Efficient prediction
View Patent ↗Disclosed are a system and method for constructing and using a predictive model to generate a prediction signal, also referred to as a classification signal when the signal indicates one of a plurality of distinct classes. In various embodiments, the disclosed technique reduces a size of a predictive Support Vector Model by extracting certain values beforehand and storing only weighting values. The technique does not sacrifice generalization performance but does significantly reduce the model size and accelerate prediction performance. The described system applies to most kernel functions, whether linear or nonlinear.
1. A method of generating a classification indicator on an electronic device based on a predictive model, the electronic device comprising a device-sensor group for sensing data related to an environment of the electronic device, the method comprising:
generating one or more respective weight vector coefficients for one or more support vectors of the predictive model on a computing device remote from the electronic device;
storing the predictive model including the one or more respective weight vector coefficients but excluding the support vectors on the electronic device;
running the predictive model on the electronic device based on the weight vector coefficients and input received from the device-sensor group;
generating a classification decision based on running the predictive model; and
generating a classification indicator to modify a behavior of the electronic device.
2. The method of claim 1 wherein the electronic device is one of a cellular phone, a smart phone, and a tablet computer.
3. The method of claim 1 wherein generating one or more respective weight vector coefficients comprises approximating a radial basis function kernel.
4. The method of claim 1 wherein the predictive model is associated with one of a linear kernel and a nonlinear kernel.
5. The method of claim 1 wherein the classification indicator represents a current state or activity of the electronic device or of a user.
6. The method of claim 1 wherein running the predictive model on the electronic device comprises running multiple predictive models on the electronic device.
7. The method of claim 1 wherein the classification decision represents a mood, activity, or environment of a user.
8. The method of claim 1 wherein the predictive model is a polynomial model.
9. The method of claim 1 wherein generating a classification indicator to modify a behavior of the electronic device comprises generating a classification indicator signaling that a user is not present at the electronic device.
10. An electronic device for generating a classification indicator based on a predictive model, the device comprising:
a device-sensor group for sensing data related to an environment of the electronic device;
a memory wherein is stored a predictive model including one or more respective weight vector coefficients for one or more support vectors of the predictive model; and
a processor configured to receive data from the device-sensor group and to run the predictive model based on the weight vector coefficients and the received data and to generate a classification indicator to modify the behavior of the electronic device.
11. The electronic device of claim 10 wherein the electronic device is one of a cellular phone, a smart phone, and a tablet computer.
12. The electronic device of claim 10 wherein the processor is further configured to generate the one or more respective weight vector coefficients by approximating a radial basis function kernel.
13. The electronic device of claim 10 wherein the predictive model is a polynomial kernel.
14. The electronic device of claim 10 wherein the classification indicator represents a current state or activity of the electronic device or of a user.
15. The electronic device of claim 10 wherein the processor is further configured to run multiple predictive models on the electronic device.
16. The electronic device of claim 10 wherein the predictive model is linear.
17. The electronic device of claim 10 wherein the predictive model is non-linear.
18. A method of making a prediction at an electronic device, the method comprising:
producing a model including a plurality of result-effective features and one or more support vectors;
generating one or more respective weight vector coefficients for the one or more support vectors;
storing the model, including the one or more respective support vector weight coefficients but not the one or more support vectors, in the electronic device; and
outputting a prediction signal from at least a portion of the electronic device.
19. The method of claim 18 wherein the electronic device is one of a cellular phone, a smart phone, and a tablet computer.
20. The method of claim 18 wherein the prediction signal represents a user mood, activity, or environment.