مقایسه عملکرد شبکه های عصبی خود سازمانده مکانی و غیرمکانی در خوشه بندی داده های اجتماعی-اقتصادی بلوک های آماری شهر اصفهان (مقاله علمی وزارت علوم)
درجه علمی: نشریه علمی (وزارت علوم)
آرشیو
چکیده
امروزه با افزایش حجم و ابعاد داده های مکانی و نیاز به درک کامل داده ها، شبکه های عصبی خودسازمانده به ابزاری استاندارد برای کار با داده های بزرگ و چند بُعدی تبدیل شده اند که می توانند در خوشه بندی، بصری سازی و انتقال داده های چند بُعدی در فضایی با ابعاد کمتر استفاده شوند. هدف از پژوهش حاضر، خوشه بندی داده های بلوک های آماری (شامل ده متغیر منتخب اجتماعی-اقتصادی مرتبط با رویکرد توسعه پایدار شهری) با شبکه های عصبی خودسازمانده بدون استفاده از پارامترهای مکانی و به کارگیری مختصات جغرافیایی بلوک های آماری به عنوان پارامتر مکانی در روند خوشه بندی و مقایسه نتایج حاصل شده، است. الگوریتم SOM رایج ترین شبکه عصبی خودسازمانده و الگوریتم Geo-SOMمکانی شده الگوریتم SOM است. روند خوشه بندی هر دو الگوریتم یکسان است و تنها تفاوت این دو الگوریتم به کا رگیری پارامترهای مکانی در روند اجرای الگوریتم Geo-SOM است. در پژوهش حاضر داده ها با الگوریتم SOM و Geo-SOM خوشه بندی شده است. نتایج نشان داد که خوشه های حاصل از دو الگوریتم به طور کامل، متفاوت است. خوشه بندی بلوک های آماری بدون توجه به خصوصیات مکانی و تنها با استفاده از معیار شباهت، منجر به خوشه های ناهمگن می شود و برعکس. با اعمال پارامترهای مکانی نه تنها از معیار شباهت، از ویژگی های مکانی داده ها نیز در فرآیند خوشه بندی استفاده می شود که این مسئله منجر به تولید خوشه های همگن می شود. ارزیابی نتایج با استفاده از ضریب سیلهوته بیانگر خوشه بندی مناسب تر الگوریتم Geo-SOM است؛ به طوری که میانگین ضریب سیلهونه برای الگوریتم SOM برابر 02/0- و برای الگوریتم Geo-SOM برابر 27/ 0است. مقایسه نتایج نشان دهنده تأثیر مثبت پارامترهای مکانی در خوشه بندی داده های اجتماعی و اقتصادی است.Comparing the Performance of Spatial and Non-Spatial Self-Organizing Neural Networks in Clustering Socio-Economic Data of Isfahan Census Blocks
The increasing volume and dimensions of spatial data have made self-organizing neural networks a prominent tool for analyzing large and multi-dimensional datasets. Clustering, an approach for extracting knowledge from big data, aims to group similar data into clusters. This research focused on clustering socio-economic data of census blocks associated with urban sustainable development using self-organizing neural networks with and without spatial parameters referred to as SOM and Geo-SOM, respectively. Both algorithms employ the same clustering process but differ in the inclusion of spatial parameters, specifically the geographic coordinates of block centroids, in the Geo-SOM algorithm. The SOM and Geo-SOM algorithms were trained and applied to cluster the data. The resulting clusters exhibited distinct dissimilarities, demonstrating that clustering census block data solely based on non-spatial attributes leads to heterogeneous and incongruent clusters, whereas incorporating spatial parameters yields homogeneous and congruent clusters. Evaluation of the results using Silhouette coefficient indicated that Geo-SOM outperformed SOM in clustering the data with average Silhouette coefficients of -0.02 and 0.27 for SOM and Geo-SOM, respectively. Comparison of the outcomes highlighted the positive impact of incorporating spatial parameters on clustering socio-economic data.Keywords: Self-organizing neural networks, SOM, Geo-SOM, Clustering, Spatial data, Census blocks, Isfahan CityIntroductionIn recent years, there has been a significant increase in the volume of spatial data available. To gain a comprehensive understanding of spatial data, it is crucial to extract meaningful knowledge from it by considering its unique characteristics. This process, known as "Knowledge Discovery in Databases" (KDDs), utilizes methods, such as Artificial Neural Networks (ANNs) to extract useful information and knowledge. Clustering, a widely used technique for large-scale data analysis, aims to group similar data into clusters while reducing the data size and ensuring that the internal differences within clusters are significantly smaller than the differences