Akademska digitalna zbirka SLovenije - logo
VSE knjižnice (vzajemna bibliografsko-kataložna baza podatkov COBIB.SI)
  • Egy új módszer az idősorok pontosabb semi-supervised osztályzására = A new approach for more accurate semi-supervised time-series classification : TDK-dolgozat
    Marussy, Kristóf
    Time series are temporal sequences of scalar- or vector-valued measurements. Tasks which can be formalised as problems concerning time series appear in numerous domains. For example, a handwritten ... word or signature may be considered as a sequence of two dimensional vectors corresponding the positions of the pen tip on the paper in consecutive moments of time. Using a sensor apparatus worn as a glove, sign language signs can be encoded similarly. In medicine, time series are a natural representation of brainwaves (EEG) and electrocardiograph (ECG) curves. There are supervised and unsupervised problems in machine learning. In supervised tasks, the training set (input data of the learning algorithm) is augmented with class labels. Based on this training set, a model is constructed that can be used for a prediction or recognition problem afterwards. The process of constructing the model is called training. In unsupervised learning, e.g. clustering, the class labels are absent (or unused in the training stage). In the semi-supervised learning protocol, only a—usually small—fraction of the training set is labeled. The labeled instances may not represent the problem domain well, thus the structure of the unlabeled training instances must be also exploited. In this work, I propose a new semi-supervised learning method for time-series. My approach is based on the instance-based learning paradigm and hierarchical clustering. The instance-based approach was selected because of its versatility: it only depends on a pairwise dissimilarity measure between instances. As dissimilarity measure, I chose dynamic time warping (DTW), which was shown to have high accuracy and performance in time-series classification (Ding et al., 2008; Keogh and C. A. Ratanamahatana, 2005). In order to assist reproduction of my work, I evaluated the algorithm on 45 publicly available data sets (Keogh et al., 2006b) from various real-word domains. In the experiments, I compared my approach against one of the most prominent state-of-the-art time-series classifiers. The results show that my approach significantly outperformed the state-of-the-art time-series classifier in terms of classification accuracy on a large fraction of the data sets To further support reproducibility, I will publish the Java-code of my algorithm.
    Vrsta gradiva - diplomsko delo ; neleposlovje za odrasle
    Založništvo in izdelava - Budapest : [K. Marussy], 2012
    Jezik - angleški
    COBISS.SI-ID - 159793411

Nobena knjižnica v sistemu COBISS.SI nima izvoda tega gradiva.
loading ...
loading ...
loading ...

Vnos na polico