Loading...
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 53589 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Soleymani Baghshah, Mahdieh
- Abstract:
- Graphs are widely used for representing structured data and analysis of them is an important area that appears in a broad domain of applications. Graph processing is of great importance in analyzing and predicting social media users' behavior, examining financial markets, detecting malware programs, and designing recombinant drugs. For example, consider a graph in which nodes and edges show the financial institutions and the financial connection between these institutions, respectively. Financial connection refers to the investment of one institute by another. Based on the graph structure, predicting trade stability and balance is extremely significant in macro decisions.In the last few years, although the use of deep networks in image and text datasets has yielded great results in various applications such as classification, the diverse structure of graphs and the different number of their nodes and edges is a significant challenge for such networks. During these years, many efforts have been performed to provide particular layers of networks which can handle graph-structured data. Nevertheless, pooling the graph's structural information in a fixed size vector and applying it for the classification by deep networks is still the favorite subject for many researchers.In this research, a deep network has been designed to classify the graphs based on their structure. For this purpose, a new tPool layer has been proposed for pooling the representation of graph nodes in a fixed size embedding space based on the graph structure. The functionality of this layer is similar to the attention mechanism and repeatedly updates the representations. Each module of the proposed network encoder is created using three graph convolutional layers and one tPool pooling layer. Furthermore, in the preprocessing stage considering the input graph structure, the value of several different centrality measures has been calculated and considered as features for graph's nodes. Finally, the proposed model performance has been evaluated by conducting several experiments. Overall, these results indicate the superiority of the proposed method in several different scenarios
- Keywords:
- Summarization ; Deep Learning ; Machine Learning ; Graph Representation ; Graph Classification ; Graph Processing
-
محتواي کتاب
- view
- 1 مقدمه
- 1-1 پیشنیازها
- 1-2 تعریف مساله
- 1-3 ویژگیها و روشهای حل مساله
- 1-4 چالشها
- 1-4.1 دادههای آموزش
- 1-4.2 دشواری طراحی شبکهای ژرف برای ورودی گراف
- 1-4.3 دشواری آموزش با افزایش ژرفا
- 1-5 اهمیت و کاربردهای مساله
- 1-6 ساختار پایاننامه
- 2 پژوهشهای پیشین
- 2-1 روشهای مبتنی بر عاملبندی ماتریسی
- 2-1.1 نگاشت ویژه لاپلاسین گراف
- 2-1.2 شباهت راس
- 2-2 روشهای مبتنی بر هسته گرافی
- 2-2.1 هستههای R-پیچشی
- 2-2.2 گشت تصادفی
- 2-2.3 الگوهای زیردرختی
- 2-2.4 گرافلتها
- 2-2.5 Weisfeiler-Lehman (WL)
- 2-2.6 هسته گرافی تباین ژرف
- 2-3 روشهای مبتنی بر یادگیری ژرف
- 2-4 تشریح سازوکار ارسال پیام
- 2-4.1 لایه ChebConv
- 2-4.2 لایه GCN
- 2-4.3 لایه Cluster-GCN
- 2-5 تشریح سازوکار خلاصهسازی
- 2-5.1 لایه DiffPool
- 2-5.2 لایه SAGPool
- 2-6 تشریح سازوکار بهنجارسازی
- 2-6.1 لایه PairNorm
- 2-6.2 لایه DropEdge
- 2-7 جمعبندی
- 2-1 روشهای مبتنی بر عاملبندی ماتریسی
- 3 راهکار پیشنهادی
- 3-1 مقدمه
- 3-2 پیشنیازها
- 3-2.1 معیار نزدیکی
- 3-2.2 معیار بینابینی
- 3-2.3 معیار ارتباطپذیری
- 3-2.4 معیار زیرگراف
- 3-3 ساختار کلی راهکار پیشنهادی
- 3-4 تشریح عملکرد پیشپردازشگر
- 3-5 تشریح ویژگی پیمانه تعبیه راسها
- 3-6 معرفی لایه خلاصهسازی tPool
- 3-7 معماری مدل پیشنهادی
- 3-8 جمعبندی
- 4 ارزیابی
- 4-1 مجموعه دادگان
- 4-2 مدلهای مورد مقایسه
- 4-3 پیکربندی
- 4-4 آزمایشها
- 4-4.1 سنجش اثر عنصرهای مختلف شبکه پیشنهادی بر کارایی
- 4-4.2 سنجش اثر تعداد راسهای انتخابی در آخرین لایه خلاصهسازی بر کارایی
- 4-5 جمعبندی
- 5 جمعبندی و کارهای آتی
- مراجع
- واژهنامه فارسی به انگلیسی
- واژهنامه انگلیسی به فارسی