Empower Your Animal Science Business with Commercial Real Estate Crm Software for Animal Science

How to create outlook signature

hello so welcome everyone to our tutorial about data science for real estate um the goals of this tutorial is to introduce real estate to data scientists essentially real estate is a fascinating industry there are very many interesting data science problems that are currently not really being addressed by data scientists in real estate because there are not that many data scientists who work in real estate actually and we want to change it and that that's why we were given that tutorial the presenters today is i'm ron deckerman the cto of cherry cherry is a real estate data aggregation platform uh with me is foster provo who doesn't really need to be introduced is a full professor at the university of new york university is really well known and very active when the data mining community gave many invited talks and have many best papers awards and he sold his startup to a big um brokerage called compass and now he is a distinguished scientist on compass and we have two more people from airbnb ali rao who is the senior data science manager at airbnb she is an economist economicalist by training and um she's managing a big rails real estate data science team at airbnb and we'll hear her later on and vania shifolski who is the cto of holmes at airbnb and before that he was the cto of pinterest and before that he built a very successful um industrial research career um the outline of our uh tutorial today is we're kind of splitting the first portion to three parts i'll start with introduce introduction to real estate and then i'll focus to into commercial real estate something that i'm quite familiar with then we'll let foster talk about residential real estate and ally talked about short-term residential which is what airbnb is uh focused on and we'll have a short break or probably not that short and in the second part of the tutorial we will have a fireside chair so the four of us will be just discussing the new frontiers of data science and real estate uh so we'll spend like about um an hour together when i'm gonna be presenting so like i'll we have to kind of introduce myself usually when someone is giving a kdd tutorial they just dive into the the um you know technical material and no one really knows who the presenter is i i don't want to be that presenter so a little bit about myself i'm definitely not a novice at kdd represented my first paper in 2000 first kdd paper in 2009 and gave my first kdt tutorial in 2011 i worked as a data scientist at linkedin back then and we were about to publish the book about scanning up machine learning so that was my tutorial in 2013 foster organized a panel about how to make money from data science and i back then i was the chief data officer of a big venture capital fund so i wasn't analyst on that panel um co-hosted kdd cup for a couple of years was the kdd uh social networking coach here for three years here's five salon me are doing some stand up at the opening ceremony of that was pretty embarrassing i would say but fun i have a best reviewer um award and overall over many years i've read and reviewed really really many kdd papers and i can say that i have never seen pretty much any kdd material that is related to real estate and this is something that really wants to change so i'll start saying that real estate is really big so like arguably this is the biggest asset class in like you know all the financial sector um no one really knows what's the size of the real estate in the us season even not talking about the world but the number that people are kind of agreeing on that it's something in order of magnitude of 100 trillion dollars so it is a really big industry and the industry is also really old after all like you know people needed a roof above their kids for hundreds of thousands of years and the real estate industry as it was shaped like about 2 000 years ago didn't really change that dramatically which is kind of really unusual and strange to think about but what ha is happening over the last uh half a year is totally shaking this industry so um you know people are talking about the their deaths of uh office space people start moving out of cities like before they were moving into this because they wanted to live next to their work and to commute west and now it's the opposite so the industry is changing really dramatically and uh it creates an enormous opportunity when the industry is so large and the change is so sudden uh we definitely can see many interesting trends that are are being created and many other trends that are being broken and that's why it's really important to work in this industry right now so what is the real estate industry after all so i'll split this in this like you know the slide of the industry map to quadrants uh first we have residential real estate and commercial real estate so residential and commercial the difference is that in residential real estate uh the owner is the individual and in commercial real estate the owner is the commercialization uh sometimes a very large organization sometimes a pretty small one uh and uh horizontally we split the domain to owner-occupied long-term rentals and short-term rentals let's focus in every quadrant and see what we have here so residential in order are occupied is just a single family home in many cases it's a house that is being owned by the person who lives in this house it's a very very big industry obviously because there are many single-family homes in in in america and other countries as well uh on the commercial side uh owner-occupied is usually a vegan organization that needs to have a an office space that you know where the organization resides where where the headquarters is and uh say like mcdonald's owns pretty much uh all the real estate of their branches um this real estate is not for investment they are not trying to make money out of it on the other side since they are holding this real estate and this real estate can go up in the price later on so it's usually profitable to have this real estate but what's even more interesting for me at least is the investment market when people have real estate in order to make income out of it so in residential long-term rental it means that it's a person who owns a house and lives in the house and owns another house and rents this house out so it's a very small operation individually