Category: Machine Learning

While everyone is focused on creating the most advanced algorithms for their predictive analytics and optimizing your team’s performance, I have not seen anything on how to manage your algorithms. A great article in Harvard Business Review – Algorithms Need Managers, Too by Michael Luca, Jon Kleinberg and Sandhil Mullainathan – does a great job of combining the two issues and providing a solution.

The authors begin by pointing out most businesses rely on predictions throughout their organization. The decisions can range from predicting a candidate’s performance and whether to hire them, what initiatives will have the highest ROI and what distribution channels will yield the most sales. Companies increasingly are using computational algorithms to make these predictions more accurate.

The issue is, if the predictions are inaccurate (and although they are computer generated, they are still predictions not facts) they can lead you into bad decisions. Netflix learned this the hard way when its algorithms for recommending movies to DVD customers did not hold when its users moved to streaming. More relevant to digital marketers, algorithms that generate high click through rates may actually bring in poor users not interested in your underlying game or product. As the authors write, “to avoid missteps, managers need to understand what algorithms do well – what questions they answer and what questions they do not.”

How algorithms can lead you amiss

An underlying issue when using algorithms is that they are different than people. They behave quite differently in two key ways:

• Algorithms are extremely literal, they do exactly what they are told and ignore any other information. While a human would understand quickly that an algorithm that gets users that generate no revenue is useless, if the algorithms was just built to maximize the number of users acquired it would continue attracting worthless users.
• Algorithms are often black boxes, they may predict accurately but not what is causing the action or why. The problem here is that you do not know when there is incomplete information or what information may be missing.

Once you realize these two limitations of algorithms, you can then develop strategies to combat these problems. The authors then provide a plan for managing algorithms better.

Be explicit about all of your goals

When initiating the creation of an algorithm, you need to understand and state everything you want the algorithm to achieve. Unlike people, algorithms do not understand the implied needs and trade-offs necessary often to optimize performance. People understand the end goal and then backward process how to best achieve that eventual goal. There are also soft goals to most initiatives, and these goals are often difficult to measure (and thus input into your algorithms). There could also be a goal of fairness, for example a bank using an algorithm to optimize loan behavior may not provide enough loans in areas where it feels a moral obligation to do so. Another example could be where you may want to optimize your business units sales but the behavior could negatively impact overall sales of your company.

The key is to be explicit about everything you hope to achieve. Ask everyone involved to list their soft goals as well as the primary objective. Ask people to be candid and up-front. With a core objective and a list of concerns in front of them, the algorithm’s designer can then build trade-offs into the algorithm. This process may entail extending the objective to include multiple outcomes, weighted by importance.

Minimize myopia

Algorithms tend to be myopic, they focus on the data at hand and that data often pertains to short-term outcomes. There can be a tension between short-term success and long-term profits and broader corporate goals. People understand this, computer algorithms do not.

The authors use the example of a consumer goods company that used an algorithm to decide to sell a fast-moving product from China in the US. While initial sales were great, they ended up suffering a high level of returns and negative customer satisfaction that impacted the brand and overall company sales. I often see this problem in the game industry, where product managers find a way to increase in-app purchases short term but it breaks player’s connection with the game and long-term results in losses.

The authors suggest that this problem can be solved at the objective-setting phase by identifying and specifying long-term goals. But when acting on an algorithm’s predictions, managers should also adjust for the extent to which the algorithm is consistent with long-term aims.

I recommend using NPS to balance out short term objectives with the long-term health of the product and company. I have written before about NPS, Net Promoter Score, which is probably the most powerful tool to measure customer satisfaction. It is also highly correlated with growth and success. By ensuring you keep your NPS high, you are providing a great way to look holistically at the success of specific initiatives.

Chose the right data inputs

Using the right data can make your algorithms much more effective. When looking at a game like Candy Crush, you can create levels by looking at when people abandon the game and decompose the levels before abandonment. However, by adding social media posts to the your data, you could get a more holistic view of which levels players are enjoying and thus build a more compelling product.

The authors also point to an example with the City of Boston. By adding Yelp reviews to what health inspectors use to determine what restaurants to inspect, they were able to maintain their exact same performance but with 40 percent fewer inspectors. Thus, the new data source had a huge impact on productivity.

The authors point to two areas of data that can improve your algorithms:

• Wider is better. Rather than focusing on more data, the amount of data you know about each customer determines the width. Leveraging comprehensive data is at the heart of prediction. As the authors write, “every additional detail you learn about an outcome is like one more clue, and it can be combined with clues you’ve already collected. Text documents are a great source of wide data, for instance; each word is a clue.”
• Diversity matters. Similar to your investment strategy, you should use data sources that are largely uncorrelated. If you use data that moves closely to your data sources, you will have the illusion of using multiple data sources but really only be looking at one angle of the data. If each data set has a unique perspective, it creates much more value and accuracy.

Understand the limitations

As with anything, it is also critical to understand the limitations of algorithms. Knowing what your algorithm will not do is equally important as understanding how it helps. Algorithms use existing data to make predictions about what might happen with a slightly different setting, population, time, or question. “In essence, you are transferring an insight from one context to another. It’s a wise practice, therefore, to list the reasons why the algorithm might not be transferable to a new problem and assess their significance,” according to the authors.

As the authors point out, “ remember that correlation still doesn’t mean causation. Suppose that an algorithm predicts that short tweets will get retweeted more often than longer ones. This does not in any way suggest that you should shorten your tweets. This is a prediction, not advice. It works as a prediction because there are many other factors that correlate with short tweets that make them effective. This is also why it fails as advice: Shortening your tweets will not necessarily change those other factors.”

Use algorithms, just use them smartly

This post is not intended for you to avoid using algorithms, it is actually the opposite goal. Algorithms are increasingly powerful and central to business success. Whether you are predicting how consumers will react with a feature, where to launch your product or who to hire, algorithms are necessary to get great results. Given the central importance of these algorithms, however, it is even more crucial to use them correctly and optimize their benefit to your company.

Key takeaways

1. Algorithms are increasingly powerful and central to business success. Given the central importance of these algorithms it is even more crucial to use them correctly and optimize their benefit to your company.<
2. Problems with algorithms result from them being literal (they do exactly what you ask) and are largely a black box (they do not explain why they are offering certain recommendations).
3. When building algorithms, be explicit about all your goals, consider the long-term implications and make sure you are using as broad data as possible.