Duration 33:49
16+
Play
Video

Learning Market Dynamics for Optimal Pricing By Sharan Srinivasan, Machine Learning Engineer, Airbnb

Sharan Srinivasan
Machine Learning Engineer at Airbnb
  • Video
  • Table of contents
  • Video
Video
Learning Market Dynamics for Optimal Pricing By Sharan Srinivasan, Machine Learning Engineer, Airbnb
Available
In cart
Free
Free
Free
Free
Free
Free
Add to favorites
31
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About speaker

Sharan Srinivasan
Machine Learning Engineer at Airbnb

Sharan Srinivasan is a machine learning engineer at Airbnb. He works on problems relating to modeling marketplace dynamics - pricing and search, optimal pricing strategies, search and location systems. Previously, Sharan worked on various data products at InSnap, a mobile hyper personalization company. He holds an MS from Stanford University, where he focused on operations research.

View the profile

About the talk

With the emergence of numerous online marketplaces in recent years, the need for efficient matching algorithms which balance supply and demand has grown significantly. Airbnb is one such marketplace which tries to match guests and hosts. Guests searching on Airbnb are often looking for the best deals that they can get on their trip. Hosts, on the other hand, look to maximize their earnings by pricing optimally. To meet these diverging objectives, Airbnb employs optimal pricing algorithms to match guests with hosts. The challenge, however, is that unlike commoditized accommodations in the hotel industry, every listing on Airbnb is unique. Airbnb homes span over a broad spectrum, from price, location, quality, to size, etc. Analogously, guests search with custom price elasticities and preferences in mind. This heterogeneity brings challenges for personalization. In this talk, we highlight some of the techniques that try to address these challenges and improve marketplace conversion. We also provide a framework to elements of machine learning with structural modeling in order to address problems related to optimal pricing. We will dive deeper into location systems that were developed with transfer learning in mind, helping address challenges of scale and dimensionality. We will also take a look at arrival lead time distributions which form a key part of demand models and pricing strategies.

Share

Awesome. So good morning, everyone. Happy Wednesday through the week. Hope everyone is doing well. My name is a B&B. So today I wanted to take a little bit of time just give you an overview of the flu shot give you an overview of how machine learning works at a B&B where where some of my work fixing and gently talk about market dynamics in the travel and accommodation space and specifically go into some algorithms and some interesting technical challenges and shattered with with a

broader audience and optimal pricing pricing has been a part of a lot of Industries specially two-sided. If this is an understanding, these market dynamics have been extremely crucial in in in optimizing for revenue and profit in the Phantom of the phases in apartment when this kind of marketplaces are prevalent online a lot of algorithms and machine learning techniques go into determining Clinton affected by market dynamics today. I'll be focusing on

what is market dynamics are and how we going and how we address. So the talk today will be split into three parts initially. I'm going to give up like an overview of every NBA what to say and then I will go into pricing and try to motivate problem statement give everyone an overview of and finally I will go into some of the pricing policy and talk about the algorithms and little bit more. Awesome. So we stopped at a B&B. I hope he is a two-sided Marketplace and in the travel and accommodation space the company is trying to as a pack of what we try to do is we try

to match. Guess what looking to make a reservation with listings or hose for looking to rent out the space for a few nights and B&B soccer fits in the middle and tries to find the best guest and tries to find for every guess the best listing among all the inventory that's available if there are a thousand listings and there are five guests were searching every day is the match every guest to the optimal listing that given a set of search parameter. On

Optimum online. So as you can imagine when I guess comes on to airbnb.com, they go through a few different experiences. So process example of how do I search ranking as well as experience? I'll be showing them prices that they really like good. They trusted platform on a Hyundai getting a sense of trust, and finally they go to all the payment process doesn't make sure and we make sure all that is. The larger goal here is to try and find for that guest

and optimal listing and optimality comes in several flavors. Can we drank our family can replace a family excetera today? I will be focusing on the pricing aspect. And the reason why this is interesting is because as as with most machine learning problems, there's no one. Answer every guest on the platform is always looking to find the best listing at the cheapest price is always trying to find the best guess at the highest possible price. Because the whole style is one stop the most revenue and the guests always wants to get a really good deal. So there's always some amount of