owned uh usually you don't have to have a big organization behind it and uh if dorsum happens to buy more and more real estate it becomes already a commercial organization and it's becoming more complex on the commercial side of the long-term rental there are many sub-sectors and the most important is multi-family here is a little bit of a confusion multi-family is a place where people live right many families live on the same building so why is it commercial not residential because it's usually owned by a commercial organization and it's a quite quite an operation to maintain this property uh and uh that's why it's still considered commercial despite the fact that people live there and the other three sectors are not uh small either but probably less of a confusion office space retail space industrial space industrial space is very interesting nowadays talk about that very little because i don't have much to say but like you know if someone wants to focus uh on real estate commercial real estate the industrial space is very interesting uh now the short-term rentals uh in the left side in the residential market we have this rental market space that airbnb be created before there was no quadrant like that at all and so airbnb kind of owns this side of the market and in the commercial space it's usually hotels and very fairly recently we started talking about co-working spaces uh we're not going to be touching that much on on that quadrant uh for uh for the for this specific tutorial and but two big chunks of the market that is like cross uh sectors one is brokerages brokerages are organizations that help you to sell to buy to rent uh and they are totally cross cross market and another section is uh technology that's where we are and uh it's also across market technology is helping pretty much everything but uh so far it's not really widely thread so proptech is is not a very big high-tech industry i'll be talking about mostly commercial real estate after this introduction i'll focus mostly on the commercial side then austria will be talking about residential and ally will focus on the rental market space of airbnb so just probably one uh like you know side note uh the market is obviously measured by the number of property and the size and the price of the property so the long tail here is residential properties are not really expensive but there are very very many of them there are about 150 million properties in the states and i'm talking about united states only but uh the head of this distribution is the commercial market but for the prices are really large we can talk about like you know if a billion dollar property uh but uh obviously there are not that many properties like that it's the saying is that it's about something like 15 percent of the properties but the size of the market in terms of the price is about uh uh one third so one third is a commercial real estate and two thirds is residential it's a very interesting um slice here which is the meat market between commercial and residential those are properties that some of them are commercial some of them are residential this is a very interesting slice that are is not really being uh researched properly and again someone wants to focus on real estate that might be a very interesting sub market to uh to work on so from now i'll focus on commercial real estate meaning that those are organizations that uh invest in real estate in order to get some income from leases you can assume that like you know imagine that you're working you're a data scientist working for a big investor like that and there are really many interesting real uh data science problems that need to be tackled so start uh saying that the real estate investment market is kind of not really up to date let's put it like that first of all it's very local meaning that uh in really investors like buy properties in the very short vicinity from each other some of them invest in just one single street in in a big city like you know in own ten buildings in a very like landmark important buildings in the city and you can have a portfolio that is like you know a billion dollar portfolio uh why is it local it's because the investors know this area they've been you know analyzing it and working with it for decades so they know really well what they're doing and that's also the reason why this market is not liquid enough so the moment in investor buys uh acquires a property they just usually hold on that property even if the property is not doing really well uh financially it's very hard for to get rid of it but because of they already know what's going on if they want to go to a different like you know part of the city i'm not even talking about a different city it's a completely uncharted territory for them they just don't know what's going on so the lack of knowledge is actually like the driving force of the real estate industry which is kind of really counterintuitive right you need in order to invest you actually need to know a lot of information you need to know what's going on with the sales prices and leases and management costs and taxes and other features and transportation like all this you need to know in order to decide whether you want to invest like you know um 100 million dollars and buying this specific property and all this data needs to be all together in one place which is not what's going on right now so let me kind of dig into some of those features that are important for a an adjustment decision uh here's a picture of the of the building where cherry's offices are and first we can like look at the neighborhood features say transportation it's really good from the transportation point of view but it's not a very nice part of the city so the crime rate is fairly high in terms of lunches yes like you know restaurants are all around no problem so like this is a if an interesting investment opportunity if you wanted to ever to buy this building so the neighborhood features are known you don't need to dig really deeply to get this information on the other hand once we are talking about the building features itself i mean structural features that's less obvious so some of them are obvious right so this building has two facets which is uh say it's increasing the management cost and insurance cost so like it's it has its disadvantages but on the other hand the building is much more visible because it's a corner building the number of elevators is important