- Research Article
- Open Access
Profile-Based Focused Crawling for Social Media-Sharing Websites
© The Author(s). 2009
- Received: 31 May 2008
- Accepted: 6 January 2009
- Published: 22 April 2009
We present a novel profile-based focused crawling system for dealing with the increasingly popular social media-sharing websites. In this system, we treat the user profiles as ranking criteria for guiding the crawling process. Furthermore, we divide a user's profile into two parts, an internal part, which comes from the user's own contribution, and an external part, which comes from the user's social contacts. In order to expand the crawling topic, a cotagging topic-discovery scheme was adopted for social media-sharing websites. In order to efficiently and effectively extract data for the focused crawling, a path string-based page classification method is first developed for identifying list pages, detail pages, and profile pages. The identification of the correct type of page is essential for our crawling, since we want to distinguish between list, profile, and detail pages in order to extract the correct information from each type of page, and subsequently estimate a reasonable ranking for each link that is encountered while crawling. Our experiments prove the robustness of our profile-based focused crawler, as well as a significant improvement in harvest ratio, compared to breadth-first and online page importance computation (OPIC) crawlers, when crawling the Flickr website for two different topics.
- User Profile
- Perceptual Group
- Document Object Model
- Profile Page
- Search Engine Result
Social media-sharing websites such as Flickr and YouTube are becoming more and more popular. These websites not only allow users to upload, maintain, and annotate media objects, such as images and videos, but also allow them to socialize with other people through contacts, groups, subscriptions, and so forth. Two types of information are generated in this process. The first type of information is the rich text, tags and multimedia data uploaded and shared on such web sites. The second type of information is the users' profile information, that can tell us what kind of interests they have. Research on how to use the first type of information has gained momentum recently. However, little attention has been paid to effectively exploit the second type of information, which are the user profiles, in order to enhance focused search on social media websites.
Prior to the social media boom, the concepts of vertical search engines and focused crawling have gradually gained popularity against popularity-based, general search engines. Compared with general search engines, topical or vertical search engines are more likely to become experts in specific topic areas, since they only focus on these areas. Although they lack the broadness that general search engines have, their depth can win them a stand in the competition.
In this paper, we explore the applicability of developing a focused crawler on social multimedia websites for an enhanced search experience. More specifically, we exploit the users' profile information from social media-sharing websites to develop a more accurate focused crawler that is expected to enhance the accuracy of multimedia search. To begin the focused crawling process, we first need to accurately identify the correct type of a page. To this end, we propose to use a Document Object Model (DOM) path string-based method for page classification. The correct identification of the right type of page not only improves the crawling efficiency by skipping undesirable types of pages, but also helps to improve the accuracy of the data extraction from these pages. In other words, the identification of the correct type of page is essential for our crawling, since we want to distinguish between list, profile, and detail pages in order to extract the right information, and subsequently estimate a reasonable ranking for each link that is encountered. In addition, we use a cotagging method for topic discovery as we think that it suits multimedia crawling more than the traditional taxonomy methods do, because it can help to discover some hidden and dynamic tag relations that may not be encoded in a rigid taxonomy (e.g., "tree" and "sky" may be related in many sets of scenery photos).
This paper is organized as follows. In Section 2, we review the related work in this area. In Section 3, we define the three types of pages that prevail on most social media-sharing websites, and discuss our focused crawling motivation. Then in Section 4, we explain the path string-based method for page classification. In Section 5, we introduce our profile-based focused crawling method. In Section 6, we discuss the cotagging topic discovery for the focused crawler. In Section 7, we wrap the previous sections to present our complete focused crawling system. In Section 8, we present our experimental results. Finally in Section 9, we make our conclusions and discuss future work.
