Agile Big Data Analytics: How to start a Big Data Analytics project

I was at a client briefing today where I was speaking and a question was asked: How do you start a big data analytics project? Do you first try to get the veracity and quality of all the data correct before you embark on such an initiative? Or are there alternative ways of kicking off such an initiative?

It was an interesting question and one I am sure many companies are faced with regardless of whether they explicitly accept it or not.

It is important to understand that every company has good data and not crib about data being not available or not usable. Indeed there is opportunity to increase the data set and the data veracity but waiting for that to happen would only delay your analytics initiative further and further - not a good idea. Here is an option to consider, with a few steps to implement it:

1. Use unsupervised machine learning algorithm - a technique which may be used to employ machine learning algorithms to work on the available data set, disregarding any data quality issues to begin with. An unsupervised machine learning algorithm would cluster the data sets into different categories, classes based on the information that is available. The clusters, classes or categories may be analyzed to reveal insight and information inherent in the data. Remember that the data set may be data in its raw or native form in a schema-less (think Hadoop) storage, in transactional systems or data warehouses, or a combination of both.
2. Present the initial findings - Polish the analytical output and make them presentable in human understandable, business-centric lingo and present it to the stakeholders. Allow them to chew on the findings as you explain the same. Remember, the good thing here is that you take no blame in your ability or inability to extract the insight that client is looking for; it is their data with no quality profiling or transformations!
3. Gather feedback from client - The feedback would most probably come back as a combination of the following:
   
   • Pleasantly surprised with the initial insight and its value.
   • The insight is not complete and requires further tuning.
4. Perform data analysis - At this point, given the type of analytical insight being sought after, it is time to look more closely at the data. The completeness and the quality of the data requires analysis and determine which subset of the data may have issues (quality, completeness among others) which requires to be fixed. This is the time when the proper data profiling and transformations will be required to put in place to go after the insight that is expected.


5. Add more data sources - Depending on the nature of insight being sought after, after the initial analysis and assessment, it be required a different data source or data set. I'd recommend not to perform a full-scale data profiling on the new data set; just bring it in! However, focus your efforts on data quality for the original data sets.
   

I hope you can detect the pattern here. The idea is to bring the data in and apply analytics to the same at the onset and derive as much value as the raw data is able to provide. Analyze the outcome of the initial analytics and then introduce the necessary data quality, profiling and transformations to the portions necessary. Repeat this process in an iterative manner till the results are as expected.

The advantage of this approach is it nature of allowing you to start very quickly on an analytics solution and then progressively iterate over it till the commensurate level of prediction accuracies are obtained.

Some may call this approach as Agile Analytic Development. I am OK with that. Ask me before using that term though. I may patent it! :-)

Difference between Descriptive, Predictive and Prescriptive Analytics

Analytics as a discipline has matured beyond its age propelled by the new era in which the human-machine-network is more Instrumented, Interconnected and Intelligent. The new era has made data more accessible than ever before which can be leveraged by the science of analytics to not only increase the accuracy of future predictions but also to up the ante by one level and start optimizing the best outcome from a set of predicted possibilities.

Analytics has a maturity curve, or more of a roadmap which starts from Descriptive Analytics and works its way up to Predictive Analytics and ultimately to Prescriptive Analytics.

Descriptive Analytics, often called after-the-fact analytics, reports on what happened, the frequency of occurrences of a certain event or action and provide drill down capabilities to get to the root cause of the problem. It provides various reporting views based on user roles; summary views for the executive dashboards, metric views for the mid-level managers and drill down root cause analysis details for the engineers and domain experts. Descriptive Analytics is rooted in what is known as traditional BI reporting.

Predictive Analytics focuses on simulating what could happen in the future, given the conditions of the recent past and forecasting the next possible events if the current trend continued for a given period of time. Predictive Analytics is rooted in building supervised and unsupervised machine learning algorithms and models.