like two elevators means that it's going to be a congestion when people want to come to the office in the morning it's four elevators meaning that it's good um and this building as you can see is mixed use so you have a retail on the bottom services right above it office spaces in the middle and some executive suites on the top uh so once you decide to invest in this building you need to know that they're actually like you know four buildings mixed together here in one building and you need to take those sectors separately and see how well they are doing and this is the most important part like when you need to invest when you think about investing in real estate you need to know how is this building doing like is it actually profitable is it the uh and the questions are related to the lease contract so how many occupants or tenants you have in this building um maybe the building is half empty uh what are the uh lease like duration and rate and all this is very important to know unfortunately if you if you don't decide to invest in this building this information is really not publicly available so it's very hard to get this information if you decided to invest in it the owner will provide all this information for you so you can make it a smart decision but we first need to kind of you know get really interested in this building and now which is even more complex obviously the price of uh the price and like you know your decision about investing in real estate depends on what's going on around this building so on the left is the same building and on in the middle on the right is those buildings that are attraction to to the one that we're considering and like something is going on there right like the like narrow stripe on the middle is actually a hotel do you know how well this hotel is is doing no idea do you care if you want to acquire this office building next door yes you actually do who would provide this information for you how well this hotel is doing no one so out of the sudden like you know you need to make a very difficult decision invest like a hundred million dollars in an office building you don't have a full visibility to what's going on so you kind of have a chicken and egg problem like you really need to be interested in investing in a specific property in order to get all this information about the property and probably information about what's going on around it but how can you be interested if you don't know this is the biggest trouble the market is actually kind of managing to solve this problem partially what they are saying is that yes you don't have a full visibility yet but we can talk about one factor that is called ecap rate that is actually giving you in like a feeling in notion of uh uh whether this investment might be kind of worse than worthy or not so a cap rate is a very simple formula you take the yearly income minus the yearly costs and you divide by the size as the price of the property that is like an invert of how many years it will take you to get your money back you invest it like a hundred million dollars you wanna eventually get it back right and you probably wanna get some more money after that so like this is the the number of years so like if the cap rate is something like three percent that essentially means that it would take you 33 years to get your money back and if the cap rate is 10 then you'll get your money back within 10 years and that's more or less the uh the range of the gear rate kept rate is like it's um people who know cap rates are very rich because it's not really easy to figure out the actual cap rate and real estate investors are talking like you know of cap rates in terms of like i've seen a building that is an eight that essentially means like it has a kept rate of eight percent which means that it's like a really really worthy investment but it's also a risky one because uh the higher the cap rate is obviously the higher the risk is like you know no one would uh pay very high rent in a building that is fairly cheap unless there is a reason for that so uh overall the real estate commercial real estate is all about cap rates and it's very important to know the cap rates and not many people actually know the gap right we as data scientists can actually estimate cap rates so the formula is very easy right so like you need to take uh leases and costs which is taxes and management costs and the sales price you have the cap rate that's so easy the good news is taxes and sales prices are publicly available in the states not in every country but in the u.s they are and there are companies that can provide you with lease information and maintenance cost information so it is technically possible to compute or estimate the cap rate obviously you will have a lot of missing data but we are data scientists we we can deal with missing data uh the bad news is that even if you have all this data to join it together would be very difficult because lease data comes per unit maintenance cost data comes per building and taxes on sales data is coming per lot and now we need to make sure we understand what those terms mean so a lot or partial it depends on a specific market uh is this real estate this property that you can buy you can sell and it's being taxed so in residential real estate it's usually piece of land on which you have the building so the piece of land and the building together is a law so uh a building can be split to units obviously in some cases um it's still one loss despite the fact that it's split to units so you need to buy the slot and rent all the units separately in some cases you might have multiple buildings on a lot in some cases one building can be split to multiple lots so each unit is a law so this is creating a lot of confusion uh even if you have access to the data like you know joining it together and figuring out what the cap rate is is very difficult because of that but um we are kind of lucky all of them lots and buildings and unions have addresses so probably instead of joining on lots and buildings and units we can add join on ads yeah it would have been easy but it's not unfortunately so addresses are not normalized in in the data real estate data so all these uh are the same address of the same building and we need to build a normalization system to figure out that that's all this is is just one single building so let's say that we even can't build a normalization system it's still not trivial because uh you