Focused crawlers were introduced in , in which three components, a classifier, a distiller, and a crawler, were combined to achieve focused crawling. A Bayes rule-based classifier was used in , which was based on both text and hyperlinks. The distillation process involves link analysis similar to hub and authority extraction-based methods. Menczer et al.  presented a comparison of different crawling strategies such as breadth-fist, best-first, PageRank, and shark-search. Pant and Srinivasan  presented a comparison of different classification schemes used in focused crawling and concluded that Naive Bayes was a weak choice when compared with support vector machines or neural networks.
In [5, 6], Aggarwal et al. presented a probabilistic model for focused crawling based on the combination of several learning methods. These learning methods include content-based learning, URL token-based learning, link-based learning, and sibling-based learning. Their assumption was that pages which share similar topics tend to link to each other. On the other hand, the work by Diligenti et al.  and by Hsu and Wu  explored using context graphs for building a focused crawling system. The two-layer context graph and Bayes rule-based probabilistic models were used in both systems.
Instead of using the page content or link context, another work by Vidal et al.  explored the page structure for focused crawling. This structure-driven method shares the same motivation with our work in trying to explore specific page-layouts or structure. In their work, each page was traversed twice: the first pass for generating the navigation pattern, and the second pass for actual crawling. In addition, some works [10, 11] for focused crawling used metasearch methods, that is, their method is based on taking advantage of current search engines. Among these two works, Zhuang et al.  used search engine results to locate the home pages of an author and then used a focused crawler to acquire missing documents of the author. Qin et al.  used the search results of several search engines to diversify the crawling seeds. It is obvious that the accuracy of the last two systems is limited by that of the seeding search engines. In , the authors used cash and credit history to simulate the page importance and implemented an Online Page Importance Computation (OPIC) strategy based on web pages' linking structure (cash flow).
Extracting tags from social media-sharing websites can be considered as extracting data from structured or semistructured websites. Research about extracting data from structured websites include RoadRunner [13, 14], which takes one HTML page as the initial wrapper, and uses Union-Free Regular Expression (UFRE) method to generalize the wrapper under mismatch. The authors in  developed the EXALG extracting system, which is mainly based on extracting Large and Frequently occurring EQuivalence classes (LFEQs) and differentiating roles of tokens using dtokens to deduce the template and extract data values. Later in , a tree similarity matching method was proposed to extract web data, where a tree edit distance method and a Partial Tree Alignment mechanism were used for aligning tags in the HTML tag tree. Research in extracting web record data has widely used a web page's structure  and a web page's visual perception patterns. In , several filter rules were proposed to extract content based on a DOM-tree. A human interaction interface was developed through which users were able to customize which type of DOM-nodes are to be filtered. While their target was for general HTML content and not for web records, they did not suit their methods to structured data record extraction. Zhao et al.  proposed using the tag tree structure and visual perception pattern to extract data from search engine results. They used several heuristics to model the visual display pattern that a search engine results page would usually look like, and combined this with the tag path. Compared with their tag path method, our path string approach keeps track of all the parent-child relationships of the DOM nodes in addition to keeping the parent-first-child-next-sibling pattern originally used in the DOM tree. We also include the node property in the path string generation process.
3.1. Popularity of Member Profile
In Section 2, we reviewed several focused crawling systems. These focused-crawling systems analyze the probability of getting pages that are in their crawling topics based on these pages' parent pages or sibling pages. In recent years, another kind of information, which is the members' profiles, started playing a prominent role in social networking and resource-sharing sites. Unfortunately, this valuable information still eludes all current focused crawling efforts. We will explore the applicability of using such information in our focused-crawling system. More specifically, to illustrate our profile-based focused crawling, we will use Flickr as an example. But our method can be easily expanded to other social networking sites, photo-sharing sites, or video-sharing sites. Hence, we refer to them as "social multimedia websites."
3.2. Typical Structure of Social Media-Sharing Websites
A list page is a page with many image/video thumbnails and their corresponding uploaders (optionally some short descriptions) displayed below each image/video. A list page can be considered as a crawling hub page, from where we start our crawling. An example list page is shown in Figure 1.
