Data science Aspirants!!
Data scientist was named as the ‘sexiest job of 21st century ‘ by glassdoor and it is no joke indeed. Since, the statistics suggest that there has been a 56% increase in the trends of this field. A data scientist, in very general terms, looks at and investigates a set of data and then comes up with different ways to answer previously posed questions. Along the way, the data scientist may consider historical data analysis to develop analytical models and dashboards that provide insights and improve decision making.For example, a data scientist for a large retailer like ‘ Crocs ‘ may look at not just past seasons’ data, but current economic and other conditions to make predictions for their upcoming season. Retail executives make use of this to improve things such as sales, revenue, marketing strategies, advertising efforts, etc. So start your journey today itself without wasting any more time . Let us first of all know some interesting facts about data science:
11 Facts about Data Science that you must know
However, as one journeys through his/her career in analytics, some truths start becoming evident over time. And while none of them are ground-shattering, they often surprise novices in the field. So, it’s worthwhile to know 11 absolute facts of data science.
1. Data is never clean
Analytics without real data is mere collection of hypotheses and theories. Data helps test them and find the right one suitable in context of end-use in hand. However, in real world data is never clean. Even in organizations which have well established data science centers for decades, data isn’t clean. Apart from missing or wrong values, one of the biggest problems refers to joining multiple datasets into coherent whole. Join key may not be consistent or granularity or format may not be suitable. And it’s not intentional. Data storage enterprises are designed and tightly integrated with front-end software and user who is generating data, and are often independently created. Data scientist enters the scene quite late, and often is just “taker” of data as-in and not part of design.
2. You will spend most of your time cleaning and preparing data
Corollary to above is that large part of your time will be spent in just cleaning and processing data for model consumption. This usually annoys people new to industries. With brilliant mind bursting with sophisticated machine learning methods, spending three-fourth of the time with just data wrangling seems waste of talent and time. Often this leads to dissatisfaction and lack of attention – errors from which can come to bite even the most fanciest of the algorithms. If you cannot do this with equanimity and focus on big picture, then perhaps you should aim for research in statistics rather than career in data science.
3. There is no full automated data science. You need to get your hands dirty
Since data is not clean and requires quite a lot of data processing, there is no ready set of scripts or buttons to push to develop analytic model. Each data and problem is different. There is no substitute for exploring data, testing models, and validating against business sense and domain experts. Depending on problem and your prior experience, you may dirty your hands less, but dirty you will. Only exception is if you get data in specific format and do the same thing over and over, but that already sounds boring, isn’t it?!
4. 95% of the tasks do not require deep learning
95% is obviously a made up number – but the idea is that most real life problems don’t require advance analytic capabilities. Solving real-world problems involves lot more understanding real-world, problem domain, decision makers and end-users, than understanding latest and greatest discovery in statistics. What moves the needle, and moves it quick, is much more valuable than what is rigorous and pure. Often, simplest models like linear regression, logistic regression, and k-Means clustering work wonders as long as problem is well formulated. Even for complex problems, simple models can provide large gains which complex models can only improve marginally. That is not to say that complicated models have no place. In fact, depending on money riding, 0.1% increase in prediction accuracy may be worth millions of dollars.
5. Big Data is just a tool
With the hype around Big Data getting louder every day, I won’t blame you for being enamored of the idea. However, key thing to remember that Big Data is just collection of tools to work with large volume of data in reasonable time and with commodity grade computer hardware. Underlying analytic problem design, modeling best practices, and scrutinizing eyes of astute analyst aren’t replaceable with Big Data. That is not to say that competency in Big Data techniques isn’t handy – it is, more so since world is moving towards Big Data and there may not “Small” Data in couple of years anymore. But tools will come and go; your machine learning experience will only persist. Big data is like analogous to AK47 rifle forpoliceman rather than flintlock carbine rifle. Sure, better tool is preferable to inferior, but being trained in policing is more important than rifle.
6. You should embrace the Bayesian approach
Data science is sequence of hypotheses testing. You have to have going-in belief which you want to prove right or wrong based on observation from data. Stronger is your going-in belief, more counter-evidence you need to prove belief wrong. That, in essence, is Bayesian approach. But while proving your hypothesis right through data is important, proving alternative hypothesis wrong is also equally important. Take this fun puzzle from New York Times to figure out how to think Bayesian.
Alternative to Bayesian thinking is to let your data tell you stories. This can be problematic because sliced and diced some way, data will always tell a story. But without a-priori belief, story may not be true in reality. This is often case of hindsight bias and poor research (and often staple of motivational and self-help books). If you want to find differences in two groups (successful business versus non-successful, athletes versus slobs, rich versus poor), you can always find some. There are hundreds of thousands of human characteristics that some will come out different just by chance. That doesn’t mean that those characteristics made someone different from others. On the other hand, if you have reasonable hypothesis about what could be causing difference, you can verify if you are right or not. In the end, either you explain results from model based on your understanding, or you modify your understandings. There is no point saying that length of nose-hair is predictive of income of person in year fifty because model says so.
7. No one cares how you did it
Consumers of data science models are decision makers and executives, and they want workable and useful model. While it’s tempting for data scientists to explain technical expertise behind the model and show-off the analytic rigor, this is often counter-productive. Your audience cares about outcome and end-use and isn’t bothered about the decision engine you have put together. In fact, complicated explanations about mathematics of model are sure way to bore your users and intimidate against use. Save your expertise with technical discussions among your data science peers.
8. Academia and business are two different worlds
This applies to almost all disciplines and analytics is no exception. Focus in academics is on discovering new methods and proving new theorems. Focus in business is on solving a problem and making money. Doesn’t matter if analytics behind the solution is fancy or not, and no one cares about that anyway. Speed is often of more essence than accuracy. Every business analytic solution should solve a real-life problem and directly or indirectly should contribute to bottom line.
9. Presentation is key
Since end-user and decision maker is often non-mathematical person, selling an analytic solution isn’t different from other sells. You can sell on quality – analytic accuracy – but you can also sell on emotions, aesthetics, story, human angle, and money. Being able to explain your method in simple terms and align with end-users’ interest is art that all data scientists who wants to make significant non-theoretical mark on world must master. At least for a while, that means, story-telling through PowerPoint should remain key weapon in your arsenal.
10. All models are wrong, but some are useful
Models, by definition, model some ‘truth’ in the world. Since world is infinitely complex (think Quantum Mechanics!), models are approximations of reality. Some models are more wrong than others, but all are wrong. However, they can be, and often are, useful since they are better than alternative of no model and no prediction. Realizing what we are aiming for and what we are competing against can be important in shaping our analytic design process – and checking our egos.
11. Just because analytic model is great doesn’t mean it will see light of day
As fun as data science is, there is more to the world than your analytical model. If you see about a third or more of your work getting implemented or used then consider yourself lucky. Notwithstanding analytic capabilities, analytic project get shelved for various reasons all the time, including, data changed, problem changed, no one interested in solution, implementation too expensive, benefit not worth the cost, someone else did it first, and solution too advanced for its time. Be calm and carry on.
I realize that perhaps there are more than 11. And perhaps some of these could be clubbed together. Point is not about counter, but about importance of internalizing these realities of industry we want to be part of. Difference companies and industries might be at different spectrum of these facts, but collectively knowing and understanding these ‘facts’ will make one a more satisfied, broad minded, and better data scientist.