Cholwich Nattee

RELATIVE POSITIONING OF STROKE-BASED CLUSTERING: A NEW APPROACH TO ONLINE HANDWRITTEN DEVANAGARI CHARACTER RECOGNITION

In this paper, we propose a new scheme for Devanagari natural handwritten character recognition. It is primarily based on spatial similarity-based stroke clustering. A feature of a stroke consists of a string of pen-tip positions and directions at every pen-tip position along the trajectory. It uses the dynamic time warping algorithm to align handwritten strokes with stored stroke templates and determine their similarity. Experiments are carried out with the help of 25 native writers and a recognition rate of approximately 95% is achieved. Our recognizer is robust to a large range of writing style and handles variation in the number of strokes, their order, shapes and sizes and similarities among classes.

A corpus-based approach for keyword identification using supervised learning techniques

This paper presents a corpus-based approach for extracting keywords from a text written in a language that has no word boundary. Based on the concept of Thai character cluster, a Thai running text is preliminarily segmented into a sequence of inseparable units, called TCCs. To enable the handling of a large-scaled text, a sorted sistring (or suffix array) is applied to calculate a number of statistics of each TCC. Using these statistics, we applied three alternative supervised machine learning techniques, naive Bayes, centroid-based and k-NN, to learn classifiers for keyword identification. Our method is evaluated using a medical text extracted from WWW. The result showed that k-NN achieves the highest performance of 79.5 % accuracy.

Spatial Similarity Based Stroke Number and Order Free Clustering

In this paper, we present an innovative approach to integrate spatial relations in stroke clustering for handwritten Devanagari character recognition. It handles strokes of any number and order, writer independently. Learnt strokes are hierarchically agglomerated via Dynamic Time Warping based on their location and their number and stored accordingly. We experimentally validate our concept by showing its ability to improve recognition performance on previously published results.

Inductive Logic Programming for Structure-Activity Relationship Studies on Large Scale Data

Inductive Logic Programming (ILP) is a combination of inductive learning and first-order logic aiming to learn first-order hypotheses from training examples. ILP has a serious bottleneck in an intractably enormous hypothesis search space. Thismakes existing approaches perform poorly on large-scale real-world datasets. In this research, we propose a technique to make the system handle an enormous search space efficiently by deriving qualitative information into search heuristics. Currently, heuristic functions used in ILP systems are based only on quantitative information, e.g. number of examples covered and length of candidates. We focus on a kind of data consisting of several parts. The approach aims to find hypotheses describing each class by using both individual and relational features of parts. The data can be found in denoting chemical compound structure for Structure-Activity Relationship studies (SAR). We apply the proposed method to extract rules describing chemical activity from their structures. The experiments are conducted on a real-world dataset. The results are compared to existing ILP methods using ten-fold cross validation.

Learning first-order rules from data with multiple parts: applications on mining chemical compound data

Inductive learning of first-order theory based on examples has serious bottleneck in the enormous hypothesis search space needed, making existing learning approaches perform poorly when compared to the propositional approach. Moreover, in order to choose the appropiate candidates, all Inductive Logic Programming (ILP) systems only use quantitive information, e.g. number of examples covered and length of rules, which is insufficient for search space having many similar candidates. This paper introduces a novel approach to improve ILP by incorporating the qualitative information into the search heuristics by focusing only on a kind of data where one instance consists of several parts, as well as relations among parts. This approach aims to find the hypothesis describing each class by using both individual and relational characteristics of parts of examples. This kind of data can be found in various domains, especially in representing chemical compound structure. Each compound is composed of atoms as parts, and bonds as relations between two atoms. We apply the proposed approach for discovering rules describing the activity of compounds from their structures from two real-world datasets: mutagenicity in nitroaromatic compounds and dopamine antagonist compounds. The results were compared to the existing method using ten-fold cross validation, and we found that the proposed method significantly produced more accurate results in prediction.

Mining chemical compound structure data using inductive logic programming

Discovering knowledge from chemical compound structure data is a challenge task in KDD. It aims to generate hypotheses describing activities or characteristics of chemical compounds from their own structures. Since each compound composes of several parts with complicated relations among them, traditional mining algorithms cannot handle this kind of data efficiently. In this research, we apply Inductive Logic Programming (ILP) for classifying chemical compounds. ILP provides comprehensibility to learning results and capability to handle more complex data consisting of their relations. Nevertheless, the bottleneck for learning first-order theory is enormous hypothesis search space which causes inefficient performance by the existing learning approaches compared to the propositional approaches. We introduces an improved ILP approach capable of handling more efficiently a kind of data called multiple-part data, i.e., one instance of data consists of several parts as well as relations among parts. The approach tries to find hypothesis describing class of each training example by using both individual and relational characteristics of its part which is similar to finding common substructures among the complex relational instances. Chemical compound data is multiple-part data. Each compound is composed of atoms as parts, and various kinds of bond as relations among atoms. We then apply the proposed algorithm for chemical compound structure by conducting experiments on two real-world datasets: mutagenicity in nitroaromatic compounds and dopamine antagonist compounds. The experiment results were compared to the previous approaches in order to show the performance of proposed approach.

Improving multiclass ILP by combining partial rules with winnow algorithm: results on classification of dopamine antagonist molecules

In this paper, we propose an approach which can improve Inductive Logic Programming in multiclass problems. This approach is based on the idea that if a whole rule cannot be applied to an example, some partial matches of the rule can be useful. The most suitable class should be the class whose important partial matches cover the example more than those from other classes. Hence, the partial matches of the rule, called partial rules, are first extracted from the original rules. Then, we utilize the idea of Winnow algorithm to weigh each partial rule. Finally, the partial rules and the weights are combined and used to classify new examples. The weights of partial rules show another aspect of the knowledge which can be discovered from the data set. In the experiments, we apply our approach to a multiclass real-world problem, classification of dopamine antagonist molecules. The experimental results show that the proposed method gives the improvement over the original rules and yields 88.58% accuracy by running 10-fold cross validation.