Estimating probabilities of events in sponsored search using adaptive models
View Patent ↗A machine-learning method for estimating probability of a click event in online advertising systems by computing and comparing an aggregated predictive model (a global model) and one or more data-wise sliced predictive models (local models). The method comprises receiving training data having a plurality of features stored in a feature set and constructing a global predictive model that estimates the probability of a click event for the processed feature set. Then, partitioning the global predictive model into one or more data-wise sliced training sets for training a local model from each of the data-wise slices, and then determining whether a particular local model estimates probability of click event for the feature set better than the global model. A given feature set may be collected from historical data, and may comprise a feature vector for a plurality of query-advertisement pairs and a corresponding indicator that represents a click on the advertisement.
1. A computer-implemented method for estimating probability of click in an online advertising system, said method comprising:
receiving, at a computer, a training set of data comprising a plurality of features in a feature set;
receiving, at a computer, a global model that estimates probability of click for said feature set;
processing, in a computer, said global model and said feature set by:
partitioning said training set of data into a plurality of slices;
training a local model for each of said slices to generate a plurality of trained local models;
determining whether each of said trained local models estimates probability of click for said feature set better than said global model;
selecting, for use in estimating probability of click, a local model from said trained local models when said local model estimates probability of click for said feature set better than said global model; and
selecting, for use in estimating probability of click, said global model when said global model performs better than said local model.
2. The method as set forth in claim 1 , wherein said feature set comprises a feature vector for a plurality of query-advertisement pairs and a binary indicator that represents a click or non-click for said query-advertisement pair.
3. The method as set forth in claim 2 , wherein said slices comprise categories of subject matter for said query-advertisement pairs.
4. The method as set forth in claim 1 , wherein said processing in a computer further comprises selecting, for use in estimating probability of click, a combination of said global model and at least one local model of said trained local models, wherein estimating probability of click comprises a uniform average of said global model and said local model.
5. The method as set forth in claim 1 , wherein said processing in a computer further comprises initializing said local model using flat initialization.
6. The method as set forth in claim 1 , wherein said processing in a computer further comprises initializing said local model using parameters of said global model.
7. The method as set forth in claim 1 , wherein said processing in a computer by partitioning said training set of data into a plurality of slices further comprises:
generating a plurality of clusters from said training data; and
partitioning said training set of data into a plurality of slices based on said clusters.
8. A computer-implemented method for estimating probability of click in a sponsored search system, said method comprising:
storing, in a computer, a plurality of query-advertisement pairs classified into a plurality of categories of subject matter;
storing a global data set for probability of click data for said query-advertisement pairs, wherein said global data set is computed using a global model that estimates probability of click for said query-advertisement pairs across all categories;
storing at least one local data set, a subset of said global data set, for probability of click data for said query-advertisement pairs that corresponds to at least one category, wherein a local data set for a category is computed using a local model that estimates probability of click for said query-advertisement pairs classified in said category;
processing, in a computer, to estimate probability of click by:
receiving a query comprising one or more terms;
determining a category for said query;
estimating probability of click for said category of said query using said global data set when said global model estimates probability of click better than said local mode for said category; and
estimating probability of click for said category of said query using said local data set when said local model for said category estimates probability of click better than said global model.
9. The method as set forth in claim 8 , further comprising selecting at least one sponsored link in response to said query comprises:
storing information about revenue associated with a plurality of candidate sponsored links; and
selecting, at least one of said sponsored links from said candidate sponsored links based on said probability of click estimated for said query and information about revenue associated with said candidate sponsored links.
10. A computer readable medium comprising a set of instructions which, when executed by a computer, cause the computer to estimate probability of click in an online advertising system, said instructions for:
receiving, at a computer, a training set of data comprising a plurality of features in a feature set;
receiving, at a computer, a global model that estimates probability of click for said feature set;
processing, in a computer, said global model and said feature set by:
partitioning said training set of data into a plurality of slices;
training a local model for each of said slices to generate a plurality of trained local models;
determining whether each of said trained local model estimates probability of click for said feature set better than said global model;
selecting, for use in estimating probability of click, a local model from said trained local models when said local model estimates probability of click for said feature set better than said global model; and
selecting, for use in estimating probability of click, said global model when said global model performs better than said local model.
11. The computer readable medium as set forth in claim 10 , wherein said feature set comprises a feature vector for a plurality of query-advertisement pairs and a binary indicator that represents a click or non-click for said query-advertisement pair.
12. The computer readable medium as set forth in claim 11 , wherein said slices comprise categories of subject matter for said query-advertisement pairs.
13. The computer readable medium as set forth in claim 10 , wherein said processing in a computer further comprises instructions for selecting, for use in estimating probability of click, a combination of said global model and at least one local model of said trained local models, wherein estimating probability of click comprises a uniform average of said global model and said local model.
14. The computer readable medium as set forth in claim 10 , wherein said processing in a computer further comprises instructions for initializing said local model using flat initialization.
15. The computer readable medium as set forth in claim 10 , wherein said processing in a computer further comprises instructions for initializing said local model using parameters of said global model.
16. A system for estimating probability of click in an online advertising system, said system comprising:
at least one data storage module; and
a server, comprising a processor and memory, coupled to said data storage modules, for receiving a training set of data comprising a plurality of features in a feature set, for receiving a global model that estimates probability of click for said feature set, and processing said global model and said feature set by:
partitioning said training set of data into a plurality of slices;
training a local model for each of said slices to generate a plurality of trained local models;
determining whether each of said trained local model estimates probability of click for said feature set better than said global model;
selecting, for use in estimating probability of click, a local model from said trained local models when said local model estimates probability of click for said feature set better than said global model; and
selecting, for use in estimating probability of click, said global model when said global model performs better than said local model.
17. The system as set forth in claim 16 , wherein said feature set comprises a feature vector for a plurality of query-advertisement pairs and a binary indicator that represents a click or non-click for said query-advertisement pair.
18. The system as set forth in claim 17 , wherein said slices comprise categories of subject matter for said query-advertisement pairs.
19. The system as set forth in claim 16 , said server for processing further comprises selecting, for use in estimating probability of click, a combination of said global model and at least one local model, wherein estimating probability of click comprises a uniform average of said global model and said local model.
20. The system as set forth in claim 16 , wherein said processing further comprises initializing said local model using flat initialization.
21. The system as set forth in claim 16 , wherein said processing further comprises initializing said local model using parameters of said global model.
22. The system as set forth in claim 16 , wherein said processing by partitioning said training set of data into a plurality of slices further for:
generating a plurality of clusters from said training data; and
partitioning said training set of data into a plurality of slices based on said clusters.