Dimensional isolation prediction in video delivery systems

Patent metadata

Field	Value
Publication number	US-10652617-B2
Application number	US-201815945675-A
Country	US
Kind code	B2
Filing date	Apr 4, 2018
Priority date	Apr 4, 2018
Publication date	May 12, 2020
Grant date	May 12, 2020

Abstract

In one embodiment, a method separates subscriber features generated from subscriber interaction with a video delivery service into feature dimensions and inputs the feature dimensions into a respective prediction network. Each prediction network is trained to output a respective dimension score. The method outputs dimension scores using parameters in the plurality of prediction networks that are trained using a variance term to control a variance of the plurality of feature dimensions and using a de-correlation term to control a correlation of the plurality of feature dimensions. The dimension scores are combined into a retention prediction score and an action is performed on the video delivery service for the subscriber based on the retention score.

First claim

What is claimed is:

1. A method comprising:
separating, by a computing device, a plurality of subscriber features generated from a subscriber's interactions with a video delivery service into a plurality of feature dimensions;
inputting, by the computing device, each of the plurality of feature dimensions into a respective prediction network of a plurality of prediction networks, wherein each prediction network in the plurality of prediction networks is trained to output a respective dimension score;
outputting, by the computing device, a plurality of dimension scores from the plurality of prediction networks using parameters in the plurality of prediction networks that are trained using a variance term to control a variance of the plurality of feature dimensions and using a de-correlation term to control a correlation of the plurality of feature dimensions, wherein the variance term includes a variance variable representing the variance and a variance parameter that controls a strength of an effect of the variance term on a cost, and wherein the variance term divides the variance parameter by the variance variable;
combining, by the computing device, the plurality of dimension scores into a retention prediction score; and
performing, by the computing device, an action on the video delivery service for the subscriber based on the retention score.

2. The method of claim 1 , wherein:
the variance term is added to a cost function that is used to train the plurality of prediction networks, and
the variance term causes a cost to increase when the variance of a dimension score decreases.

3. The method of claim 1 , wherein the variance parameter is tuned to a value for each of the plurality of dimension scores output by the plurality of prediction networks.

4. The method of claim 1 , wherein:
the de-correlation term is added to a cost function that is used to train the plurality of prediction networks, and
the de-correlation term causes a cost to increase when the correlation between dimension scores increases.

5. The method of claim 1 , wherein:
a first de-correlation term is used for all active subscribers, and
a second de-correlation term is used for a portion of the plurality of dimensions.

6. The method of claim 1 , wherein the plurality of prediction networks are trained by:
processing a feature dimension using a plurality of drop out layers; and
processing the feature dimension using a plurality of activation layers.

7. The method of claim 1 , wherein outputting the plurality of dimension scores comprises:
determining whether a flag field for a feature dimension indicates data for the feature dimension is good or bad;
encoding the flag field as a first number when the data is good; and
encoding the flag field as a random number when the data is bad.

8. The method of claim 1 , wherein training the prediction networks comprises:
randomly removing a portion of the plurality of dimension scores during the training of the prediction networks.

9. The method of claim 1 , wherein training the prediction networks comprises:
disabling a portion of the plurality of dimension scores during the training of the prediction networks of inactive subscribers.

10. The method of claim 1 , wherein the plurality of feature dimensions are generated based on correlations between features.

11. A method comprising:
separating, by a computing device, a plurality of subscriber features generated from a subscriber's interactions with a video delivery service into a plurality of feature dimensions;
inputting, by the computing device, each of the plurality of feature dimensions into a respective prediction network of a plurality of prediction networks, wherein each prediction network in the plurality of prediction networks is trained to output a respective dimension score;
outputting, by the computing device, a plurality of dimension scores from the plurality of prediction networks using parameters in the plurality of prediction networks that are trained using a variance term to control a variance of the plurality of feature dimensions and using a de-correlation term to control a correlation of the plurality of feature dimensions, wherein the de-correlation term adds a sum over off-diagonal elements of a correlation matrix to a cost function that is minimized;
combining, by the computing device, the plurality of dimension scores into a retention prediction score; and
performing, by the computing device, an action on the video delivery service for the subscriber based on the retention score.

12. The method of claim 11 , wherein the variance term includes a variance variable representing the variance and a variance parameter that controls a strength of an effect of the variance term on the cost.

13. The method of claim 12 , wherein the variance term divides the variance parameter by the variance variable.

14. A method comprising:
separating, by a computing device, a plurality of subscriber features generated from a subscriber's interactions with a video delivery service into a plurality of feature dimensions;
inputting, by the computing device, each of the plurality of feature dimensions into a respective prediction network of a plurality of prediction networks, wherein each prediction network in the plurality of prediction networks is trained to output a respective dimension score;
outputting, by the computing device, a plurality of dimension scores from the plurality of prediction networks using parameters in the plurality of prediction networks that are trained using a variance term to control a variance of the plurality of feature dimensions and using a de-correlation term to control a correlation of the plurality of feature dimensions;
combining, by the computing device, the plurality of dimension scores into a retention prediction score, wherein combining the plurality of dimension scores into the retention prediction score comprises multiplying the plurality of dimension scores to generate the retention prediction score; and
performing, by the computing device, an action on the video delivery service for the subscriber based on the retention score.

15. A non-transitory computer-readable storage medium containing instructions, that when executed, control a computer system to be configured for:
separating a plurality of subscriber features generated from a subscriber's interactions with a video delivery service into a plurality of feature dimensions;
inputting each of the plurality of feature dimensions into a respective prediction network of a plurality of prediction networks, wherein each prediction network in the plurality of prediction networks is trained to output a respective dimension score;
outputting a plurality of dimension scores from the plurality of prediction networks using parameters in the plurality of prediction networks that are trained using a variance term to control a variance of the plurality of feature dimensions and using a decorrelation term to control a correlation of the plurality of feature dimensions, wherein:
a first de-correlation term is used for all active subscribers, and
a second de-correlation term is used for a portion of the plurality of dimensions;
combining the plurality of dimension scores into a retention prediction score; and
performing an action on the video delivery service for the subscriber based on the retention score.

16. A method comprising:
outputting, by a computing device, a plurality of dimension scores from a plurality of prediction networks using a plurality of subscriber features generated from a subscriber's interaction

Assignees

• Hulu Llc

Inventors

• Becker Nathan
• Zhou Colin
• Holcombe Matthew
• Phadnis Atul Arun
• Li Hang
• Subramanian Sridhar Srinivasa
• Huff Kristen

Classifications

• H04N21/252

  Processing of multiple end-users' preferences to derive collaborative data · CPC title

• H04N21/4662 Primary

  characterized by learning algorithms · CPC title

• H04N21/6582

  Data stored in the client, e.g. viewing habits, hardware capabilities, credit card number (arrangements where receivers interact with the broadcast H04H20/38) · CPC title

• H04N21/44213

  Monitoring of end-user related data (arrangements for monitoring the users' behaviour or opinions in broadcast systems H04H60/33) · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title