A detail page is a page with only one major image/video and a list of detailed description text such as title, uploader, and tags around it. A detail page can be considered as a crawling target page, which is our final crawling destination. An example detail page is shown in Figure 2.
A profile page is a page that describes a media uploader's information. Typical information contained in such a page includes the uploader's image/video sets, tags, groups, and contacts, and so forth. Further, such information can be divided into two categories: inner properties, which describe the uploader's own contributions, such as the uploader's photo tags, sets, collections, and videos, and inter properties, which describe the uploader's networking with other uploaders, such as the uploader's friends, contacts, groups, and subscribers. We will use information extracted from profile pages to guide our focused crawling process.
3.3. Profile-Based Focused Crawling
Our motivation while crawling is to be able to assess the importance of each outlink or detail page link before we actually retrieve that detail page given a list page and a crawling topic. For the case of Figure 3, suppose that we are going to crawl for the topic flowers, then we would intuitively rank the first detail page link, which links to a real flower, higher than the second detail page link, which links to a walking girl and happened to be also tagged as "flower." The only information available for us to use is the photo thumbnails and the photo uploaders such as "U-EET" and "haggard37." Processing the content photo thumbnails to recognize which one is more conceptually related to the concept of real flowers poses a challenging task. Hence, we will explore the photo uploader information to differentiate between different concepts. Luckily, most social media-sharing websites keep track of each member's profile. As shown in Figure 3, a member's profile contains the member's collections, sets, tags, archives, and so forth. If we process all this information first, we can have a preliminary estimate of which type of photos the member would mainly upload and maintain. We can then selectively follow the detail page links based on the corresponding uploader profiles extracted.
Before we actually do the crawling, we need to identify the type of a page. In this section, we will discuss our page classification strategy based on the DOM path string method. Using this method, we are able to identify whether a page is a list page, detail page, profile page, or none of the above.
4.1. DOM Tree Path String
In Figure 4, the whole DOM tree can be seen as a document node, whose child is the element node <html>, which further has two children <head> and <body>, both element nodes, and so on. The element nodes are all marked with in Figure 4. At the bottom of the tree, there are a couple of text nodes. In the DOM structure model, the text nodes are not allowed to have children, so they are always the leaf nodes of the DOM tree. There are other types of nodes such as CDATASection nodes and comment nodes that can be leaf nodes. Element nodes can also be leaf nodes. Element nodes may have properties. For example, "<tr class="people">" is an Element Node "<tr>" with property "class="people"." Readers may refer to http://www.w3.org/DOM/ for a more detailed specification.
A path string of a node is the string concatenation from the node's immediate parent all the way to the tree root. If a node in the path has properties, then all the display properties should also be included in the concatenation. We use "-" to concatenate a property name and "/" to concatenate a property value.
For example, in Figure 4, the path strings for "John," "Doe," and for "Alaska" are " ."
Note that when we concatenate the property DOM node into path strings, we only concatenate the display property. A display property is a property that has an effect on the node's outside appearance when viewed in a browser. Such properties include "font size," "align," "class," and so forth. Some properties such as "href," "src," and "id" are not display properties as they generally do not affect the appearance of the node. Thus, including them in the path string will make the path string overspecified. For this reason, we will not concatenate these properties in the path string generation process.
A path string node value (PSNV) pair is a pair of two text strings, the path string ps, and the node value whose path string is ps. For example, in Figure 4, " " and "John" are a PSNV pair.
A perceptual group of a web page is a group of text components that look similar in the page layout. For example, "Sets," "Tags," "Map," and so on, in the profile page of Figure 3 are in the same perceptual group; and "U-EET" and "haggard37" are in the same perceptual group in the list page in Figure 1.
4.2. DOM Path String Observations
Path string efficiency. First, when we extract path strings from the DOM tree, we save a significant amount of space, since we do not need to save a path string for every text node. For example, we only need one path string to represent all different "tags" in a detail page shown in Figure 2, as all these "tags" share the same path string. Second, transforming the tree structure into linear string representation will reduce the computational cost.
