Diese Präsentation wurde erfolgreich gemeldet.
Wir verwenden Ihre LinkedIn Profilangaben und Informationen zu Ihren Aktivitäten, um Anzeigen zu personalisieren und Ihnen relevantere Inhalte anzuzeigen. Sie können Ihre Anzeigeneinstellungen jederzeit ändern.

Prediction machines

30 Aufrufe

Veröffentlicht am

How to understand the impact of Machine learning on businesses like Booking.

Veröffentlicht in: Daten & Analysen
  • Als Erste(r) kommentieren

Prediction machines

  1. 1. Agrawal & al. • Simple economic model • Machine learning is making Prediction cheaper • Components of prediction
 will become more valuable • Judgment & Action • Many conferences on video
 https://www.youtube.com/ playlist?list=PL-h49ivT1t6wC- zoj1EeLuFcnZ-uBRVYW 1 This presentation is about Machine learning, but I will not talk about machine learning at all. This is about what’s around machine learning, and will be affected by it. What we can predict, even if we don’t know what models will be like in the future, will happen to companies who try to use machine learning. This presentation is mainly taken from the recent book by three economists, Agrawal, Gans & Goldfarb. They have worked on how technology change the structure of the economy for twenty years now. They have made very interesting arguments in the past, notably around the internet “new economy”. The idea at the time was that the rules had changed dramatically. They argued: the rules haven’t changed, but some things, things computer could do, had become cheaper. Dramatic price changes have unexpected direct impact, but they do something very predictable: the increase the value of their complements. That’s how they predicted the dot-com boom twenty years ago. That’s how they can predict some of the future impact of machine learning now. They argument is very similar. Machine learning doesn’t change the rules of economy, it does one thing: it makes predictions cheaper. Because of that, complements of prediction, judgment and action, become more valuable. That concept of complements is key and we’ll look into it in mode details. Before I dig into that, the authors have talked about their own work far more and far better than I can, so if you have the time or if I’m not clear, please look into their own presentations. They made at least a dozen, from a two minute TV appearance to an hour-and-a-half talk with questions at Google. I’ve listed them all on that link.
  2. 2. Decision model Customer calls
 sounds unhappy Want to cancel A price rebate might keep
 them on board Suggest a
 price rebate Free insurance
 too, but less so Customer accepts Booking still
 make a profit 2 The key concept that they introduce is a decision model. Everything we do is part of a decision process. That process has a four steps. - First, you are given a context, an input, say, a customer call; he is unhappy and want to cancel an order; - Then you have to make predictions: you actually probably make several. You try to estimate if that customer is price-sensitive, if they want more from you. You anticipate their reaction to some actions you can take. The definition that they give of a prediction is: estimating something you don’t know (the outcome) based on something you do know (the input). - You them optimise for the consequence and take an action, based on those prediction. In out case, a commercial gesture, a price rebate. - Customer’s reaction defined the outcome: whether you were able to keep them on-board, how much that rebate ate your profit. Does that make sense so far? Should be fairly straightforward: they are just labelling things that are trivial.
  3. 3. Decision model Booking
 made at a loss Customer remains engaged 3 The truth is: there is a fifth element in that process: your judgement, how you evaluate the outcome. You don’t just need that prediction, you also need to evaluate how valuable is each outcome to make a decision. You need to say how valuable, in terms of profit, is to have a customer remain on board, if you are OK with loosing money for that.
  4. 4. Decision model A price rebate might keep
 them on board Free insurance
 too, but less so 4 The sixth aspect of that model is that your prediction doesn’t come from nowhere: every time you go through that process, you remember what happened. Those predictions can then use details in the input with more nuance. This trains your ability to make more accurate predictions. Those are the six components of a their decision model.
 Nothing too technical so far?
