Adriana Birlutiu

Learning from multiple annotators with Gaussian processes

In many supervised learning tasks it can be costly or infeasible to obtain objective, reliable labels. We may, however, be able to obtain a large number of subjective, possibly noisy, labels from multiple annotators. Typically, annotators have different levels of expertise (i.e., novice, expert) and there is considerable diagreement among annotators. We present a Gaussian process (GP) approach to regression with multiple labels but no absolute gold standard. The GP framework provides a principled non-parametric framework that can automatically estimate the reliability of individual annotators from data without the need of prior knowledge. Experimental results show that the proposed GP multi-annotator model outperforms models that either average the training data or weigh individually learned single-annotator models.

Transcriptional plasticity of a soil arthropod across different ecological conditions.

Ecological functional genomics, dealing with the responses of organisms to their natural environment is confronted with a complex pattern of variation and a large number of confounding environmental factors. For gene expression studies to provide meaningful information on conditions deviating from normal, a baseline or normal operating range (NOR) response needs to be established which indicates how an organism’s transcriptome reacts to naturally varying ecological factors. Here we determine the transcriptional plasticity of a soil arthropod, Folsomia candida, exposed to various natural environments, as part of a first attempt in establishing such a NOR. Animals were exposed to 26 different field soils after which gene expression levels were measured. The main factor found to regulate gene expression was soil‐type (sand or clay). Cell homeostasis and DNA replication were affected in collembolans exposed to sandy soil, indicating general stress. Multivariate analysis identified soil fertility as the main factor influencing gene expression. Regarding land‐use, only forest soils showed an expression pattern deviating from the others. No significant effect of land‐use, agricultural practice or soil type on fitness was observed, but arsenic concentration was negatively correlated with reproductive output. In conclusion, transcriptional responses remained within a limited range across the different land‐uses but were significantly affected by soil‐type. This may be caused by the contrasting soil physicochemical properties to which F. candida strongly responds. The broad range of conditions over which this soil‐living detritivore is able to survive and reproduce, indicates a strategy of high plasticity, which comes with extensive gene expression regulation.

Multi-task preference learning with an application to hearing aid personalization

We present an EM-algorithm for the problem of learning preferences with semiparametric models derived from Gaussian processes in the context of multi-task learning. We validate our approach on an audiological data set and show that predictive results for sound quality perception of hearing-impaired subjects, in the context of pairwise comparison experiments, can be improved using a hierarchical model.

Efficiently learning the preferences of people

This paper presents a framework for optimizing the preference learning process. In many real-world applications in which preference learning is involved the available training data is scarce and obtaining labeled training data is expensive. Fortunately in many of the preference learning situations data is available from multiple subjects. We use the multi-task formalism to enhance the individual training data by making use of the preference information learned from other subjects. Furthermore, since obtaining labels is expensive, we optimally choose which data to ask a subject for labelling to obtain the most of information about her/his preferences. This paradigm—called active learning—has hardly been studied in a multi-task formalism. We propose an alternative for the standard criteria in active learning which actively chooses queries by making use of the available preference data from other subjects. The advantage of this alternative is the reduced computation costs and reduced time subjects are involved. We validate empirically our approach on three real-world data sets involving the preferences of people.

Expectation Propagation for Rating Players in Sports Competitions

Rating players in sports competitions based on game results is one example of paired comparison data analysis. Since an exact Bayesian treatment is intractable, several techniques for approximate inference have been proposed in the literature. In this paper we compare several variants of expectation propagation (EP). EP generalizes assumed density filtering (ADF) by iteratively improving the approximations that are made in the filtering step of ADF. Furthermore, we distinguish between two variants of EP: EP-Correlated, which takes into account the correlations between the strengths of the players and EP-Independent, which ignores those correlations. We evaluate the different approaches on a large tennis dataset to find that EP does significantly better than ADF (iterative improvement indeed helps) and EP-Correlated does significantly better than EP-Independent (correlations do matter).

