Liviu Ciortuz

Malware detection using machine learning

We propose a versatile framework in which one can employ different machine learning algorithms to successfully distinguish between malware files and clean files, while aiming to minimise the number of false positives. In this paper we present the ideas behind our framework by working firstly with cascade one-sided perceptrons and secondly with cascade kernelized one-sided perceptrons. After having been successfully tested on medium-size datasets of malware and clean files, the ideas behind this framework were submitted to a scaling-up process that enable us to work with very large datasets of malware and clean files.

Fluid Construction Grammar and Feature Constraint Logics

Fluid Construction Grammars (FCGs) are a flavor of Construction Grammars, which are themselves unification-based grammars.

Learning attribute values in typed-unification grammars: on generalised rule reduction

We present Generalised Reduction (GR), a learning technique for generalising attribute/feature values in typed-unification grammars. GR eliminates as many as possible of the feature constraints (FCs) from the type feature structures (FSs) while applying a criterion of preserving the parsing results on a given, training corpus. For parsing with GR-restricted rule FSs, and for checking the correctness of obtained parses on other corpora, one may use a new form FS unification which we call two-step unification to speed up parsing. We report results on a large-scale HPSG grammar for English.

On two classes of feature paths in large-scale unification grammars

We investigate two related techniques, Quick Check and Generalised Reduction, that contribute significantly to speeding up parsing with large-scale typed-unification grammars. The techniques take advantage of the properties of two particular classes of feature paths. Quick check is concerned with paths that most often lead to unification failure, whereas generalised reduction takes advantage of paths that do not (or only seldom) contribute to unification failure. Both sets of paths are obtained empirically by parsing a training corpus. We experiment with the two techniques, using a compilation-based parsing system on a large-scale grammar of English. The combined improvement in parsing speed we obtained is 56%.

FCGlight: A System for Studying the Evolution of Natural Language

We defined FCGlight, a refined version of the Fluid Construction Grammar (FCG), which is a formalism for studying the evolution of the natural language. We picked a core subset of FCG, and expressed it in the semantic framework of the Order-Sorted Features (OSF) logic. This allows for efficient processing, and also gives FCG a solid formal background for further analysis and improvement. Inspired from the conception of LIGHT[5], a system for natural language processing with large scale unification grammars, we developed a prototype system which implements FCGlight and can conduct language evolution experiments in a multi-agent population. We proved the functionalities of this system by running a experiment which models the evolution of the Russian verb aspect.

Using Base Pairing Probabilities for MiRNA Recognition

We designed a new SVM for microRNA identification, whose novelty consist in the fact that many of its features incorporate the base-pairing probabilities provided by McCaskills algorithm. Comparisons with other SVMs for microRNA identification prove that our SVM obtains competitive results. One of the advantages of our approach is that it makes no use of so-called normalised features which are based on sequence shuffling, which is a sensitive issue from the biological point of view. This also makes our approach much less time consuming.

A Hybrid Genetic Programming and Boosting Technique for Learning Kernel Functions from Training Data

This paper proposes a technique for learning kernel functions that can be used in non-linear SVM classification. The technique uses genetic programming to evolve kernel functions as additive or multiplicative combinations of linear, polynomial and RBF kernels, while a procedure inspired from InfoBoost helps the evolved kernels concentrate on the most difficult objects to classify. The kernels obtained at each boosting round participate in the training of non-linear SVMs which are combined, along with their confidence coefficients, into a final classifier. We compared on several data sets the performance of the kernels obtained in this manner with the performance of classic RBF kernels and of kernels evolved using a pure GP method, and we concluded that the boosted GP kernels are generally better.

MiRNA features for automated classification

We present a system for miRNA classification that implements a wide variety of miRNA features found in literature: structural, thermodynamical, information-theoretical, statistical, and comparative. A total of 1485 features are computed and various tests are performed. The classifier of choice used is Random Forests, which is also employed along with various feature selection strategies to determine the most salient features and increase automate classification performance.

Combining Old-Fashioned Computer Go with Monte Carlo Go

In this paper we discuss the idea of combining old-fashioned computer Go with Monte Carlo Go. We introduce an analyze-after approach to random simulations. We also briefly present the other features of our present Monte Carlo implementation with upper confidence trees. We then explain our approach to adding this implementation as a module in the GNU Go 3.6 engine, and finally show some preliminary results of the entire work, ideas for future work and conclusions.