Research on node ranking in peer - To - peer networks

Lời cảm ơn Lời đầu tiên em xin bày tỏ lòng biết ơn sâu sắc tới TS. Nguyễn Hoài Sơn, các thầy đã hướng dẫn và là nguồn cảm hứng cho quá trình nghiên cứu của em. Em xin bày tỏ lòng biết ơn tới các thầy, cô giáo trong Khoa Công nghệ thông tin - Trường Đại học Công nghệ - ĐHQGHN. Các thầy cô đã dạy bảo, chỉ dẫn chúng em và luôn tạo điều kiện tốt nhất cho chúng em học tập trong suốt quá trình học đại học đặc biệt là trong thời gian làm khoá luận tốt nghiệp. Chapter 1 Table of Contents Abstract Error! Bookmark not defined. List images . 5 List tables . 7 Chapter 1: Peer to Peer and Ranking Problem 5 1.1. Peer to Peer . 5 1.1.1. Peer to Peer overview . 5 1.1.2. Architecture of Peer to Peer Systems Error! Bookmark not defined.7 1.1.3. Distributed hash tables . 8 1.2. Ranking in Peer to Peer networks . 9 1.2.1. Introduction Error! Bookmark not defined. 1.2.2. Ranking Roles Error! Bookmark not defined. 1.2.3. Research’s important objects . Error! Bookmark not defined. Chapter 2: Ranking on DHT Peer to Peer Networks . 11 2.1. Chord Protocol 11 2.2. Pagerank 12 2.2.1. Description . 12 2.2.2. Algorithms . 13 2.3. Distributed Computing . 17 2.2.1. Introduction 17 2.2.2. Algorithms . Error! Bookmark not defined. 2.4 if-idf . 18 Chapter 3: Building a new algorithm for ranking in chord networksError! Bookmark not define 3.1. Targets and Missions of Research Error! Bookmark not defined. 3.2. Idea Error! Bookmark not defined. iii Research on Node Ranking – Peer to Peer Hoàng Cường 3.2.1. Major problems to exploit . Error! Bookmark not defined. 3.2.2. Ranking Idea Error! Bookmark not defined. Chapter 4: Ranking on Details Error! Bookmark not defined. 4.1. Ranking algorithm Error! Bookmark not defined. 4.2. Ranking’s features Error! Bookmark not defined. Chapter 5: Evaluation . 50 Chapter 6: Related Work 52 Chapter 7: Contributions and future work 53 References . 54

pdf59 trang | Chia sẻ: banmai | Lượt xem: 2140 | Lượt tải: 0download
Bạn đang xem trước 20 trang tài liệu Research on node ranking in peer - To - peer networks, để xem tài liệu hoàn chỉnh bạn click vào nút DOWNLOAD ở trên
The computer program which runs in the distributional system said that a distributed program, the distribution programming writes such program' s process. And the distributed computing mentions the use distributional system explanation estimate question. In the distributed computing, the question is divided many responsibilities, the computer explains everybody. 20 Research on Node Ranking – Peer to Peer …. Hoàng Cường Image 2.3: Distributed Nodes Graph example We pass use computer’s hope automation; s many responsibilities held responsible with answer the type: We hope to ask the question, and the computer should cause the answer. In the computer science theoretically, is called the estimate question like this voluntarily. It is estimated that the question has each template including the instance is an explanation officially together. The example is the question which we asked that and the explanation is anticipates the answer to these questions. (How does the theory computer science seek needs to understand the estimate question possibly through use that the complex theory solution computer (the computability theory) and high efficiency computation). In the tradition, said the question perhaps through the use solution computer, if perhaps we design all concrete instances are correct explanation algorithm causes. Perhaps such algorithm possibly implements the computer program which runs in an general calculator: Studies from the input question instance's holiday eye, carries out some computation, and causes the explanation to adopt the product. Formalism for example random access ' perhaps the s machine or the universal Turing machine use the achievement to carry out such algorithm continuously general calculator' s abstraction model. In many computer situations, consistent and distributed computing area research similar question or execution interaction process system computer: Which estimate question how can solve in such network and the high efficiency place? However, it is not obvious in concurrent or the distributional system situation, “solves the problem is all meanings” 2.4 Computing PageRank in a distributed system 21 Research on Node Ranking – Peer to Peer …. Hoàng Cường Lectured the net graph in distribution system's recent research work to divide into messes up the website or the domain case. The net is molded takes many messes up the network server. Is divided in net's ultra link two categories, the internal cut-off link and the mutual server link. The internal server link is between the page link in the server, and these links use in calculating on each server's place PageRank intermediate vector. The mutual server link is between the page link with the different server, and they use in calculating ServerRank. ServerRank surveys the different network server's relative importance. The server which submits is being merged finally from many network server's result causes an arrangement ultra link name list. The ranking algebra proposed that deals with the ranking in the different granularity level, is utilized possibly also in gathering the place ranking and the stand ranking obtains the global ranking. Has in one disperses the system fully in the PageRank approximation work, each of the same generation is autonomous, and perhaps of the same generation mutually overlaps. Was proposing the JXP algorithm, each of the same generation calculates the place PageRank score, then meets other of the same generations and increases it gradually through the exchange information willfully about the global net graph knowledge, then recomputation in place of the same generation's PageRank score. This conference and the recomputation process is duplicated, collects the enough information until of the same generation. If of the same generation meets the sufficient number of times exchange information finally, JXP score polymerization to the real global PageRank score. Supposes is each page of out degree in global graph awareness. However, these operations are providing the approximation the focal point are the global graph, in centralized system or distribution system. 