際際滷

際際滷Share a Scribd company logo

Global Online Marketplaces Summit 2019

Download as PPTX, PDF
Gabriel Gruber
Gabriel Gruber

1. The document discusses AI at OLX Group and Properati, covering topics like how Properati uses AI to build a price valuation tool. 2. It provides an overview of AI concepts like linear regression, neural networks, and how Properati collects property listing data to train models that predict price based on attributes. 3. Key takeaways are that AI is powered by simple math repeated many times on large datasets, it is taught with labeled data rather than programmed, and transfer learning can save time by reusing pre-trained models.

1 of 35
Download to read offline
Gabriel Gruber AI in RE 07-June-19 The best way to your next home
AGENDA 1. OLX Group+Properati 2. AI 101 3. AI at OLX Group 4. AI at Properati 5. AI takeaways
01. OLX Group and Properati
Almost 1 year ago ... 01 PROPERATI
Properati became the New kids on the block at OLX Group OLX GROUP TODAY: THE WORLD'S #1 CLASSIFIEDS BUSINESS HORIZONTALS REAL ESTATE VERTICALS OTHER VERTICALS CARS VERTICALS Global USA Russia UAE Africa and Philippines Russia Portugal Poland Romania, Egypt Heavy machinery, Global Services Poland Poland South Africa Romania Portugal CONVENIENT TRANSACTIONS Cars Global Cars UAE Cars UAE Latin America 02 PROPERATI Furniture France Jobs India
Argentina Uruguay Colombia Ecuador Per炭 We launched 3 new countries, so far :) 03 PROPERATI
01 PROPERATI OUR MISSION To empower buyers so they can have the best way to their next home through tech tools and relevant data. In parallel, we seek to help sellers achieve a more efficient sales process by enabling the best service to their potential buyers. 04 PROPERATI
As Seen On TV 05 PROPERATI
02. AI 101
AI vs ML What's the difference between Machine Learning and Artificial Intelligence? If it is written in Python, it's probably Machine Learning. If it is written in PowerPoint, it's probably AI :) 06 PROPERATI
07 AI challenges for all tech companies PROPERATI Data scale. Scale in infrastructure and engineering capacity to process all that data. The ability to apply AI to solve specific customer problems.
08 AI advantage for the tech giants PROPERATI AI looks tailor-made for the incumbent tech giants. They have quickly moved to put AI at the center of their strategies. Aware of its massive potential, they have started to invest appreciably in AI talent, data infrastructure and even their own dedicated chips.
09 Warren Buffett卒s castle moat in the age of AI PROPERATI AIAI AI
10 But there are some AI opportunities for the rest of us PROPERATI While the tech powerhouses certainly have most of the data now, their Achilles heel may be a lack of deep domain expertise. Many new winners can be created by applying AI to distinct problems in the titans blind spots. Integrating AI very tightly into your business processes should also allow companies to compete with the giants previously thought to be invulnerable.
04. AI at OLX Group
11 AI at OLX Group PROPERATI
12 AI at OLX Group PROPERATI
13 AI at OLX Group PROPERATI
01 PROPERATIPROPERATI11 05. AI at Properati
14 Con AI help us answer this question in LATAM ? PROPERATI
15 Let's have a look at Properati today PROPERATI Listings = Datasets Statistics Price valuator tool
16 We have years of listings history that we can use as datasets PROPERATI latitude longitude total_surface covered_surface Price (cop $) 4.731 -74.036 112 112 5.500.000 4.694 -74.078 36 36 2.500.000 4.707 -74.069 124 124 3.000.000 4.696 -74.036 100 100 2.586.000 4.609 -74.067 80 80 2.100.000 Attributes Response (target)
