System and method for estimating user ratings from user behavior and providing recommendations

• US 20060041548A1
• Filed: 07/22/2005
• Published: 02/23/2006
• Est. Priority Date: 07/23/2004
• Status: Active Grant

• Current Assignee: Jeffrey Parsons, Paul Ralph
• Original Assignee: Jeffrey Parsons, Paul Ralph
• Inventors: Parsons, Jeffrey, Ralph, Paul
• Granted Patent: US 7,756,879 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 16/954 Navigation, e.g. using categorised browsing
• CPC Class Codes: G06Q 30/02 Marketing; Price estimation or determination; Fundraising

Abstract

A recommendation system and method for generating a set of recommended items for a user who browses items from an item data store. The recommendation system estimates ratings for at least some of the items previously viewed by the user, and then constructs a representation of the user'"'"'s implicit ratings of attributes of these items. These implicit attribute ratings are then used to estimate the user'"'"'s ratings for unseen items in the item data store.

Claims

1. A recommendation system for providing a recommendation set of recommended items from an item data store to a user who uses an interface to browse an item set made from items in the item data store, the recommendation system comprising:

a) a ratings estimation module that receives item property data including item attribute data for items in the item set, the ratings estimation module being connected to the interface for receiving training set data including user-specific information and being adapted to generate a training set based on at least one item browsed by the user and, based on at least one user-specific property and one or more item attributes, the ratings estimation module generates an item rating matrix by estimating user ratings for the items in the training set; and

,

b) a recommendation module connected to the ratings estimation module for receiving the item rating matrix and the item set, wherein the recommendation module estimates user ratings for unseen items in the item set and generates the recommendation set based on the estimated ratings.

2. The recommendation system of claim 1, wherein the at least one user-specific property includes viewing time of an item in the item set.

3. The recommendation system of claim 1, wherein the at least one user-specific property further includes at least one of the age of the user, the gender of the user, degree of experience of the user in the same or a similar setting, and the mental state of the user.

4. The recommendation system of claim 1, wherein the item attributes include at least one of item complexity, presentation complexity, presentation length, novelty, familiarity, surprisingness, and position in the training set.

5. The recommendation system of claim 4, wherein the item attributes further include a measure of at least one of the item attributes as perceived by the user.

6. The recommendation system of claim 1, wherein the ratings estimation module includes:

a) a training set generator for generating the training set based on the training set data and the item property data for at least some of the items viewed by the user;

b) a data pre-processing module connected to the training set generator for receiving the training set and generating pre-processed training set data by applying zero or more statistical transformations; and

,

c) a user ratings estimator connected to the data pre-processing module for receiving the pre-processed training set data and generating the item rating matrix.

7. The recommendation system of claim 6, wherein the user ratings estimator generates an entry in the item rating matrix for a training set item e based on a combination of item attribute values for the training set item e and rating predictor variables.

8. The recommendation system of claim 6, wherein the user rating estimator generates an entry in the item rating matrix for a training set item e according to:

$\mathrm{Item}<br><br>\text{\hspace{1em}<br><br>}<br><br>{\mathrm{Rating}}_e=\frac{c+a_1<br><br>x_1+a_2<br><br>x_2+\mathrm{\dots <br><br>}+a_n<br><br>x_n}{3}$

where c is a constant, each a_i is a coefficient corresponding to a rating predictor variable, and each x_i is a value of the i^th rating predictor variable and where the coefficients a_i are generated according to a model based on a combination of item attributes and user-specific properties.

9. The recommendation system of claim 8, wherein the model can be based on one of:

linear regression, discriminant analysis, canonical correlation analysis, and structural equations modeling.

10. The recommendation system of claim 1, wherein the recommendation module includes:

a) a data pre-processing module for receiving the items in the item set and generating pre-processed items by applying zero ore more statistical transformations;

b) a user profile generator connected to the data pre-processing module for receiving the item rating matrix and generating a user profile having some subset of attributes of the pre-processed items along with at least one of an inferred user rating and explicit user rating, the inferred user rating being derived for some or all of the attributes from ratings in the item rating matrix; and

,

c) an item set rating estimator connected to the user profile generator for receiving the user profile and the pre-processed items for generating estimated user ratings for items in the item set and generating the recommendation set based on the estimated user ratings for the items.

11. The recommendation system of claim 10, wherein the item set rating estimator generates the user rating for an item e according to

${\mathrm{Rating}}_e=\frac{\sum _{}^{}}{}$

i = 1 

R 

⁢

r i ⁢

w i ∑

i = 1 

W 

⁢

w i

in which r_i is an attribute rating of the i^th attribute, w_i is relative importance of the i^th attribute, |R| is the number of attributes in the user profile and |W| is the number of weights.

12. The recommendation system of claim 10, wherein the item attributes include at least one of text attributes and numeric attributes.

13. The recommendation system of claim 12, wherein at least one of the numeric attributes in the user profile is generated according to

${\mathrm{UQ}}_i=\frac{\sum _{}^{}}{}$

j = 1 

Tp 

⁢

r j ⁢

q ij ∑

J = 1 

Tp 

⁢

r j

in which r_j is the rating of the j^th positive item in the training set, q_ij is the value of the i^th numeric attribute for the j^th positive training set item and |Tp| is the number of positive items and wherein a positive item is defined as an item having an attribute value greater than a predefined threshold value.

14. The recommendation system of claim 10, wherein for at least some numeric attributes, the user profile generator selects a subset of items in the training set to generate the user profile, the subset being selected based on items in the training set having a rating greater than a predefined threshold value.

