Automated valuation model using a siamese network

1. A method comprising:

receiving, by a server, subject real-estate property listing information associated with a subject real-estate property;

identifying a plurality of comparable real-estate property listings based on attributes of the subject real-estate property listing information;

processing the subject real-estate property listing information together with the plurality of comparable real-estate property listings using a trained machine learning technique to predict a value for the subject real-estate property, the trained machine learning technique being trained to jointly establish a relationship between weights assigned to a set of training comparable real-estate property listings and value adjustments of the set of training comparable real-estate property listings and a value of a real-estate property of interest, the trained machine learning technique being trained by performing training operations including:

selecting, as the real-estate property of interest, a first comparable real-estate property listing from a set of comparable real-estate property listings;

computing a first quantity comprising an estimate of a relative price difference between the real-estate property of interest and a given comparable real-estate property listing;

computing a second quantity comprising a weight characterizing a relative strength of the given comparable real-estate property listing;

adjusting values of a collection of the training comparable real-estate property listings based on the first quantity and the second quantity;

computing a regression loss based on a point estimate of the value of the real-estate property of interest and a ground truth value for the real-estate property of interest; and

updating parameters of the machine learning technique based on the computed regression loss; and

performing an action with respect to the subject real-estate property based on the predicted value of the subject real-estate property.

2. The method of claim 1 , wherein the trained machine learning technique is further trained to jointly optimize mean and uncertainty predictions for the set of training comparable real-estate properties.

3. The method of claim 1 , wherein the trained machine learning technique comprises a Siamese Network.

4. The method of claim 1 , wherein the weights assigned to the set of training comparable real-estate properties represent how much a value of a respective one of the set of training comparable real-estate properties influences the value of the real-estate property of interest, and wherein the value adjustments represent how much more or less expensive the respective one of the set of training comparable real-estate properties is relative to the real-estate property of interest.

5. The method of claim 1 , wherein the trained machine learning technique is configured to output the first quantity and the second quantity.

6. The method of claim 5 , further comprising:

adjusting values of the plurality of comparable real-estate property listings based on the first quantity associated with each of the plurality of comparable real-estate property listings; and

modifying the adjusted values of the plurality of comparable real-estate property listings respectively by the second quantity of each of the plurality of comparable real-estate property listings to compute a point estimate of the value of the subject real-estate property.

7. The method of claim 1 , wherein the training operations are performed before receiving the subject real-estate property listing information associated with a subject real-estate property, the method comprises:

accessing the set of training comparable real-estate property listings;

selecting, as the real-estate property of interest, a first comparable real-estate property listing from the set of comparable real-estate property listings; and

identifying the collection of the training comparable real-estate property listings that have attributes that match attributes of the real-estate property of interest.

8. The method of claim 7 , further comprising repeating the training operations for each of the comparable real-estate properties in the set of comparable real-estate properties, wherein the set of training comparable real-estate property listings is stored as a three-dimensional tensor in which, a first dimension of the three-dimensional tensor represents a plurality of subject properties, a second dimension of the three-dimensional tensor represents a plurality of comparable properties associated with each of the plurality of subject properties, and a third dimension of the three-dimensional tensor represents a plurality of features of the plurality of subject properties and the plurality of comparable properties.

9. The method of claim 7 , further comprising repeating the training operations periodically and in response to receiving new training comparable real-estate property listings.

10. The method of claim 1 , wherein the machine learning technique is trained to compute the weights assigned to the set of training comparable real-estate properties and the value adjustments of the set of training comparable real-estate property listings jointly based on a same set of parameters.

11. The method of claim 1 , further comprising:

predicting, by the trained machine learning technique, a distribution of values for the subject real-estate property, the distribution of values comprising the predicted value.

12. The method of claim 11 , wherein the predicted value comprises a median value for the subject real-estate property, and wherein the distribution of values further comprises at least one of a 10 th percentile value, a 25 th percentile value, a 75 th percentile value or a 99 th percentile value.

