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

• US 7,756,879 B2
• Filed: 07/22/2005
• Issued: 07/13/2010
• Est. Priority Date: 07/23/2004
• Status: Expired due to Fees

• Current Assignee: Jeffrey Parsons, Paul Ralph
• Original Assignee: Jeffrey Parsons, Paul Ralph
• Inventors: Parsons, Jeffrey, Ralph, Paul
• Primary Examiner(s): Vital; Pierre M
• Assistant Examiner(s): Vo; Truong V
• Application Number:
  US11/186,951
• Publication Number:
  US 20060041548A1
• Time in Patent Office: 1,817 Days
• Field of Search: 707/5
• US Class Current: 707/748
• 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, the recommendation system comprising:

a memory storing the item data store;

an interface adapted to allow the user to browse an item set made from items in the item data store;

a processor connected to the memory and the interface, the processor comprising;

a ratings estimation module for;

receiving item property data comprising at least one item attribute for each item in the item set, wherein the at least one item attribute is a property of the item;

receiving training set data including user-specific information, wherein the user-specific information comprises at least one user-specific property that relates only to the user, and wherein the at least one user-specific property includes viewing time of at least one item browsed by the user, wherein the viewing time is an individual viewing time the user viewed the specific item; and

generating a training set of items comprising the at least one item browsed by the user; and

generating an item rating matrix by estimating user ratings for the items in the training set based on at least one rating predictor variable;

wherein each rating predictor variable comprises one of the at least one user-specific property and the at least one item attribute;

wherein an entry in the item rating matrix for a training set item e is generated according to;

Item Rating_e=f(x₁, x₂, . . . , x_n)

 

wherein each x_i is a value of the i^th rating predictor variable, n is the number of rating predictor variables, and f is one of a non-linear combination and a linear combination of the rating predictor variables;

wherein at least one of the rating predictor variables is the viewing time; and

,

a recommendation module connected to the ratings estimation module for receiving the item rating matrix and the item set, wherein the recommendation module (i) estimates user ratings for unseen items in the item set based on the item rating matrix without using information about any other user and (ii) generates the recommendation set based on the estimated ratings.

2. 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.

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

4. The recommendation system of claim 3, wherein the at least one item attribute further includes a measure of the at least one item attribute as perceived by the user.

5. 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.

6. The recommendation system of claim 5, wherein the user ratings estimator generates an entry in the item rating matrix for a training set item e based on a combination of the at least one item attribute and at least one rating predictor variable of the training set item e.

7. The recommendation system of claim 5, wherein the user rating estimator generates item rating entries for use by the recommendation module to generate the recommendation set, wherein the user rating estimator generates an entry in the item rating matrix for a training set item e according to:

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

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 the at least one item attribute and the at least one user-specific property.

8. The recommendation system of claim 7, wherein the model is based on one of:

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

9. 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 or 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 unseen items in the item set and generating the recommendation set based on the estimated user ratings for the unseen items.

10. The recommendation system of claim 9, wherein the item set rating estimator generates estimated user ratings for unseen items in the item set and the recommendation module uses the estimated user ratings to generate the recommendation set, wherein the item set rating estimator generates the estimated 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 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.

11. The recommendation system of claim 9, wherein the at least one item attribute includes at least one of a text attribute and a numeric attribute.

12. The recommendation system of claim 11, wherein the numeric attribute 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.

13. The recommendation system of claim 9, 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.

14. The recommendation system of claim 9, wherein the item set rating estimator generates the recommendation set by selecting n_rec unseen items having the highest estimated user ratings where n_rec is an integer value greater than or equal to one.

15. The recommendation system of claim 9, wherein the item set rating estimator generates the recommendation set by selecting the unseen items having an estimated user rating that is greater than a threshold value.

16. The recommendation system of claim 9, 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.

17. 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 unseen items having a certain attribute and does not recommend these unseen items.

18. 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.

19. 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.

20. 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 comprising at least one item attribute for each item in the item set, wherein the at least one item attribute is a property of the item, and training set data including user specific information, wherein the user specific information comprises at least one user-specific property that relates only to the user, and wherein the at least one user-specific property includes viewing time of at least one item browsed by the user, wherein the viewing time is an individual viewing time the user viewed the specific item;

b) generating a training set of items comprising the at least one item browsed by the user;

c) generating an item rating matrix by estimating user ratings for the items in the training set based on at least one rating predictor variable;

wherein each rating predictor variable comprises one of the at least one user-specific property and the at least one item attribute;

wherein an entry in the item rating matrix for a training set item e is generated according to;

Item Rating_e=f(x₁, x₂, . . . , x_n)

wherein each x_i is a value of the i^th rating predictor variable, n is the number of rating predictor variables, and f is one of a non-linear combination and a linear combination of the rating predictor variables;

wherein at least one of the rating predictor variables is the viewing time;

d) estimating user ratings for unseen items in the item set based on the item rating matrix without using information about any other user; and

,

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

21. The method of claim 20, 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.

22. The method of claim 20, wherein the at least one item attribute includes at least one of item complexity, presentation complexity, presentation length, novelty, familiarity, surprisingness, and position in the training set.

23. The method of claim 22, wherein the at least one item attribute further includes a measure of the at least one item attribute as perceived by the user.

24. The method of claim 20, wherein generating the item rating matrix includes:

a) generating pre-processed training set data 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 the at least one item attribute and at least one rating predictor variable of the training set item e.

25. The method of claim 20, wherein generating the item rating matrix includes:

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

,

b) generating at least one item rating entry in the item rating matrix for a training set item e, the at least one item rating entry being used to generate the recommendation set, wherein the at least one item rating entry is generated according to

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

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 the at least one item attribute and the at least one user-specific property.

26. The method of claim 20, 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.

27. The method of claim 21, wherein the method further includes generating estimated user ratings for unseen items in the item set and using the estimated user ratings to generate the recommendation set, wherein estimating the user rating for an unseen item e is done 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.

28. The method of claim 26, wherein the method further includes characterizing the at least one item attribute by at least one of a text attribute and a numeric attribute.

29. The method of claim 28, wherein for estimating user ratings for unseen items in the item set, the method includes generating the numeric attribute 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 j^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.

30. The method of claim 26, 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.

31. The method of claim 26, wherein the method includes generating the recommendation set by selecting n_rec unseen items having the highest estimated user ratings where n_rec is an integer value greater than or equal to one.

32. The method of claim 26, wherein the method includes generating the recommendation set by selecting the unseen items having an estimated user rating that is greater than a threshold value.

33. The method of claim 26 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.

34. The method of claim 20, wherein the method further includes processing the items in the item data store to generate the item property data.