17 Let's say that the price is a function of the total surface (chart) PROPERATI residual a is the slope, price per unit of surface b is the intercept, the price at zero surface
18 This is called linear regression model PROPERATI There is a plethora of options for defining the function f the relationship between the price and the attributes of a property. There is one that: was invented 100 years ago, is still used is simple and awesome It is the linear regression. We are going to pick just one attribute for the moment, the total surface. a and b are parameters of the model. A linear regression with only 1 variable is not optimal for price valuation
19 So let卒s do a multiple linear regression PROPERATI We use multiple attributes to estimate the price. total surface # rooms etc. Each variable is accompanied by a coefficient.
20 Now let卒s go from multiple regression to a neuron PROPERATI total surface # rooms price sum Ps: for simplicity, we have left aside activation functions and non-linearities. product Coefficients are also called weights in this context.
And finally from one neuron to a neural network total surface # rooms price This is what it looks like when some neurons, from now on units, are connected in a feed-forward manner: Inputs Outputs Information propagates from inputs to outputs. Layer A column of units is called a layer. Units within a layer do not connect. A unit is connected to all units of adjacent layers. 21 PROPERATI
And finally from one neuron to a neural network total surface # rooms price This is just like a linear regression. Each connection is weighted, so this model has more parameters. 22 PROPERATI
And finally from one neuron to a neural network total surface # rooms price This is another linear regression. These regressions do not predict the price but a value that is related to it. 23 PROPERATI
And finally from one neuron to a neural network total surface # rooms price And one more linear regression. This regression does predict the price but from some new attributes that are constructed from the originals. 24 PROPERATI
Neural Networks 101 The family of functions that represent feed-forward multilayer networks is somewhat more complex than the family of linear regression; mainly because those functions are compositions of linear regressions. Neural networks usually have much more parameters than simpler models like the regression example Still both type of models can be trained using gradient descent; neural networks additionally require the backpropagation algorithm. More parameters mean more power to capture the underlying patterns in the data. 25 PROPERATI
26 AI + UI/UX and we have our price valuator tool ! PROPERATI
06. AI takeaways
27 AI key takeaways PROPERATI AI is just simple math repeated trillions of times. AI is not programmed, it is taught with labeled data. Biases in the data are transferred to AI what you get out is only as good as the labeled data you put in. Data quantity becomes an advantage when applying deep neural networks. Transfer learning can save a lot of time the intelligence in a neural network is just a long list of numbers, which can be transferred to another empty neural network. In practice theres often a trade-off between feature engineering and model fine- tuning, computation resources. Machines are not really intelligent, they dont have real understanding (yet).
The best way to your next home Thanks!