Decision tree models for developing molecular classifiers for cancer diagnosis

The aim of this study is to propose a methodology for developing intelligent systems for cancer diagnosis and evaluate it on bladder cancer. Owing to recent advances in high-throughput experiments, large data repositories are now freely available for use. However, the process of extracting information from these data and transforming it into clinically useful knowledge needs to be improved. Consequently, the research focus is shifting from merely data production towards developing methods to manage and analyze it. In this study, we build classification models that are able to discriminate between normal and cancer samples based on the molecular biomarkers discovered. We focus on transparent and interpretable models for data analysis. We built molecular classifiers using decision tree models in combination with boosting and cross-validation to distinguish between normal and malign samples. The approach is designed to avoid overfitting and overoptimistic results. We perform experimental evaluation on a data set related to the urothelial carcinoma of the bladder. We identify a set of tumor microRNAs biomarkers, which integrated in an ensemble of decision tree classifiers, can discriminate between normal and cancer samples with the best published accuracy.

Domain Generalization Based on Transfer Component Analysis

This paper investigates domain generalization: How to use knowledge acquired from related domains and apply it to new domains? Transfer Component Analysis (TCA) learns a shared subspace by minimizing the dissimilarities across domains, while maximally preserving the data variance. We propose Multi-TCA, an extension of TCA to multiple domains as well as Multi-SSTCA, which is an extension of TCA for semi-supervised learning. In addition to the original application of TCA for domain adaptation problems, we show that Multi-TCA can also be applied for domain generalization. Multi-TCA and Multi-SSTCA are evaluated on two publicly available datasets with the tasks of landmine detection and Parkinson telemonitoring. Experimental results demonstrate that Multi-TCA can improve predictive performance on previously unseen domains.

A Bayesian framework for combining protein and network topology information for predicting protein-protein interactions

Computational methods for predicting protein-protein interactions are important tools that can complement high-throughput technologies and guide biologists in designing new laboratory experiments. The proteins and the interactions between them can be described by a network which is characterized by several topological properties. Information about proteins and interactions between them, in combination with knowledge about topological properties of the network, can be used for developing computational methods that can accurately predict unknown protein-protein interactions. This paper presents a supervised learning framework based on Bayesian inference for combining two types of information: i) network topology information, and ii) information related to proteins and the interactions between them. The motivation of our model is that by combining these two types of information one can achieve a better accuracy in predicting protein-protein interactions, than by using models constructed from these two types of information independently.

A kernel-based framework for intra-fractional respiratory motion estimation in radiation therapy

In radiation therapy, tumor tracking allows to adjust the beam such that it follows the respiration-induced tumor motion. However, most clinical approaches rely on implanted fiducial markers to locate the tumor and, thus, only provide sparse information. Motion models have been investigated to estimate dense internal displacement fields from an external surrogate signal, such as range imaging. With increasing surrogate complexity, we propose a respiratory motion estimation framework based on kernel ridge regression to cope with high-dimensional domains. This approach was validated on five patient datasets, consisting of a planning 4DCT and a follow-up 4DCT for each patient. Mean residual error was at best 2.73 ± 0.25 mm, but varied greatly between patients.

Multi-Domain Transfer Component Analysis for Domain Generalization

This paper presents the domain generalization methods Multi-Domain Transfer Component Analysis (Multi-TCA) and Multi-Domain Semi-Supervised Transfer Component Analysis (Multi-SSTCA) which are extensions of the domain adaptation method Transfer Component Analysis to multiple domains. Multi-TCA learns a shared subspace by minimizing the dissimilarities across domains, while maximally preserving the data variance. The proposed methods are compared to other state-of-the-art methods on three public datasets and on a real-world case study on climate control in residential buildings. Experimental results demonstrate that Multi-TCA and Multi-SSTCA can improve predictive performance on previously unseen domains. We perform sensitivity analysis on model parameters and evaluate different kernel distances, which facilitate further improvements in predictive performance.