2.5. tf-idf The tf–idf weight (term frequency–inverse document frequency) is a weight often used in information retrieval and text mining. This weight is a statistical measure used to evaluate how important a word is to a document in a collection or corpus. The importance increases proportionally to the number of times a word appears in the document but is offset by the frequency of the word in the corpus. Variations of the tf– idf weighting scheme are often used by search engines as a central tool in scoring and ranking a document's relevance given a user query. One of the simplest ranking functions is estimated by summing the tf-idf for each query term; many more sophisticated ranking functions are variants of this simple model. Motivation Suppose we have a set of English text documents and wish to determine which document is most relevant to the query "the brown cow." A simple way to start out is by eliminating documents that do not contain all three words "the," "brown," and "cow," but this still leaves many documents. To further distinguish them, we might count the number of times each term occurs in each document and sum them all together; the number of times a term occurs in a document is called its term frequency. 22 Research on Node Ranking – Peer to Peer …. Hoàng Cường However, because the term "the" is so common, this will tend to incorrectly emphasize documents which happen to use the word "the" more, without giving enough weight to the more meaningful terms "brown" and "cow". Also the term "the" is not a good keyword to distinguish relevant and non-relevant documents and terms like "brown" and "cow" that occur rarely are good keywords to distinguish relevant documents from the non-relevant documents. Hence an inverse document frequency factor is incorporated which diminishes the weight of terms that occur very frequently in the collection and increases the weight of terms that occur rarely. Mathematical details The term count in the given document is simply the number of times a given term appears in that document. This count is usually normalized to prevent a bias towards longer documents (which may have a higher term count regardless of the actual importance of that term in the document) to give a measure of the importance of the term ti within the particular document dj. Thus we have the term frequency, defined as follows. where ni,j is the number of occurrences of the considered term (ti) in document dj, and the denominator is the sum of number of occurrences of all terms in document dj. The inverse document frequency is a measure of the general importance of the term (obtained by dividing the total number of documents by the number of documents containing the term, and then taking the logarithm of that quotient). with • | D | : total number of documents in the corpus • : number of documents where the term ti appears (that is ). If the term is not in the corpus, this will lead to a division-by-zero. It is therefore common to use Then A high weight in tf–idf is reached by a high term frequency (in the given document) and a low document frequency of the term in the whole collection of documents; the weights hence tend to filter out common terms. The tf-idf value for a term will always be greater than or equal to zero. 23 Research on Node Ranking – Peer to Peer …. Hoàng Cường Example Consider a document containing 100 words wherein the word cow appears 3 times. Following the previously defined formulas, the term frequency (TF) for cow is then 0.03 (3 / 100). Now, assume we have 10 million documents and cow appears in one thousand of these. Then, the inverse document frequency is calculated as log(10 000 000 / 1 000) = 4. The TF-IDF score is the product of these quantities: 0.03 × 4 = 0.12. Chapter 3: Building a new algorithm for ranking nodes in chord networks 3.1 Targets and Missions of Research The P2P deployment is the building distributed search network. Proposed that the system support discovers with retrieval all results, but lacks the essential information to arrange them. User, however, is mainly to most is related and is not all most possible results to be interested. The use random sampling, we expand the well-known information retrieval ranking algorithm class they may apply like this in this distributed establishment. How do we analyze the ceiling our method, and the quota our system scale along with the document, the system size, the inquiry document to ties the mapping correctly (uniform to non-uniform) and the type increases the digit (rarely to universal deadline). Our analysis and the imitation indicated that a) these extensions are the high efficiency, and possibly calls with a ceiling to the large-scale system, and b) uses the result accuracy which and the centralized implementation the distributed ranking obtains is comparable. 3.2 Idea: 24 Research on Node Ranking – Peer to Peer …. Hoàng Cường Table 3.2.1: The Pagerank converge and HITS converge When I write a small code, I see that’s the converges of two algorithms: HITS- Pagerank. HITS’s better converge more than Pagerank. It’s can be see as the experiment’s image.. experiment test with Graph 1000 nodes. The blue line is the converge ( iterators ) of Pagerank The red line is the converge ( iterators ) of HITS Table 3.2.2: The Pagerank converge increasing to fast 25 Research on Node Ranking – Peer to Peer …. Hoàng Cường I also see that’s the converges of the algorithm: Pagerank. It’s can be see as the experiment’s image that’s the time to calculate converge of Pagerank algorithm. It goes to much faster. It can’t be done in peer to peer network ( the fast increase of time is more than the fast increase of graph’s size a lot ) The When I write a small code, I see that’s the converges of two algorithms: HITS-Pagerank. HITS’s better converge more than Pagerank. It’s can be see as the experiment’s image.. The Graph with 1000 nodes. The blue line is the converge ( iterators ) of Pagerank The red line is the converge ( iterators ) of HITS Image taken from the experimental results, the convergence of the Pagerank algorithm. 