Methods, systems, and media for providing direct and hybrid data acquisition approaches
Methods, systems, and media for providing direct and hybrid data acquisition approaches are provided. In accordance with some embodiments of the disclosed subject matter, a method of data acquisition for construction of classification models that incorporates multiple human reviewing resources is provided, the method comprising: receiving a cost structure for constructing a classification model using a data set; instructing human reviewing resources to search through the data set and select instances of a class that satisfy a criterion, providing the human reviewing resources with a definition of the class; training the classification model with the instances from the human reviewing resources; determining when an expected gain for performing additional searches by the human reviewing resources as a function of the cost structure is lower than a given threshold; and, instructing the human reviewing resources searching through the set to label one or more examples from the data set.
1. A method for data acquisition for construction of classification models, the method comprising:
(a) receiving, using a hardware processor, a budget and a cost structure for constructing a classification model using a data set that includes positive and negative examples of a class of interest, wherein the classification model is set to a guided learning mode that receives selected instances of the class of interest that satisfy at least one criterion from the data set and wherein the cost structure includes a cost of a search performed by a human reviewer;
(b) in response to the classification model being set to the guided learning mode, transmitting, using the hardware processor, a definition of the class of interest over the Internet for review by a plurality of human reviewers;
(c) in response to the classification model being set to the guided learning mode, transmitting, using the hardware processor, instructions over the Internet to the plurality of human reviewers to search through the data set and select one or more instances of the class that satisfy at least one criterion;
(d) receiving, using the hardware processor, an indication over the Internet that the one or more instances from the data set have been selected by at least one of the plurality of human reviewers;
(e) training the classification model in the guided learning mode, using the hardware processor, with the one or more selected instances;
(f) estimating, using the hardware processor, a performance of the classification model after training the classification model with the one or more selected instances to construct a learning curve;
(g) determining, using the hardware processor, a rate of change in the estimated performance of the classification model as a function of the cost of at least one of the plurality of human reviewers performing an additional search in the guided learning mode, wherein the rate of change in the estimated performance of the classification model is determined based on a slope of the learning curve;
(h) repeating (b)-(g) until it is determined at (g) that the rate of change in the estimated performance as a function of the cost of the at least one of the plurality of human reviewers performing the additional search in the guided learning mode is lower than a first predetermined threshold;
(i) switching, using the hardware processor, the classification model from the guided learning mode to an active learning mode, which includes machine learning, that receives labelled instances from portions of the data set selected as being useful to the classification model in response to determining that the rate of change in the estimated performance as a function of the cost of the at least one of the plurality of human reviewers performing the additional search in the guided learning mode is lower than the first predetermined threshold at (h), until the estimated performance of the classification model is greater than a second predetermined threshold or the budget for constructing the classifier has been exhausted, wherein the plurality of human reviewers include human reviewers that are available to label one or more instances of the class of interest during training of the classification model using the active learning mode and human reviewers that are available to search for the one or more instances of the class of interest in response to the instruction transmitted at (c), and wherein the training of the classification model using the active learning mode is discontinued in response to determining that the rate of change in the estimated performance as a function of a cost of labeling performed by the plurality of human reviewers is lower than a third predetermined threshold;
(j) receiving, using the hardware processor, identifying information of a web page to be classified by the trained classification model;
(k) classifying, using the hardware processor, the web page using the trained classification model to determine whether the web page is a member of the class of interest; and
(l) transmitting information indicating whether the web page is a member of the class of interest to an advertiser, over the Internet, in response to receiving a request for classification information about the web page.
2. The method of claim 1 , further comprising:
allocating, using the hardware processor, a portion of the budget to the plurality of human reviewers available for searching and a remaining portion of the budget to the plurality of human reviewers available for labeling.
3. The method of claim 1 , wherein the at least one criterion includes a criterion that the one or more instances to be selected are to be positive examples of the class of interest.
4. The method of claim 1 , wherein a subset of the data set is selected for presentation to the plurality of human reviewers by at least one of: uncertainty sampling and boosted disagreement with query-by-committee.
5. The method of claim 1 , wherein the data set includes online resources containing pointers to the class of interest, and wherein the instructions to search through the data set further comprise instructions to query the online resources for examples of the class of interest that meet the at least one criterion.