have a notion of alternative addresses so say 401 seventh avenue and four or three seventh avenue are those two separate buildings of the same yes it's the same building because in the u.s the house number has a range and if the house now if two numbers are within this range it's the same building some uh street num names have aliases so 77 and fashion avenue is actually the same and here are two more addresses and you're looking at other like well those are definitely unrelated addresses right no unfortunately they are if a building is a corner building in many cases it has it has an address of an avenue and an address of a street so all these are addresses of the same property and uh like visually it's not recognizable you need to build a an address standardization system to actually figure these things out and that's what the data scientists and real estate focus at least in some cases okay let's say that we managed to join the input data and estimate cap rates now we know that like you know this is the area where the cap rates are high enough which means that we are interested in this specific area we want to india and now whom should we talk to so it looks like that's a simple question right like we should talk to whoever is selling well in commercial real estate it's not really the case because if someone is selling a say an office building and the building is on the market for quite a while that essentially means that something is wrong with that building in many cases the commercial real estate investor would just acquire the building and hold on it that like you know it's a rare case when someone wants to sell it or something something bad is happening so if you really want to buy a a good building that nothing is really happening with you need to create the deal flow like you need to find out who the owner is and contact the owner with with a particular offer so the question is whom should you contact turns out that is not easy in real estate either so buildings are owned by tiny companies tiny llps that are absolutely obscure so like um say the chrysler building is owned by 405 lexington and no one really knows what this company is there is a large a and well-known company behind the sllc but on the papers the owner is this small llc no one knows what this llc is all about and we need to figure out who is actually behind those tiny uh masking entities companies that are being used to mask the real owner from the the market so this is not an a difficult and not an easy task at all and to solve it we need to build a knowledge crowd and actually at cherry we have built a knowledge graph that contains properties and addresses and people and companies and other organizations all connected together based on productions that they they have been making currently we have about half a billion nodes and about one and one and a half billion edges in the graph and it covers the entire u.s market so all the properties that are in the united states uh commercial or residential are in this graph so it's very important to build a comprehensive knowledge graph that covers all the markets because all of them are very connected to each other so like you cannot build a knowledge craft for commercial real estate and not take into account the residential part at all this is not going to work so obviously i guess you guys are familiar with knowledge crafts the knowledge graph has the topicality a topical locality feature which means that things that are related to each other will be close to each other in the ground and they are organized in like what with what we call semantic neighborhoods like we need to separate like you know geographic neighborhoods and realistic semantic neighborhoods and the graph which means that those are nodes that are close to each other and they are connected to each other and if we traverse traverse the the knowledge graph we actually can reveal hidden relationships and that's where we're hitting here's an example a small portion of that knowledge graph on the far left we have a real well-known real estate investor called donald trump and uh donald trump is connected to people and companies and properties those companies are connected to addresses and other companies and other people and the relationship between donald trump and the person called paul davies is can be revealed using the the traversal of the knowledge craft and this is not a trivial relationship if you say google that those two names don lump and paul davies you will just find the different paul davies not the one that that we are talking about here so this how we can use the knowledge graph and throughout this part of this part of my with the tutorial i'll talk about the knowledge graph quite a lot so we use the knowledge graph to do owner unmasking start with the property and reverse the graph until you reach some type of a real well-known real estate investor which is also a node in the graph and that is most probably the owner behind this specific uh property uh two um comments about knowledge graphs uh usually knowledge graphs are open-ended meaning that like you know if we didn't find the relationship in two nodes maybe it exists but we haven't found it in real estate the knowledge graph is actually a closed world meaning that if say silverstein properties is not connected to trump tower that essentially means that they just don't have anything to do together so this is a very uh interesting and important feature to use we can assume that the knowledge graph is kind of modeling the real estate ecosystem as is and if there are no connections between connections between two nodes that essentially would mean there is nothing that is in common between them many real world graphs are scale free and that makes uh uh data scientist life difficult because there are those bridge nodes that connect completely unrelated parts of the craft and the good news is that the real estate knowledge graph is not skill-free because if you think about it there is some type of a geographic grounding to the knowledge graph that is being built because those are properties and people who invest in those properties they have this local characteristics of the geography that they are working in and that's why the the graph is not scary which is a good news like when we're talking about semantic neighborhoods which don't have bridge nodes we actually can really well model what's going on in this specific locality the trouble is