The Information Retrieval Collection (IRC)

Tags

Data Mining, information retrieval, ir, miner, minerazzi

A new miner is available at Minerazzi.com: The Information Retrieval Collection (http://www.minerazzi.com/irc).

What you can do with it?

Use this topic-specific search engine to find information retrieval resources such as research work (articles, conference proceedings, seminal work,…) and lecture material (books, lecture notes, tutorials,…) on algorithms and models like

Singular Value Decomposition
Principal Component Analysis
Latent Dirichlet Allocation
Latent Semantic Analysis and Indexing
Vector Support Machines
Vector Space Theory and Models
Inverse Document Frequency
Probabilistic Models
Markov Hidden Models
K-Means
Cluster Analysis
Search Engines (of course)
etc…

Some of the resources come from Nist’s TREC proceedings, AIRWeb, WICOW, Cornell’s ECommons, SIGIR, SIAM, Wikipedia, and work from top-level information retrieval researchers.

More Power to the People! 

You can also use the recrawling power of Minerazzi to build your own customized collection or to enhance a third-party collection.

Building topic-specific collections, the easy way

Tags

Data Mining, microindexing, minerazzi, mining

We have improved the Minerazzi platform (http://www.minerazzi.com) by adding new features.

That includes an internal filter for deduplicating urls, which is currently being tested. The interface is also a lot cleaner.

We found that with the recrawling features of Minerazzi, it is extremely easy to build topic-specific collections from sites like Wikipedia.

Once one gets a Wikipedia record in a search result, recrawling links simplifies discovery of relevant URLs which can then be exported to a personal collection project. Cool and nice!

Feel the power of micro-indexing

That’s the great thing about microindexing. It allows you to do things you cannot do with a static, monster-sized index. That is, you start with a tiny index and explore its reach as you use it to navigate the Web. It is the reach of an index and not its size what really matters to a data miner gathering topic-specific resources.

Unlike with traditional search engines, wherein you search a huge index of cached records (often irrelevant or with outdated information), you go into a discovery journey of current records as they currently exist online! You start with a seed of highly relevant URLs (the micro-index) and then start mining and building while searching. In my book, that’s a nice search paradigm.

On other matters, the final touches of the Information Retrieval Collection (IRC) are being implemented.

So far, IRC includes records from NIST’s TREC conference proceedings, the Gerard Salton Collection from Cornell University ECommons database, Wikipedia, IR books, and lecture material or research articles from top level IR researchers.

Lessons learned from building an IR collection

Tags

information retrieval, ir, tagging, temporary tags

We are currently building the Information Retrieval Collection (IRC) with the Minerazzi platform. URLs pointing to resources like articles from scholarly journals and teaching material from scholarly web pages are crawled and indexed.

So far we have found that a non trivial amount of said URLs point to teaching material (lecture notes, tutorials,..) in a markup format with title tags or file names that poorly describe what their content is about, then difficulting indexing. Some don’t even have any useful information in their html head section. Something similar, occurs with the URLs of .pdf and .ps files: we often found no useful file names.

For article journals, this is understandable as that could be the result of editorial policies; not so for teaching materials, though.

Saddly, the taste is that either university webmasters or the scholars who wrote the resources seem to be sloppy or go by the “do as I teach, no as I do” rule.

In any case, the above documents have descriptors that are too short or poorly written to help a human or robot with the indexing. One can do better by going with the anchor text displayed in the documents, but again, that workaround relies on the content quality of said text.

We are working on a partial solution to the problem using temporary tagging instead of extracting summaries after full-text indexing. It is an interesting trick for dynamically building collections, but again it is not a perfect solution.

Improving the Data Structures and Algorithms Collection

Tags

Algorithms, Data Structures, dsac, minerazz, miners

We have almost doubled the index of the Data Structures and Algorithms (DSAC) miner. In addition, we are moving to indexing books relevant to this miner. Some changes were also made to the output of the search result pages, so these are now less cluttered.

Essentially, users now have the option of viewing sections of the search results. We did not do this before because of collisions with div ids.

These changes to the serps apply to all current miners built with the Minerazzi platform, too. Enjoy it.

Unveiling Link Honey Pots with Minerazzi

Tags

Algorithms, Data Structures, dsac, honey pots, minerazzi, spam

In Web Spam Taxonomy, Gyongyi and Garcia-Molina, describe several web spam techniques, one being honey pots.

They describe these as “a set of pages that provide some useful resource …, but that also have (hidden) links to the target spam page(s).

To target non-human visitors (e.g., web crawlers), said links could be placed in HTML elements that are made invisible (e.g. division tags with a display:none CSS rule) and then with rel=”dofollow” in the anchor tags.

These types of tricks can be easily unveiled with the recrawling feature of Minerazzi.

For instance, searching for [ heap sort ] in http://www.minerazzi.com/dsac retrieves several records, one being http://www.aihorizon.com/resources/sourcecode/trees/heap_h.htm.

Clicking the “Recrawl it” link retrieves several URLs, one being http://www.aihorizon.com/index.htm.

Clicking again the “Recrawling the Web” Recrawling that link quickly reveals that said index page has several hidden links to porn sites. Looking at the source code of that URL shows that the above adversarial technique was used.

Minerazzi: Allowing Users to Recrawl Search Results

Tags

Crawls, Data Mining, Recrawling, Search

Effectively immediately Minerazzi (http://www.minerazzi.com) allows users to recursively recrawl search results.

Why is recrawling so important?

The purpose of allowing users to recrawl URLs is to expose them to new content, to involve them in learning through discovery. To turn their searches into a mining activity. This makes more sense than limiting their search experience to inspecting zillion of cached records from a search engine index. The problem with the latter is that frequently those records are either outdated or irrelevant, not to mention that in that scenario the users are simply passive expectators.

Allowing users to recrawl search results has many advantages and possibilities. For instance, users can use the discovered URLs to build curated collections, self-guide investigative work, or gather link intelligence from sites, directories, blogs, forums, or social networks. In general, recrawling allows users to discover hidden paths to fresh, new, or rich content.

Considering that the total number of primary and secondary URLs defines the reach of a microindex, in theory recrawling should result into an endless reach.

At this time, we do not recrawl .css and .pdf files, but we recrawl the most common file formats (.php, .asp, .aspx, .html, .htm, .js, etc). However, if the content of a file is dynamic, obfuscated, or poorly coded more likely it will return garbage or nothing.

Having said that, we invite you to try the recrawling experience with our public miners.