outperforms standard kNN in problems where the monotonicity constraints are applicable.

Exceptional Model Mining

Finding subsets of a dataset that somehow deviate from the norm, i.e. where something interesting is going on, is a classical Data Mining task. In traditional local pattern mining methods, such deviations are measured in terms of a relatively high occurrence (frequent itemset mining), or an unusual distribution for one designated target attribute (common use of subgroup discovery). These, however, do not encompass all forms of “interesting”. To capture a more general notion of interestingness in subsets of a dataset, we develop Exceptional Model Mining (EMM). This is a supervised local pattern mining framework, where several target attributes are selected, and a model over these targets is chosen to be the target concept. Then, we strive to find subgroups: subsets of the dataset that can be described by a few conditions on single attributes. Such subgroups are deemed interesting when the model over the targets on the subgroup is substantially different from the model on the whole dataset. For instance, we can find subgroups where two target attributes have an unusual correlation, a classifier has a deviating predictive performance, or a Bayesian network fitted on several target attributes has an exceptional structure. We give an algorithmic solution for the EMM framework, and analyze its computational complexity. We also discuss some illustrative applications of EMM instances, including using the Bayesian network model to identify meteorological conditions under which food chains are displaced, and using a regression model to find the subset of households in the Chinese province of Hunan that do not follow the general economic law of demand.

Subgroup Discovery Meets Bayesian Networks -- An Exceptional Model Mining Approach

Whenever a dataset has multiple discrete target variables, we want our algorithms to consider not only the variables themselves, but also the interdependencies between them. We propose to use these interdependencies to quantify the quality of subgroups, by integrating Bayesian networks with the Exceptional Model Mining framework. Within this framework, candidate subgroups are generated. For each candidate, we fit a Bayesian network on the target variables. Then we compare the network’s structure to the structure of the Bayesian network fitted on the whole dataset. To perform this comparison, we define an edit distance-based distance metric that is appropriate for Bayesian networks. We show interesting subgroups that we experimentally found with our method on datasets from music theory, semantic scene classification, biology and zoogeography.

Benefits of a short, practical questionnaire to measure subjective perception of nasal appearance after aesthetic rhinoplasty.

Background: The authors tested a short, practically designed questionnaire to assess changes in subjective perception of nasal appearance in patients before and after aesthetic rhinoplasty. Methods: A prospective cohort study was conducted in a group of 121 patients who desired aesthetic rhinoplasty and were operated on by one surgeon. The questionnaire contained five questions (E1-E5) based on a five-point Likert scale and a visual analogue scale (range, 0 to 10). Two questions were designed as trick questions to help the surgeon screen for signs of body dysmorphic disorder. Results: All patients rated the appearance of their nose as improved after surgery. The visual analogue scale revealed a Gaussian curve of normal distribution (range, 0.5 to 10) around a significant improvement (mean, 4.36 points, p = 0.018). Also, question E1, question E2, and the sum of questions E1 through E5 showed a statistically significant improvement after surgery (p = 1.74 × 10–36, p = 4.29 × 10–33, and p = 9.23 × 10–31, respectively). The authors found a linear relationship between preoperative score on the trick questions and postoperative increase in visual analogue scale score. Test-retest reliability could be investigated in 74 of 121 patients (61 percent) and showed a positive correlation between postoperative (1 year after surgery) and repostoperative response (2 to 4 years after surgery). Conclusions: The authors concluded that a surgeon performing aesthetic rhinoplasty can benefit from using this questionnaire. It is simple, takes no more than 2 minutes to complete, and provides helpful subjective information regarding patients’ preoperative nasal appearance and postoperative surgical outcome. CLINICAL QUESTION/LEVEL OF EVIDENCE: Therapeutic, IV.

Exploiting False Discoveries -- Statistical Validation of Patterns and Quality Measures in Subgroup Discovery

Subgroup discovery suffers from the multiple comparisons problem: we search through a large space, hence whenever we report a set of discoveries, this set will generally contain false discoveries. We propose a method to compare subgroups found through subgroup discovery with a statistical model we build for these false discoveries. We determine how much the subgroups we find deviate from the model, and hence statistically validate the found subgroups. Furthermore we propose to use this subgroup validation to objectively compare quality measures used in subgroup discovery, by determining how much the top subgroups we find with each measure deviate from the statistical model generated with that measure. We thus aim to determine how good individual measures are in selecting significant findings. We invoke our method to experimentally compare popular quality measures in several subgroup discovery settings.