to build the knowledge graph obviously and because like you know you need to create nodes for entities and the names of those entity are not normalized here on the left side you have the same person on the right side you have the same address but the graph is completely disconnected in the real world it's even worse because it is actually connected but connected kind of very sparsely when you look at that graph you're like ah that's a good graph since the graph is so large it's very hard to figure out that the graph is actually pretty garbage and a lot of work needs to be done to connect all these nodes on the left together and all these nodes on the right together as well there are very many ways to write the same name of the same person um here's the aaron zieglemann quite a real estate hero he used to own a lot of real estate in the 80s in new york and there are very many ways to make a typo in this guy's name 25 000 what is this number that's the number of ways to write the name of jp morgan chase bank so we found 25 000 different ways to write the name of that bank that would include all the subsidiaries all the branches all the typos everything wherever you can making mistakes someone will definitely make a mistake um so as you can see it's uh like cleaning the data is a very important part of this the work in uh every data scientist essentially but in commercial real estate especially i'll give you a few examples i'll try to be quick here say those two names we taught the system already that it's the name of the mark morgan's uh jp morgan chase bank despite the fact that they have absolutely nothing in common we know that it's the same organization but what about this name that looks like a person whose first name is chase last name is morgan and the middle initial is in probably that's not a bank right but what if you're looking at something like that n a is actually a very good indicator that it's a bank and so on the right side it looks like it's more like a bank and then you're looking at the left side you're like huh that looks like a bank too that's probably not a person here's another problem like tishman spire is a really well-known real estate investor and the name is very unique so whatever indicator words like llc or corp you add to the end to of tishman spire it's still going to be the same fish transpire but what if you take a very generic name of atypical uh a company that owns just one property say one three main street if it's llc or corp is this the same company or not probably not what if the indicator work uses company or say co right core can be corp or company so it's becoming really ambiguous so you have no idea what is referring to here's another example meryl streep you write it in any order of words it's still going to be the same person right what about david kim if you write the david king's name in in a different order you end up with king david and that looks more like a hotel or more maybe a restaurant it's already not a person just because the order of words is different here i i just came up with some type of a very rare last name and the first name is john so on the right on the left it's probably the same person right despite the fact that on on the right it's written with the typo but what about the if the name of jon that's a female name right so it's just one uh character of a the string distance is one but those are probably different people and what if it's joanne you know like [Music] maybe since both of those are female last name for first names maybe this is the same person actually and the last example is uh the decision about how to normalize the data can be very cultural so i happen to be a hebrew speaker and i know that the name on the bottom left is an organization and the name of the bottom right is a person but i just happen to be a hebrew speaker i don't speak chinese i don't speak spanish and like i'm pretty much i'm pretty sure that problems like them being in in other languages as well so it's a very difficult task to normalize this data to build the knowledge graph but once the knowledge graph is built we can apply owner and masking and we can find out who the owners or properties are so let's say that we are commercial real estate investors we decided that we want to invest in a specific property and we know whom to talk to because we use donor unmasking now what about the price it's kind of an existential question right like you what is the price at all so in some cases especially in residential real estate is not really hard if those two buildings were sold for 200k what's the price of the building in the middle right i think it's a fairly simple answer right but what if you have a big city and it's very diverse obviously there is a building that costs 200 million or a building that costs 20 million and what is the price for chrysler building if you have this information it's very hard to figure out probably like i would say completely impossible so the price is a big big trouble and there are really many evaluation models models that kind of try to estimate the price of a specific property a simple one is obviously based on cap rates so if we happen to know the cap rate if we know that the cap rate is say six percent you can take the data of the property meaning the lease data and the management cost data and the tax data and you can compute the sales price just based on on the formula of the cap rate that's very simple right the problem is that no one knows cap rate so like it's very hard to use this formula just because the left side always in most cases unknown so you have many other types of evaluation models to apply in this situation and the most common is comps comms is comparable properties those are properties that are similar to the one that we're considering and also probably recently sold so like you know the sales price of the comp if you know the comps actually you can very easily estimate the price of the property that you care about you basically take so in real estate you usually take about a talk about the interval price uh price intervals like you know from the cheapest to the most expensive if you just plot those prices of comps and somewhere in the middle most probably is going to be the price of building but like you know as you can understand this is a very simplistic model that uh in many cases uh doesn't work we can come up with a little bit more complex uh uh evaluation models based on comms which