1000 nodes network simulation, 2000 nodes, 4000 nodes. (Execution time calculation) Light blue line is the only way to calculate the time convergence of the Pagerank algorithm is at the network node 4000. Dark blue line is the only way to calculate the time convergence of the Pagerank algorithm is at the network node 2000. The red dot is the only way to calculate the time convergence of the Pagerank algorithm is at the network node 1000. Computing time increases when the number of exponential increase network node Want to calculate all the nodes in the network takes several performance (larger network node -> lose greater efficiency (greater than performance node added) Table 3.2.3: Pagerank convergence are not steady when Epsilon small Image taken from the experimental results of the convergence Pagerank algorithm. Network Simulation 1000 node. (Execution time calculation) 26 Research on Node Ranking – Peer to Peer …. Hoàng Cường Dark blue line includes the red dot is the only way to calculate the time convergence of the Pagerank algorithm is at the network node 1000 under the different conditions randomly. With a little error Epsilon, convergence of k Pagerank is completely stable. (Change gap) 3.2.4: HITS convergence ( steady+ take lots of time more than Pagerank) Image taken from the experimental results the convergence of HITS algorithm. Network simulation is 1000 nodes. (Terms Iterator going to converge). Dark blue line includes the red dot is the only way to calculate the time convergence of HITS algorithm is at 1000 node network under the different conditions randomly. With a little error Epsilon, convergence of HITS stable than PageRank. Idea: Combined HITS and Pagerank. HITS method n nodes filter out content authentication best results) and used to calculate the Pagerank that n nodes. (n very little compared to the total number of network nodes) Advantages of this new approach: • Exact result • Accurate • Computing easier. • Easy feasible • faster Let’s going to see what’s happened to get a system which has some features like that in deeply later. Basically, let’s analyze a simple example in Google search engine: 27 Research on Node Ranking – Peer to Peer …. Hoàng Cường Image 3.2: Google almost is not exact In the results we can see that the topic results. They’re different, and no relate. And so some are exact, but some results are not exactly ( many ). 3.2 Ranking Idea: Customized semantic query answering, personalized search (Google is the first of the crucial search engines to take personalized results on a massive scale. Weighing a number of factors including but not limited to user history, bookmarks, community behavior and site click-through rate and stickiness, Google is providing results that are specific to what they believe you are searching for. Currently this service is only available to those who are logged into their Google account), focused crawlers and localized search engines frequently focus on ranking the Nodes contained within a sub-graph of the global Peer to Peer Nodes graph. The claim for these utilizations is to estimate PageRank-style scores expeditiously on the sub-graph, i.e., the ranking must manifest the global link structure of the Peer to Peer Node graph exactly ( which means calculating returns true global pagerank-scored order ) but it must do so without bearing the high overhead affiliated to global computation. We propose a framework of an exact unraveling and an analogous solution for computing ranking on a sub- graph. without running PageRank on the whole Nodes Graph. So, Approaching sub- graph and multiple sub-graphs respectively is the main discovery in this ranking research. 28 Research on Node Ranking – Peer to Peer …. Hoàng Cường Image 3.2.2: Intersect idea The Rank_local_idea algorithm is an rigorous scientific method with the supposing that the scores of “outside” Nodes are known. We analyse that the Rank_local_idea values for nodes in the sub-graph converge. Since the PageRank- style values of out nodes may not expectedly be achievable, we introduce the Pagerank_local algorithm to evaluate score results for the sub-graph. We use graph EG ( Outside Graph ) to symbolize a major graph. Both Rank_local_idea and Pagerank_local put on the set of outside nodes with an outside node graph EG and magnify the sub-graph with links to graph EG. They also modify the PageRank-style transition matrix ( The transition possibility distribution be symbolizeed by a matrix ) with respect to graph EG. We analyze the L1 distance between Rank_local_idea scores and Pagerank_local scores of the sub-graph and show that it is within a constant factor of the L1 distance of the outside nodes (e.g., the true PageRank scores and uniform scores assumed by Pagerank_local). We compare Pagerank_local and a stochastic complementation approach (SC), a current best solution for this problem, on different types of sub-graphs. Pagerank_local has similar or superior performance to SC and typically improves on the runtime performance of SC by an order of magnitude or better. We demonstrate that Pagerank_local provides a good approximation to PageRank for a variety of sub-graphs. 3.3 Idea finding a subset Many assignments have to find salient and diverse items to satisfy the user’s information need. This problem has been addressed in document summarization, information retrieval, and various data mining applications. In this paper, we present a new method that can select salient and diverse items from many information contents. We formulate the mission as a graph summarization problem: given a weighted graph, how can we find a subset of salient and diverse vertices to summarize the graph, subject to some pre-specified constraints. We present a linear programming based approximate algorithm to optimize an objective function which takes into account both salience and diversity. The method was applied to two missions: 29 Research on Node Ranking – Peer to Peer …. Hoàng Cường • multi-document summarization, which brings out salient sentences from sentence graphs; • mining symbolizeative experts in the data-mining community from co-author graphs. In comparison to the state-of-the-art graphed based methods, our method produces superior results in these applications. 