  5. 5. Economic complements • Coffee & milk • Hotels & flights & rides • Websites & Search engine 5 The last thing that they mention is the idea of economic complements. When things are cheaper, you buy more of them. That’s simple. But you typically don’t use product in isolation. You use them in combination with others. So if the price of, say, coffee goes down, you might buy more of it. But if you always put milk in your coffee, you will probably buy more milk too. This means that, if the price of coffee goes down, you buy more milk. That makes milk a complement of coffee. There are many cases with complements. Actually, almost every aspect of the economy has them. One that you will be familiar with: accommodation and travel are complements, so if the price of flights go down, people travel more and book more hotel rooms. Same thing for taxi rides, or car rental. That idea of complements is what is behind the GEM strategy. One big example is web search engines. They made finding a website very easy, cheap on your time. So when we started having cheap search engines, websites became very valuable. There were websites for decades before the 1990s, but people couldn’t really find them, so there were niches. But, starting from 1995, they became a must-have for any company, hence the dot-com boom.
  6. 6. Cheaper arithmetic • Computer do arithmetic • Dramatic cost reduction 6 But that last example is part of a wider process: it’s not just search engines, but a whole class of operation that has been becoming cheaper for the last 50 years, much cheaper: computers. The trust is, computers do just that: compute. They do nothing but arithmetic. Unless you care about adding things, that shouldn’t really impact you. And for the first decades, it didn’t really impact many people outside of those that had complex arithmetic problems to solve: accountants, or people trying to launch canon balls, or missiles.
  7. 7. Augmented by
 Cheaper arithmetic • Accounting & reporting • More crucial job: planning • Forecasts then automated 7 But things changed significantly, first for accountants. In the 50s, they had to learn how to add very effectively because that was the key part of their job: add all the expense reports, and the bills, and the sales, and the salaries to figure out how much money they company had left. They literally trained by picking up the phone book and adding phone numbers. That was boring and computers took that part of the job over very early, around the 60s. Luckily for the accountants, they was one part that computers didn’t do really well at that time: present that information to decision-makers, make informed comments, predict. So they moved from adding numbers to explaining financials. And that’s why now you have a class of people who talk about cashflow, risk, forecast and are not adding up phone numbers anymore: they computer does that for them. They were able to focus on more interesting parts: - principle, optimisations - predicting consequences They were augmented by computers.
  8. 8. Replaced by
 Cheaper arithmetic • Digital cameras • Photo-editing: lasso, filters • Transcription • Spelling, grammar, style • Natural language processing • Voice as signal processing 8 But not every industry is so lucky: silver-film based photography was progressively
 replaced by cameras based digital-sensors. Why replaced and not augmented? Because cheap sensors and cheap arithmetic was able to replace all the features of expensive cameras, not leaving them with a lot of things to do. All the laborious development process could be done in a fraction of a second by computers who treated images like a set of numbers, and could blur an image by averaging colours, or sharpen it by computing differences, finding a contour of a person with more arithmetic. Rapidly computers could do a lot more with images as numbers than cameras and professional human editors could. Some editors were augmented. Some were replaced. Same thing for text processing: a lot of the tedious work around transcribing, correcting, translating a text, but even understanding, summarising, converting a voice to a written text, all that were things that computers, treating words and sounds like numbers, could do faster, cheaper, and even better.
  9. 9. New complements • Photo-based intimacy • User-created marketing • Filters & 3D-masks 9 And it didn’t stop there. Now that images are numbers, we can: - send them over the phone faster, and share them - Computers can process your social activity & predict who you care about 
 >> that leads to News Feed, Messenger (filtered), Snap & Stories - detect who could be influential to your customer base
 >> and computers finds people who to solicit to start and influencer-based campaign - Computers can even process the shape of your face 
 >> and augment that with beautifying filters, 3D masks, etc. That makes photos more valuable, because all those are complement activity.
  10. 10. Cheaper predictions 10 So we now have three key ideas: - Decisions are made of different steps, including prediction. - Complements mean that the price of one thing drives the demand for another. - Everything arithmetic is becoming cheaper very fast.
