Causal Models in Practice at Lyft with Sean Taylor

800 800 The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)

Today we’re joined by Sean Taylor, Staff Data Scientist at Lyft Rideshare Labs.

We cover a lot of ground with Sean, starting with his recent decision to step away from his previous role as the labs director to take a more hands-on role, and what inspired that change. We also discuss his research at Rideshare Labs, where they take a more “moonshot” approach to solving the typical problems like forecasting and planning, marketplace experimentation, and decision making, and how his statistical approach manifests itself in his work.

Finally, we spend quite a bit of time exploring the role of causality in the work at rideshare labs, including how systems like the aforementioned forecasting system are designed around causal models, if driving model development is more effective using business metrics, challenges associated with hierarchical modeling, and much much more.

Connect with Sean!


Join Forces!

  • Abba Taha

    Mr. Sean Taylor says: 22.27 that: “every model is a causal model” then he gives a few examples and says again that: “those all are causal questions.” Honestly, when was the last time Sean has read a respectable research on causality. Sorry to say, Sean you have no clue and you are just throwing stuff at us: people who are watching the webcast so that you seem smart.

    Please Sean be careful when you answer technical questions to technical audience like us.

    Note: I bet you Sam you will not dare to put this comment!

    • sam


      If you look around the site you’ll see that I don’t have any problem publishing controversial comments that add to the conversation.

      That said, I have to say I think your comment is lacking in substance and I wish you would have elaborated on it more. He is clearly laughing when he said “every model is a causal model,” then he says “or at least they should be” and even goes on to acknowledge he’s trying to be provocative by adding “…it’s a very strong perspective.” And you also don’t give any clue as to which of the examples Sean references as causal questions are not in fact causal questions and what in the research informs your view.

      Without additional details it sounds like name calling and disparaging a well-respected practitioner (who never claimed to be a researcher by the way) for unknown reasons, which is unfortunate.

      If you have any substance behind your comment I dare you to post it 🙂

      • Abba Taha

        Hello Sam and thanks for the reply,

        @ 22:44 Mr. Sean says:
        “Which driver we will dispatch to as a rider given your request to ride is a decision that we make and there are counter factuals around that decision. We could have dispatched that driver or other driver or we could have not dispatched a driver at all, because we do not have enough of them or because we need to allocate them in a scare way. THESE ARE ALL CAUSAL QUESTIONS and that is a micro level decision.”

        In my humble opinion these questions are not causal and are not answered by causal inference methods. Rather they are general decision-related questions that are usually answered by a combination of cost optimization and resource allocation methods. Similar for example, to questions often found in the airline industry – should we overbook this flight and if so how many extra seats we should overbook based on historical data and the number of no-show passengers as well as the cost/loss of over/under booking the flight.

        Causality and causal inference are totally different. They are based on finding in the first place causal relationships (cause-effect) then either doing an experiment or analyzing available data to confirm the strength of the relationships.

        Some of the famous causal questions: Does smoking increase chances of cancer? Does the COVID AstraZeneca vaccine cause cases of rare blood clots?


        • Sean Taylor

          Hey Abba,

          I agree the questions I’m discussing are not traditionally examples covered in causal inference, but they fit neatly into Pearl’s structural modeling framework and also the potential outcome framework first proposed by Rubin. I think you may be thinking about causal inference too narrowly. In the business context, we absolutely need to estimate the effects of choices we make, and making these choices are causes in both the causality and causal inference senses. The estimates produced by our causal models are *inputs* to decision problems, and yes ultimately we need to do some decision-making (either by humans or algorithmically) that is informed by our models, so there is always some an optimization step. But optimization requires a model of the world, and that model of the world should have causal interpretation if you’re going to be changing something.

          In some applications, like the one you mentioned about how much to overbook a flight, that optimization step is the focus. But there are consequences to overbooking that may need to be accounted for through causal models, for instance bumping a passenger onto a later flight is a negative experience that may cause them to book with your airline less often in the future. Estimating that effect would require a causal model.

          The famous causal questions you list are just special cases — they happen to be in medicine, are tied to specific interventions, and are binary hypotheses. They are just scratching the surface of the kinds of questions answered by causal inference. I’m expanding that to covering interventions we make in the business I model, where I care about specific estimates of their effects (not hypothesis testing) because we need to make nuanced tradeoffs among a variety of alternative courses of action.

          I appreciate the healthy criticism here, it’s valid to ask me to justify such a strong claim, but I think it’s pretty unfair to claim I have no clue.


Leave a Reply

Your email address will not be published.