Path string differentiability. Using our path string definition, it is not hard to verify that text nodes "flowers," "canna," and "lily" in Figure 2 share the same path string. Interestingly, they share a similar appearance when displayed to users as an HTML page, thus, we say that they are in the same perceptual group. Moreover, their display property (perceptual group) is different from that of "U-EET," "haggard37," and so on, in Figure 1, which have different path strings. Generally, different path strings correspond to different perceptual groups as the back-end DOM tree structure decides the front-end page layout. In other words, there is a unique mapping between path strings and perceptual groups. At the same time, it is not hard to notice that different types of pages contain different types of perceptual groups. List pages generally contain the perceptual group of uploader names, while detail pages usually contain the perceptual group of a list of tags, and their respective path strings are different. These observations have encouraged us to use path strings to identify different types of pages, and the identification of the types of pages is essential for our crawling, since we want to distinguish between profile and detail pages in order to extract the right ranking for a link.
4.3. Page Classification Using Path String
4.3.1. Extracting Schema Path String Node Value Pairs
Algorithm 1: Deduce schema PSNV pairs.
Input: N Pages for schema extraction
Schema = All PSNV-pairs of Page 1.
for i = 2 to N
do Temp = All PSNV-pairs of Page i
Schema = intersection(Schema, Temp)
In code 1, we adopt a simple way for identifying schema data and real data. That is, if the data value and its PSNV pair occur in every page, we identify them as a schema pair, otherwise it is considered a real data pair. The for loop of line 2–4 performs a simple intersection operation on the pages, while line 5 returns the schema. Note that this is a simple and intuitive way of generating schema. It can be extended by using a threshold value. Then, if a certain PSNV pair occurs in at least a certain percent of all the pages, it will be identified as schema data.
4.3.2. Classifying Pages Based on Real Data Path Strings
Noting that the same types of pages have the same perceptual groups and further the same path strings, we can use whether a page contains a certain set of path strings to decide whether this page belongs to a certain type of pages. For example, as we already know that all list pages contain the path string that corresponds to uploader names, and almost all detail pages contain the path string that corresponds to tags, we can then use these two different types of path strings to identify list pages and detail pages. Algorithm 2 gives the procedure of extracting characteristic path strings for pages of a given type.
Algorithm 2: Extracting a page type's path strings.
Input: N Pages of the same type
for page type path strings extraction
Set = All Path Strings of Page 1 - Schema PSs.
for i = 2 to N
do Temp = All PSs of Page i - Schema PSs
Set = intersection(Set, Temp)
Now, that we are able to identify the correct page type using the path string method, we are equipped with the right tool to start extracting the correct information from each type of page that we encounter while crawling, in particular, profile pages. In this section, we discuss our profile-based crawling system. The basic idea is that from an uploader's profile, we can gain a rough understanding of the topic of interest of the uploader. Thus, when we encounter a media object such as an image or video link of that uploader, we can use this prior knowledge which may relate to whether the image or video belongs to our crawling topic in order to decide whether to follow that link. By doing this, we are able to avoid the cost of extracting the actual detail page for each media object to know whether that page belongs to our crawling topic. To this end, we further divide a user profile into two components, an inner profile and an inter profile.
5.1. Ranking from the Inner Profile
The inner profile is an uploader's own property. It comes from the uploader's general description of the media that they uploaded, which can roughly identify the type of this uploader. For instance, a "nature" fan would generally upload more images and thus generate more annotations about nature; an animal lover would have more terms about animals, dogs, pets, and so on, in their profile dictionary. For the case of the Flickr photo-sharing site, an uploader's inner profile terms come from the names of their "collections," "sets," and "tags." As another example, for the YouTube video-sharing site, an uploader's inner profile comes from their "videos," "favorites," "playlists," and so on. It is easy to generalize this concept to most other multimedia sharing websites.
