With so much data in the world, it’s time to figure out how much of it is unstructured -- that which a human needs to look at in order to understand it best -- and what to do about it.
IT research firm IDC has estimated 7.9 zettabytes of digital data in the world by 2015, and I think the biggest chunk of it will come from social media generated on mobile platforms and driven from email. Intel estimates that at least 2.5 billion people will be online by 2015, generating more and more data and requiring more resources for storing and processing the data, as reported in InformationWeek. Such outlooks have led evangelist analysts to gush over unstructured data's potential; Google's Avinash Kaushik, for example, publicly claimed to have “orgasms over big, unstructured data.”
How to structure unstructured data
Many of us are really only getting started with unstructured data, looking for ways to get started and trying to figure out how best to handle it all. Actually, we need to ask ourselves if we should even bother to try working with it, as many previous attempts to add structure robotically have been disappointing to say the least, and fail at least much of the time. After all, dealing with and automating processes around structured data is tough enough!
Here I've put together a few things you can do with or relative to unstructured data:
Distribute the data in the cloud -- just store more of it and hope you can see useful patterns in the data with advanced big-data analytics and predictive analytics platforms.
Develop more powerful analytics engines to analyze the data, most of which will be in the cloud, in real time
Merge as much data as you can into large data files, a lesson learned by Team Obama in preparing for the 2012 election recently; merging several different databases and cleaning the data made developing predictions and gleaning insights easier.
Clean the data -- this assumes unstructured data is dirty, or not useful for analysis in its current state. You can purge duplicate information, ensure consistency in the naming of entities, and empty and sparse datasets, for example. Consider checking out Saleforce Data.com's Social Key, which ties customer data records to social media accounts and online content by those accounts. Perhaps Salesforce is on to something here; the cost of cleaning data might be shared, as the data a company is able to clean for its own use also goes back (the part that can be shared) to the overall Data.com repository in Salesforce’s cloud.
Working with unstructured data won't be easy -- but it will be necessary.
What advice do you have for working with unstructured data? Share below.
webmetricsguru, - One big challenge i see is analytics of sentiment oriented data even though the area has grown a great deal in the course of this year -- new apps and all. This kind of data at least for now may still need to be reviewed closer because its difficult to automate it in one given pattern without locking out new sentiments that are outside of that original set. However as intelligent systems learn this data, it will cut down what we have to look at even in sentiment analytics.
If I understand you correctly, that's the bane of the PR / Marcom industry - that you can actually look at bunch of verbatim (maybe that's ok for 10-30) but what happens when you have thousands or more?
I think a discussion on just what cleaning data is and how to to best do it would be good for AllAnalytics.com personally. I'd like to see what we come up with, and I bet a lot of others would too.
"If we can find the essential information or pattern, we might not need to look at most of it"
I see. I suppose that those hand-written patterns can just be domain specific and will be difficult to generalize. I agree that extracting the most useful patterns might be enough in most cases, as it difficult to think of all possible patterns. One of drawback of such model is that human patterns are often low-recall, even if precision is high.
It is true that some of the points you mentioned are debatable - like "Distribute the data in the cloud". But they are valid points to take into account when dealing with unstructured data. To the question how to clean unstructured data? I think that it depends on the shape and the model that has been defined.
What I meant is that currently, people usually end up needing to look at the data to understand it (because it is un structured information) and attempts to use software to understand it, in my opinion, won't work, at least not today.
What you can do, I think, and maybe our friends here can confirm or argue this, is cut down on what we have to look at.
If we can find the essential information or pattern, we might not need to look at most of it - and hopefully the software created can help surface that information, and maybe that's the best we can hope for (big data hype or not).
At any rate, this is an interesting discussion and I don't have all the answers - but I am wondering just what they are.
"I define it as something a human needs to look at to fully process"
I still don't get it. Do you mean that there is the need for human intervention to figure out whether the data is unstructured or not? Won't that be time consuming and practically impossible for human to go through all the instances of the data due to it size? Maybe it is not what you mean?
While 97% of insurers say that insurance fraud has increased or remained the same in the past two years, most of those companies report benefits from anti-fraud technology in limiting the impact of fraud, including higher quality referrals, the ability to uncover organized fraud, and improve efficiency for investigators.