Online trained object property estimator
This disclosure describes systems and methods for using an estimator to produce values for dependent variables of streaming objects based on values of independent variables of the objects. The systems and methods may include continuously tuning the estimator based on any objects received with pre-populated values for the dependent variables.
1. A method for estimating values of unknown features of a series of objects, the objects being represented as digital feature vectors, at least some of such digital feature vectors including a plurality of X variables having corresponding values populated on the respective digital feature vector before the estimating as a plurality of populated X-variables, and, for at least a portion of the series of objects, a Y-variable having an unknown value prior to the estimating, the method comprising operations performed by an estimator, the estimator comprising hardware logic and/or software logic executing via processing circuitry, the operations comprising:
maintaining, by the estimator on a non-transitory digital memory, an array of models, wherein
each model of the array of models comprises a Y-variable value and a plurality of X-variable values corresponding to the respective Y-variable value, and
the array of models is organized according to Y-variable values of the models; and
for each object of at least a portion of the series of objects,
computing, by the estimator, an estimated value of the Y-variable of the respective object by identifying, from the array of models, a set of closest matching models based on a measure of difference between the values of at least a subset of the X-variables of the respective object and a corresponding subset of the X-variable values of at least a portion of the models of the array of models, and
producing the estimated value of the Y-variable of the respective object based at least in part on a Y-variable value of one or more models of the set of closest matching models,
wherein said maintaining includes:
ascertaining and populating an actual value corresponding to the estimated value of the Y-variable of the respective object to make it a training object,
as a real-time operation interleaved with the estimating values of the unknown features of the series of objects, inserting the training object within the series of objects, and
updating a model of the array of models corresponding to the Y-variable value of the training object, based at least in part on the X-variable values of the training object and of existing X-variable values of the model.
2. The method of claim 1 , wherein:
the array of models comprises a set of model banks arranged by their associated object classifications, and
the identifying, from the array of models, a set of closest matching models, furthermore involves:
determining, for the respective object, at least one candidate classification from a plurality of classifications, and
for each given one of one or more model banks associated with the at least one candidate classification, finding, from object models in the given bank, one or more closest matching models, each associated with said classification, for the respective object.
3. The method of claim 1 , wherein the updating forms updated X-variable values for the model corresponding to the Y-variable value of the training object by computing, for any given X-variable, a weighted average of the respective X-variable value of the training object and the respective existing X-variable value of the model.
4. The method of claim 3 , wherein for the computing of the weighted average, the weight factor for the existing X-variable values of the model is
(i) increased based on occurrences of the estimated value of the Y-variable matching the corresponding ascertained actual value, and/or
(ii) decreased based on occurrences of the estimated value of the Y-variable not matching the corresponding ascertained actual value.
5. The method of claim 1 , wherein the updating of a model of the array of models corresponding to the Y-variable value of a given training object involves, in case said array of models does not include any model corresponding to the training object, creating a new object model in the array based on variable values of that training object.
6. The method of claim 5 , where the model array is considered to not include a model corresponding to a given training object in case a vector distance measure between the given training object and any of the existing object models in the array is above a configured threshold distance.
7. The method of claim 5 , where the model array is considered to not include a model corresponding to a given training object in case the array of models does not include any model corresponding to the Y-variable value of the training object.
8. The method of claim 1 , wherein the operations further comprise outputting, by the estimator, a given object of the series of objects as an output object, wherein a value of the Y-variable of the digital feature vector of the output object is set to the estimated value for the given object.
9. The method of claim 1 , wherein, in case a given one of the objects in the series has one or more of the X-variables unpopulated on its respective digital feature vector before the estimating, said X-variables referred to as unpopulated variables, the operations further comprise outputting, by the estimator, such given object of the series of objects as an output object, wherein a value of at least one of said unpopulated variables of the output object is set to an estimated value based at least in part based on variable values a model of the array of models
(i) that corresponds to the estimated value of the Y-variable, and/or
(ii) that is closest in vector distance to the given object when considering its populated X-variables.