land and that's where and that's where the optimization comes in because we really want to find a doctor optimal price point which not only maximizes revenue for the host but also helps keep it take into account guest preferences and get it ready for the weekend. So that was a very quick overview of Airbnb iron. I would like to do it now is a sort of a going to what I mean by pricing and motivate problem statement. I'm so pricing an Airbnb essentially is a very host facing a picture of her. So what you see on the left

is is a typical host calendar which is which is surfacing prices that the host has set for every night. So you can you can see you in a few places here that the hostess set a $44 on a Sunday in 51 on a Wednesday excetra. And these prices are sometimes Sick by the horse, but are often helpful understanding of the market ecosystem. The only know about what's going on with the existing which is a lot of information, but they don't have outside of Dallas. In

order to help them price every night on their calendar. We at the pricing theme provide them with Optimus with the price recommendations that help optimize something this something could be profit Revenue occupancy, whatever but the point is that we try to help them optimized for their for their requirement. Not only accounting for how good the listing is but also plan to the marketing mix outside of the pricing product is to provide a recommendation for every night on the calendar

and do this for every listing on the backbone. I do this every day. So as you can so as you can imagine the largest eyes with any kind of theoretical pricing problem, you always have a supply curve that sort of increases with price the more if I can tell if I if I go tell people to Residence Inn Tahoe that that they can fetch every-night in Tahoe can fetch a $30,000 then the number of listings in Tahoe that I go to be hosting on Airbnb is going to it's going to stay but the number of guests were looking to book that same distance isn't so there's

always a supply-demand. There's always a supply to Montreal and we would like to find a price that optimizes on this front yard, which is the green zone there. And that's essentially what the optimal price range would be. So if we quickly take a minute to think about what kind of market dynamics really affect this kind of pricing problem, we can split it up into three different domains. So we have special factors that influence it so things like where your listing is really affects how much you price a listing a special

factors come in liquid coming macro or micro flavors of macro would be like administrator Baltic countries. Zip codes are smaller neighborhood Piper Lopez leave and we also have intrinsic properties of the listing so things that help with your listing is or how big it is what kind of reviews of products that are also influence. how to influence what kind of price you can separate So today I will largely be talking about and as you can imagine all these three factors thought of interplay with each other. You can never have one without the other and neither can live by

the Today will be focusing on lead time, which is a very important temporal aspect of things. The reason why we are so that's that's a high level overview of the pricing problem so that they take away from this from the previous section was that the goal of the problem of the goal of the pricing problem is essentially to find that one optimal price that you want to stick for January 1st of 2021, which maximizes your Revenue but also you want a higher price as possible so that you definitely get booked for the best

Revenue that you can possibly fetch but you don't want it to high because if you have it too high then then gets an offer to come in and so, how can we doing? So we bear in mind that leave time Dynamix play a very important role? What are the ten Dynamics I'm going to go into that in the next. So a banquet primer only time today on December 1st, and I was looking to book a trip with a check and date of New Year's Eve. In that case delete I'm here would be 30 days. The gate on which I want to take the trip - today would be the least time to the

trip. I've just come into the platform as guests coming to the book at various different and the lead time of essentially refers to the time between the date of booking and the trip check it with a check-in of New Year's Eve being booked 30 days in advance would mean a 30-day If you're going to book on the same day, so if I was sitting through they looking for today at check-in then the lead time would be zero. Jets will continue to make bookings for New York you as time

progresses going towards New Year's Eve will right up to the end this booking process reflects the influence of the month and week and can be treated as a stochastic arrival process. As honest as in a standard queuing Theory formulation and disc response and the corresponding distribution of bookings over lifetime. Would be termed as the meat and distribution of booking. The Logical question that anyone would think coffee is Wymore delete everything what's what's their lead time? How is it connected to pricing on the

left shows how the man would look like a different times of the year. So as you can imagine that I can we keep our guns there and then on New Year's Eve, we see a spike because everyone would like to take a trip for New Year's Eve, and then it sort of goes back to normal and there's also some seasonality of addition and remodel with demand and Supply Signal Peak Center. The main reason that we want to model the time is because even if we choose a check-in date, even if he chews on a December 21st as a check in there. The kind of demand that comes into December 21st

