SSML 2012 Abstracts

Full Papers Short Papers
Full Papers
Paper Nr: 1
Title:

ISSUES WITH PARTIALLY MATCHING FEATURE FUNCTIONS IN CONDITIONAL EXPONENTIAL MODELS
Authors:

Carsten Elfers, Hartmut Messerschmidt and Otthein Herzog
Abstract: Conditional Exponential Models (CEM) are effectively used in several machine learning approaches, e.g., in Conditional Random Fields. Their feature functions are typically either satisfied or not. This paper presents a way to use partially matching feature functions which are satisfied to some degree and corresponding issues while training. Using partially matching feature functions improves the inference accuracy in domains with sparse reference data and avoids overfitting. Unfortunately, the typically used Maximum Likelihood training includes some issues for using partially matching feature functions. In this context three problems (inequality of influence, unlimited weight boundaries and local optima in parameter space) with Improved Iterative Scaling (a popular training algorithm for Conditional Exponential Models) using such feature functions are stated and solved.
Download
Paper Nr: 12
Title:

PREDICTION FOR CONTROL DELAY ON REINFORCEMENT LEARNING
Authors:

Junya Saito, Kazuyuki Narisawa and Ayumi Shinohara
Abstract: This paper addresses reinforcement learning problems with constant control delay, both for known case and unknown case. First, we propose an algorithm for known delay, which is a simple extension of the model-free learning algorithm introduced by (Schuitema et al., 2010). We extend it to predict current states explicitly, and empirically show that it is more efficient than existing algorithms. Next, we consider the case that the delay is unknown but its maximum value is bounded. We propose an algorithm using accuracy of prediction of states for this case. We show that the algorithm performs as efficient as the one which knows the real delay.
Download
Short Papers
Paper Nr: 8
Title:

COMBINING GENE EXPRESSION AND CLINICAL DATA TO INCREASE PERFORMANCE OF PROGNOSTIC BREAST CANCER MODELS
Authors:

Jana Šilhavá and Pavel Smrž
Abstract: Microarray class prediction is an important application of gene expression data in biomedical research. Combining gene expression data with other relevant data may add valuable information and can generate more accurate prognostic predictions. In this paper, we combine gene expression data with clinical data. We use logistic regression models that can be built through various regularized techniques. Generalized linear models enables combining of these models with different structure of data. Our two suggested approaches are evaluated with publicly available breast cancer data sets. Based on the results, our approaches have a positive effect on prediction performances and are not computationally intensive.
Download
Paper Nr: 10
Title:

USING GENETIC ALGORITHMS WITH LEXICAL CHAINS FOR AUTOMATIC TEXT SUMMARIZATION
Authors:

Mine Berker and Tunga Güngör
Abstract: Automatic text summarization takes an input text and extracts the most important content in the text. Determining the importance depends on several factors. In this paper, we combine two different approaches that have been used in text summarization. The first one is using genetic algorithms to learn the patterns in the documents that lead to the summaries. The other one is using lexical chains as a representation of the lexical cohesion that exists in the text. We propose a novel approach that incorporates lexical chains into the model as a feature and learns the feature weights by genetic algorithms. The experiments showed that combining different types of features and also including lexical chains outperform the classical approaches.
Download