Smart exploration methods for mitigating item cold-start problem in collaborative filtering recommendation systems

Patent metadata

Field	Value
Publication number	US-9910898-B2
Application number	US-201514857258-A
Country	US
Kind code	B2
Filing date	Sep 17, 2015
Priority date	Sep 17, 2015
Publication date	Mar 6, 2018
Grant date	Mar 6, 2018

Abstract

Systems and methods for building a latent item vector and item bias for a new item in a collaborative filtering system are disclosed. The method includes dividing incoming users into intervals with each interval having a learning phase and a selection phase. The learning phase scores each incoming user according to a best estimate latent vector and bias and saves the highest score. In the selection each incoming user is scored and a user exceeding the highest score is selected. The best estimate latent vector and bias is then updated based on the user's vector and bias, and the user's interaction with the item. The updated best estimate latent vector is then used in further intervals for learning and selecting users.

First claim

The invention claimed is:

1. A method for building a latent item vector and item bias for a new item in a collaborative filtering recommendation system, wherein each user has a vector representing a user's characteristics and a bias representing the likelihood that the user will be interested in an arbitrary item, the method comprising:
determining a first quantity of users to use as a training group and a second quantity of users to use as a recommendation group;
receiving a first plurality of users with each user received sequentially;
scoring each of the first plurality of users based on each user's vector and bias, and an best estimate latent vector to find a maximum individual score for the first plurality of users;
receiving a second plurality of users with each user received sequentially;
scoring each of the second plurality of users as they arrive with each user's vector and bias, and the best estimate latent vector;
in response to an event selected from the group consisting of receiving a user of the second plurality of users with a score greater than the maximum individual score and receiving a final user from among the second plurality of users, selecting the user for presenting the item; and
recording the selected user's interaction with the item and updating the best estimate latent vector, wherein the updating is based on the selected user's vector, bias, and the selected user's interaction with the item.

2. The method of claim 1 , wherein calculating each score comprises adding the user's bias to a dot product of the user's vector and the best estimate latent vector.

3. The method of claim 1 , wherein the best estimate latent vector is updated to a vector and bias that give a minimum squared error between a predicted interaction and the selected user's interaction.

4. The method of claim 3 , further comprising:
determining a third quantity of users to use as a second training group and a fourth quantity of users to use as a second recommendation group;
receiving a third plurality of users with each user received sequentially;
scoring each of the third plurality of users based on each user's vector and bias, and the updated best estimate latent vector to find a second maximum individual score for the third plurality of users;
receiving a fourth plurality of users with each user received sequentially;
scoring each of the fourth plurality of users as they arrive with each user's vector and bias, and the updated best estimate latent vector;
in response to an event selected from the group consisting of receiving a second user of the fourth plurality of users with a score greater than the second maximum individual score and receiving a second final user from among the fourth plurality of users, selecting a second user for presenting the item; and
recording the selected second user's interaction with the item and updating the best estimate latent vector and bias, wherein the updating is based on the second user's vector, bias, and the selected user's interaction with the item.

5. The method of claim 1 , wherein the training group and the selection group each have the same quantity of users.

6. The method of claim 1 , wherein the updating the best estimate latent vector is initially a zero vector.

7. A method for building a latent item vector and item bias for a new item in a collaborative filtering recommendation system, wherein each user has a vector representing a user's characteristics and a bias representing the likelihood that the user will be interested in an arbitrary item, the method comprising:
receiving a quantity of users for selection as a test user;
determining an interval for each of the quantity of users for selection as a test user;
dividing each interval into a training phase and a selection phase; and
for each interval:
sequentially receiving a first plurality of users during the training phase;
scoring each of the first plurality of users based on each user's vector and bias, and an best estimate latent vector to find a maximum individual score for the first plurality of users;
sequentially receiving a second plurality of users during the selection phase;
scoring each of the second plurality of users as they arrive with each user's vector and bias, and the best estimate latent vector;
in response to an event selected from the group consisting of receiving a user of the second plurality of users with a score greater than the maximum individual score and receiving a final user from among the second plurality of users, selecting the user for presenting the item; and
recording the selected user's interaction with the item and updating the best estimate latent vector and bias to an updated best estimate latent vector and bias that gives a minimum squared error between predicted interactions of previously selected users and the actual interactions of the previously selected users, wherein the updating is based on the previously selected users' vectors, biases, and the previously selected users' actual interactions with the item.

8. The method of claim 7 , wherein calculating each score comprises adding the user's bias to a dot product of the user's vector and the best estimate latent vector.

9. The method of claim 7 , wherein each interval contains the same number of users.

10. The method of claim 7 , wherein the training phase and the selection phase have the same number of users.

11. The method of claim 7 , wherein the best estimate latent vector is initially a zero vector.

12. The method of claim 7 , wherein the interaction comprises rating the item.

13. The method of claim 7 , wherein the method builds a latent item vector and item bias for a plurality of new items.

14. The method of claim 13 , wherein each of the users is scored for each of the plurality of items and the maximum score is the maximum for all of the items.

15. The method of claim 13 , wherein each of the users is scored for each of the plurality of items and a maximum score is saved for each item.

16. The method of claim 13 , wherein each of the users is scored for each of the plurality of items and the combination of a user and item having the highest score is selected.

17. A system for building a latent item vector and item bias for a new item in a collaborative filtering recommendation system, wherein each user has a vector representing a user's characteristics and a bias representing the likelihood that the user will be interested in an arbitrary item, the system comprising:
a processor;
memory storing computer executable instructions, that when implemented by the processor cause the system to perform functions comprising:
receiving a quantity of users for selection as a test user;
determining an interval for each of the quantity of users for selection as a test user;
dividing each interval into a training phase and a selection phase; and
for each interval:
sequentially receiving a first plurality of users during the training phase;
scoring each of the first plurality of users based on each user's vector and bias, and an best estimate latent vector to find a maximum individual score for the first plurality of users;
sequentially receiving a second plurality of users during the selection phase;
scoring each of the second plurality of users as they arrive with each user's vector and bias, and the best estimate latent vector;
in response to an event selected from the group consisting of receiving a user of the second plurality of users with a score greater than the maximum individual score and receiving a final user from among the second plurali

Assignees

• Yahoo Holdings Inc

Inventors

• Somekh Oren Shlomo
• Aharon Michal
• Golan Shahar
• Avigdor Elgrabli Noa
• Cohen Dana Drachsler

Classifications

• G06N7/01

  Probabilistic graphical models, e.g. probabilistic networks · CPC title

• G06N5/04

  Inference or reasoning models · CPC title

• G06F17/3053 Primary

  Physics · mapped topic

• G06F17/30867

  Physics · mapped topic

• G06N99/005

  Physics · mapped topic