varies as we progress towards in other words as lead time gets shorter or longer the demand for the checking bit changes significantly. This demand also changes with supper with the with the area that has special characteristics that we mentioned before it also varies with a kind of listing that you have and also takes into account. A night in pewter example, you're an interesting one is that if we think about a high demand night, for example New Year's Eve or east of the Monday night

generally tend to get booked out far far in advance 90 days in advance a hundred days in advance. In these kind of maca to get expect all listings to be booked out far ahead in time, which means our pricing policy needs to reflect the fact that bookings are going to come in really early and we'll need to change a price and somehow in accordance to the Ender Portal this kind of process so that we can fetch a host optimal revenue and give our guests the best deal as ugliest person on the other hand. If you're going to have a really low demand

night or what we're looking at an extended amount constrained Market in such a scenario for a few days in advance. So we may want to keep our prices as low as possible and then sort of dip them down so that we can This is the reason we want to model each time. We want to know when demand is going to come in for a second night so that we can adjust our prices accordingly. Get some more concrete empirical example of water distribution for San Francisco with some of the numbers

represent what fraction of bookings for a certain night that said December 4th, December 1st. Comes in at delete time specified by the x-axis example in the region. It means the most density of bookings at 7 daily time would be safe 5% of 6% A formalized is a little bit better. If x of T distribution that we looking at we want to find the probability. That a certain listing is going to get booked at the lead time of Safe 10 days given that it is going to be booked in other words if I know that he's going to

get booked. And I know that the probability of of that listing getting booked the safe 30% Then I want this to do that 40% in order to finally arrived at the lead time and what we are in love with you what we called the arrival process. The boiler is the model that mass density X. Not only did we want to model if we want to we want this kind of distribution for every listing for every night on the calendar and put every time going into that night. So if you remember the calendar view that showed you

want to price every Square on the calendar and these prices need to change every day because competition function except So we'll pay I'm a machine-learning engineer. So what do you do that? I could can you direct directly go to machine learning like use user normal conventional brute-force approach approach Esther finest and identity brekkie. I mean we could we could directly if we could capture all the training later that we think it's indicative of the elite and distribution demand-supply competition of listing after we can have a label and we can go ahead

and predict bee stings didactic. but it comes with a bunch of pallets. Is not a is what we're trying to protect and using a straight-up machine learning approach soccer breaks down in a couple of different someone machine learning and general doesn't account. Like if we just directly predict an outcome variable that doesn't really account for the probabilistic. What we want is to say that has a distribution of lead time in this kind of manner. We don't

want to see that December 31st has an average lifetime of 30 days December 1st has an average lifetime of 20 days. We want to see if we want a density curve over the lifetime access for every night in the future and machine learning models in general would not account for this. A second important factor is the sparsity of bookings. Every listing for every night in the future can get booked once and only once a spot datapoint. If you really want to use this one day a point to arrive at a distribution for that particular night, especially for

us and we want to account for Spotify and massage it in such a way that we can actually solve solve this formulation even though High dimensional so unlike more commoditized the same. I'm discreet while this is why this is Fantastic Four. Just giving an extremely large slew of options. If it's a it's a modeling nightmare because it with every living being different for each other but no consistency the features and I created it make every listing eunuch an extremely high dimensional data

point. We had the scale of the problem. We have ordered of a million listings on the top form and each of these listings say we want to Model A Thousand Nights into the future prices for a Thousand Nights in the future, which is like 3 years and I'm for every night. Let's everyone remodel a hundred days in advance. Then would really easily looking at at least a million predictions every single day and a trillion predictions for the machine learning problem while it's doable with the with Advanced with with extremely good infrastructure that we

have today. It's it it's like Africans impractical went down in skin. So what do we do? Is there something is is there some other consideration that we can think of just looking at that distribution above like the first thing that stood out to us was that hey, there are two things going on demand is the king with Lifetime by and large. There is some kind of there is some kind of connector is not always the thing but it is a unimodal bimodal pick something and the other thing that I

