Methods for training saturation-compensating predictors of affective response to stimuli

• US 8,863,619 B2
• Filed: 06/25/2011
• Issued: 10/21/2014
• Est. Priority Date: 05/11/2011
• Status: Active Grant

• Current Assignee: Affectomatics ltd.
• Original Assignee: Anat Thieberger Ben-Haim, Ari M. Frank, Gil Thieberger
• Inventors: Frank, Ari M., Thieberger, Gil, Thieberger Ben-Haim, Anat
• Primary Examiner(s): Gaffin, Jeffrey A
• Assistant Examiner(s): FINK, THOMAS ANDREW
• Application Number:
  US13/168,965
• Publication Number:
  US 20120290514A1
• Time in Patent Office: 1,214 Days
• Field of Search: None
• US Class Current: 76/12
• CPC Class Codes: G06F 3/013 Eye tracking input arrangements G06F3/015 takes precedence
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/02 Knowledge representation; Symbolic representation

Abstract

Described herein are methods for training a machine learning-based predictor of affective response to stimuli. The methods involve receiving samples comprising temporal windows of token instances to which a user was exposed, and target values representing affective response annotations of the user in response to the temporal windows of token instances. This data is used for the training of the predictor along with values indicative of the number of the token instances in the temporal windows of token instances, which are used to compensate for non-linear effects resulting from saturation of the user.

Claims

1. A method comprising:

receiving samples comprising temporal windows of token instances to which a user was exposed;

the token instances spread over at least a day, and a subset of the token instances originating from a same source and having overlapping instantiation periods;

receiving target values representing affective response annotations of the user in response to the temporal windows of token instances; and

training an emotional response predictor on input data comprising;

the samples, the corresponding target values, and values indicative of the number of token instances in the temporal windows of token instances;

wherein the emotional response predictor utilizes the values indicative of the number of the token instances to compensate for non-linear effects resulting from saturation of the user due to the user being exposed to the token instances in the temporal windows of token instances including the token instances originating from the same source and having overlapping instantiation periods.

2. The method of claim 1, further comprising using the emotional response predictor for predicting emotional state of the user after being exposed to a sample comprising a new temporal window of token instances.

3. The method of claim 1, wherein the token instances are spread over a week during which the user is in different situations, and the emotional response predictor captures situation dependent variations in the affective response of the user.

4. The method of claim 3, wherein training the emotional response predictor involves training at least two different machine learning-based emotional response predictors on data collected over periods where the user was in different situations.

5. The method of claim 1, wherein the training sets parameters of a regression model utilized by the emotional response predictor to make its predictions.

6. The method of claim 1, wherein the training sets weights of a neural network utilized by the emotional response predictor to make its predictions.

7. The method of claim 1, wherein the training sets parameters of a support vector machine for regression utilized by the emotional response predictor to make its predictions.

8. The method of claim 1, wherein the training sets parameters of a maximum entropy model classifier utilized by the emotional response predictor to make its predictions.

9. The method of claim 1, wherein the training sets parameters of a random forest model utilized by the emotional response predictor to make its predictions.

10. The method of claim 1, wherein the training sets parameters of a Naive Bayes classifier utilized by the emotional response predictor to make its predictions.

11. The method of claim 1, wherein the training comprises training multiple predictors;

and further comprising combining the predictions of the multiple predictors using an ensemble method.

12. The method of claim 1, further comprising receiving the token instances and a baseline value for the user;

and generating the samples from the token instances and the baseline value for the user.

13. The method of claim 1, further comprising receiving the token instances and initial user measurement channel values representing state of the user at beginning of some of the temporal windows of token instances;

and generating the samples from the token instances and the initial user measurement channel values.

14. The method of claim 1, further comprising receiving the token instances and initial affective response values representing state of the user at beginning of some of the temporal windows of token instances;

and generating the samples from the token instances and the affective response value for the user.

15. The method of claim 1, further comprising receiving additional samples comprising temporal windows of token instances that do not have corresponding target values, and training the emotional response predictor utilizing a semi-supervised training method.

16. A method comprising:

receiving samples comprising temporal windows of token instances, to which a user was exposed, comprising representations of elements extracted from digital media content;

the token instances spread over at least a day, and a subset of the token instances originating from a same digital media source and having overlapping instantiation periods;

receiving target values corresponding to the temporal windows of token instances;

the target values representing affective response annotations of the user in response to the temporal windows of token instances;

training a machine learning-based predictor to predict values representing affective response of the user to being exposed to token instances;

wherein the training is performed on data comprising the samples, the corresponding target values, and values indicative of the number of token instances in the temporal windows of token instances; and

wherein the machine learning-based predictor utilizes the values indicative of the number of the token instances to compensate for non-linear effects resulting from saturation of the user due to the user being exposed to the token instances in the temporal windows of token instances including the token instances originating from the same digital media source and having overlapping instantiation periods.

17. The method of claim 16, wherein the target values are affective response annotations expressed as emotional responses, and the machine learning-based predictor is a machine learning-based emotional response predictor.

18. The method of claim 16, wherein the target values are derived from user measurement channels of the user, and the machine learning-based predictor expresses the predicted response to tokens in terms of values of the user measurement channel of the user.

19. The method of claim 16, further comprising receiving the token instances and a baseline value for the user;

and generating the samples from the token instances and the baseline value for the user.

20. A method comprising:

receiving samples comprising temporal windows of token instances to which a user was exposed;

the token instances spread over at least a day, and a subset of the token instances originating from same source and having overlapping instantiation periods;

receiving target values representing a response of the user to the temporal windows of token instances expressed as values of a measurement channel of the user; and

training a machine learning-based predictor on input data comprising;

the samples, the corresponding target values, and values indicative of the number of token instances in the temporal windows of token instances;

wherein the predictor utilizes the values indicative of the number of the token instances to compensate for non-linear effects resulting from saturation of the user due to the user being exposed to the token instances in the temporal windows of token instances including the token instances originating from the same source and having overlapping instantiation periods.