Semi-Supervised Threshold Queries on Pharmacogenomics Time Sequences

Abstract

The analysis of time series data is of capital importance for pharmacogenomics since the experimental evaluations are usually based on observations of time dependent reactions or behaviors of organisms. Thus, data mining in time series databases is an important instrument towards understanding the effects of drugs on individuals. However, the complex nature of time series poses a big challenge for effective and efficient data mining. In this paper, we focus on the detection of temporal dependencies between different time series: we introduce the novel analysis concept of threshold queries and its semi-supervised extension which supports the parameter setting by applying training datasets. Basically, threshold queries report those time series exceeding an user-defined query threshold at certain time frames. For semi-supervised threshold queries the corresponding threshold is automatically adjusted to the characteristics of the data set, the training dataset, respectively. In order to support threshold queries efficiently, we present a new efficient access method which uses the fact that only partial information of the time series is required at query time. In an extensive experimental evaluation we demonstrate the performance of our solution and show that semi-supervised threshold queries applied to gene expression data are very worthwhile.

Time Series Analysis Using the Concept of Adaptable Threshold Similarity

“…Our approach is different to that proposed in [4] in the following aspect. We propose a novel approach to compute the separation score for a given threshold τ using the concept of support vector machines that provides an optimal solution for the task of learning the best suitable threshold for the succeeding clustering step rather than using simple heuristics.…”

“…We will review and extend the basic definitions in Section 3. The only method to adopt an "optimal" threshold for threshold similarity from a small training set so far has been proposed in [4]. If the training set contains instances of m ≥ 2 classes C i , 1 ≤ i ≤ m, the best threshold of a query for a specific class C i is evaluated by computing the pairwise silhouette width of C i to all other classes C j , i = j.…”

“…In addition, we show how our new concepts can be integrated into a complete framework for semi-supervised cluster analysis of time series databases using the concept of threshold similarity. Last but not least, we present an experimental evaluation where we show the superiority of our approach in comparison to [4] by evaluating the quality of the learned threshold using the entire database objects rather than some sample queries (as in [4]). We further demonstrate in our experiments that the concept of threshold similarity is better suitable in some applications than using the Euclidean distance measure and that our method is able to gain new insights from a database even in the case that the training set is incomplete, i.e.…”

“…In fact, we measure the separation score of a broad range of possible thresholds. One key difference of our approach to [4] is the computation of this separation score. In [4], the separability is measured using the minimum of all pairwise silhouette widths of each pair of classes C i and C j in T .…”

“…adapted threshold) is performed in order to detect novel and important patterns. So far, there is only one approach to "optimally" adapt the threshold for a threshold similarity analysis of time series [4]. However, this approach uses a very simple and -in most casesnot very accurate method to judge the quality of a given threshold for the analysis task.…”