Search

Clustering is an important technique in data analysis which can reveal hidden patterns and unknown relationships in the data. A common problem in clustering is the proper choice of parameter settings. To tackle this, automated algorithm configuration is available which can automatically find the best parameter settings. In practice, however, many of our today’s data sources are data streams due to the widespread deployment of sensors, the internet-of-things or (social) media. Stream clustering aims to tackle this challenge by identifying, tracking and updating clusters over time. Unfortunately, none of the existing approaches for automated algorithm configuration are directly applicable to the streaming scenario. In this paper, we explore the possibility of automated algorithm configuration for stream clustering algorithms using an ensemble of different configurations. In first experiments, we demonstrate that our approach is able to automatically find superior configurations and refine them over time.

Word embeddings are a useful tool for extracting knowledge from the free-form text contained in electronic health records, but it has become commonplace to train such word embeddings on data that do not accurately reflect how language is used in a healthcare context. We use prediction of medical codes as an example application to compare the accuracy of word embeddings trained on health corpora to those trained on more general collections of text. It is shown that both an increase in embedding dimensionality and an increase in the volume of health-related training data improves prediction accuracy. We also present a comparison to the traditional bag-of-words feature representation, demonstrating that in many cases, this conceptually simple method for representing text results in superior accuracy to that of word embeddings.

We present an empirical evaluation of machine learning algorithms in cross-domain few-shot learning based on a fixed pre-trained feature extractor. Experiments were performed in five target domains (CropDisease, EuroSAT, Food101, ISIC and ChestX) and using two feature extractors: a ResNet10 model trained on a subset of ImageNet known as miniImageNet and a ResNet152 model trained on the ILSVRC 2012 subset of ImageNet. Commonly used machine learning algorithms including logistic regression, support vector machines, random forests, nearest neighbour classification, naive Bayes, and linear and quadratic discriminant analysis were evaluated on the extracted feature vectors. We also evaluated classification accuracy when subjecting the feature vectors to normalisation using p-norms. Algorithms originally developed for the classification of gene expression data—the nearest shrunken centroid algorithm and LDA ensembles obtained with random projections—were also included in the experiments, in addition to a cosine similarity classifier that has recently proved popular in few-shot learning. The results enable us to identify algorithms, normalisation methods and pre-trained feature extractors that perform well in cross-domain few-shot learning. We show that the cosine similarity classifier and \(\ell ^2\)-regularised 1-vs-rest logistic regression are generally the best-performing algorithms. We also show that algorithms such as LDA yield consistently higher accuracy when applied to \(\ell ^2\)-normalised feature vectors. In addition, all classifiers generally perform better when extracting feature vectors using the ResNet152 model instead of the ResNet10 model.

Machine learning has become one of the most important tools in data analysis. However, selecting the most appropriate machine learning algorithm and tuning its hyperparameters to their optimal values remains a difficult task. This is even more difficult for streaming applications where automated approaches are often not available to help during algorithm selection and configuration. This paper proposes the first approach for automated algorithm selection and configuration of stream clustering algorithms. We train an ensemble of different stream clustering algorithms and configurations in parallel and use the best performing configuration to obtain a clustering solution. By drawing new configurations from better performing ones, we are able to improve the ensemble performance over time. In large experiments on real and artificial data we show how our ensemble approach can improve upon default configurations and can also compete with a-posteriori algorithm configuration. Our approach is considerably faster than a-posteriori approaches and applicable in real-time. In addition, it is not limited to stream clustering and can be generalised to all streaming applications, including stream classification and regression.

Deep neural networks produce state-of-the-art results when trained on a large number of labeled examples but tend to overfit when small amounts of labeled examples are used for training. Creating a large number of labeled examples requires considerable resources, time, and effort. If labeling new data is not feasible, so-called semi-supervised learning can achieve better generalisation than purely supervised learning by employing unlabeled instances as well as labeled ones. The work presented in this paper is motivated by the observation that transfer learning provides the opportunity to potentially further improve performance by exploiting models pretrained on a similar domain. More specifically, we explore the use of transfer learning when performing semi-supervised learning using self-learning. The main contribution is an empirical evaluation of transfer learning using different combinations of similarity metric learning methods and label propagation algorithms in semi-supervised learning. We find that transfer learning always substantially improves the model's accuracy when few labeled examples are available, regardless of the type of loss used for training the neural network. This finding is obtained by performing extensive experiments on the SVHN, CIFAR10, and Plant Village image classification datasets and applying pretrained weights from Imagenet for transfer learning.

Neural networks have been successfully used as classification models yielding state-of-the-art results when trained on a large number of labeled samples. These models, however, are more difficult to train successfully for semi-supervised problems where small amounts of labeled instances are available along with a large number of unlabeled instances. This work explores a new training method for semi-supervised learning that is based on similarity function learning using a Siamese network to obtain a suitable embedding. The learned representations are discriminative in Euclidean space, and hence can be used for labeling unlabeled instances using a nearest-neighbor classifier. Confident predictions of unlabeled instances are used as true labels for retraining the Siamese network on the expanded training set. This process is applied iteratively. We perform an empirical study of this iterative self-training algorithm. For improving unlabeled predictions, local learning with global consistency [22] is also evaluated.

Effective regularisation of neural networks is essential to combat overfitting due to the large number of parameters involved. We present an empirical analogue to the Lipschitz constant of a feed-forward neural network, which we refer to as the maximum gain. We hypothesise that constraining the gain of a network will have a regularising effect, similar to how constraining the Lipschitz constant of a network has been shown to improve generalisation. A simple algorithm is provided that involves rescaling the weight matrix of each layer after each parameter update. We conduct a series of studies on common benchmark datasets, and also a novel dataset that we introduce to enable easier significance testing for experiments using convolutional networks. Performance on these datasets compares favourably with other common regularisation techniques. Data related to this paper is available at: https://www.cs.waikato.ac.nz/~ml/sins10/.