3.4: Finding ranking factors Image 3.4: Factor Percent 30 Research on Node Ranking – Peer to Peer …. Hoàng Cường Chapter 4: Ranking algorithm details 4 .1: Idea The explosion of files sharing available on the Peer to Peer Networks has made the ranking of peer to peer systems an high-priced but unavoidable component. The more users try to search, the more system overloads. Since hyper-bandwidth-links from one node to another usually points to an “Authorization” or “Commendation”, bandwidth-link analysis plays a essential role in elect the “importance” of nodes in network. PageRank and HITS are two foremost approaches in this area. PageRank iteratively estimates the score of a page based on the scores of its parent Image 4.1: Bandwidth is the key of ranking trusted Data as above equation 31 Research on Node Ranking – Peer to Peer …. Hoàng Cường HITS separates the role of each node into a hub or authority. In the HITS algorithm, the first step is to retrieve the set of results to the search query. The computation is performed only on this result set, not across all Graph Nodes. Authority and hub values are defined in terms of one another in a mutual recursion. An authority value is computed as the sum of the scaled hub values that point to that page. A hub value is the sum of the scaled authority values of the pages it points to. Some implementations also consider the relevance of the linked nodes. The algorithm performs a series of iterations, each consisting of two basic steps: • Authority Update: Update each node's Authority score to be equal to the sum of the Hub Scores of each node that points to it. That is, a node is given a high authority score by being linked to by nodes that are recognized as Hubs for information. • Hub Update: Update each node's Hub Score to be equal to the sum of the Authority Scores of each node that it points to. That is, a node is given a high hub score by linking to nodes that are considered to be authorities on the subject. The hub score estimates the value of its bandwidth-links to other nodes and the authority score estimates the importance of the node. These algorithms are expensive because of the number of node data/objects involved in the computation. On 15 November 2008, The Pirate Bay announced that it had reached over 25 million unique peers. And according to the Pirate Bay statistics, as of December 2009, The Pirate Bay has over 4 million registered users, although registration is not necessary to download torrents. To make ranking feasible, and to manifest the diversity of node users' information needs, peer to peer networking applications such as semantic search ("Semantic search seeks to improve search accuracy by understanding searcher intent and the contextual meaning of terms as they appear in the searchable dataspace, whether on the Web or within a closed system, to generate more relevant results. Author Seth Grimes lists "11 approaches that join semantics to search", and Hildebrand et al. provide an overview that lists semantic search systems and identifies other uses of semantics in the search process.), focused crawlers (A focused crawler or topical crawler is a web crawler that attempts to download only web data that are relevant to a pre-defined topic or set of topics. Topical crawling generally assumes that only the topic is given, while focused crawling also assumes that some labeled examples of relevant and not relevant data are available. ), localized search engines, and personalized search have emerged. They all have a common objective to rank a sub-graph. The first intriguing application is a focused crawler, also called a topical crawler. A focused crawler is interested in collecting a subset of the node data that are related to a specific topic. Compared to a standard crawler which can easily get lost and waste resources, a focused crawler acquires relevant data using a Best First Search; it selects links based on their scores. In contrast to focused crawlers which are 32 Research on Node Ranking – Peer to Peer …. Hoàng Cường topic specific, a localized search engine indexes a subset of node data that are within a specific domain. The node piece retrieved by the focused crawler (or localized search engine) is a sub-graph of the global node graph. Only PageRank scores for local data in the sub-graph are of interest to users. Users submit queries to the sub-graph collected by a focused crawler and local query answers are returned to the user. The ranking on this local graph, however, should reflect the link structure of all node data. Another interesting scenario is semantic ranking. ObjectRank creates a schema graph to model the semantic connections between entity sets, e.g., authors or conferences. The semantic connections are associated with an authority transfer assignment which can be arbitrarily set by a domain expert based on her interpretation of the domain. Figure 2 shows an authority transfer schema graph for DBLP. Building on previous work on how to model contextual information for desktop search and how to implement semantically rich information exchange in social networks. Peer-Sensitive ObjectRank for ranking resources on the desktop. The algorithm takes into account different trust values for each peer, generalizing previous biasing PageRank algorithms. The ObjectRank system applies the random walk model, the effectiveness of which is proven by Google's PageRank, to keyword search in databases modeled as labeled graphs. The system ranks the database objects with respect to the user- provided keywords. The PageRank technique assigns to each page p a score based on the score of the pages pointing to p. Hence, pages pointed by many high-score pages receive a high score as well. Alternatively, the score of p is equal to the possibility that a random surfer, starting from a random page, will be at p at a specific time. For more information on PageRank see the original PageRank paper. While ObjectRank is flexible and allows the tuning of ObjectRank scores by a domain expert, it leads to computational challenges if a search engine has to consider all possible combinations of keywords and authority transfer assignments. Recent research on reformulating ObjectRank scores based on individual user feedback and a graph exploration framework for the biological node highlights the optimization challenges of query answering and ranking for the semantic Node. If we can model a sub-graph to contain the subset of data associated with the entity sets of interest to some domain expert, we can then define the ObjectRank problem as a problem of ranking a sub-graph. This problem, too, is to exploit existing PageRank scores for other regions of the graph that are not of interest to the domain expert, and whose scores may also remain largely unchanged. Figure 3 shows an example where a sub-graph is associated with an authority transfer assignment and the outside Node are beyond the focus of the domain expert. 