Back to results
US 10,897,649 B1Granted
Mature themes prediction for online content
AssigneeAmazon Technologies, Inc.
InventorsVernon Germano, Kripa Kanchana Sivakumar, Xiang Hao, Emily Evon McAninly
Life of this patent
Estimating the term…
Abstract
Methods and apparatus are described that relate to the use of machine learning techniques to accurately predict ratings for various types of content across multiple mature themes.
Claims18
A computer-implemented method, comprising:
- for each of a plurality of training video titles, labeling the training video title with a plurality of mature theme ratings, each mature theme rating corresponding to one of a plurality of mature themes and representing an intensity for the corresponding mature theme for the training video title, the mature themes including violence and sexuality;
- generating a feature set for each training video title, each feature set including visual features derived from images of the corresponding training video title using computer vision techniques, each feature set also including caption features derived from caption data of the corresponding training video title;
- generating a machine learning model for each of the mature themes using the feature sets for the training video titles and the mature theme ratings;
- identifying an unrated video title;
- generating a feature set for the unrated video title, the feature set for the unrated video title including visual features derived from images of the unrated video title using computer vision techniques, the feature set for the unrated video title also including caption features derived from caption data of the unrated video title;
- generating a plurality of mature theme ratings for the unrated video title by comparing the feature set for the unrated video title with the feature sets for a subset of the training video titles and using the machine learning model for each of the mature themes, each mature theme rating for the unrated video title representing an intensity for a corresponding one of the mature themes for the unrated video title;
- and
- recommending the unrated video title to a subset of a plurality of users based on the mature theme ratings associated with the unrated video title.
5 more elements
A computer-implemented method, comprising:
- selecting a plurality of machine learning models, each of the machine learning models corresponding to one of a plurality of mature themes, the plurality of mature themes including violence and sexuality, the machine learning models representing feature sets for a plurality of first content items, each feature set including visual features derived from images of the corresponding first content item and caption features derived from caption data of the corresponding first content item;
- identifying a second content item;
- generating a feature set for the second content item that includes visual features derived from images of the second content item and caption features derived from caption data of the second content item;
- generating a plurality of ratings for the second content item by comparing the feature set for the second content item with the feature sets for the first content items and using the machine learning model for each of the mature themes, each of the ratings corresponding to one of the mature themes and representing an intensity for the corresponding mature theme, the ratings including a violence rating and a sexuality rating;
- and
- using each of the ratings explicitly in connection with distribution of the second content item.
3 more elements
A computer program product, comprising one or more non-transitory, computer-readable media having computer program instructions stored therein, the computer program instructions being configured such that, when executed by one or more computing devices, the computer program instructions cause the one or more computing devices to:
- select a plurality of machine learning models, each of the machine learning models corresponding to one of a plurality of mature themes, the plurality of mature themes including violence and sexuality, the machine learning models representing feature sets for a plurality of first content items, each feature set including visual features derived from images of the corresponding first content item and caption features derived from caption data of the corresponding first content item;
- identify a second content item;
- generating a feature set for the second content item that includes visual features derived from images of the second content item and caption features derived from caption data of the second content item;
- generate a plurality of ratings for the second content item by comparing the feature set for the second content item with the feature sets for the first content items and using the machine learning model for each of the mature themes, each of the ratings corresponding to one of the mature themes and representing an intensity for the corresponding mature theme, the ratings including a violence rating and a sexuality rating;
- and
- use each of the ratings explicitly in connection with distribution of the second content item.
3 more elements
Dependencies are inferred from the claim text.