Strategy for Data Scientists
One of the key ways military strategy has been taught in the US is according to the formula ‘Strategy = ends + ways + means’, with future military leaders using this summary to analyse past commanders’ campaigns.
Using this formulation provides an ability to see analyse how strategies are varied according to the objectives.
Art Lykke offers some examples of military objectives in his article discussing ways and means. They include ‘defend a homeland’, ‘restore territory’ and ‘deter aggression’. It is intuitive that, for example, ‘deter aggression’ could be achieved by projecting an appearance of powerful forces without firing a live round, but ‘restore territory’ will require incursion into an area currently held by the enemy.
For data scientists, it is instructive to pay attention to the difference between objectives and the means to achieve them. The number of times the word ‘visualisation’ is used in articles for data scientists (referring to tools for producing graphs — a means of communication) compared to the word ‘communication’, or, even more so, words like ‘convince’ or ‘persuade’ shows the focus on ways and means amongst data scientists over the ends they serve. Even more telling are articles, like this, where data science tools — the means to execute a data science strategy - are presented without a discussion of the objective they could be used to achieve.
Frank Harrell, in Regression Modeling Strategies, offers a number of strategies that effectively imply two orthogonal axes — one measuring predictive ability from none to superior, and the other measuring ability to explain the role of inputs from none to superior. He also suggests complementary strategies, where effectively separate models are created for prediction and inference.
This analysis provides low level objectives (‘build a model that predicts well even if it can’t be understood’ or ‘build a model that can be understood, even if other models could be better predictors’), but the higher order objectives aren’t in view. Note that Harrel’s text offers fully developed strategies, in that after presenting objectives, he then outlines a method to achieve those objectives.
This highlights another way that miltary strategies have elements in common with Data Science strategies — it has been recognised that military means are often used in support of political objectives. This was starkly shown in the famous exchange at the end of the Vietnam War between US Colonel Summers and North Vietnamese Colonel Tu — when Colonel Summers observed ‘ You know, you never defeated us on the battlefield’, Colonel Tu retorted ‘True, but also irrelevant’.
In designing strategies for Data Science, the danger is always there that a well executed model that performs in excess of its designers most optimisic expectations, completely fails to increase sales or profits, or reduce costs, or contribute in any other meaningful to the business’s goals.
More insidiously, a Data Science strategy can be successful initially, before becoming an albatross due to political blowback, or failing due to a loss of user confidence.
In the tech world, Facebook are an example of a company that risk having an experience analgous to the USA’s Vietnam experience — they have achieved their ‘military’ objectives but are now at risk politically. This week Facebook and Mark Zuckerberg have been in the news as Zuckerberg appears before the US Congress.
Zuckerberg made reference to developing an AI system to detect hate speech within a 10 year time frame. His objective here is to win back the trust of both users of Facebook, and that of other stakeholders, such as the lawmakers. However, while this strategy has the features of both a clear objective and means of achieving that objective, those means incompletely meet the objective — developing the technology to screen out hate speech on its own won’t win back people’s trust. To do that Facebook will need a complete strategy to ensure people believe they can Facebook can protect them from hate speech — a communication or persuasion strategy.
Data Science teams everywhere need to fully understand their objectives in order to fully develop an effective strategy. That will often mean applying a defensive strategy — one that anticipates threats to its own success. Certainly, the meaning of success will need closer attention and deeper thought.
Check my book in progress The Lazy Data Scientist out on leanpub!