13. The method of claim 1 , further comprising:

receiving input requesting that the predicted value for the subject real-estate property be for a specified quantile above a median; and

generating a display comprising data representing at least one of weights or value adjustments associated with the plurality of comparable real-estate property listings based on the predicted value for the subject real-estate property being the specified quantile above the median.

14. The method of claim 1 , further comprising generating, by the trained machine learning technique, a display comprising data representing a distribution of values for the subject real-estate property and changes to weights associated with the plurality of comparable real-estate property listings relative to the distribution of values, the display comprising information on how the weights change as different values along the distribution of values are selected.

15. A system comprising:

a memory that stores instructions; and

one or more processors on a server configured by the instructions to perform operations comprising:

receiving, by a server, subject real-estate property listing information associated with a subject real-estate property;

identifying a plurality of comparable real-estate property listings based on attributes of the subject real-estate property listing information;

processing the subject real-estate property listing information together with the plurality of comparable real-estate property listings using a trained machine learning technique to predict a value for the subject real-estate property, the trained machine learning technique being trained to jointly establish a relationship between weights assigned to a set of training comparable real-estate property listings and value adjustments of the set of training comparable real-estate property listings and a value of a real-estate property of interest, the trained machine learning technique being trained by performing training operations including:

selecting, as the real-estate property of interest, a first comparable real-estate property listing from a set of comparable real-estate property listings;

computing a first quantity comprising an estimate of a relative price difference between the real-estate property of interest and a given comparable real-estate property listing;

computing a second quantity comprising a weight characterizing a relative strength of the given comparable real-estate property listing;

adjusting values of a collection of the training comparable real-estate property listings based on the first quantity and the second quantity;

computing a regression loss based on a point estimate of the value of the real-estate property of interest and a ground truth value for the real-estate property of interest; and

updating parameters of the machine learning technique based on the computed regression loss; and

performing an action with respect to the subject real-estate property based on the predicted value of the subject real-estate property.

16. The system of claim 15 , wherein the trained machine learning technique is further trained to jointly optimize mean and uncertainty predictions for the set of training comparable real-estate properties.

17. The system of claim 15 , wherein the trained machine learning technique comprises a Siamese Network.

18. The system of claim 15 , wherein the weights assigned to the set of training comparable real-estate properties represent how much a value of a respective one of the set of training comparable real-estate properties influences the value of the real-estate property of interest, and wherein the value adjustments represent how much more or less expensive the respective one of the set of training comparable real-estate properties is relative to the real-estate property of interest.

19. The system of claim 15 , wherein the trained machine learning technique is configured to output the first quantity and the second quantity.

20. A non-transitory computer-readable medium comprising instructions stored thereon that are executable by at least one processor to cause a computing device to perform operations comprising:

receiving, by a server, subject real-estate property listing information associated with a subject real-estate property;

identifying a plurality of comparable real-estate property listings based on attributes of the subject real-estate property listing information;

processing the subject real-estate property listing information together with the plurality of comparable real-estate property listings using a trained machine learning technique to predict a value for the subject real-estate property, the trained machine learning technique being trained to jointly establish a relationship between weights assigned to a set of training comparable real-estate property listings and value adjustments of the set of training comparable real-estate property listings and a value of a real-estate property of interest, the trained machine learning technique being trained by performing training operations including:

selecting, as the real-estate property of interest, a first comparable real-estate property listing from a set of comparable real-estate property listings;

computing a first quantity comprising an estimate of a relative price difference between the real-estate property of interest and a given comparable real-estate property listing;

computing a second quantity comprising a weight characterizing a relative strength of the given comparable real-estate property listing;

adjusting values of a collection of the training comparable real-estate property listings based on the first quantity and the second quantity;

computing a regression loss based on a point estimate of the value of the real-estate property of interest and a ground truth value for the real-estate property of interest; and

updating parameters of the machine learning technique based on the computed regression loss; and

performing an action with respect to the subject real-estate property based on the predicted value of the subject real-estate property.