33 Research on Node Ranking – Peer to Peer …. Hoàng Cường Another application that involves ranking a sub-graph is peer-to-peer networks. The advent of peer-to-peer(P2P) technology has further encouraged data information retrieval by jump on distributed computing power, storage, and connectivity. A distributed or decentralized system has multiple peers or servers, each of which puts in storage its own sub-graph of the Peer to Peer nodes. A user may take queries on one peer and ranked query answers that are available locally are presented to the user. The ranking relies upon the context of the query. A final scenario is a absorption of the constant change of the Peer to Peer nodes - Peer to Peer node data. The ranking of data needs to be updated frequently, especially for the sub-graph of the nodes that experiences the most change. This sub- graph can be either a set of dangling files that crawlers have not as yet crawled, referred to as the web “frontier”, or the set of data that are most affected by updates. It is desirable that any strategy to update the ranking of this sub-graph exploits existing PageRank scores for other regions of the graph which may remain largely unchanged. In response to these many motivating applications, we address the problem of computing ranking scores for a sub-graph. For ease of presentation, and to compare with existing approaches, we use the PageRank metric for explanation and experiments. However, our general approaches can be applied to estimate ObjectRank scores as well. We note that current ranking techniques (to be discussed in the next section) either bear the cost of a global computation to get an accurate ranking, or they have to solve another potentially difficult problem: to determine a relevant super-graph of web data that impact the rank of the sub-graph. Our challenge is to obtain an accurate ranking that reflects the global link structure of the Web graph and to do so without bearing the high overhead associated with a global PageRank computation or having to solve the difficult problem of identifying a relevant supergraph. We would also like to exploit pre-estimated 34 Research on Node Ranking – Peer to Peer …. Hoàng Cường PageRank scores for outside data if and when they are available and appropriate for use. We propose a framework based on an exact and an approximate solution to estimate PageRank on a sub-graph. The Rank_local_idea algorithm is an exact solution. It assumes that the PageRank scores of outside data are known. We prove that the Rank_local_idea scores for data in the sub-graph converge to the true PageRank scores. Since the PageRank scores of outside data may not typically be available, we propose the Pagerank_local algorithm to estimate PageRank scores for the sub-graph. Both Rank_local_idea and Pagerank_local symbolize the set of outside data with an outside node graph EG and extend the sub-graph with links to graph EG. They also modify the PageRank transition matrix with respect to (the links to) graph EG. The Rank_local_idea and Pagerank_local framework formalizes the problem of ranking a sub-graph. It allows us to model multiple scenarios where ranking a sub- graph is important. Rank_local_idea can be used to model scenarios where PageRank scores of the global graph are known a priori and can potentially be re-used. This includes the case where the sub-graph symbolizes the data that have been updated, or the sub-graph symbolizes the data that contain all the semantic types of interest to a domain expert for a personalized or semantic ranking such as ObjectRank. Pagerank_local can be applied in general to all these problems, when we do not know the PageRank scores of outside data. We compare our approach with the stochastic complementation (SC) approach. SC builds a super-graph by carefully examining candidate outside data and adding them into the super-graph if adding this page has a significant influence on the PageRank scores of the sub-graph. Our approach in contrast models the outside data using a node graph EG, and it can be used in situations when a super-graph cannot be obtained easily. Our approach also avoids the cost of a global computation. The Pagerank_local computation is also much cheaper than SC since SC bears the high cost of constructing the super-graph. We experimentally study the effect of size and type of the sub-graphs on the accuracy of Pagerank_local. We study several types of sub-graphs including domain specific sub-graphs, topic specific sub-graphs, and sub-graphs gathered by a Breadth First Search crawler. We compare our results with SC and two baseline ranking algorithms; one was discussed in [18], and the other is local PageRank on the sub- graph (ignoring the outside data). We show that Pagerank_local has similar or superior ranking accuracy to SC and typically its runtime performance is an order of magnitude better than SC. Pagerank_local also outperforms the two baseline algorithms on ranking accuracy. Our contributions are as follows: • We define an efficient algorithm, Rank_local_idea, to estimate PageRank scores for a sub-graph when PageRank scores of the outside data are known. The random walk defined by Rank_local_idea utilizes these scores. • We prove that the Rank_local_idea scores converge to the true PageRank scores for all local data in the sub-graph, and the Rank_local_idea score for the outside node graph EG converges to the sum of true PageRank scores for all outside data. 35 Research on Node Ranking – Peer to Peer …. Hoàng Cường • When PageRank scores of outside data are not known, we define an efficient algorithm Pagerank_local. We provide important properties of Pagerank_local scores. • We show through empirical results that the Pagerank_local ranking accuracy is similar (sometimes superior) to the best competitor SC, and it overwhelmingly outperforms the runtime efficiency of SC. 4.2 Ranking - RANK_LOCAL_IDEA