6. A system for data acquisition for construction of classification models, the system comprising:
a processor that:
(a) receives a budget and a cost structure for constructing a classification model using a data set that includes positive and negative examples of a class of interest, wherein the classification model is set to a guided learning mode that receives selected instances of the class of interest that satisfy at least one criterion from the data set and wherein the cost structure includes a cost of a search performed by a human reviewer;
(b) in response to the classification model being set to the guided learning mode, transmits a definition of the class of interest over the Internet for review by a plurality of human reviewers;
(c) in response to the classification model being set to the guided learning mode, transmits instructions to the plurality of human reviewers over the Internet to search through the data set and select one or more instances of the class that satisfy at least one criterion;
(d) receives an indication over the Internet that the one or more instances from the data set have been selected by at least one of the plurality of human reviewers;
(e) trains the classification model in the guided learning mode with the one or more selected instances;
(f) estimates a performance of the classification model after training the classification model with the one or more selected instances to construct a learning curve;
(g) determines a rate of change in the estimated performance of the classification model as a function of the cost of at least one of the plurality of human reviewers performing an additional search in the guided learning mode, wherein the rate of change in the estimated performance of the classification model is determined based on a slope of the learning curve;
(h) repeats (a)-(g) until it is determined at (g) that the rate of change in the estimated performance as a function of the cost of the at least one of the plurality of human reviewers performing the additional search in the guided learning mode is lower than a first predetermined threshold;
(i) switches the classification model from the guided learning mode to an active learning mode, which includes machine learning, that receives labelled instances from portions of the data set selected as being useful to the classification model in response to determining that the rate of change in the estimated performance as a function of the cost of the at least one of the plurality of human reviewers performing the additional search in the guided learning mode is lower than the first predetermined threshold at (h), until the estimated performance of the classification model is greater than a second predetermined threshold or the budget for constructing the classifier has been exhausted, wherein the plurality of human reviewers include human reviewers that are available to label one or more instances of the class of interest during training of the classification model using the active learning mode and human reviewers that are available to search for the one or more instances of the class of interest in response to the instruction transmitted at (c), and wherein the training of the classification model using the active learning mode is discontinued in response to determining that the rate of change in the estimated performance as a function of a cost of labeling performed by the plurality of human reviewers is lower than a third predetermined threshold;
(j) receives identifying information of a web page to be classified by the trained classification model;
(k) classifies the web page using the trained classification model to determine whether the web page is a member of the class of interest; and
(l) transmitting information indicating whether the web page is a member of the class of interest to an advertiser, over the Internet, in response to receiving a request for classification information of the web page.
7. The system of claim 6 , wherein the processor is further configured to:
allocate a portion of the budget to the plurality of human reviewers available for searching and a remaining portion of the budget to the plurality of human reviewers available for labeling.
8. The system of claim 6 , wherein the at least one criterion includes a criterion that one or more instances to be selected are to be positive examples of the class of interest.
9. The system of claim 6 , wherein a subset of the data set is selected for presentation to the plurality of human reviewers by at least one of: uncertainty sampling and boosted disagreement with query-by-committee.
10. The system of claim 6 , wherein the data set includes online resources containing pointers to the class of interest, and wherein the processor is further configured to instruct the plurality of human reviewers to query the online resources for examples of the class of interest that meet the at least one criterion.
11. A non-transitory computer-readable medium containing computer-executable instructions that, when executed by a processor, cause the processor to perform a method for data acquisition for construction of classification models, the method comprising:
(a) receiving a budget and a cost structure for constructing a classification model using a data set that includes positive and negative examples of a class of interest, wherein the classification model is set to a guided learning mode that receives selected instances of the class of interest that satisfy at least one criterion from the data set and wherein the cost structure includes a cost of a search performed by a human reviewer;
(b) in response to the classification model being set to the guided learning mode, transmitting a definition of the class of interest over the Internet to a plurality of human reviewers;
(c) in response to the classification model being set to the guided learning mode, transmitting instructions to the plurality of human reviewers to search through the data set and select one or more instances of the class that satisfy at least one criterion;
(d) receiving an indication over the Internet that the one or more instances from the data set have been selected by at least one of the plurality of human reviewers;
(e) training the classification model in the guided learning mode with the one or more selected instances;
(f) estimating a performance of the classification model after training the classification model with the one or more selected instances to construct a learning curve;
(g) determining a rate of change in the estimated performance of the classification model as a function of the cost of at least one of the plurality of human reviewers performing an additional search in the guided learning mode, wherein the rate of change in the estimated performance of the classification model is determined based on a slope of the learning curve;
(h) repeating (b)-(g) until it is determined at (g) that the rate of change in the estimated performance as a function of the cost of the at least one of the plurality of human reviewers performing the additional search in the guided learning mode is lower than a first predetermined threshold;
(i) switching, using the hardware processor, the classification model from the guided learning mode to an active learning mode, which includes machine learning, that receives labelled instances from portions of the data set selected as being useful to the classification model in response to determining that the rate of change in the estimated performance as a function of the cost of the at least one of the plurality of human reviewers performing the additional search in the guided learning mode is lower than the first predetermined threshold at (h), until the estimated performance of the classification model is greater than a second predetermined threshold or the budget for constructing the classifier has been exhausted, wherein the plurality of human reviewers include human reviewers that are available to label one or more instances of the class of interest during training of the classification model using the active learning mode and human reviewers that are available to search for the one or more instances of the class of interest in response to the instruction transmitted at (c), and wherein the training of the classification model using the active learning mode is discontinued in response to determining that the rate of change in the estimated performance as a function of a cost of labeling performed by the plurality of human reviewers is lower than a third predetermined threshold;
(j) receiving identifying information of a web page to be classified by the trained classification model;
(k) classifying the web page using the trained classification model to determine whether the web page is a member of the class of interest; and
(l) transmitting information indicating whether the web page is a member of the class of interest to an advertiser, over the Internet, in response to receiving a request for classification information about the web page.