Prescriptive Analytics builds on top of Predictive Analytics and focuses on evaluating the various possible outcomes from predictive models and coming up with the best possible outcome by employing optimization algorithms. Such algorithms are also capable of factoring in the effects of variability. Prescriptive Analytics leverages stochastic optimization algorithms and models.

It is imperative to realize that there is no short cut for any enterprise to achieve the highest maturity levels in Analytics (i.e. Prescriptive Analytics) without developing a solid and sound foundation of descriptive analytics followed by predictive analytics.

Enterprises need also to realize that, just by virtue of being in the new era of instrumented, interconnected and intelligent human-machine-network does not give them a free ticket to accessing the data; the data that is required for analytics to be useful. A solid foundation of data access with key focus on ensuring the veracity and viscosity of the data is of superior quality is the very first step to reap the benefits of modern day analytic processing.

Data Virtualization - Virtualize more than Consolidate

Data consolidation continues to be a persistent IT challenge, a source of constant frustration and IT spend. The days of a full-time IT spend on continuous data consolidation on a ever moving target of data sources and data types should be over. Well, even if it is not that strong as "over", at a minimum enterprises should be seriously considering other alternatives. This is where Data Virtualization comes to the party!

Data virtualization focuses on an abstracted layer which provides the necessary hooks to take a business-centric query and deconstruct the same into a set of atomic queries. Each such atomic query, focuses on a sub-set of data elements/types (from the original business-centric query) and determines which data source(s) to go against to retrieve the data. Each such atomic query is executed by the Data Virtualization layer and the returned data sets are then processed (joined) to form the final consolidated result set which is then made ready to be returned as a result to the business-centric query. The mode of data return can be standard SQL, Web Services or any other format which is standard enough to be consumable by business and/or enterprise applications.

The technology is real today. It is only important that enterprises take a close look at Data Virtualization and consider leveraging the same as a part of their overall enterprise data architecture strategy.

And yes, the technologies today can virtualize across both structured and unstructured data spread across databases and schema-less file systems.

The Genesis of Big Data

Big Data, big data, big data! It is the hype that has taken the IT industry by storm. The term Big Data which has been formed from a combination of two of the simplest words - big and data, has, with their combination have had a profound impact. Enterprises are intrigued by Big Data and all of them feel that there is something in in it for them.

There is no doubt that data is grown and that too grown in huge proportions. If you think about it from a different angle, this data was already there. What has changed is that technology has now allowed enterprises to get access to this huge ocean of data. The fundamental shift is that enterprises traditionally had access to the structured data sets which were primarily generated from business transactions and internal business process executions. Such data were primarily resident in databases, data warehouses where they were captured in a well structured form. However, with the new era of social computing, of data feeds from myriad of sources that are external to the enterprise, the enterprise is all of a sudden exposed to the internet of things which were traditionally not under their control. The industry has come to the realization that such data has a profound impact to the way business are and will be run in the future.

The ability to capture customer sentiments, their desires, feelings, product feedback and intentions in real time, as they happen and be able to influence the next business action or decision is going to provide that competitive advantage which has the potential to make of break product brands and improve our lifestyle through real time up to date decision support system. Some examples of the following may be:

• Capturing a customer segment's negative sentiment and take prompt decisions to take corrective action
• Predicting customer movements e.g. commenting on making a move from one mobile carrier to another based on bad experiences
• Providing location-based product offers e.g. offering a $2 off on a subway sandwich if she is driving by a Subway sandwich store
• Informing rush hour travelers on the optimum route to take to their destination based on real-time data feeds from traffic surveillance cameras
• ...

the list is endless and each industry can come up with their own such list of untapped potential.

This non-traditional data does not follow the norms of database structures and designs; they are typically in the form of semi-structured textual data in social networks like Facebook, Twitter, LinkedIn or unstructured data from audio and video feeds.
The popular belief is that the combination of the semi-structured and unstructured data sets forms around 80% of the world's current data. The enterprises have realized that their future business decisions have been traditionally developed based on only 20% of the data (the structured forms) and the rest 80% is untapped!