5.2. Ranking from the Inter Profile
In contrast to the inner profile which gives an uploader's standalone property, strictly related to their media objects, we note that an uploader in a typical social media-sharing website, tends to also socialize with other uploaders on the same site. Thus, we may benefit from using this social networking information to rank a profile. For instance, a user who is a big fan of one topic, will tend to have friends, contacts, groups, or subscriptions, and so forth, that are related to that topic. Through social networking, different uploaders form a graph. However, this graph is typically very sparse, since most uploaders tend to have a limited number of social contacts. Hence, it is hard to conduct a systematic analysis on such a sparse graph. In this paper, we will use a simple method, in which we accumulate an uploader's social contacts' inner ranks to estimate the uploader's inter rank.
5.3. Combining Inner Rank and Inter Rank
where is the th image thumbnail link and is the th user that corresponds to the th image thumbnail link. and are calculated using (1) and (2), respectively. We could further normalize to obtain probability scores, however, this will not be not needed, since they are only used for ranking links.
where is the number of pictures cotagged by both tag T and tag T1, and and are the number of pictures tagged by tag T and tag T1, respectively. Suppose that tag T belongs to the crawling topic, then gives the score of whether T1 also belongs to the crawling topic. When is bigger than a preset threshold, we will count T1 as belonging to the crawling topic.
In order to make the crawling topic tags more robust, we further use the following strategies.
only one image if multiple images are tagged with an identical set of tags. This is usually because an uploader may use the same set of tags to tag a group of images that they uploaded to save some time.
the top cotagging tags, start a new round of cotagging discovery. This process is depicted in Figure 7. Then use the expanded cluster of high frequency co-occurring tags as the final crawling topic.
We developed a two-stage crawling process that includes a cotagging topic discovery stage and a profile-based focused crawling stage. Both of these stages use the page classifier extensively to avoid unnecessary crawling of undesired types of pages and to correctly extract the right information from the right page. The details of crawling are explained in Sections 7.1 and 7.2.
7.1. Cotagging Topic Discovery Stage
The first stage of our profile-based focused system is the cotagging topic discovery stage. In this stage, we collect images that are tagged with the initial topic tag, record their cotags, process the final cotagging set, and extract the most frequent co-occurring ones. Figure 8 gives the diagram of the working process of this stage, and Algorithm 3 gives the detailed steps.
Algorithm 3: Stage one: cotagging topic discovery.
Input: Initial Crawling Topic Tag, T
Set Queue Q = empty
for i = 1 to n
do Enqueue p_i into Q
while Q! = Empty
do page p = Dequeue Q
if p = List Page
if o_i = Detail Page Link
then Enqueue o_i to Q
else if o_i = Profile Page Link
then discard o_i
else if p = Detail Page
then extract tags data from p
analyze the tags to get the most frequent
In Algorithm 3, lines (4)–(14) do the actual crawling work. The page classifier described in Section 4 is used in line (6) to decide whether a page is a list page or a detail page. We already know that in social media-sharing websites, list pages have outlinks to detail pages and profile pages, and we name such links detail page links and profile page links, respectively. It is usually easy to differentiate them because in the DOM tree structure, detail page links generally have image thumbnails as their children, while profile page links have text nodes, which are usually the uploader names, as their children. Combined with our path string method, we can efficiently identify such outlinks. In lines (11)-(12), by not following profile page links, we save a significant amount of effort and storage space. Since we are not following profile page links, the classification result for page in line (6) would not be a profile page. Lines (15)-(16) do the cotagging analysis and line (17) returns the expanded topic tags.
