This paper is focused on modern approaches to machine learning, most of which are as yet used infrequently or not at all in chemoinformatics. Machine learning methods are characterized in terms of the “modes of statistical inference” and “modeling levels” nomenclature and by considering different facets of the modeling with respect to input/ouput matching, data types, models duality, and models inference. Particular attention is paid to new approaches and concepts that may provide efficient solutions of common problems in chemoinformatics: improvement of predictive performance of structure–property (activity) models, generation of structures possessing desirable properties, model applicability domain, modeling of properties with functional endpoints (e.g., phase diagrams and dose–response curves), and accounting for multiple molecular species (e.g., conformers or tautomers).
J. Chem. Inf. Model., 2012, 52 (6), pp 1413–1437

Two inductive knowledge transfer approaches - multitask learning (MTL) and Feature Net (FN) - have been used to build predictive neural networks (ASNN) and PLS models for 11 types of tissue-air partition coefficients (TAPC). Unlike conventional single-task learning (STL) modeling focused only on a single target property without any relations to other properties, in the framework of inductive transfer approach, the individual models are viewed as nodes in the network of interrelated models built in parallel (MTL) or sequentially (FN). It has been demonstrated that MTL and FN techniques are extremely useful in structure-property modeling on small and structurally diverse data sets, when conventional STL modeling is unable to produce any predictive model. The predictive STL individual models were obtained for 4 out of 11 TAPC whereas application of inductive knowledge transfer techniques resulted in models for 9 TAPC. Differences in prediction performances of the models as a function of the machine-learning method, and of the number of properties simultaneously involved in the learning, has been discussed.
J. Chem. Inf. Model., 2009, V. 49, No. 1, P. 133-144. DOI: 10.1021/ci8002914

This tutorial demonstrates performance of ensemble learning methods applied to classification and regression problems.
In: 2nd Chemoinformatics Strasbourg Summer School, Obernai, France, 2010

Last years witnessed the growth of demand for road monitoring systems based on image or video analysis. These systems usually consist of a survey vehicle equipped with photo and video cameras, laser scanners and other instruments. Sensors mounted on the van collect different types of data while the vehicle goes along the road. Recorded video can be geographically referenced with the help of global positioning systems. Road monitoring systems require special software for data processing. This paper addresses the problem of video analysis automation, and particularly the pavement monitoring functionality of such mobile laboratories. We show that computer vision methods applied to this problem help to reduce amount of manual labour during data analysis. Our method transforms video collected by mobile laboratory into rectified geo-referenced images of road pavement surface, and allows mapping of lane marking and road pavement defects with minimum user interaction. In our work the mapping workflow consists of two stages: off-line and online stage. In order to reduce user effort during error correction we take advantage of hierarchical image segmentation, which helps to delete false detections or mark missing objects with just a few clicks. Through continuous training of detection algorithm with the help of operator input error rate of automatic detection decreases; thus minimal input is required for accurate mapping. Experiments on real-world road data show effectiveness of our approach.

Системы семантической сегментации изображений применяются в различных областях. Часто в основе таких систем лежат методы машинного обучения, поэтому для настройки системы обычно требуется разметить большую базу изображений, что представляет собой довольно трудоёмкий процесс. Мы поставили перед собой цель облегчить его, сделав процесс обучения интерактивным. В данной работе предлагается использовать при разметке базы изображений классификатор, обученный на уже размеченной части данных, таким образом, вместо полной разметки изображения пользователю необходимо лишь исправить ошибки классификации. В качестве алгоритма онлайн обучения предлагается использовать разработанный ранее алгоритм On-line Costing.

В данной работе рассматривается задача онлайнового обучения с учетом разных штрафов за ошибки на разных классах. Алгоритм онлайн-бэггинг часто применяется в областях, требующих обрабатывать поступающие данные "на лету". Доказано, что классификатор, полученный с помощью онлайн-бэггинга, асимптотически сходится к классификатору, полученному с помощью стандартного бэггинга, если они обучаются на одинаково распределённых данных. Тем не менее, классический алгоритм предполагает равноценность ошибок на разных классах, что недопустимо, например, при несбалансированном распределении классов. В статье предложен подход, позволяющий устанавливать произвольные штрафы за ошибки на разных классах.

The functional annotation of amino acid sequences is one of the most important problems in bioinformatics. Different programs have been successfully applied for recognition of some functional classes; nevertheless, many functional groups still cannot be predicted with the required accuracy. We developed a new method for protein function recognition using the original approach of sequence description. Each sequence of the training set is compared with the query sequence, and the local similarity scores are calculated for the query sequence positions and used as input data for the original classifier. The method was tested using leave-one-out cross-validation for three data sets covering 58 enzyme classes. Two tested sets including noncrossing functional classes were recognized with high accuracy at various levels of classification hierarchy. The majority of these classes were predicted with 100% accuracy, showing a зrediction ability comparable with the HMMer method and an accuracy superior to the SVM-Prot program. When the tested set was composed of intersected classes of ligand specificity, the prediction accuracy was less; however, the accuracy increased as the size of the predicted class expanded. The proposed method can be used for both predicting protein functional class and selecting the functionally significant sites in a sequence.
J. Bioinform. Computat. Biol., 2008, 6 (4), 709-725.