System and method for creating a preference profile from shared images

1. A method, comprising:

for a subject category of interest having a plurality N of sub-categories, where N is an integer greater than one, training over a first set of images a plurality of image classifiers to yield a plurality of pre-trained classifiers (C 1 . . . C N ), where each individual one of the pre-trained classifiers is trained to recognize one sub-category of the plurality of sub-categories;

obtaining a second set of images (i=1 . . . M) from at least one social media web site that a person has shared/posted/pinned or otherwise expressed an interest in, where M represents a total number of images in the second set of images;

applying at least some of the pre-trained classifiers (C 1 . . . C N ) to each image i in the second set of images;

obtaining a classification score vector {right arrow over (S)} i =[s 1 . . . s N ] for image i(i=1 . . . M), where s j indicates a classification score output from a classifier C j by applying it to image i;

aggregating the M classification score vectors over all of the images of the second set of images;

deriving a data construct that is descriptive of a preference profile of the person with respect to the sub-categories of the category of interest as a probability distribution over the N sub-categories; and

characterizing by how much the person's profile differs from an average profile of a group of persons by assigning the person to a segment, where P a represents a probability vector for the person, and calculating an average P segment _ i for each segment i, where a degree to which the person is a member of the segment i is a dot product of the two vectors:

P a• P segment _ i /(| P a | |P segment _ i |).

2. The method of claim 1 , where the category is fashion, and where the sub-categories are each a particular fashion style.

3. The method of claim 1 , where aggregating the M classification score vectors comprises finding an average of all M classification score vectors:

S

⇀

=

∑

i

=

1

M

⁢

S

⇀

i

.

4. The method of claim 1 , where aggregating the M classification score vectors comprises finding a maximum of all M classification score vectors:

S

⇀

=

Max

i

⁡

[

S

⇀

i

]

.

5. The method of claim 1 , where aggregating the M classification score vectors comprises thresholding each classification score to obtain a binary output and counting the number of classification score vectors falling into each sub-category.

6. The method of claim 1 , where deriving the profile of the person comprises converting confidence scores in the aggregated classification score vector into a probability, where for each sub-category the person's probability of preferring that sub-category equals the percentage of that person's second set of images containing that sub-category.

7. A method, comprising:

obtaining from an online social media site a plurality of instances of images of objects associated with a person;

analyzing with a data processor the plurality of instances of the images with a plurality of predetermined style classifiers to obtain a score for each image for each style classifier; and

determining with the data processor, based on the obtained scores, a likely preference of the person for a particular style of object; where

the method further comprises an initial step performed, for a subject category of interest having a plurality N of sub-categories, where N is an integer greater than one, training over a first set of images a plurality of image classifiers to yield a plurality of pre-trained classifiers (C 1 . . . C N ), where each individual one of the pre-trained classifiers is trained to recognize one sub-category of the plurality of sub-categories;

where the step of obtaining comprises obtaining a second set of images (i=1 . . . M) from at least one social media website that a person has shared/posted/pinned or otherwise expressed an interest in, where M represents a total number of images in the second set of images;

and where the steps of analyzing and determining comprise,

applying at least some of the pre-trained classifiers (C 1 . . . C N ) to each image i in the second set of images;

obtaining a classification score vector {right arrow over (S)} i =[s 1 . . . s N ] for image i(i=1 . . . M), where s j indicates a classification score output from a classifier C j by applying it to image i;

aggregating the M classification score vectors over all of the images of the second set of images;

deriving a data construct that is descriptive of a preference profile of the person with respect to the sub-categories of the category of interest as a probability distribution over the N sub-categories; and

characterizing by how much the person's profile differs from an average profile of a group of persons by assigning the person to a segment, where P a represents a probability vector for the person, and calculating an average P segment _ i for each segment i, where a degree to which the person is a member of the segment i is a dot product of the two vectors:

P a• P segment _ i /(| P a | |P segment _ i |).

8. The method of claim 7 , where aggregating the M classification score vectors comprises one of:

finding an average of all M classification score vectors:

S

⇀

=

∑

i

=

1

M

⁢

S

⇀

i

;

and

finding a maximum of all M classification score vectors:

S

⇀

=

Max

i

⁡

[

S

⇀

i

]

.

9. The method of claim 7 , where aggregating the M classification score vectors comprises thresholding each classification score to obtain a binary output and counting the number of classification score vectors falling into each sub-category.

10. The method of claim 7 , where deriving the profile of the person comprises converting confidence scores in the aggregated classification score vector into a probability, where for each sub-category the person's probability of preferring that sub-category equals the percentage of that person's second set of images containing that sub-category.