خوشه بندي گروهي و وزن دهي به ويژگي در داده سري زماني

A time series is a collection of data observations obtained by repeating measurements over time. Time series clustering is the act of organizing time points into clusters based on similarity an‎d dissimilarity criteria. Clustering is an unsupervised learning approach that is applied to maximize the similarity of data within a cluster an‎d minimize the similarity of data between different clusters. In machine learning, an ensemble is a machine learning system consisting of a set of separate models. The output of these models is combined using a decision-making strategy to produce an optimal solution for the problem. A clustering ensemble is formed to combine the clustering models into a single model to achieve higher efficiency. In this research, frameworks are designed for center-based enmseble clustering an‎d deep learning-based ensemble clustering. They use specific methods an‎d algorithms to analyze features in annual time series data. These frameworks aim to cluster data points from different years in a unified way. First, the original time series data are converted into features, an‎d then the new feature space is partitioned to reduce the computational load an‎d facilitate processing. Center-based clustering frameworks leverage feature weighting algorithms to increase the influence of important features in clustering. Deep learning based clustering framework is proposed to perform clustering based on Self-Organizing Map (SOM) neural network. This framework applies shape-based an‎d feature-based clustering approaches to different views of data. In the shape-based clustering approach, the Dynamic Time Warping similarity measure is combined with SOM to accelerate the convergence process of partitions, an‎d in the feature-based clustering approach, feature extraction techniques are used to han‎dle the correlation among features. In addition, prediction models compatible with the proposed frameworks are presented. They use the obtained clustering results to predict the class label of new observations. eva‎luations are performed using historical weather datasets from Open ML an‎d Kaggle platforms.Temperature an‎d humidity data are partitioned an‎d generated feature space is eva‎luated. Homogeneity, Completeness, an‎d V-measure metrics are used to compute the clustering accuracy of the frameworks. In order to measure the efficiency of the proposed models in terms of the classification accuracy, the partitioned feature space is splitted into train an‎d test sets. The results indicate the ability of the proposed frameworks to generate stable an‎d high-quality clusters.

مسعودرضا هاشمي , زينب زالي

مازيار پالهنگ , زينب مالكي , حسين كارشناس نجف آبادي