This 4 fold increase of data, its sheer Volume and Variety based on the entire gamut of semi-structured and unstructured data is going to be a force to reckon with. Throw in the fact that the rate at which the non-traditional data is created is staggering and uncontrolled i.e. its Velocity has not been dealt with before. The IT industry has come to the realization that based on the sheer Volume, Variety and Velocity of the untapped data it is something very Big - a phenomena which our traditional technologies and infrastructure were incapable of handle. Big, in this context means that it is beyond the current comprehension. Data that is so Big that we have not seen before and have had a need to handle and process it. That is the genesis of Big Data!

Technologies are catching up with the 3 V's and we have started to realize that what was Big Data a year or two back may not be that big anymore now i.e. we are capable of handling it. It is important to understand that the term Big Data is temporal which means that what is big today may not and will not be 'big' tomorrow. Another important concept has unfolded recently. Although we have been able to come to terms with the volume, variety and velocity of the data, with this huge influx of data enterprises are faced with yet another challenge - how do I know that the data that I am gathering from non-conventional sources are indeed authentic and truthful? How do I believe in the Veracity of the data? So, add yet another V and now we get the 4 V's of Big Data Volume, Variety, Velocity, Veracity .

I strongly feel that the focus for us going forward is not about whether the data is big, medium of small; it is more about what we can do with the data. How can we empower the business with the next best decision and that with a level of confidence which can empower the executives to make decisions with confidence. Think about it! Let's talk about it another day.

Hadoop Usage Patterns

There has been enough talk about how exactly to use Hadoop in an enterprise. Many companies have jumped into the Big Data bandwagon with their first step typically being to get a Hadoop installation. The more serious players have started to realize some of the real values of Hadoop beyond the usual hype and have put their minds in using it to their business advantage. The more serious usage of Hadoop has resulted in the emergence of a few patterns. Although it is too early to stamp the 'best practice' authority seal on the emerging usage patterns, the usages have some merits which warrants some discussion.

The following are some of the emerging themes and patterns of Hadoop usage:

• As a data dumping ground - Enterprises today, who have not had their current state analytics leverage the entire gamut of data set, primarily the unstructured type, are provisioning all their ingested data from various unstructured and semi-structured sources into a Hadoop file system. Storage is not too much of an issue as commodity hardware is becoming the usage norm to provision very large data sets. Such large data sets once provisioned are made available for any subsequent analytics to be performed on the same.
• High speed processing - Traditional data warehouses were not built to support data analysis or querying on data volumes which breaks the terabyte or petabyte barriers. The SQL queries would still work but the time taken for them to return the result sets would typically be in inordinately large (e.g. the order of days). Although the data warehouse technologies are catching up one has to acknowledge the fact that the fundamental design premise of data warehouse were not optimized to handle such ultra large data sets. Data in a Hadoop file system can be processed at very high speeds. The MapReduce technology enables programmers to write massively parallel processing logic which makes the same queries, which takes in the order of days to return results, to execute in a matter of minutes or hours. This multiple order of magnitude improvement cannot get unnoticed and enterprises are using this usage patterns much more regularly and consistently.
• Storage of only the relevant unstructured data elements - Lot of enterprises have already started to ingest data from non-conventional (read it as - non-structured) data sources. However, they are aware that not all ingested data is of relevance to business decision making. In such scenarios, enterprises are deploying stream computing pre-processing before data is stored in a Hadoop file system. In these cases, data filtering algorithms are used on the real time data ingest. Such algorithms work on the deconstructed data sets and filter out the data elements which bear no importance to any analytical processing for the enterprise. The data elements which pass the processing filters are provisioned into Hadoop. Smart enterprises are keeping their storage and data management and maintenance costs down by employing such real time data filtering technologies.
• Perform data analytics on the entire data set - Assuming that the most relevant data elements have all their data points stored in Hadoop, enterprises are now running analytics on top of the huge data set. It is quite natural that such analytics (on the data set volume which was not available before) is yielding more insights into patterns which were hitherto unknown or untapped. Patterns are also evolving on how further insights can be developed when such unstructured data elements are correlated with the structured data which already resides in the data warehouses. Once such unstructured data elements are identified by employing sophisticated mathematical and statistical models, the identified data elements are further processed (cleansed, quality-checked, etc.) and then passed on to be stored in the data warehouses. This pattern has the advantage of adding only those data elements to the data warehouse which are enabling enterprises to develop better, more robust predictive models from the transactional data records in the data warehouses.