APPROACH Image 4.2: Eigenvalue We officially define the Rank_local_idea algorithm to estimate PageRank scores for a local graph. Our approach is passional by research on collapsing matrices with the same eigenvector. (If x is an eigenvector of the linear transformation A with eigenvalue λ, then any scalar multiple αx is also an eigenvector of A with the same eigenvalue. Similarly if more than one eigenvector shares the same eigenvalue λ, any linear combination of these eigenvectors will itself be an eigenvector with eigenvalue λ.) Rank_local_idea performs a random walk( Often, the walk is in discrete time, and indexed by the natural numbers, as in . However, some walks take their steps at random times, and in that case the position Xt is defined for the continuum of times ) on a modified local graph called the extended local graph, where an outside node of graph EG is added to the local graph. graph EG symbolizes the set of data that are not local. 36 Research on Node Ranking – Peer to Peer …. Hoàng Cường Image 4.2.2: Random walk The transition matrix probabilities of Rank_local_idea are derived from the transition matrix of PageRank for the global graph. Rank_local_idea assumes that the PageRank scores of all outside data in graph EG are known. This assumption will be relaxed in the next section where we present an approximate solution. Consider two graphs; a global graph of size N, and a local graph of size n. The local graph is a sub- graph of the global graph. The data in the local graph are called local data while data in the global graph and that are not in the local graph are called outside data. The goal is to provide the true PageRank for the local graph without running PageRank on the global graph. List the symbols used to define our algorithms. EG : outside nodes, the artificial node symbolizeing all outside nodes. G1 A sub-graph of the graph nodes with n nodes Gg The global nodes graph with N nodes. Ge The extended local graph with n + 1 nodes The Rank_local_idea algorithm There are edges between the graph and the local node edge outside knot according to the global character. However, this kind of explanation is impossible to distinguish like sees at a link instance or between the place page and between exterior data many links in the below example: 37 Research on Node Ranking – Peer to Peer …. Hoàng Cường Let Figure 4 be a global graph. Node A,B,C, and D are local data, and node X, Y and Z in the cloud are outside data. Figure 5 provides an example of adding an outside node to symbolize the exterior data The edges added from the local data to the outside node do not need the strategy to modify the primitive PageRank transition matrix to reflect each such edge to symbolize in global graph many edges. When computing the standard Pagerank algorithm on this graph, the possibility continuance from a node is proportional to the inverse of its out-degree. node C which has 3 incoming edges from the outside data is treated similarly to node D which has only 1 incoming edge from the outside data. Intuitively, however, we should expect a higher possibility of following links from the outside data to node C. Similarly, the possibility of following links from page A to graph EG is 1=3. This too is lower than the transition possibility based on the global graph. Rank_local_idea addresses this shortcoming with the following solution: The first step is to add an outside node graph EG to the sub-graph to symbolize all outside data. The second step is to construct the extended local graph Ge, the graph EG enriched graph of size n+ 1. There is an edge from graph EG to a local page in Ge if there is an edge from an outside page to that local page. The same hold for edges out of local data. Similarly, there is an edge from graph EG to graph EG if there is an edge between outside data. The next step is to define a transition matrix A_deal_matrix and a personalization vector Pideal. Let’s consider a little about the jargon “personalization vector”. At each time step, with possibility (1 −α), a surfer visiting any node will jump to a random Web Node (rather than following an outlink). The destination of the random jump is chosen according to the possibility distribution given in v. For this reason, we refer to v as the personalization vector Example of personalization vector value in Pagerank formula: 38 Research on Node Ranking – Peer to Peer …. Hoàng Cường The details will be discussed next section. Finally, a random walk is performed on Ge. The Rank_local_idea vector Rank_ideal is defined as follows: Algorithm Rank_local_idea (Gl, Gg) 1. Add outside node graph EG to G1. Image 4.2.4: (n+1) graph nodes 2. Create edge associated with the graph graph EG and get Ge. 3. Assign values to P_ideal and A_ideal. 4. Perform a random walk on the extended local graph according to Formula. Although swarming scales well to tolerate flash crowds for popular content, it is less useful for unpopular content. Peers arriving after the initial rush might find the content unavailable and need to wait for the arrival of a seed in order to complete their downloads. The seed arrival, in turn, may take long to happen, since maintaining seeds for unpopular content entails high bandwidth and administrative costs, which runs counter to the goals of publishers that value BitTorrent as a cheap alternative to a client-server approach. A strategy adopted by many publishers which significantly increases availability of unpopular content consists of bundling multiple files in a single swarm. So it’s only way as the key of ranking system – bandwith analyze. A_deal_matrix and Pideal We define an (n+1) x (n+1) transition matrix A_deal_matrix and a length (n+1) personalization vector Pideal. Let A symbolise the N x N transition matrix for PageRank on the global graph. Entry A[i,j] has the value of the inverse out-degree of page i, if there is an edge (i; j); the value is the expectation of a random walk following this edge from i. Without lack the generality, we assume the local data to be the first close n data in A and the outside data are indexed from n + 1 to N in A. Assume that the PageRank scores for all outside data are known. The values are PageRank (node n+1, node n+2, …, node N ) = { R[n + 1];R[n + 2]; … ;R[N] } 39 Research on Node Ranking – Peer to Peer …. Hoàng Cường (graph Ge có các node từ 1-> n node, và outside graph có (n+1 -> N) node) respectively. Let A_ideal is defined as follows, based on the entries in the original PageRank transition matrix A: Next we explain the elements in A_deal_matrix. These values are as follows: 1. The n x n sub-matrix at upper left is identical to the equating factors in transition matrix A for the global graph. They symbolize the possibility of transition between edges in the local graph. 