15. The recommendation system of claim 10, wherein the item set rating estimator generates the recommendation set by selecting the n_rec items having the highest estimated user ratings where n_rec is an integer value greater than or equal to one.

16. The recommendation system of claim 10, wherein the item set rating estimator generates the recommendation set by selecting the items having an estimated user rating that is greater than a threshold value.

17. The recommendation system of claim 10, wherein the item set rating estimator places a given unseen item in the recommendation set only when the estimated user rating of the given unseen item exceeds the estimated user ratings of all items previously viewed by the user.

18. The recommendation system of claim 1, wherein the recommendation module, based on the user-specific information, is adapted for inferring that the user is not interested in items having a certain attribute and does not recommend these items.

19. The recommendation system of claim 1, wherein the recommendation system includes the item data store, and the item data store is preprocessed to provide the item property data.

20. The recommendation system of claim 1, wherein the item data store is separate from the recommendation system, and the recommendation system further includes an item attribute data generator for processing the items in the item data store to generate the item property data.

21. A method for providing a recommendation set of recommended items from an item data store to a user who uses an interface to browse an item set made from items in the item data store, the method including:

a) obtaining item property data including item attribute data for items in the item set and training set data including user specific property information;

b) generating a training set based on at least one item browsed by the user and attribute values for the at least one item;

c) generating an item rating matrix by estimating user ratings for items in the training set based on at least one user specific property and one or more item attributes;

d) estimating user ratings for unseen items in the item set based on the item rating matrix; and

,

e) generating the recommendation set based on the estimated ratings.

22. The method of claim 21, wherein the at least one user-specific property includes viewing time of an item in the item set.

23. The method of claim 21, wherein for the at least one user-specific property, the method includes obtaining at least one of the age of the user, the gender of the user, the degree of experience of the user in the same or a similar setting, and the mental state of the user.

24. The method of claim 21, wherein the item attributes include at least one of item complexity, presentation complexity, presentation length, novelty, familiarity, surprisingness, and position in the training set.

25. The method of claim 24, wherein the item attributes further include a measure of at least one of the attributes as perceived by the user.

26. The method of claim 21, wherein generating the item rating matrix includes:

a) generating pre-processed training items by applying zero or more statistical transformations to the training set; and

,

b) generating an entry in the item rating matrix for a training set item e based on a combination of item attribute values for the training set item e and rating predictor variables.

27. The method of claim 21, wherein generating the item rating matrix includes:

a) generating pre-processed training items by applying zero or more statistical transformations to the training set items; and

,

b) generating an entry in the item rating matrix for a training set item e according to

$\mathrm{Item}<br><br>\text{\hspace{1em}<br><br>}<br><br>{\mathrm{Rating}}_e=\frac{c+a_1<br><br>x_1+a_2<br><br>x_2+\mathrm{\dots <br><br>}+a_n<br><br>x_n}{3}$

where c is a constant, each a_i is a coefficient corresponding to a rating predictor variable, and each x_i is a value of the i^th rating predictor variable and where the coefficients a_i are generated according to a model based on a combination of item attributes and user-specific properties.

28. The method of claim 21, wherein estimating user ratings for unseen items includes:

a) generating pre-processed items by applying zero or more statistical transformation to the unseen items in the item set;

b) generating a user profile having some subset of attributes of the pre-processed items along with at least one of an inferred user rating and explicit user rating, the inferred user rating being derived for some or all of the attributes from ratings in the item rating matrix; and

,

c) generating the estimated user ratings for some or all of the unseen items in the item set based on the estimated user ratings for the items.

29. The method of claim 23, wherein the method further includes estimating the user rating for an unseen item e according to

${\mathrm{Rating}}_e=\frac{\sum _{}^{}}{}$

i = 1 

R 

⁢

r i ⁢

w i ∑

i = 1 

W 

⁢

w i

in which r_i is a rating of the i^th attribute, w_i is relative importance of the i^th attribute, |R| is the number of attributes in the user profile and |W| is the number weights.

30. The method of claim 28, wherein the method further includes characterizing item attributes by at least one of text attributes and numeric attributes.

31. The method of claim 30, wherein the method includes generating at least one of the numeric attributes in the user profile according to

${\mathrm{UQ}}_i=\frac{\sum _{}^{}}{}$

j = 1 

Tp 

⁢

r j ⁢

q ij ∑

J = 1 

Tp 

⁢

r j

in which r_j is the rating of the i^th positive item in the training set, q_ij is the value of the i^th numeric attribute for the j^th positive item in the training set and |Tp| is the number of positive items and wherein a positive item is defined as an item having an attribute value greater than a predefined threshold value.

32. The method of claim 28, wherein for at least some numeric attributes, the method includes selecting a subset of items in the training set to generate the user profile, the subset being selected based on items in the training set having a rating greater than a predefined threshold value.

33. The method of claim 28, wherein the method includes generating the recommendation set by selecting n_rec items having the highest estimated user ratings where n_rec is an integer value greater than or equal to one.

34. The method of claim 28, wherein the method includes generating the recommendation set by selecting the items having an estimated user rating that is greater than a threshold value.

35. The method of claim 28 where the method includes placing a given unseen item in the recommendation set only when the estimated user rating of the given unseen item exceeds the ratings of all items previously viewed by the user.

36. The method of claim 21, wherein the method further includes processing the items in the item data store to generate the item property data.