Other usage patterns for Hadoop will emerge and some will be hardened with repeatable success to be imprinted as best practices. Till then, we continue to innovate a multitude of ways to get the best usage out of Hadoop.

I encourage you to add to this repertoire of Hadoop usage patterns and let us create a compendium for the usage of Hadoop for Big Data Analytics!

Stream Computing (Streams) versus Complex Event Processing (CEP)

There is a general notion around IT professionals that stream computing (a.k.a. Streams) is just another buzz term for the traditional complex event processing (CEP). Although there are conceptual similarities between the Streams and CEP and acknowledging the fact that both of them fall under the analytical discipline of 'Continuous Intelligence', there are a few fundamental differences which put them into different leagues.

CEP is primarily used for analysis/analytics on discrete business events. Events are correlated in time using simple IF/THEN/ELSE logic. The events need not be of a single type or category. The data encapsulated in the business events are primarily structured in their form. The common CEP engines support modest data rates or around 10K messages/second with a latency typically in the 'seconds' range. The maximum data processing rates can scale up to around 100K events/second.

Streams on the other hand is designed to handle processing rates which are an order of magnitude higher than CEP. It can handle around millions of events per second with built-in linear scalability. Streams data sources are typically of a single event type e.g. camera feeds from traffic signals, sensor data generated from a pipeline or medical device, and so on. Streams is designed to handle the full gamut of unstructured data and contrary to IF/THEN/ELSE based logic in CEP, it is capable of performing advanced analytics on the data set. Examples of advanced analytics are only limited by the power of the mathematical and statistical models. Fast Fourier Transforms, Holt Winters algorithm, time series analysis algorithms would be some real world examples.

To summarize, although both Streams and CEP fall under the category of 'Continuous Intelligence', keep the following image in mind when any of your colleagues engage in the discussion:

Enjoy!

Combining Data At-Rest Analytics with Data In-Motion Analytics

Take a look at this short video first.

The video provides a sneak peek at the 'Art of the Possible' - how traditional analytics based on existing data in data warehouses and data marts can be combined with real time analytics based on streaming data feeds to develop a closed look continuous feedback improvement system.

The structured data set residing in the traditional data warehouses and the marts account for only ~20% of the worlds data. The rest 80% is the world of unstructured, ambiguous, naturally unrelated data set.

The fundamental premise of combining at-rest analytics with that of in-motion analytics is the following:
1. Leverage the wealth of existing data to develop statistical models which can be used to detect patterns on unknown data as well as predict future outcomes with a high degree of confidence and certainty.
2. Deploy such parametrized models to a stream computing environment where data comes in real time and is primarily unstructured i.e. textual data, video, audio and any other form of unstructured data feeds.
3. Allow the real time data, typically as single records or small data sets captured in short time windows, to be fed as parameters to the predictive models.
4. Allow the models to track the real time data feeds and provide real time predictions.
5. If the models cannot detect patterns and the amount of 'unknowns' rise over a certain threshold, then trigger a mechanism to recalibrate the original statistical model.
6. The statistical models will ideally leverage not only the existing data in the warehouses and marts, but also leverage the more current data and other relevant and related data from other sources. The expectations is the model to be able to predict more events and detect more patterns.
7. Deploy the recalibrated model back into the streaming computing environment and expect the models to detect more events that are happening in real time.



In a subsequent blog, when I find some time, I will explain how, what you saw in the video was implemented using a set of products and techniques.

Stay tuned!

And, as usual, drop me a note with any questions or topics on Big Data that you want to discuss.