7.2. Profile-Based Focused Crawling Stage
In the second stage, which is the actual crawling stage, we use the information acquired from the first stage to guide our focused crawler. For this stage, depending on the system's scale, we can choose to store the member profiles either on disk or in main memory. The system diagram is shown in Figure 9, and the process detail is shown in Algorithm 4. In Algorithm 4, similar to the cotagging stage, we classify page in line (6). The difference is that, since we are not pruning profile page links in lines (13)-(14) and we follow them to get the user profile information, we will encounter the profile page branch in the classification result for line (6), as shown in lines (17)-(18). Another difference is how we handle detail page links, as shown in lines (10)–(12). In this stage, we check whether a detail page link's user profile rank according to the crawling topic. If the rank is higher than a preset threshold, , we will follow that detail page link, otherwise, we will discard it. Note that in this process, we need to check whether a user's profile rank is available or not, which can be done easily by setting a rank available flag, and we omit this implementation detail in the algorithm. In lines (17)-(18), we process profile pages and extract profile data. Another issue is deciding when to calculate the user profile rank since the profiles are accumulated from multiple pages. We can set a fixed time interval to conduct the calculation or use different threads to do the job, which is another implementation detail that we will skip here.
Algorithm 4: Stage two: profile-based focused crawling.
Output: Crawled Detail Pages
Queue Q = empty
for i = 1 to n
while Q ! = Empty
do page p = Dequeue Q
if p = List Page
else if p = Detail Page
then Extract Tags Data from p
else if p = Profile Page
then Extract Prof Data from p
else if p = Other Type Page
then ignore p
Return Detail Pages Tags Data
8.1. Path String-Based Page Classification
Path string differentiation.
Top cotagging tags for the topic "flowers."
Top cotagging tags for "nyc."
8.2. Topic Discovery through Cotagging
We tested two topics for the cotagging topic discovery process using Flickr photo-sharing site. In the first test, we used the starting tag "flowers," and we collected 3601 images whose tags contain the keyword flowers. From this 3601-image tag set, we found the following tags that occur in the top cotagging list (after removing a few noise tags such as "nikon," that are easy to identify since they correspond to camera properties and not media object properties).
In the second round of tests, we used the starting tag "nyc," and after collecting 3567 images whose tag sets contain "nyc," we obtained the following expanded topic tag set.
We can see that these results are reasonable. We then used these two sets of crawling topics for the following focused crawling experiments.
8.3. Profile-Based Focused Crawling
where is the number of relevant pages (belonging to the crawl topic) and is the number of total pages crawled. To calculate the harvest ratio, we need a method to calculate the relevancy of the crawled pages. If the crawled page contains any of the tags that belong to the crawl topic, we would consider this page as relevant, otherwise it will be considered as irrelevant. For comparison, we compared our focused crawling strategy with the breadth-first crawler.
In the next set of experiments, we compared our profile-based focused crawler with that of the OPIC crawler  for both the topic "nyc" and "flower."
We also performed experiments to compare the detail page capture ratio between profile-based focused crawling and OPIC-based crawling.
We can see that in both cases, the detail page capture ratio is higher for the profile-based focused crawler than for the purely OPIC-based crawler.
We presented a profile-based focused crawler, which ranks users with more topic-relevant media objects higher during crawling. To further differentiate profiles while taking into account the special characteristics of social media sites, we have introduced and used the notions of the inner profile and inter profile. We have used cotagging in a first stage, for automated crawling topic discovery, and thus build a consistent set of tags for a given topic. In both the cotagging topic discovery process and the profile-based focused crawling process, we used a path string-based page classification scheme in order to allow us to extract the correct type of information from each page type, and in order to correctly calculate the profile ranks for a given topic. Our experimental results confirmed the effectiveness of our profile-based focused crawling system from the perspective of harvest ratio and robustness. In the future, we would like to deploy the proposed focused crawling on a real system for real-time vertical social media search.