 And you know what is arithmetic? Machine learning is arithmetic. That makes the cost of prediction drop like a stone. What does that mean? Things complement to predictions, things that work with predictions become more valuable. Input, training are valuable, but their cost is also going down, also because of technology. Judgment and actions, on the other hand, remain harder to automate: you still need to think about the best objective for your company. You still need to implement a solution to a customer’s problem. But now, if you can tailor that problem, finding the right customer for that tailored solution is easier. If you are good at negotiating and understanding nuanced objectives, prediction machines can compute the best outcome for that objective more easily. What about people making prediction themselves? Doctors who make diagnostic. Like for photo cameras, it can mean either that existing prediction process can be replaced, or they can be augmented. But we know for sure that complements become more valuable.
  11. 11. Cheaper decisions Cheaper Cheaper Cheaper Cheaper 11
  12. 12. Applications Speculation 12 What does that mean, practically?
  13. 13. Thought experiment: Improve prediction • Amazon prediction • 20% accurate: Shop & ship — display suggestion • 80% accurate: Ship & shop — optimise returns • BookingGo • Pre-emptive changes to booking? • Flight delay 13 It means that you will be able to get away with more daring actions. How? The authors imagine that you have a quality of prediction, say the precision. If I predict that you are interested in 10 objects, you will buy 2 of them. That a 20% prediction. That’s the prediction that Amazon has, roughly: the authors reckon that they by about 20% of the objects that they see as recommended when they log on. Make a though experiment and imagine that Amazon can increase that number. They process data better, and they can get to an 80% precision. That’s impressive and their landing page become really compelling. You want to buy all that. But why wait? With a certain level of precision, maybe 80%, maybe 95%, Amazon could stop waiting for you to log in, and, instead of risking that you buy elsewhere, they could send you a box with all the things they think you will like, and you keep what you want. If the precision is 99%, that’s generally the whole box. Rather than let you shop on-line and ship what you want, they could ship what they predict you will want, and let you shop at home. To make that cheaper, they will need to improve how returns work. Actually, if they know they need to go there, they could improve the cost of returns now, organise rounds, and if the cost of returns is low enough, they could start doing She & shop even earlier, because the 10% or 2% of errors, don’t cost so much to return. These ideas, these transformations are true for all businesses, and for instance, BookingGo could react if they predict with certainly that your flight would be delayed, and suggest to change your reservations accordingly.
  14. 14. Google “AI first” • Voice Assistants need 98% accuracy • Priority #1: Improving prediction, gather training data • at the expense of performance, switch to mobile • learn customers’ preferences & how to serve it 14 One company that already feels that way is Google. They have understood that increasing the precision of their model is essential, so they have decided to make it their priority. They invest in interactions with the customers that need that precision, like Voice Assistants. Without a great precision, they wouldn’t know what you mean, which Alice to call, where to order your favourite coffee, so they really need to make the prediction really good. To do that, gathering data about what you like, what you do, where you are becomes priority number 1. All the efforts they put into adapting their products to mobile is now pushed away, to the benefit of having an international that is seamless because it relies on better predictions—even if that means leapfrogging mobile entirely.
  15. 15. Jobs, flows & tasks • Not jobs: Workflows of tasks • ML takes over some tasks Thomas Lockton Web Replatforming Current website flows 15 We want to do the same. We want to be able to understand customers, we want to find where we can predict who they are, what they want, what we can tell about them. And to do that, we need to apply that framework of input, output, predictions, list all the actions that we could take, identify what the customer told us, and train with that goal in mind. We need to not think of big, vague jobs but brake those down into small tasks, small decisions, and know which prediction informs our actions for each. That’s why, when I see a giant graph like the one the website flow that Tom shared last month, I think this is really promising. This means that we are able to identify where we could use a better prediction, where there is a branch that machine learning can help with. The bulk of the work is to describe what we do, so that we can identify where automation could help. That’s why I’m confident that there is a lot that we could do with machine learning.