Automatically predicting relevant contexts for media items

Patent metadata

Field	Value
Publication number	US-10936653-B2
Application number	US-201715720880-A
Country	US
Kind code	B2
Filing date	Sep 29, 2017
Priority date	Jun 2, 2017
Publication date	Mar 2, 2021
Grant date	Mar 2, 2021

Abstract

The present technology pertains to automatically contextlabeling media items with relevant contexts, and further for algorithmically generating high quality playlists built around a context that are personalized to a profile of an account. This is accomplished by combining data from observed playlists, and data representing intrinsic properties of media items to predict contexts for media items.

First claim

The invention claimed is:

1. At least one non-transitory computer readable medium comprising instructions stored on the computer readable medium that when executed cause a computing system to:
obtain a plurality of media items from a plurality of user-defined playlists from user accounts of a media service;
determine a plurality of contexts based on playlist titles of the plurality of user-defined playlists;
identify context associations between media items and the plurality of contexts;
generate a behavioral representation of the plurality of media items based on the context associations, wherein the behavioral representation is based on occurrences of media items appearing in common playlists;
generate an intrinsic representation of the plurality of media items based on properties of each of the plurality of media items;
form a combination representation from the behavioral representation and the plurality of intrinsic representations;
and
select media items for inclusion in a candidate pool for an input context, the selected media items being associated with the input context in the combination.

2. The at least one non-transitory computer readable medium of claim 1 wherein the instructions are effective to cause the computing system to:
receive the input context from a client device associated with a user account of the user accounts, the input context originating from a request made by a user of the device, wherein the device has processed the request to determine the input context.

3. The at least one non-transitory computer readable medium of claim 1 wherein the instructions are effective to cause the computing system to:
determine media items from the candidate pool that are compatible with a media item profile of a user account of the user accounts, for which a playlist is being created; and
create the playlist from the media items that are compatible with the media item profile of the account.

4. The at least one non-transitory computer readable medium of claim 3 wherein the instructions to create a playlist include instructions to:
apply a sequencing model to the media items compatible with the media item profile to determine media items that can be combined into an acceptable sequence for inclusion in the playlist.

5. The at least one non-transitory computer readable medium of claim 1 wherein the instructions to determine the collection of account defined contexts include instructions to:
aggregate playlist titles from accounts of a media service;
remove playlist titles for which context is not ascertainable; and
select the playlist titles that remain that appear in greater than a threshold number of accounts of the media service to yield a collection of account defined contexts.

6. The at least one non-transitory computer readable medium of claim 1 wherein the instructions are effective to cause the computing system to:
create a behavioral representation for media items available from a media service, wherein the instructions to create the behavioral representation include instructions to cause the computing system to:
perform clustering on the media items available from the media service, wherein the media items are clustered according to a likelihood that respective media items appear together in a playlist;
derive the behavioral representation for each media item based on its respective arrangement within clusters with respect to other media items.

7. The at least one non-transitory computer readable medium of claim 1 wherein the instructions are effective to cause the computing system to:
create an intrinsic property representation for media items available from a media service, wherein the instructions to create the intrinsic property representation include instructions to cause the computing system to:
analyze the media items available from the media service to determine physical properties of each media item;
analyze the media items available from the media service to determine semantic characteristics of each media item;
combine the physical properties with the semantic characteristics to result in an intrinsic property representation.

8. The at least one non-transitory computer readable medium of claim 7 , wherein the instructions to analyze the media items available from the media service to determine physical properties of each media item include instructions to:
analyze each media item for timbre properties;
analyze each media item for rhythm properties; and
output the physical properties representing the timbre properties and the rhythm properties.

9. The at least one non-transitory computer readable medium of claim 8 , wherein the instructions to analyze each media item for timbre properties includes analyzing each media item in temporal segments with a convolutional neural network to create a representation of timbre properties, and
wherein the instructions to analyze each media item for rhythm properties includes instructions to analyze the representation of timbre properties over time with the convolutional neural network to yield the rhythm properties.

10. The at least one non-transitory computer readable medium of claim 1 , wherein the instructions are effective to cause the computing system to:
train a learning model to insert media items into the combination, the learning model being trained by optimizing media item insertions to:

minimize a distance between pairs of account defined contexts and media items associated with the account defined contexts,
minimize a distance between pairs of media items having similar respective behavioral and intrinsic representations,
minimize a distance between pairs of media items having similar respective behavioral representations, and
minimize a distance between pairs of media items having similar respective intrinsic property representations, which results in an arranged context embedding space.

11. The at least one non-transitory computer readable medium of claim 10 , wherein the instructions to include media items not associated with the account defined contexts into the context embedding space includes instructions to:
insert a combined behavioral and intrinsic representations for the respective media items into the combination using the trained learning model;
insert the behavioral representations for the respective media items into the combination using the trained learning model;
insert the intrinsic property representations for the respective media items into the combination using the trained learning model.

12. The at least one non-transitory computer readable medium of claim 10 , wherein the instructions are effective to cause the computing system to:
output media-item-context associations based on a distance between representations of a respective media item and an account defined context in the combination.

13. The at least one non-transitory computer readable medium of claim 12 , wherein the instructions to select media items for inclusion in the candidate pool includes selecting media items having a media-item-context association with the input context greater than a threshold, and wherein the instructions are effective to cause the computing system to:
select media items that are compatible with a media item profile of an account from the candidate pool for inclusion in a playlist targeted to the input context.

14. A system comprising:
one or more processors; and
at least one non-transitory computer readable medium comprising instructions stored on the computer readable medium that when executed by the one or more processors cause the system to:
obtain a plurality of media items from

Assignees

• Apple Inc

Inventors

• Levy Mark H
• Shenoy Arvind S
• Mauch Matthias
• Cartoon Daniel

Classifications

• G06N5/01

  Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/084 Primary

  Backpropagation, e.g. using gradient descent · CPC title