would I would have this kind of knowledge in my mind and then go about watching the process. This was something that immediately stood out to us. So this sort of taking a step back and zooming out we can arrive at two different schools after we can either go to what is the most conventional machine learning or the most structural approach might be saying this is how the leap and it looks like and we want to model it explicitly in terms of defilement. If you go for the

most importantly it doesn't model the underlying data generation. It doesn't model bookings are coming into the platform on the other time. If I impose structure to the problem, we get much higher predictive power. We can build our own custom model and at the same time it does. It doesn't matter what we did. Is there try and Mary both of these supposed to start for the final final solution. So what I'm going to do next is sort of going to explaining how we predict this kind of arrival process in a given that I'm motivated the

problem and it's done in three different steps. These three steps amalgamate the two schools of thought that I outlined before structural modeling as well as conventional machine or deep learning accounts for these four problems that are stated accounting Republic outcomes, sparsity dimensionality and skin problems by combining these two different is what I'm going to do in the future is what I'm going to go to in the next few slides. Quickly jumping into it step one. We have a lot of listings in the

tackle. Every listing is different from what other invention So the first thing that we wanted to do was address this problem of extremely high dimensionality and scale and we did this by Matt by Abba Technics record as Supply folding chair Trojan 18 these listings by what's your price and quality in geography affect demand and what we can probably do clusters such that each month cluster is characteristic of a circle on profile. Maybe these are the times for example, if I take

them maybe I would get a bunch of listings that are extremely residential and some of them are exposed to extremely suitable for University Housing or some that are very close to Beach. If we're able to supply this kind of demand pull this kind of Supply, then what we can get at the end is clusters of listings that share very similar offenses. Meaning if I was a guest looking to make a booking on Airbnb then I would very likely look within the Scutter if they're 100% in the class though. I'd make a choice Within These

hundred percent in the woods. The pundit listings are competing for the same guest Visa be myself. Very likely just want to listen to Alternative share the same to your special attributes because people want to go to one place. They have a place in mind if I'm going with a party of 5 Oil Change and we also for modeling poses wanted Supply balancing, which means if I want to break out machine learning models on clusters of the things we want these clusters of PID and also be a sort of similar sizes so that we can use this not only for modeling problems, but also put on

Let It Go, Finally, we wanted her article attributes meaning we want to look at clusters hyper local neighborhood as well as you can imagine changes at a market. That was maybe I season for something slow season for San Francisco, but maybe it doesn't the nips conference that going on at Stanford hyper local geography. So we'd really like your company. With these in mind restock off use a how do we go about to Plastering these listings? I'm not going to go to too much of the technical details of it.

But what are some general intuition in the interest of time when guests coming to the platform did view our listings in sequence before they go about can make a book review listing 1 2 3 4 5 6 7 1 books lifting one and guests to books listening to there's some sense of correlation definition of sessions or just want to just to sort of tell me that listings 1 2 and 3 attack the same kind of just listing for 567 guests and what would like to do? Is combined

listings in such a matter? So that the listings in one cluster cater to the same kind of guessed. Perfect, baby listing 1 2 3 which facing in 4 5 6 7 where University? We take a standard NLP problem to this kind of approach so you can think of words listings being words guess possessions mean sentences and what we want to find is embeddings of words by Isabel is so that we can take a high dimensional listing and encode that into say an eighth dimensional Vector. We move down

from millions of listings to maybe in the high tower season and what we can do that after its partition. We have we have an internally developed a recursive partitioning 3, which takes listings in the whole world and part of brexit down here to farm tractors, and these clusters can be viewed as a zoom level example. Stick it into a child of left and right and then done this because of the beta stage. So now we've got to do some listings of pastas with Tackle dimensionality by encoding listings. Does embedding hierarchy. Bye. Bye by forming this recursive partitioning.

