【摘要】：隨著互聯(lián)網(wǎng)的普及,網(wǎng)絡(luò)數(shù)據(jù)呈現(xiàn)爆炸式增長(zhǎng),如何從中找出有用的信息成為研究者們關(guān)注的焦點(diǎn)。作為鏈接挖掘的一個(gè)重要分支,鏈接預(yù)測(cè)通過(guò)對(duì)已知網(wǎng)絡(luò)進(jìn)行建模分析,挖掘節(jié)點(diǎn)與鏈接形成的關(guān)系,從而為人們提供有價(jià)值的信息�，F(xiàn)實(shí)網(wǎng)絡(luò)具有動(dòng)態(tài)性和稀疏性的特點(diǎn)。一方面,網(wǎng)絡(luò)是動(dòng)態(tài)發(fā)展的,隨著時(shí)間的推移,網(wǎng)絡(luò)中節(jié)點(diǎn)和鏈接的數(shù)量都在不斷更新。節(jié)點(diǎn)之間形成鏈接的時(shí)間信息通�？梢苑磻�(yīng)出節(jié)點(diǎn)之間的潛在關(guān)系。采用鏈接形成的時(shí)間屬性,對(duì)預(yù)測(cè)將來(lái)新鏈接的形成有著重要的意義。另一方面,大規(guī)模網(wǎng)絡(luò)中節(jié)點(diǎn)的數(shù)量很大,但節(jié)點(diǎn)間形成鏈接的數(shù)目卻相對(duì)較少,導(dǎo)致網(wǎng)絡(luò)異常稀疏。如果能從中選取一些具有代表性且對(duì)分類器有較大改進(jìn)的節(jié)點(diǎn)對(duì)樣本,則可以緩解訓(xùn)練壓力且保持較好的預(yù)測(cè)效果。針對(duì)以上問(wèn)題,本課題主要在以下三個(gè)方面展開(kāi)研究:1、對(duì)現(xiàn)有的鏈接預(yù)測(cè)方法進(jìn)行綜述�？偨Y(jié)了近年來(lái)一些著名的鏈接預(yù)測(cè)方法,提出了目前鏈接預(yù)測(cè)任務(wù)存在的瓶頸和挑戰(zhàn),重點(diǎn)介紹動(dòng)態(tài)網(wǎng)絡(luò)鏈接預(yù)測(cè)方法的研究現(xiàn)狀。2、針對(duì)網(wǎng)絡(luò)動(dòng)態(tài)性的特點(diǎn),提出了一種基于集成學(xué)習(xí)的動(dòng)態(tài)鏈接預(yù)測(cè)方法,稱為Dynamic。傳統(tǒng)的鏈接預(yù)測(cè)方法大多針對(duì)網(wǎng)絡(luò)的靜態(tài)結(jié)構(gòu)預(yù)測(cè)隱含的鏈接,而忽視了網(wǎng)絡(luò)在動(dòng)態(tài)演變過(guò)程中的潛在信息。本課題提出的方法使用機(jī)器學(xué)習(xí)技術(shù)對(duì)網(wǎng)絡(luò)結(jié)構(gòu)特征的變化進(jìn)行訓(xùn)練,學(xué)習(xí)每種結(jié)構(gòu)特征的變化并得到一個(gè)分類器,最終采用集成學(xué)習(xí)方法將每個(gè)分類器加權(quán)得到預(yù)測(cè)結(jié)果。3、針對(duì)網(wǎng)絡(luò)稀疏性的特點(diǎn),本課題在網(wǎng)絡(luò)進(jìn)化及鏈接預(yù)測(cè)過(guò)程中引入主動(dòng)學(xué)習(xí)范式,提出一種新的基于主動(dòng)學(xué)習(xí)的動(dòng)態(tài)鏈接預(yù)測(cè)方法,稱為DynActive。該方法為網(wǎng)絡(luò)中每個(gè)結(jié)構(gòu)特征的變化序列都生成一個(gè)分類器,再利用這些分類器對(duì)每個(gè)未連接的節(jié)點(diǎn)對(duì)進(jìn)行評(píng)分把預(yù)測(cè)結(jié)果差異較大的節(jié)點(diǎn)對(duì)樣本交由用戶判別,一旦獲取真實(shí)的標(biāo)記,系統(tǒng)采用更新的訓(xùn)練集重新訓(xùn)練各分類器并整合到最終的模型。在三個(gè)合著者網(wǎng)絡(luò)數(shù)據(jù)集的實(shí)驗(yàn)結(jié)果證明,在動(dòng)態(tài)網(wǎng)絡(luò)鏈接預(yù)測(cè)方法中引入集成學(xué)習(xí)和主動(dòng)學(xué)習(xí),AUC指標(biāo)均得到了顯著提高。
[Abstract]:With the popularity of the Internet, the network data shows explosive growth, how to find out useful information from it has become the focus of researchers' attention. As an important branch of link mining, link prediction provides valuable information for people by modeling and analyzing known networks and mining the relationship between nodes and links. The real network is dynamic and sparse. On the one hand, the network is dynamic, with the passage of time, the number of nodes and links in the network is constantly updated. The time information of the link between nodes can usually reflect the potential relationship between nodes. It is of great significance to predict the formation of new links in the future by using the time attribute of link formation. On the other hand, the number of nodes in large-scale networks is very large, but the number of links between nodes is relatively small, resulting in the network is extremely sparse. If we can select some representative node pairs which can improve the classifier greatly, the training pressure can be alleviated and the prediction effect can be maintained. In view of the above problems, this paper mainly focuses on the following three aspects: 1, the existing link prediction methods are summarized. This paper summarizes some famous link prediction methods in recent years, puts forward the bottleneck and challenge existing in the link prediction task at present, and emphatically introduces the research status of dynamic network link prediction methods. 2, aiming at the characteristics of network dynamics, This paper presents a dynamic link prediction method based on integrated learning, which is called Dynamic.. Most of the traditional link prediction methods focus on the hidden links in the static structure prediction of the network, but ignore the potential information of the network in the process of dynamic evolution. The method proposed in this paper uses machine learning technology to train the changes of network structure features, to learn the changes of each structure feature and to obtain a classifier. Finally, the ensemble learning method is used to weight each classifier to get the prediction result. 3, according to the characteristics of network sparsity, this paper introduces the active learning paradigm in the process of network evolution and link prediction. A new dynamic link prediction method based on active learning is proposed, which is called DynActive.. The proposed method generates a classifier for each sequence of structural features in the network, and then uses these classifiers to grade each unconnected pair of nodes to give the samples of node pairs whose prediction results are different from each other to be judged by the user. Once the real markers are obtained, the system retrains the classifiers with the updated training set and integrates them into the final model. The experimental results of three co-authors' network data sets show that the AUC index is significantly improved by the introduction of integrated learning and active learning into the dynamic network link prediction method.
【學(xué)位授予單位】：南京郵電大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2017
【分類號(hào)】：TP181