10. The method of claim 1 , wherein the populated X-variables include further a set of synthesized X-variables formed based at least in part on values of received X-variables of the digital feature vector of the respective object.
11. A system for estimating values of unknown features of a series of objects, the objects being represented as digital feature vectors, at least some of such digital feature vectors including a plurality of X variables having corresponding values populated on the respective digital feature vector before the estimating as a plurality of populated X-variables, and, for at least a portion of the series of objects, a Y-variable having an unknown value prior to the estimating, the system comprising an estimator comprising hardware logic and/or software logic stored on a non-transitory digital medium and configured for execution via processing circuitry, the estimator performing operations comprising:
maintaining, by the estimator on a non-transitory digital memory, an array of models, wherein
each model of the array of models comprises a Y-variable value and a plurality of X-variable values corresponding to the respective Y-variable value, and
the array of models is organized according to Y-variable values of the models; and
for each object of at least a portion of the series of objects,
computing, by the estimator, an estimated value of the Y-variable of the respective object by identifying, from the array of models, a set of closest matching models based on a measure of difference between the values of at least a subset of the X-variables of the respective object and a corresponding subset of the X-variable values of at least a portion of the models of the array of models, and
producing the estimated value of the Y-variable of the respective object based at least in part on a Y-variable value of one or more models of the set of closest matching models,
wherein said maintaining includes:
ascertaining and populating an actual value corresponding to the estimated value of the Y-variable of the respective object to make it a training object,
as a real-time operation interleaved with the estimating values of the unknown features of the series of objects, inserting the training object within the series of objects, and
updating a model of the array of models corresponding to the Y-variable value of the training object, based at least in part on the X-variable values of the training object and of existing X-variable values of the model.
12. The system of claim 11 , wherein:
the array of models comprises a set of model banks arranged by their associated object classifications, and
the identifying, from the array of models, a set of closest matching models, furthermore involves:
determining, for the respective object, at least one candidate classification from a plurality of classifications, and
for each given one of one or more model banks associated with the at least one candidate classification, finding, from object models in the given bank, one or more closest matching models, each associated with said classification, for the respective object.
13. The system of claim 11 , wherein the updating forms updated X-variable values for the model corresponding to the Y-variable value of the training object by computing, for any given X-variable, a weighted average of the respective X-variable value of the training object and the respective existing X-variable value of the model.
14. The system of claim 13 , wherein for the computing of the weighted average, the weight factor for the existing X-variable values of the model is
(iii) increased based on occurrences of the estimated value of the Y-variable matching the corresponding ascertained actual value, and/or
(iv) decreased based on occurrences of the estimated value of the Y-variable not matching the corresponding ascertained actual value.
15. The system of claim 11 , wherein the updating of a model of the array of models corresponding to the Y-variable value of a given training object involves, in case said array of models does not include any model corresponding to the training object, creating a new object model in the array based on variable values of that training object.
16. The system of claim 15 , where the model array is considered to not include a model corresponding to a given training object in case a vector distance measure between the given training object and any of the existing object models in the array is above a configured threshold distance.
17. The system of claim 15 , where the model array is considered to not include a model corresponding to a given training object in case the array of models does not include any model corresponding to the Y-variable value of the training object.
18. The system of claim 15 , wherein the operations further comprise outputting, by the estimator, a given object of the series of objects as an output object, wherein a value of the Y-variable of the digital feature vector of the output object is set to the estimated value for the given object.
19. The system of claim 15 , wherein, in case a given one of the objects in the series has one or more of the X-variables unpopulated on its respective digital feature vector before the estimating, said X-variables referred to as unpopulated variables, the operations further comprise outputting, by the estimator, such given object of the series of objects as an output object, wherein a value of at least one of said unpopulated variables of the output object is set to an estimated value based at least in part based on variable values a model of the array of models
(iii) that corresponds to the estimated value of the Y-variable, and/or
(iv) that is closest in vector distance to the given object when considering its populated X-variables.
20. The system of claim 11 , wherein the populated X-variables include further a set of synthesized X-variables formed based at least in part on values of received X-variables of the digital feature vector of the respective object.