ML does not just add another feature as an after-thought but is an entirely new architecture that compels us to completely rethink our relationships with our products and how we conceptualize, design, and develop them.
What makes an ML-first Product Manager (PM)? How do you develop the mindset, the skills, the principles, and the experiences to become an ML-first PM? You can also build the next big ML-first idea by learning the craft, developing intuition, and understanding the levers.
The world around us is changing with Machine Learning. Users’ relationships with products are changing. Expectations that users have from our products are changing drastically.
Product Managers have to rethink their entire approach from conceptualizing, designing, and developing products in an ML-first world, which is very different from the rule-based approach of building products.
Some of us who went through the cycle of mobile-first products understand that a new paradigm comes with an entirely new set of opportunities, assumptions, and constraints. PMs who learned to retrofit their desktop-based products for the mobile-first world survived okay, but, those who thrived looked at this new world from an entirely new lens, e.g., Uber, Instagram, Snapchat, and Facebook.
ML is powering completely new experiences that were not possible earlier. e.g., Alexa, Siri, and Google Assistant are creating voice-based accessible experiences that were sci-fi just a few years ago. TikTok has a feed that is entirely curated by ML. It gives creators a platform, which genuinely tries to make the discovery of content democratic. Anybody can get that explosive virality for their short-form videos that were not possible earlier.
If you are embarking on becoming an ML-first PM, here is some advice for you.
Develop your intuition
Unlearn the past
The traditional way of developing software products was rule-based and procedural. It works great with a very well defined problem, including a finite space of input data and a very well defined output. ML algorithms work very differently. You need to unlearn some of the instincts that you most likely developed over all these years, building rule-based systems.
For developing a strong intuition for ML applications, you need to think in a way that includes looking at the data very differently. As Mukesh, who is Applying ML techniques to build next-generation personalization, search, discovery, navigation, and Alexa voice experiences at Amazon, explained in this post, you need to develop an eye to determine which problems are a good fit for applying ML.
As you go through living your life, keep your eyes and senses open to appreciate products where ML creates successful experiences, or, contexts where ML can be applied effectively.
As you experience digital products like Netflix, Amazon, Facebook, and Google Search, look for signs where the experience is hyper-personalized for you or announces “Recommendations for you.” Pay closer attention to how the experience evolves as it learns and gets better.
Learn to trust the machine
It is natural to insert our biases into the objective function and parameters while tuning. It takes time to start trusting the machine and to let it do its job to learn and improve over time with the data and its real inferences.
Learn to think in loops
Human beings are in a constant loop of learning. We do that from our birth and continue throughout our lifetime. ML algorithms and their data have lives of their own, and we need to develop intuition to think in completing the circles of data collection, curation, experimentation, and deploying inferences.
Learn the craft
There are masters of the trade like Tomer Cohen, from LinkedIn, who are great teachers for the basics of the craft. He recently released a course titled “Becoming an AI-first product leader.” Tomer led the development of, amongst many other AI-driven experiences LinkedIn, the very relevant news-feed.
Udacity has an Artificial Intelligence Nanodegree program. Peter Norvig and Sebastian Thrun are the best in the industry to teach the basics of Artificial Intelligence and Machine Learning. It takes substantial efforts to go through this nanodegree, but it explains the engineering concepts in a way that is suitable for Product Managers as well.
Understand the levers
There are primarily three things that, as a product owner, you need to decide.
Lever#1: Objective function
An objective function is a goal that you specify for your program.
Understand how to select a useful objective function that your ML algorithm tries to achieve.
Be mindful of the unintended consequences of your decisions to choose a specific objective function. e.g.
Be aware of the biases that creep into the objective function.
Great objective functions are not singular, but a composite of many objectives rolled into one. An example is for Amazon; it is not merely enough to optimize the number of orders placed. The additional objective there is also to make sure there are minimal returns, or customers sign up for Amazon Prime.
Lever #2: Types of algorithm
There are different types of algorithms to be applied to different types of problems. There are three types of algorithms that you need to learn.
Supervised learning algorithms
Supervised learning includes providing a well-defined data set which is pre-tagged to the algorithm. It requires the availability of clean training data, which is curated and tagged beforehand.
We can group supervised learning into classification or regression problems, e.g., classification for images into different species of animals or regression values such as weight or price forecast.
Unsupervised learning algorithms
Unsupervised algorithms learn from input data that is not pre-labeled. It is closer to how human beings learn.
In this case, you only have your input data and no corresponding training data. The goal here is to come up with a model that translates patterns in your input data without any interference.
Unsupervised learning algorithms are relevant when the training data is not available or very expensive to generate. An example of unsupervised learning includes recommendation algorithms, e.g., people who buy X also tend to buy Y.
Reinforcement learning algorithms
The primary difference in reinforcement learning algorithms from the previous two is the presence of actions that the machines take and then learn from trial and error as the desired outcomes are optimized.
An example of reinforcement learning is AlphaGo Zero, that learned the game as it played against itself.
Lever #3: Data collection and curation
As specified in our previous post, understand how data wrangling is critical for your ML project.
One critical kind of data available to software product owners is the user activity stream. Every action by the user collected and curated provides a powerful way to understand and personalize preferences for the users.
Be careful about not overlooking negative signals from the users, which are equally important as the positive signals.
The chicken-and-egg problem of coming up with data sets is real. You need to be very creative about how to overcome the cold-start problem and build your seed dataset.
Your ML product is just as good as the data fed into it. If you input bad data, you get bad results.
Key Takeaways to Remember
- Develop your intuition for greenfield ML opportunities by unlearning the rule-based past, being curious about ML-based applications, learning to trust the machine, and thinking to build loops.
- Learn the craft by doing courses like Udacity’s AI nanodegree.
- Understand the levers that you have at your disposal with a deep understanding of objectives, algorithms, and data.
Let us know how you evolved as an ML-First PM. We plan to elaborate on some of these tactics in-depth in our upcoming posts.