Now, we have clusters. What do we do with them? Get us some examples of interesting fastest diesel car mode for a Visa Moto visualization on the left most you can see a set of listings that are very close to South Lake Tahoe and perfect for the summer. Getaway some that are very close to the squaw region and something else. Like ya'll say examples of what is Justice on the beach in Copacabana, or Hi, Rocky would look something like this. So for example, once we do that partitioning

Florida what we get is all listings in Florida, just South Beach and then and then smaller neighborhoods within South Beach. So that's part one. We've taken all these things and made them in the fastest most efficient and distribution and this distribution has a psychic pattern to it. So what we can do is convince this information this information that we have to the model and say hey when I'm predicting distribution, I would like to parameter. Is it in a certain manner? When we looked at this distribution and petekey what we noted

was just because who came into the back form hide in terrible times that were exponentially distributed in some Lanka. I'm sure a lot of you know where I'm going with this if internal rival times are exponentially distributed the number of booking that you get in a unit interval of time would be possible in Lambda and the time to the event which is the time to the booking from the check-in date would be, in some Al Bundy. Okay, so if we knew this then what we can do is impose this kind of parametric functional form on top of our distribution

and what we get at. The end is the FBI part of me price by Alpha and betta with just say that this arrival process is in fact, in the equations below. Great, we did this but we also noted that the recycle cartons are Enrique formulation. Somehow we can all we need to do is consider this process as a waveform. Now Bigfoot defended apply our usual find the frequencies. And once we do this weekend modern detective cartoon harmonic, we find suitable as scop being a road sign a waiver

as well as the Omega Phi and and the amplitude of the other psychic 105 parameters that we've no importa fraction form. Life becomes a little something now. Once we do this is a final piece of the puzzle. We have trust us. We have parametric functional forms. We can go ahead and predict for each cluster and for every night on the calendar. What is 5 parameters would look like Hazard machine learning problem. This is far simpler than taking every listing every lead time every night on the calendar and

predicting values to that. I'm going to fall asleep because it took me a little bit of time to digest. Panera Bread reducing dimensionality significance by doing so well at least improving on our on our scale by a factor of a thousand what we get at the end is what every cluster what we can do is build out a trained with with predictors that we think are everything are predictable. This process great label. Would be the empedocles fitted sheet and distribution are emphatically fitted Alpha B.

What is this process now as a maximum likelihood problem where we're trying to find the part of me does this pipe out of meters that maximize the likelihood of the arrival process being as such and once we do this weekend place that we see on the left and the right. So the red distribution is the empirical version by example, so the left distribution is the empirical distribution. The red one and the green represents the defeated distribution of the arrival process

and especially on the skin. This is largely how we go bye-bye. So instead of taking the smallest unit of prediction, which is listing night the time we obligated and do it in bulk. It shouldn't sorry to interrupt you. We are running out of time all this comes for For A Cause right? So for example, besides pricing, we also have some excellent sight advantages to this example people who travel to La want to know how far in advance do you need to book? So we give them this kind of information in addition. We know that for example on the right

side. Hey. Hey 22% of guests are making bookings are very likely to have made booking spice up by a certain time. So maybe you should decrease or increase your price. We get these insights in addition to the actual price and we love giving these kind of insects do a horse and again in order to maximize conversion. final thoughts of perspective imposing the sky explicit relationship combining start traditional models really helps most importantly model 3 regeneration

mechanism paper trying to impose on top data with dealing with imperfect which I'm sure a lot of you have PS. I am dealing with some cated data talk to the morning can really help bolster the machine. That's all I had for today. Thank you so much for the offer Union. I will now.

Cackle comments for the website

Buy this talk

Access to the talk “Learning Market Dynamics for Optimal Pricing By Sharan Srinivasan, Machine Learning Engineer, Airbnb”
Available
In cart
Free
Free
Free
Free
Free
Free

Ticket

Get access to all videos “Global Artificial Intelligence Virtual Conference”
Available
In cart
Free
Free
Free
Free
Free
Free
Ticket

Similar talks

Sudip Chakrabarti
Venture Investor at Madrona
+ 4 speakers
Ian Foley
Venture Partner at Level Ventures
+ 4 speakers
Anupam Rastogi
Partner at Emergent Ventures
+ 4 speakers
Vik Sasi
Partner at Dreamers VC
+ 4 speakers
Dennis Liu
Investor at Ford
+ 4 speakers
Available
In cart
Free
Free
Free
Free
Free
Free

Buy this video

Video

Access to the talk “Learning Market Dynamics for Optimal Pricing By Sharan Srinivasan, Machine Learning Engineer, Airbnb”
Available
In cart
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
561 conferences
22100 speakers
8257 hours of content