would be if you know what the com are comparable properties are you can compute cap rates based on the comps data so you you know the price of the building because the building was sold recently you know the the uh financial um features of that building so you can compute cap rate and you can use the cap rate for estimating the price of your building because you you know the cap rate and you know the financial features of the building that you care about so this is working the only problem is that what is the comp how can you figure out that the property is a comparable we use the knowledge graph for that usually um the real estate industry is so much not data driven like you know a person could deal with real estate would say here's the comp right like here look at that that's a comp that that's pretty much how the comp is being defined obviously very kind of you know uh hand wavy and rudimentary and the right approach is probably at least the approach that we're using is to use the knowledge graph so if you have properties that have similar features and they also are within the same semantic neighborhood in the knowledge graph or they're in similar semantic neighborhoods in the knowledge crowd then you can ensure that those are comps once you want to find the comp i you really need to traverse the knowledge graph right and with the little bit of supervision you can say that like you know once you're traversing uh the graph from x to y and you know that x and y are comparisons because you have some label data what you can see is that this process uh explodes really fast because the graph is well connected right like if you go like on the third hop from x you already end up with millions of properties but with a little bit of supervision you can see that the most important path between x and z all goes through one node sorry between x and y goes through one node which is x and you can learn the attention vector for the problem of finding counts so like the attention mechanism here can be used and it creates a very interesting model by the way i'll be happy to talk about that uh offline there are other ways of estimating uh price of a building this one discounted cash flow dc dcf is quite involved from the financial point of view and whoever has taken a financial math class would probably learn this model so the model is based on the fact of how much money you want to spend now to get some cash flow in the future and so you estimate how much money property would make and you kind of discount that that amount by how much money you want to invest now to get this cash flow and then since obviously you have to have a a close horizon here because otherwise this number would be infinite you would say that in 10 years you're gonna sell this property and you're going to make some profits out of this sale as well and you need to take this profit into account into the model like again how much money would you spend now to get that much over yearly income plus this profit from sales so this is a more complex model but it's also more precise if you knew how to estimate uh your income in the future and also the problem with this model is that it doesn't take any features of the building into account that's why there is a completely different model that is called a replacement cost model which takes only the building features into account so you say how much money would it take me now to build exactly the same building that i want to buy it's all based on construction cost data so like there are huge tables of like if you have this type of air conditioning it's going to cost you that amount if you have a different type it's going to cost you that amount so you sum all these expenses together you will end up with some some amount and this is going to be the lower bound on the price of the building because obviously based on that you are getting an empty building that doesn't have tenants which essentially means that it's not making any income that essentially means that you need to work hard to actually uh bring the tenants to the building and uh that costs money so you end up with air number but that's usually used as the lower bound of all the prices and you are combining all those valuations together into what's called the football field like the price of the building would probably go from the replacement cost amount that you calculated on the left to the asking price which is usually above everything else and you need to come up with a number somewhere in the middle maybe probably in this case closer to the gcf model performance but like you know it really depends on uh how you look at that so anyhow in real estate people usually look at the price intervals there are many uh automated valuation models that are built by data scientists unfortunately those are simple models like you just built a feature vector for a property and you're just throwing it to to a regressor and that's pretty much what those avms are all about nothing i've never seen anything like substantially more complain that and this is a big opportunity obviously like we can't use all the machinery that has been created in the data science domain for the over the last like 10 years to actually build a better model and since those models are so simple actually they don't work well in practice so the result of that the prediction of the price in commercial real estate is pretty bad because like you remember like you know if you if you want to estimate the price of the chrysler building it's very hard to build a model uh a machine learning model for that so we have an internal jockey joker's cherry that is saying that there are only two factors actually that affect the price of a property which is the buyer and the seller and if you don't take those two factors into account it's very hard to assess the value of the property because the seller will sell the same building for different prices to different buyers it's a negotiation process right so probably a better approach for estimating the price of the building is first to use owner and masking to figure out who the owner is and we need to use link prediction in the knowledge graph to figure out who might be potential buyer again don't forget that we're talking about commercial real estate in which there are not that many big players like you know if it's a big organization that currently owns the building it will probably sell that building to another big organization well known so