2. The n x 1 sub-matrix at upper right symbolizes the possibility continuance from a local page to the node graph EG. We note that the possibility of reaching graph EG is the sum of the possibility of reaching any outside node from the local node. For local node k, the value is 3. The 1 x n sub-matrix at lower left represents to the possibility continuance from graph EG to local data. For local node k, the value is 4. The entry at the lower right corner denotes the possibility continuance from graph EG to graph EG. The last row has entries that are each a weighted sum of probabilities summed over all outside data. The weight is determined by the PageRank score of the outside node. This is a key feature of A_deal_matrix and will be discussed next. We define A_deal_matrix formally as follows: A_deal_matrix = Q1AQ2 40 Research on Node Ranking – Peer to Peer …. Hoàng Cường where Q1 is an (n+1) x N matrix and Q2 is an N x (n+1) matrix. Let Q2 be an N x (n + 1) matrix as follows: where In is an n x n identity matrix, B is an n x 1 0-matrix, C is a (N - n) x n 0- matrix, and D is a (N -n) x 1 matrix with all 1's. The effect of AQ2 on the ranking vector is to aggregate the authority continuance from local data to all outside data, which indicates the authority goes to graph EG. Let Q1 be the following (n + 1) x N matrix: where In is an n x n identity matrix, CT is an n x (N - n) 0-matrix and BT is a 1 x n 0-matrix. The matrix of interest is E, a 1 x (N - n) matrix. It considers the PageRank scores for all outside data. Recall that Sum is the sum of PageRank scores for all outside data Then, E can be expressed as follows: Let’s take an example: 41 Research on Node Ranking – Peer to Peer …. Hoàng Cường Image 4.2.5: Graph example - 6 nodes the transition Matrix Ranking value by using Pagerank-Formula: 42 Research on Node Ranking – Peer to Peer …. Hoàng Cường Q1 is an N x ( n + 1) matrix: Assume N = 6, n = 4; In matrix Q2 Matrix 43 Research on Node Ranking – Peer to Peer …. Hoàng Cường Tương tự: Giả sử 2 node outside là node 5 và node 6. N = 6, n = 4, 2 node outside. R[5] = 0.206 R[6] = 0.2862 Sum = 0.206 + 0.2862 = 0.4922 Matrix E E = ( , = (0.418529053, 0.581470947) A_deal_matrix = Q2AQ1 = 44 Research on Node Ranking – Peer to Peer …. Hoàng Cường = [ 0. 0.5 0.5 0. 0. ] [ 0 0 0 0 0 ] [ 1/3 1/3 0 0 1/3 ] [ 0. 0 0 0. 1. ] [ 0. 0 0 0.79073547 0.20926453] Xét lại công thức : Với R[5] = 0.206 R[6] = 0.2862 A[5][5] = 0 45 Research on Node Ranking – Peer to Peer …. Hoàng Cường A[5][6] =1/2 A[6][5] = 0 A[6][6] = 0 A[5][4] = 1/2 A[6][4] = 1 Sum = 0.4922 = [ 0. 0.5 0.5 0 0. = 0+0 ] [ 0 0 0 0 0 = 0 + 0 ] [ 1/3 1/3 0 0 1/3 = 1/3+0 ] [ 0. 0 0 0 1 = 1/2 + 1/2 ] [ 0=0+0 0=0+0 0=0+0 0.79073547 0.20926453 ] 0.79073547 = (0.206*0.5+0.2826)/0.4922 (đúng) 0.20926453 = (0.206*0 + 0.206*1/2 + 0.2826*0+ 0.2826*0)/0.4922 ( đúng) Và ma trận A 46 Research on Node Ranking – Peer to Peer …. Hoàng Cường Image 4.2.6: Multiplication result example The idea of multiplying the values of entries in A with the two matrices Q1 and Q2, where Q1 derived from the ranking vector for outside data, is key to the approach of A_deal_matrix. It has the effect of distributing the possibility continuance from the outside nodes, in a manner that is proportional to the importance of each of the outside data in the original PageRank vector. Recall that the personalization vector in the original PageRank is defined as a uniform vector Instead, for Rank_local_idea we define the personalization vector Pideal according to the number of outside data and total number of data in the graph. More specifically, the i-th entry of Pideal, Pideal[i] can be expressed as follows: Summary, according to the equalization: 47 Research on Node Ranking – Peer to Peer …. Hoàng Cường Convergence of Rank_local_idea Let Rank_ideal be the final ranking vector of Rank_local_idea, where the first n elements are scores for local data and the (n+1)-th element is the score for the outside node graph EG. We show that the scores of first n elements are identical to the true PageRank scores. Theorem 1: In Rank_ideal, scores for the first n data converge to the true PageRank scores. The score for the (n + 1) th element, graph EG, converges to the sum of true PageRank scores for all outside data. Proof: Let R be the true PageRank vector such that R is the converged stationary distribution for A. Let R’ = be a vector with n + 1 entries. We also know that R = . It is obvious that R’[i] = R[i] for first n elements and We will show that R’ is the Rank_local_idea vector. We know that Next consider a left multiply with to obtain the following: Since A_deal_matrix is stochastic and Markov Chain defined by Rank_local_idea is irreducible and aperiodic, there is a unique stationary distribution for A_deal_matrix. Therefore, R0 = Rank_ideal. The Rank_local_idea algorithm addresses several applications. One is where some sub-graph of the Web graph has been updated. A second case is when the personalized authority transfer is limited to the sub-graph. In these cases, the knowledge of PageRank scores can be potentially relied on to estimate new ranking scores. 