- Chakrabarti S, van den Berg M, Dom B: Focused crawling: a new approach to topic-specific web resource discovery. Computer Networks 1999,31(11–16):1623-1640.View ArticleGoogle Scholar
- Chakrabarti S, Dom B, Indyk P: Enhanced hypertext categorization using hyperlinks. Proceedings of the ACM SIGMOD International Conference on Management of Data, June 1998, Seattle, Wash, USA 307-318.Google Scholar
- Menczer F, Pant G, Srinivasan P: Topical web crawlers: evaluating adaptive algorithms. ACM Transactions on Internet Technology 2004,4(4):378-419. 10.1145/1031114.1031117View ArticleGoogle Scholar
- Pant G, Srinivasan P: Learning to crawl: comparing classification schemes. ACM Transactions on Information Systems 2005,23(4):430-462. 10.1145/1095872.1095875View ArticleGoogle Scholar
- Aggarwal CC, Al-Garawi F, Yu PS: On the design of a learning crawler for topical resource discovery. ACM Transactions on Information Systems 2001,19(3):286-309. 10.1145/502115.502119View ArticleGoogle Scholar
- Aggarwal CC, Al-Garawi F, Yu PS: Intelligent crawling on the world wide web with arbitrary predicates. Proceedings of the 10th International Conference on World Wide Web (WWW '01), May 2001, Hong Kong 96-105.View ArticleGoogle Scholar
- Diligenti M, Coetzee F, Lawrence S, Giles CL, Gori M: Focused crawling using context graphs. Proceedings of the 26th International Conference on Very Large Data Bases (VLDB '00), September 2000, Cairo, Egypt 527-534.Google Scholar
- Hsu C-C, Wu F: Topic-specific crawling on the web with the measurements of the relevancy context graph. Information Systems 2006,31(4-5):232-246. 10.1016/j.is.2005.02.007View ArticleGoogle Scholar
- Vidal MLA, da Silva AS, de Moura ES, Cavalcanti JMB: Structure-driven crawler generation by example. Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '06), August 2006, Seatttle, Wash, USA 292-299.View ArticleGoogle Scholar
- Zhuang Z, Wagle R, Giles CL: What's there and what's not?: focused crawling for missing documents in digital libraries. Proceedings of the 5th ACM/IEEE Joint Conference on Digital Libraries (JCDL '05), June 2005, Denver, Colo, USA 301-310.View ArticleGoogle Scholar
- Qin J, Zhou Y, Chau M: Building domain-specific web collections for scientific digital libraries: a meta-search enhanced focused crawling method. Proceedings of the 4th ACM/IEEE Joint Conference on Digital Libraries (JCDL '04), June 2004, Tucson, Ariz, USA 135-141.Google Scholar
- Abiteboul S, Preda M, Cobena G: Adaptive on-line page importance computation. Proceedings of the 12th International Conference on World Wide Web (WWW '03), May 2003, Budapest, Hungary 280-290.Google Scholar
- Crescenzi V, Mecca G, Merialdo P: Roadrunner: towards automatic data extraction from large web sites. Proceedings of the 27th International Conference on Very Large Data Bases (VLDB '01), September 2001, Roma, Italy 109-118.Google Scholar
- Grumbach S, Mecca G: In search of the lost schema. Proceedings of the 7th International Conference on Database Theory (ICDT '99), January 1999, Jerusalem, Israel, Lecture Notes in Computer Science 1540: 314-331.Google Scholar
- Arasu A, Garcia-Molina H, University S: Extracting structured data from web pages. Proceedings of the ACM SIGMOD International Conference on Management of Data, June 2003, San Diego, Calif, USA 337-348.Google Scholar
- Zhai Y, Liu B: Web data extraction based on partial tree alignment. Proceedings of the 14th International Conference on World Wide Web (WWW '05), May 2005, Chiba, Japan 76-85.View ArticleGoogle Scholar
- Li Z, Ng WK, Sun A: Web data extraction based on structural similarity. Knowledge and Information Systems 2005,8(4):438-461. 10.1007/s10115-004-0188-zView ArticleGoogle Scholar
- Gupta S, Kaiser G, Neistadt D, Grimm P: Dom-based content extraction of html documents. Proceedings of the 12th International Conference on World Wide Web (WWW '03), May 2003, Budapest, Hungary 207-214.Google Scholar
- Zhao H, Meng W, Yu C: Mining templates from search result records of search engines. Proceedings of the 13th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '07), August 2007, San Jose, Calif, USA 884-893.View ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.