Neural networks are known to be very sensitive to the initial weights. There has been a lot of research on initialization that aims to stabilize the training process. However, very little research has studied the relationship between initialization and generalization. We demonstrate that poorly initialized model will lead to lower test accuracy. We propose a supervised pretraining technique that helps improve generalization on binary classification problems. The experimental results on four UCI datasets show that the proposed pretraining leads to higher test accuracy compared to the he_normal initialization when the training set is small. In further experiments on synthetic data, the improvement on test accuracy using the proposed pretraining reaches more than 30% when the data has high dimensionality and noisy features. Code related to this paper is available at: https://github.com/superRookie007/supervised_pretraining.

A system of nested dichotomies (NDs) is a method of decomposing a multiclass problem into a collection of binary problems. Such a system recursively applies binary splits to divide the set of classes into two subsets, and trains a binary classifier for each split. Many methods have been proposed to perform this split, each with various advantages and disadvantages. In this paper, we present a simple, general method for improving the predictive performance of NDs produced by any subset selection techniques that employ randomness to construct the subsets. We provide a theoretical expectation for performance improvements, as well as empirical results showing that our method improves the root mean squared error of NDs, regardless of whether they are employed as an individual model or in an ensemble setting.

Nested dichotomies (NDs) are used as a method of transforming a multiclass classification problem into a series of binary problems. A tree structure is induced that recursively splits the set of classes into subsets, and a binary classification model learns to discriminate between the two subsets of classes at each node. In this paper, we demonstrate that these NDs typically exhibit poor probability calibration, even when the binary base models are well-calibrated. We also show that this problem is exacerbated when the binary models are poorly calibrated. We discuss the effectiveness of different calibration strategies and show that accuracy and log-loss can be significantly improved by calibrating both the internal base models and the full ND structure, especially when the number of classes is high.

A system of nested dichotomies is a method of decomposing a multi-class problem into a collection of binary problems. Such a system recursively applies binary splits to divide the set of classes into two subsets, and trains a binary classifier for each split. Although ensembles of nested dichotomies with random structure have been shown to perform well in practice, using a more sophisticated class subset selection method can be used to improve classification accuracy. We investigate an approach to this problem called random-pair selection, and evaluate its effectiveness compared to other published methods of subset selection. We show that our method outperforms other methods in many cases when forming ensembles of nested dichotomies, and is at least on par in all other cases. The software related to this paper is available at https://svn.cms.waikato.ac.nz/svn/weka/trunk/packages/internal/ensemblesOfNestedDichotomies/.

The ubiquity of data streams has been encouraging the development of new incremental and adaptive learning algorithms. Data stream learners must be fast, memory-bounded, but mainly, tailored to adapt to possible changes in the data distribution, a phenomenon named concept drift. Recently, several works have shown the impact of a so far nearly neglected type of drifcccct: feature drifts. Feature drifts occur whenever a subset of features becomes, or ceases to be, relevant to the learning task. In this paper we (i) provide insights into how the relevance of features can be tracked as a stream progresses according to information theoretical Symmetrical Uncertainty; and (ii) how it can be used to boost two learning schemes: Naive Bayesian and k-Nearest Neighbor. Furthermore, we investigate the usage of these two new dynamically weighted learners as prediction models in the leaves of the Hoeffding Adaptive Tree classifier. Results show improvements in accuracy (an average of 10.69 % for k-Nearest Neighbor, 6.23 % for Naive Bayes and 4.42 % for Hoeffding Adaptive Trees) in both synthetic and real-world datasets at the expense of a bounded increase in both memory consumption and processing time.

The Pairwise Meta-Rules (PMR) method proposed in [18] has been shown to improve the predictive performances of several meta-learning algorithms for the algorithm ranking problem. Given m target objects (e.g., algorithms), the training complexity of the PMR method with respect to m is quadratic: \(\binom{m}{2} = m \times (m - 1) / 2\). This is usually not a problem when m is moderate, such as when ranking 20 different learning algorithms. However, for problems with a much larger m, such as the meta-learning-based parameter ranking problem, where m can be 100+, the PMR method is less efficient. In this paper, we propose a novel method named Hierarchical Meta-Rules (HMR), which is based on the theory of orthogonal contrasts. The proposed HMR method has a linear training complexity with respect to m, providing a way of dealing with a large number of objects that the PMR method cannot handle efficiently. Our experimental results demonstrate the benefit of the new method in the context of meta-learning.

Multi-instance learning is a generalisation of attribute-value learning where examples for learning consist of labeled bags (i.e. multi-sets) of instances. This learning setting is more computationally challenging than attribute-value learning and a natural fit for important application areas of machine learning such as classification of molecules and image classification. One approach to solve multi-instance learning problems is to apply propositionalisation, where bags of data are converted into vectors of attribute-value pairs so that a standard propositional (i.e. attribute-value) learning algorithm can be applied. This approach is attractive because of the large number of propositional learning algorithms that have been developed and can thus be applied to the propositionalised data. In this paper, we empirically investigate a variant of an existing propositionalisation method called TLC. TLC uses a single decision tree to obtain propositionalised data. Our variant applies a random forest instead and is motivated by the potential increase in robustness that this may yield. We present results on synthetic and real-world data from the above two application domains showing that it indeed yields increased classification accuracy when applying boosting and support vector machines to classify the propositionalised data.