like what we need to do is to predict who the buyer might be and once we have this prediction edge between the seller and the buyer we can build in a very very specific abm avm model evaluation model for for this specific year and this is probably the future of uh avms now the biggest question is like that like uh okay we can't predict this this is going to be the buyer but why this seller will want to sell it all uh and this is a bigger very big question so there are a few real estate uh types of real estate players one is a developer who is building uh basically constructing the the building and then selling it another one is a flipper who buys the property um kind of you know adds a little bit of repairs here and there and sells it but the vast majority of real estate investors they just buy property and hold on it because this property is producing income why should they sell it right and they sell a property only in the case of a distress distress is a very important question especially given the fact that we are talking about uh our times when you know when kavit stroke really badly and uh there are many aspects of this distress that we need to take into account one is obviously the structural so like you know the roof is bad for in this building this is the simpler type of distress if the owner is drawn the owner will invest money fix the rule right there is the second type which is occupancy distress the building is okay but the tenant just left the building is empty or half empty how do you deal with that so this is not that impossible either because you know you use your marketing skills you make this building full again but there is this third type of distress which is more difficult to deal with and that's the tenant distress the building is okay the tenant lives in this building but the tenant doesn't have money to pay for some reason and this is a very common case nowadays with uh kovite 19. so the tenant doesn't want to leave and the owner doesn't want the tenant to leave but there is no cash transaction here so like how to fix that problem it's very difficult and the fourth type of distress is actually the owner of the strength so the owner has some type of a problem probably from other properties but the owner has like you know owns very good structurally and tenant-wise properties that the owner probably wants to sell in order to make their financial situation better and this is an opportunity so if we can't detect different types of distresses based on the knowledge graph we actually can make predictions what is going to happen with the real estate and which type of transactions will happen so we can use say a label propagation and a knowledge graph we can say we know that this is the properties that is currently in distress or this is an owner that is currently in distress how this distress is being propagated throughout the knowledge graph is something that needs to be estimated when it stops how it is being re-weighted over the uh a semantic neighborhood in the knowledge graph this is an interesting question to answer and we don't have good answer to that so if you don't mind since i started five minutes later i'll spend another five minutes to discuss uh the maintenance of the knowledge craft so we used a lot of data normalization to build the knowledge graph but obviously we cannot do it perfectly still in the knowledge graph you would have entities that are represented by multiple nodes and we need to merge those nodes together if those are three nodes on the left that are that they have the same semantic neighborhood and the nodes have similar names we need to merge them together into one and that's the entity resolution as a knowledge graph problem that is very important uh the trouble is that if we are talking about similar names of nodes in similar semantic neighborhoods of those nodes we kind of are talking about the quadratic algorithm right so run over all the nodes and run for each node run over all the nodes again to find those similar nodes right a quadratic algorithm in the knowledge graph of half a billion nodes is not going to work so you need to come up with a different way which is to start with a neighbor so if you started with a node and looked for every other node that might be similar that's not going to work but if you start with the neighbor that has x and z's in there uh semantic neighborhood then what you need to do is to run over pairs of x and z's with x and y sorry within the same semantic neighborhood of z that creates the algorithm that is feasible computationally and you can find uh pairs of nodes x and v that should be merged together because they are actually the same entity and i entity disambiguation is a complement of entity resolution so an entity resolution you're saying there are multiple nodes on the graph that refer to the same entity in entity disability that's the opposite you have one node in the graph that mistakenly represents multiple entities just because it's not the same name here's an example jose gomez is a very common name and if you have one node in uh the knowledge graph for host ergonomics it's connecting parts of the network that are completely unrelated to each other and this is a mistake and if you need to traverse the graph those mistakes might lead to really wrong decisions so what we need to do is we need to split this node to power multiple nodes we need to success one possible means two cosecants and then the graph is going to be correct how do we do that we take the semantic neighborhood of that node jose columbus in our case and we delete this node and we buy the connected component algorithm and that essentially will split the semantic networks to multiple chunks and each chunk will represent a specific instance of the same name that we mistakenly put as one now and that's how we play them so to conclude i gave an introduction to the commercial side of the real estate um there are many real data science challenges there are with you um data scientists who work on commercial real estate and it's a very fascinating topic and since it's currently changing uh upside down because of covid this is the best opportunity for us to jump in and to work in commercial real estate and to help the industry survive and get substantially better in the future and that concludes my part of the tutorial saying cherries hiring i think is uh kind of obvious if you guys are interested in joining cherry please contact me