Data structures for collaborative filtering systems

• US 8,560,528 B2
• Filed: 03/17/2010
• Issued: 10/15/2013
• Est. Priority Date: 03/17/2010
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Herbrich, Ralf, Bachrach, Yoram
• Primary Examiner(s): Boccio, Vincent F
• Application Number:
  US12/725,862
• Publication Number:
  US 20110231405A1
• Time in Patent Office: 1,308 Days
• Field of Search: 707/722, 707/732, 707/748, 707/749, 707/754, 707/758, 707/733, 707/734, 707/723, 715/811
• US Class Current: 707/722
• CPC Class Codes: G06F 16/24573 using data annotations, e.g. user-defined metadata

Abstract

Data structures for collaborative filtering systems are described. In an embodiment sketches which extremely concisely represent a list of items that a user has rated are created and stored for use by a collaborative filtering system to recommend items. For example, the sketches are created by using several versions of a cryptographic hash function to permute the item list and store a minimal value from each permutation in the sketch together with a user rating. In examples the sketches are used to compute estimates of similarity measures between pairs of users such as rank correlations including Spearman'"'"'s Rho and Kendall'"'"'s Tau. For example, the similarity measures are used by a collaborative filtering system to accurately and efficiently recommend items to users. For example the sketches are so concise that massive amounts of data can be taken into account in order to give high quality recommendations in a practical manner.

Claims

1. A computer-implemented method comprising:

accessing an item list that includes one or more items that have been rated by a user; and

creating and storing a sketch of the item list, the sketch being a data structure storing a concise description of the item list, wherein creating each sketch includes;

selecting a hash function from a plurality of hash functions of a type;

using the hash function to generate a permutation of the item list, the permutation including a plurality of hashed values, each of the plurality of hashed values corresponding to at least one item in the item list;

storing a minimum value of the permutation in the sketch, the minimum value being a minimum of the plurality of hashed values;

repeatedly generating other permutations of the item list, using other hash functions selected from the plurality of hash functions of the type;

storing other minimum values of the other permutations in the sketch; and

storing item ratings in the sketch such that there is one stored item rating associated with each stored minimum value, the one stored item rating being a rating, made by the user, associated with an item represented by the stored minimum value.

2. A method as claimed in claim 1 wherein the hash functions are cryptographic hash functions.

3. A method as claimed in claim 1 wherein the hash functions are randomly selected from a min-wise independent family of hashes.

4. A method as claimed in claim 1 wherein the hash functions are keyed hash message authentication code versions of a cryptographic hash function.

5. A method as claimed in claim 1 which further comprises:

receiving a target user identifier;

arranging a processor to identify pairs of users that include the target user; and

for each of the identified pairs, computing an estimate of a rank correlation by using the sketches associated with the identified pairs of users.

6. A method as claimed in claim 5 which further comprises:

providing the rank correlation estimates to a collaborative filtering system; and

arranging the collaborative filtering system to generate recommendations of items for the target user identifier on the basis of the rank correlation estimates.

7. A method as claimed in claim 5 wherein the rank correlation is a rank correlation of the form

8. A method as claimed in claim 5 wherein the rank correlation includes at least one of Kendall'"'"'s Tau and Spearman'"'"'s Rho.

9. A method as claimed in claim 1 wherein the item lists are accessed from secondary storage or as a data stream and wherein the sketches are stored in RAM.

10. A method comprising:

accessing a sketch for each user of a pair of users, each sketch being a data structure holding a description of a list of items rated by a user comprising a plurality of item identifiers and a rating for each item identifier;

each sketch being smaller than its associated item list, each sketch created based on a plurality of permutations of the list of items, and each permutation generated using a hash function selected from a plurality of hash functions of a same type, a size of the sketch being determined based on an amount of memory available;

arranging a processor to identify sketch collisions between the sketches where item identifiers at corresponding positions in the sketches are the same; and

arranging the processor to examine the ratings of each item that occurs on a sketch collision and to use those ratings to compute an estimate of the rank correlation.

11. A method as claimed in claim 10 wherein the processor is further arranged to compute an estimate of Spearman'"'"'s Rho by using the examined ratings to compute Spearman'"'"'s Rho random variables and combining those.

12. A method as claimed in claim 10 wherein the processor is further arranged to find the proportion of sketch collisions which are agreements and use that proportion to compute an estimate of Kendall'"'"'s Tau.

13. A method as claimed in claim 12 wherein the processor is further arranged to identify sketch collisions which are agreements by finding two adjacent collisions in a pair of sketches, for those collisions identifying the highest rating in each sketch, comparing the items having the identified ratings and identifying a sketch collisions agreement if the items are the same.

14. A system comprising:

a memory holding a plurality of sketches, each sketch being a data structure holding a description of a list of items rated by a user comprising a plurality of item identifiers and a rating for each item identifier;

each sketch being smaller than its associated item list, each sketch created based on a plurality of permutations of the list of items, and each permutation generated using a hash function selected from a plurality of hash functions of a same type; and

a processor arranged to;

identify sketch collisions between pairs of the sketches where item identifiers at corresponding positions in the sketches are the same;

examine the ratings of items that occur on sketch collisions and to use those ratings to compute estimates of a rank correlation between pairs of users; and

predict the rating a target user would give to an unexamined item using at least some of the rank correlation estimates.

15. A system as claimed in claim 14 wherein the processor is further arranged to make the prediction using the ratings given to the unexamined item by other users, weighted according to the rank correlation estimates.

16. A system as claimed in claim 14 wherein the processor is further arranged to compute estimates of a rank correlation which is any rank correlation of the form

17. A system as claimed in claim 14 wherein the processor is further arranged to compute estimates of a rank correlation which is selected from any of Kendall'"'"'s Tau and Spearman'"'"'s Rho.

18. A system as claimed in claim 14 further comprising a sketch creation engine arranged to create each sketch by using a hash function to generate a permutation of the item list, storing a minimum value of the permutation in the sketch, repeatedly generating permutations and storing a minimum value in the sketch for more hash functions, and also storing item ratings in the sketch such that there is one stored item rating associated with each stored minimum value.

19. A system as claimed in claim 18 wherein the hash functions are keyed hash message authentication code versions of a cryptographic hash function.

20. A system as claimed in claim 14 wherein the processor is further arranged to generate recommendations of items for the target user using the predictions.