3 ms·

Data Science is a poor term (unfortunately, the right one, statistics or statistical modeling, was deemed to be "unsexy"), but the expertise is certainly very r

Data Science is a poor term (unfortunately, the right one, statistics or statistical modeling, was deemed to be "unsexy"), but the expertise is certainly very relevant in today's data-driven businesses. Think predictions, forecasting, clustering, A/B experiments, recommender systems. Yes, some parts of what's commonly called Data Science can also be called Machine Learning, then maybe Deep Learning can be called AI, and in the end it seems that whatever choice we make with definitions, it is largely a mess anyway. And, on the top of that, the definitions are situational. There is certainly a lot of busy work in DS like there is a lot of busy work in any non-critical pockets of operations, software engineering included. I don't know what "over-optimization" means. If it means that, for instance, that the accuracy provided by the models is greater than the business needs, sometimes that's true and I agree. But are other parts of the business at fault for not being, broadly speaking, responsive enough or it is an "over-optimization" problem?

jollybean 4y ago
'Over optimization' means that a lot recommendation systems, A/B tests, forecasts in business are nearly irrelevant, or rubbish i.e. a lot of cost for very marginal value add. Such a small margin that it's within variance of what might even be considered a positive outcome. There are few business that have the kind scale in which a lot of that will be hugely useful (at high degrees of sophistication), let alone, afford the teams of researchers. Most of the optimizations are low hanging fruit which can be achieved via some kind of basic algorithm. I believe you could nuke Snowflake and all their giant lakes of historical data and the world would not skip a beat.
- borroka 4y ago
  This is take I don't agree with on a real problem, which is the usefulness of data science in modern business, and for which processes in the business. I try to explain. Recommender systems add tremendous value, millions when not billions in revenue, to any retail business, streaming business. Is it by chance that all the online retailers had to set up a recommender system? A/B tests are not needed? If A/B testing and actions taken looking the results would not be useful, don't you think that business would know it? Forecasts in business are nearly irrelevant, or rubbish, you say. Like in the sense that in complex systems is irrelevant to know what's going to happen next, that is better just to play it along at best? Having worked in companies in media, retail, healthcare, and telco for which forecasting is crucial for planning the use of resources, negotiation of contracts etc., this comment is off-base. Going from abstract to practical, I have worked recently with one telco giant on one of the services they offer. Their forecast of subscribers done by drawing a straight line that follows the latest data is making them leave tens of millions on the table each year. I proposed to make some A/B testing on their "coupon" strategy and they were not interested, preferring to send a "coupon" to everybody and their grandfather because they did not want to study for a week how to set up an experiment. It is mostly laziness. Clearly to know if A/B testing is working or not (say increase engagement, revenue etc.), one has to try it. But they don't, because they are mentally lazy. One thing I agree with is that most legacy business are not ready to adopt a data-driven modern way of doing business, due to stiff processes, old-timers that would like to go on with their business like they were doing 30 years ago, and general fear of changing. The fact that the legacy businesses are still "working" may give the impression that all these data-driven "baloney" is subtracting rather than adding to the business. Why do we need forecasts, they say, I cannot believe that these people want to use data, isn't our expertise enough? Don't you see we are the still among the 5 top companies that do what we do? But their position in the market is not due to sound business practices, but due to their dominant position in the market. But then the dominant position goes away and the old-timers finally recognize that A/B testing, forecasting etc. combined with domain-knowledge are a plus for the business. Maybe you don't need 1,000 DS, maybe you need 20 (practitioners, not "researchers"), that depends on the business. Clearly the mom and pop business does not need any data science.

jollybean 4y ago

'Over optimization' means that a lot recommendation systems, A/B tests, forecasts in business are nearly irrelevant, or rubbish i.e. a lot of cost for very marginal value add. Such a small margin that it's within variance of what might even be considered a positive outcome. There are few business that have the kind scale in which a lot of that will be hugely useful (at high degrees of sophistication), let alone, afford the teams of researchers. Most of the optimizations are low hanging fruit which can be achieved via some kind of basic algorithm. I believe you could nuke Snowflake and all their giant lakes of historical data and the world would not skip a beat.

borroka 4y ago
This is take I don't agree with on a real problem, which is the usefulness of data science in modern business, and for which processes in the business. I try to explain. Recommender systems add tremendous value, millions when not billions in revenue, to any retail business, streaming business. Is it by chance that all the online retailers had to set up a recommender system? A/B tests are not needed? If A/B testing and actions taken looking the results would not be useful, don't you think that business would know it? Forecasts in business are nearly irrelevant, or rubbish, you say. Like in the sense that in complex systems is irrelevant to know what's going to happen next, that is better just to play it along at best? Having worked in companies in media, retail, healthcare, and telco for which forecasting is crucial for planning the use of resources, negotiation of contracts etc., this comment is off-base. Going from abstract to practical, I have worked recently with one telco giant on one of the services they offer. Their forecast of subscribers done by drawing a straight line that follows the latest data is making them leave tens of millions on the table each year. I proposed to make some A/B testing on their "coupon" strategy and they were not interested, preferring to send a "coupon" to everybody and their grandfather because they did not want to study for a week how to set up an experiment. It is mostly laziness. Clearly to know if A/B testing is working or not (say increase engagement, revenue etc.), one has to try it. But they don't, because they are mentally lazy. One thing I agree with is that most legacy business are not ready to adopt a data-driven modern way of doing business, due to stiff processes, old-timers that would like to go on with their business like they were doing 30 years ago, and general fear of changing. The fact that the legacy businesses are still "working" may give the impression that all these data-driven "baloney" is subtracting rather than adding to the business. Why do we need forecasts, they say, I cannot believe that these people want to use data, isn't our expertise enough? Don't you see we are the still among the 5 top companies that do what we do? But their position in the market is not due to sound business practices, but due to their dominant position in the market. But then the dominant position goes away and the old-timers finally recognize that A/B testing, forecasting etc. combined with domain-knowledge are a plus for the business. Maybe you don't need 1,000 DS, maybe you need 20 (practitioners, not "researchers"), that depends on the business. Clearly the mom and pop business does not need any data science.