between clusters. However, clustering algorithms for spatial data differ from those for non-spatial data. Therefore, this research focused on using self-organizing neural networks to cluster socio-economic data related to sustainable urban development while comparing the performance of two algorithms, SOM and Geo-SOM, for this task. MethodologyThis research utilized socio-economic data of Isfahan census blocks (2015) obtained from Iranian Statistics Center. The dataset consisted of 13,362 statistical blocks, each characterized by socio-economic variables. The SOM and Geo-SOM algorithms were employed to cluster the socio-economic data of census blocks and their results were compared. The self-organizing map is a single-layer feed-forward neural network that organizes output neurons in a low-dimensional topological structure. The SOM effectively maps high-dimensional data onto two-dimensional feature maps, forming networks of units or neurons. On the other hand, the Geo-SOM algorithm is an adaptation of the self-organizing map that incorporates a competitive learning process, focusing more on spatial aspects by restricting the search for the Best Matching Unit (BMU) to the geographical proximity of input patterns. Visualization tools, such as the U-matrix, Component Plane, and Hits Maps are employed to analyze the output of SOM and Geo-SOM algorithms. The network topology in this study was set to 7 x 7 after assessing multiple U-matrix sizes to identify clusters, resulting in a network with 49 neurons. Evaluation of the clustering results was performed using Silhouette coefficient, which measured similarity of the data within clusters and their dissimilarity between clusters. Research Findings The results obtained from this study revealed a fundamental distinction between the SOM and Geo-SOM algorithms, primarily stemming from their respective approaches to mapping input vectors onto network neurons, which in turn led to disparate clustering outcomes. Analysis of the component planes for both algorithms demonstrated notable variations in the relative weight assigned to each variable during the mapping process. Specifically, within the SOM algorithm, higher values of the variable of married population ratio tended to be associated with neurons corresponding to Cluster No. 4, while lower values exhibited a preference for neurons corresponding to Cluster No. 1. In contrast, the Geo-SOM algorithm assigned higher values of this variable to Clusters No. 8 and 9, with lower values predominantly allocated to Clusters No. 2 and 10.The SOM algorithm yielded intertwined, heterogeneous, and scattered clusters when applied to the census block data. In contrast, the Geo-SOM algorithm achieved the classification of census block data into homogeneous and congruous clusters through the utilization of spatial parameters. Notably, the Geo-SOM algorithm exhibited a higher clustering quality as indicated by its average Silhouette coefficient of 0.27, surpassing the SOM algorithm's average Silhouette coefficient of -0.02. Discussion of Results & ConclusionThis research employed two self-organizing map algorithms, SOM and Geo-SOM, to cluster census block data with the incorporation of spatial parameters. By integrating the geographic coordinates of census block centroids with other non-spatial inputs, we observed the positive impact of spatial parameters in the clustering process within the Geo-SOM algorithm. Comparatively, Geo-SOM exhibited superior performance, yielding cohesive and meaningful clusters of statistical blocks based on their similarities and spatial characteristics. The findings underscored the potential of the Geo-SOM algorithm in delineating homogeneous regions and identifying similar blocks within real-world datasets. We can conclude that utilization of this algorithm holds practical value for facilitating urban planning endeavors as it enables identification of homogeneous areas aligned with sustainable urban development principles and selected variables.
تبلیغات