Different slopes for different folks: mining for exceptional regression models with cook's distance

Exceptional Model Mining (EMM) is an exploratory data analysis technique that can be regarded as a generalization of subgroup discovery. In EMM we look for subgroups of the data for which a model fitted to the subgroup differs substantially from the same model fitted to the entire dataset. In this paper we develop methods to mine for exceptional regression models. We propose a measure for the exceptionality of regression models (Cooks distance), and explore the possibilities to avoid having to fit the regression model to each candidate subgroup. The algorithm is evaluated on a number of real life datasets. These datasets are also used to illustrate the results of the algorithm. We find interesting subgroups with deviating models on datasets from several different domains. We also show that under certain circumstances one can forego fitting regression models on up to 40% of the subgroups, and these 40% are the relatively expensive regression models to compute.

Split Hump Technique for Reduction of the Overprojected Nasal Dorsum: A Statistical Analysis on Subjective Body Image in Relation to Nasal Appearance and Nasal Patency in 97 Patients Undergoing Aesthetic Rhinoplasty

OBJECTIVES To describe the split hump technique (SHT) and to examine its effectiveness for correction of an overprojected nasal dorsum in patients undergoing aesthetic rhinoplasty. METHODS This prospective study included 97 patients. Objective assessment was performed using a short, practical questionnaire. Investigation focused on nasal patency and the patient perception of body image in relation to nasal appearance using 5-point Likert scale questions and visual analog scales. RESULTS Use of the SHT resulted in a significant improvement in nasal patency and aesthetic nasal perception. Sum functional question scores decreased from 9.154 to 6.351 and aesthetic question scores from 13.897 to 6.825 (P < .001 for both). Mean aesthetic visual analog scale scores improved in all patients, from 3.346 to 7.782 (P < .001). Graphic illustration of this improvement revealed a gaussian curve of normal distribution around a mean (SD) improvement of 4.48 (1.93). CONCLUSIONS Traditional en bloc humpectomy maneuvers are frequently combined with spreader graft use to avoid postoperative inferomedial repositioning of the upper lateral cartilages and inverted-V deformity. The SHT for correction of the overprojected dorsum creates a paradigm change in this patient group. The transverse segments of the upper lateral cartilages are saved and repositioned instead of being resected as a part of an en bloc osseocartilaginous composite hump resection in a transverse plane. Several modifications of the SHT enable the surgeon to deproject the nose while keeping sufficient strength in the keystone area and augmenting dorsal width. Using statistical analysis of subjective patient data, we could prove a broad acceptance and appreciation for the SHT.

Multilayer perceptron for label ranking

Label Ranking problems are receiving increasing attention in machine learning. The goal is to predict not just a single value from a finite set of labels, but rather the permutation of that set that applies to a new example (e.g., the ranking of a set of financial analysts in terms of the quality of their recommendations). In this paper, we adapt a multilayer perceptron algorithm for label ranking. We focus on the adaptation of the Back-Propagation (BP) mechanism. Six approaches are proposed to estimate the error signal that is propagated by BP. The methods are discussed and empirically evaluated on a set of benchmark problems.

Understanding Where Your Classifier Does (Not) Work -- The SCaPE Model Class for EMM

FACT, the First G-APD Cherenkov Telescope, detects air showers induced by high-energetic cosmic particles. It is desirable to classify a shower as being induced by a gamma ray or a background particle. Generally, it is nontrivial to get any feedback on the real-life training task, but we can attempt to understand how our classifier works by investigating its performance on Monte Carlo simulated data. To this end, in this paper we develop the SCaPE (Soft Classifier Performance Evaluation) model class for Exceptional Model Mining, which is a Local Pattern Mining framework devoted to highlighting unusual interplay between multiple targets. In our Monte Carlo simulated data, we take as targets the computed classifier probabilities and the binary column containing the ground truth: which kind of particle induced the corresponding shower. Using a newly developed quality measure based on ranking loss, the SCaPE model class highlights subspaces of the search space where the classifier performs particularly well or poorly. These subspaces arrive in terms of conditions on attributes of the data, hence they come in a language a domain expert understands, which should aid him in understanding where his/her classifier does (not) work. Found subgroups highlight subspaces whose difficulty for classification is corroborated by astrophysical interpretation, as well as subspaces that warrant further investigation.

Multi-label lego -- enhancing multi-label classifiers with local patterns

The straightforward approach to multi-label classification is based on decomposition, which essentially treats all labels independently and ignores interactions between labels. We propose to enhance multi-label classifiers with features constructed from local patterns representing explicitly such interdependencies. An Exceptional Model Mining instance is employed to find local patterns representing parts of the data where the conditional dependence relations between the labels are exceptional. We construct binary features from these patterns that can be interpreted as partial solutions to local complexities in the data. These features are then used as input for multi-label classifiers. We experimentally show that using such constructed features can improve the classification performance of decompositive multi-label learning techniques.