48 Research on Node Ranking – Peer to Peer …. Hoàng Cường THE PAGERANK_LOCAL ALGORITHM Unlike the previous scenario where PageRank values for outside data are known, we now consider scenarios where the PageRank scores are not known a priori. To cover this state, our framework has an approximate solution Pagerank_local. The key difference is that for Pagerank_local, the algorithm is not able to differentiate the (previously weighted) contribution of authority from each individual outside page (since these PageRank scores are unknown). Instead, Pagerank_local will consider the authority continuance from outside data assuming they are equally important. We analyze the L1 distance between Rank_local_idea scores and Pagerank_local scores of the sub-graph and reveal that it is within a constant factor of L1 distance between the true PageRank scores and uniform scores of the outside data. We will show through experiments that Pagerank_local is a good approximation. A. The Pagerank_local algorithm The Pagerank_local vector Rapprox is defined as follows Pagerank_local adopts the same personalization vector as Rank_local_idea. It however, defines its own transition matrix Matrix_A_approx. B. Matrix_A_approx definition Matrix_A_approx is an (n + 1) x (n + 1) matrix. It is defined as follows: For example( the above graph) The matrix A 49 Research on Node Ranking – Peer to Peer …. Hoàng Cường New matrix Approx: [ 0 0.5 0.5 0 0 ] [ 0 0 0 0 0 ] = [1/3 1/3 0 0 1/3 ] [ 0 0 0 0.5 1 ] [ 0 0 0 0.75 1/4] Calculating local-pagerank Choose alpha = 0.85 ( according to Pagerank ) At iterator 0: Rapprox = [1/6, 1/6, 1/6, 1/6, 1/3] Pideal = [1/6, 1/6, 1/6, 1/6, 1/3] At iterator 1: Rapprox = 0.85* *R approx + 0.25* [0.25, 0.25, 0.25, 0.25, 0.5] … Program results after 10 iterators: 50 Research on Node Ranking – Peer to Peer …. Hoàng Cường Image 4.2.7 : Multiplication result example at iterators We Can be clearly seen: Order ( node 1-> 4) And order Are the same. Matrix_A_approx is different from A_deal_matrix in the last row, since Rank_local_idea does not utilize knowledge about PageRank scores of outside data in the first n rows. For the first n entries in the last row, the value symbolizes the (average) possibility continuance accumulated from (N - n) outside data to each local page. The last entry in this n-th row of the matrix is the (average) possibility continuance from outside data to other outside data. Similar to A_deal_matrix = Q1AQ2, Matrix_A_approx can be formally defined as Matrix_A_approx = Q’ 1AQ2, where the vector E is replaced by a vector Eapprox in Q’ 51 Research on Node Ranking – Peer to Peer …. Hoàng Cường In approx, the values at the last row are as follows: 1) For the first n values, (1 <= k <= n), the possibility from graph EG to a local page k is assigned the summation of continuance from all outside data to k, divided by the number of outside data. For local page k, it is 2) For the (n+1)-th value, the possibility for the self-loop edge is determined by the total authority continuance among outside data, divided by the number of outside data. Given the global graph example in Figure 4, the probabilities assigned by Matrix_A_approx are shown in Figure 6. We provide some examples of edge weight calculation following these rules. According to rule 1, the authority continuance on edge AB, AC, CB, BD, CD, DA are the outdegree inverse. Since A points to page X, Z, the authority continuance on edge (A;graph EG) is 1/2. The authority continuance on edge (graph EG;C) The self-loop edge authority continuance will be An advantageous quality about Pagerank_local is that it is suitable to adopt precomputation for various sub-graphs. With the same global graph, Approx can be figured out easily from the difference between the local values and the global values. This is especially beneficial for applications where there are multiple sub-graphs. 52 Research on Node Ranking – Peer to Peer …. Hoàng Cường 53 Pagerank_local scores converge to a unique vector R approx. There are two reasons. First, the transition matrix is a column stochastic matrix, as the sum of each column is 1. Second, since we complement the random walk with jumps from dangling data, the Markov Chain we defined is irreducible and aperiodic. RANK satisfies the two conditions of being irreducible and aperiodic of the E rgodic Theorem for Markov chains. Next we will vestigate how close is Rapprox to Rank_ideal, which we have shown to be the true ageRank scores for local data. in P Research on Node Ranking – Peer to Peer …. Hoàng Cường 54 Chapter 5 Evaluation The results show that our approach is robust to ranking, and converge fast, feasible. I evaluate our ranking technique in a simulator using real document sets from the small program simulation – peer to peer networks. As below is some results. Table 5.1: the number of iterators which local_pagerank converges It’s Depending on the number of nodes need to rank. But is limited. Research on Node Ranking – Peer to Peer …. Hoàng Cường 55 Chapter 6. Related Work raditional centralized pproaches, and search strategies over structured P2P networks. Our paper builds on prior work on efficient lookup and storage schemes. We assume the existence of a lookup protocol provided by the underlying system. Such lookup protocols have been studied in detail both in a structured setting (e.g., Chord, Pastry, Kademlia, Viceroy and Skipnet). Providing a useful search facility has been an important area of research. Prior work in searching can broadly be classified into two categories: t a Research on Node Ranking – Peer to Peer …. Hoàng Cường 56 Chapter 7. Conclusions and Future Work h s based on asynchronous iteration • ur hile our ys oaches. We plan to investigate these approaches in our future work. 7.1 Contributions In this paper, we have presented a distributed algorithm for ranking searc results. Our solution demonstrates that distributed ranking is feasible with little network overhead. Some of contributions: • Distributed computation of Pagerank o Application in P2P systems o Application on Internet scale Very large scale asynchronous iteration computation There are several areas worthy of further investigation. Performance could potentially be improved by mechanisms such as relevance feedback and caching. O analysis could be extended to account for popularity-based replication. W s tem provides a first solution to distributed ranking, other appr Research on Node Ranking – Peer to Peer …. Hoàng Cường 57 References 1. Google. Google search engine. Sergey Brin, Lawrence Page. “The Anatomy of a Large-Scale Hypertextual 2. Web Search Engine,” In Proc. 7th Int. World Wide Web and val. R etrieval, pp. 334-342. 4. Wikipedia. Conf., 1998. 3. Amy N. Langville & Carl D. Meyer (2006) Google's PageRank Beyond: The Science of Search Engine Rankings information retrie esearch and Development in Information R

Các file đính kèm theo tài liệu này:

  • pdfKLTN_HoangCuong_06020044.pdf
Tài liệu liên quan