More Related Content

Global Online Marketplaces Summit 2019

  • 1. Gabriel Gruber AI in RE 07-June-19 The best way to your next home
  • 2. AGENDA 1. OLX Group+Properati 2. AI 101 3. AI at OLX Group 4. AI at Properati 5. AI takeaways
  • 4. Almost 1 year ago ... 01 PROPERATI
  • 5. Properati became the New kids on the block at OLX Group OLX GROUP TODAY: THE WORLD'S #1 CLASSIFIEDS BUSINESS HORIZONTALS REAL ESTATE VERTICALS OTHER VERTICALS CARS VERTICALS Global USA Russia UAE Africa and Philippines Russia Portugal Poland Romania, Egypt Heavy machinery, Global Services Poland Poland South Africa Romania Portugal CONVENIENT TRANSACTIONS Cars Global Cars UAE Cars UAE Latin America 02 PROPERATI Furniture France Jobs India
  • 6. Argentina Uruguay Colombia Ecuador Per炭 We launched 3 new countries, so far :) 03 PROPERATI
  • 7. 01 PROPERATI OUR MISSION To empower buyers so they can have the best way to their next home through tech tools and relevant data. In parallel, we seek to help sellers achieve a more efficient sales process by enabling the best service to their potential buyers. 04 PROPERATI
  • 8. As Seen On TV 05 PROPERATI
  • 10. AI vs ML What's the difference between Machine Learning and Artificial Intelligence? If it is written in Python, it's probably Machine Learning. If it is written in PowerPoint, it's probably AI :) 06 PROPERATI
  • 11. 07 AI challenges for all tech companies PROPERATI Data scale. Scale in infrastructure and engineering capacity to process all that data. The ability to apply AI to solve specific customer problems.
  • 12. 08 AI advantage for the tech giants PROPERATI AI looks tailor-made for the incumbent tech giants. They have quickly moved to put AI at the center of their strategies. Aware of its massive potential, they have started to invest appreciably in AI talent, data infrastructure and even their own dedicated chips.
  • 13. 09 Warren Buffett卒s castle moat in the age of AI PROPERATI AIAI AI
  • 14. 10 But there are some AI opportunities for the rest of us PROPERATI While the tech powerhouses certainly have most of the data now, their Achilles heel may be a lack of deep domain expertise. Many new winners can be created by applying AI to distinct problems in the titans blind spots. Integrating AI very tightly into your business processes should also allow companies to compete with the giants previously thought to be invulnerable.
  • 15. 04. AI at OLX Group
  • 16. 11 AI at OLX Group PROPERATI
  • 17. 12 AI at OLX Group PROPERATI
  • 18. 13 AI at OLX Group PROPERATI
  • 20. 14 Con AI help us answer this question in LATAM ? PROPERATI
  • 21. 15 Let's have a look at Properati today PROPERATI Listings = Datasets Statistics Price valuator tool
  • 22. 16 We have years of listings history that we can use as datasets PROPERATI latitude longitude total_surface covered_surface Price (cop $) 4.731 -74.036 112 112 5.500.000 4.694 -74.078 36 36 2.500.000 4.707 -74.069 124 124 3.000.000 4.696 -74.036 100 100 2.586.000 4.609 -74.067 80 80 2.100.000 Attributes Response (target)
  • 23. 17 Let's say that the price is a function of the total surface (chart) PROPERATI residual a is the slope, price per unit of surface b is the intercept, the price at zero surface
  • 24. 18 This is called linear regression model PROPERATI There is a plethora of options for defining the function f the relationship between the price and the attributes of a property. There is one that: was invented 100 years ago, is still used is simple and awesome It is the linear regression. We are going to pick just one attribute for the moment, the total surface. a and b are parameters of the model. A linear regression with only 1 variable is not optimal for price valuation
  • 25. 19 So let卒s do a multiple linear regression PROPERATI We use multiple attributes to estimate the price. total surface # rooms etc. Each variable is accompanied by a coefficient.
  • 26. 20 Now let卒s go from multiple regression to a neuron PROPERATI total surface # rooms price sum Ps: for simplicity, we have left aside activation functions and non-linearities. product Coefficients are also called weights in this context.
  • 27. And finally from one neuron to a neural network total surface # rooms price This is what it looks like when some neurons, from now on units, are connected in a feed-forward manner: Inputs Outputs Information propagates from inputs to outputs. Layer A column of units is called a layer. Units within a layer do not connect. A unit is connected to all units of adjacent layers. 21 PROPERATI
  • 28. And finally from one neuron to a neural network total surface # rooms price This is just like a linear regression. Each connection is weighted, so this model has more parameters. 22 PROPERATI
  • 29. And finally from one neuron to a neural network total surface # rooms price This is another linear regression. These regressions do not predict the price but a value that is related to it. 23 PROPERATI
  • 30. And finally from one neuron to a neural network total surface # rooms price And one more linear regression. This regression does predict the price but from some new attributes that are constructed from the originals. 24 PROPERATI
  • 31. Neural Networks 101 The family of functions that represent feed-forward multilayer networks is somewhat more complex than the family of linear regression; mainly because those functions are compositions of linear regressions. Neural networks usually have much more parameters than simpler models like the regression example Still both type of models can be trained using gradient descent; neural networks additionally require the backpropagation algorithm. More parameters mean more power to capture the underlying patterns in the data. 25 PROPERATI
  • 32. 26 AI + UI/UX and we have our price valuator tool ! PROPERATI
  • 34. 27 AI key takeaways PROPERATI AI is just simple math repeated trillions of times. AI is not programmed, it is taught with labeled data. Biases in the data are transferred to AI what you get out is only as good as the labeled data you put in. Data quantity becomes an advantage when applying deep neural networks. Transfer learning can save a lot of time the intelligence in a neural network is just a long list of numbers, which can be transferred to another empty neural network. In practice theres often a trade-off between feature engineering and model fine- tuning, computation resources. Machines are not really intelligent, they dont have real understanding (yet).
  • 35. The best way to your next home Thanks!