METHOD AND SYSTEM OF OPTIMIZING A RANKED LIST OF RECOMMENDED ITEMS

• US 20140188865A1
• Filed: 12/28/2012
• Published: 07/03/2014
• Est. Priority Date: 12/28/2012
• Status: Active Grant

• Current Assignee: Telefonica SA
• Original Assignee: Telefonica SA
• Inventors: KARATZOGLOU, Alexandros, BALTRUNAS, Linas
• Granted Patent: US 8,935,303 B2
• Time in Patent Office: Days
• US Class Current: 707/732
• CPC Class Codes: G06Q 30/0282 Rating or review of business operators or products

Abstract

A method and system of optimizing a ranked list (5) of recommended items that is based in a multidimensional data set (2) comprising context-aware information about the of a plurality of users and a plurality of items. A mathematical recommendation model (3) is trained with the multidimensional data set (2) by applying a smooth objective function that allows the use of fast optimising algorithm and that quantifies the relevance of the ranked lists provided by an optimization algorithm.

Claims

1. A method of optimizing an output ranked list (5) of recommended items given an input user, an input item list, and an input context, comprising:

providing a multidimensional data set (2) that comprises information of interactions from a plurality of users (6) with a plurality of items (7) and in a plurality of contexts (B);

computing a mathematical recommendation model (3) by optimizing an objective function over the multidimensional data set (2), the recommendation model comprising a score value for each combination of user, item and context;

and computing the output ranked list (5) by applying the computed recommendation model to the input user, input item list and input context;

wherein that the recommendation model (3) further comprises a ranked list of recommended items for each user and context, being each ranked list determined by sorting the scores of the plurality of items (7) for each user and context; and

in that the objective function is a smooth function that quantifies a relevance of the recommended items of each ranked list of the recommendation model (3), calculated over at least some of the plurality of users (6) and over at least some of the plurality of contexts (8).

2. The method of claim 1 further comprising factorizing the recommendation model (3) into as many two-dimensional matrices (9) as dimensions of the multidimensional data set (2).

3. The method of claim 1 wherein the multidimensional data set (2) is a binary dataset that indicates whether each user has interacted with each item in each context.

4. The method of claim 1 wherein the data set (2) comprises information about a plurality of context variables, and the information of interactions of each user and each item with all the context variables is comprised in a single entry of the multidimensional data set (2).

5. The method of claim 1 wherein the objective function is biased according to a rank of each item in the ranked list.

6. The method of claim 5 wherein the objective function is a smooth function that approximates a mean average precision function.

7. The method of claim 6 wherein the objective function approximates (a) a binary condition relative to the ranks in the ranked list of two given items, by a continuous and differentiable function relative to the scores of two given items in the recommendation model, (b) a rank in the ranked list of a given item by a continuous and differentiable function relative to a score of the given item in the recommendation model, or (c) both (a) and (b).

9. The method of claim 1 wherein the recommendation model (3) is optimized with a gradient ascend algorithm.

10. The method of claim 1 wherein the step of computing the recommendation model (3) further comprises selecting a subset of data from the multidimensional data set (2), being the objective function optimized over the subset of data.

11. The method of claim 10 wherein the subset of data for a given user and a given context is formed by all the items the given user has interacted with in the given context, and by a subset of the items the given user has not interacted with in the given context.

12. The method of claim 11 wherein the subset of items the given user has not interacted with, is a plurality of items with a score greater than a given threshold.

13. The method of claim 11 wherein the chosen subset of items the given user has not interacted with is a plurality of items with a score greater than the lowest score among all the items the given user has interacted with in the given context.

14. The method of claim 1 wherein the interactions whose information is comprised in the multidimensional data set (2) are selected from a group of implicit feedbacks comprising:

a click, a mouse movements, a purchase, an installation of an application, a browsing history, an usage history and a search pattern.

15. The method of claim 1 wherein the plurality of contexts comprise at least one context variable selected from a group comprising:

time, date, season, location, activity, weather, emotional state, social network, user device, company and speed.

16. The method of claim 1 wherein said calculation over at least some of the plurality of users (6) and over at least some of the plurality of contexts (8) is an average over at least some of the plurality of users (6) and over at least some of the plurality of contexts (8).

21. A computer program comprising computer program code means adapted to perform the steps of the method according to claim 1 when said program is run on a computer, a digital signal processor, a field-programmable gate array, an application-specific integrated circuit, a micro-processor, a micro-controller, or any other form of programmable hardware.

22. A digital data storage medium storing a computer program product comprising instructions causing a computer executing the program, to perform all steps of a method according to any of claim 1.

8. (canceled)

17. A system of optimizing an output ranked list (5) of recommended items given an input user and an input item list comprising context awareness means (18) adapted to determine an input context of the user;

an interface (15) adapted to show information and receive commands from the input user; and

computing means (21) adapted to;

provide a multidimensional data set (2) that comprises information of interactions from a plurality of users (6) with a plurality of items (7) and in a plurality of contexts (8);

compute a mathematical recommendation model (3) by optimizing an objective function over the multidimensional data set (2), the recommendation model comprising a score value for each combination of user, item and context;

and compute the output ranked list (5) by applying the computed recommendation model to the input user, input item list and input context;

wherein the recommendation model (3) further comprises a ranked list of recommended items for each user and context, being each ranked list determined by sorting the scores of the plurality of items (7) for each user and context; and

in that the objective function is a smooth function that quantifies a relevance of the recommended items of each ranked list of the recommendation model (3), calculated over at least some of the plurality of users (6) and over at least some of the plurality of contexts (8).

18. The system according to claim 17 wherein the context awareness means (18) are comprised in a user device (12), the computing means (21) are comprised in a remote server (13), and the system further comprises connection means (17) adapted to connect the user device (12) and the remote server (13).

19. The system according to claim 17 wherein the remote server further comprises (a) a context processor (18) adapted to retrieve context information about the interactions between the plurality of users (6) and the plurality of items (7), (b) a purchasing means (19) adapted to receive purchasing orders from the user through the interface and to send said purchasing orders to an external provider (23), or (c) both (a) and (b).

20. (canceled)