【參考文獻(xiàn)】

相關(guān)期刊論文 前7條

1 高維和;史玨琳;;全球城市文化資源配置力評(píng)價(jià)指標(biāo)體系研究及五大城市實(shí)證評(píng)析[J];上海經(jīng)濟(jì)研究;2015年05期

2 李玉華;肖海嶺;李棟才;李瑞軒;;基于鏈接重要性的動(dòng)態(tài)鏈接預(yù)測(cè)方法研究[J];計(jì)算機(jī)研究與發(fā)展;2011年S3期

3 許偉明;王瑞祥;曹旭明;;地源熱泵冷熱水機(jī)組性能的ARMA模型[J];北京建筑工程學(xué)院學(xué)報(bào);2011年02期

4 胡正平;高文濤;;基于改進(jìn)加權(quán)壓縮近鄰與最近邊界規(guī)則SVM訓(xùn)練樣本約減選擇算法[J];燕山大學(xué)學(xué)報(bào);2010年05期

5 竇慧麗;劉好德;吳志周;楊曉光;;基于小波分析和ARIMA模型的交通流預(yù)測(cè)方法[J];同濟(jì)大學(xué)學(xué)報(bào)(自然科學(xué)版);2009年04期

6 王義民;于興杰;暢建霞;黃強(qiáng);;基于BP神經(jīng)網(wǎng)絡(luò)馬爾科夫模型的徑流量預(yù)測(cè)[J];武漢大學(xué)學(xué)報(bào)(工學(xué)版);2008年05期

7 龍軍;殷建平;祝恩;趙文濤;;主動(dòng)學(xué)習(xí)研究綜述[J];計(jì)算機(jī)研究與發(fā)展;2008年S1期

，

本文編號(hào)：2440362