Generating a model based on input
Described herein are techniques for generating a model based on input. In an example, a first model can be generated based on decomposing a rating matrix into a product of two matrices U and V, where U represents correlations between users and a plurality of factors and V represents correlations between items and the plurality of factors. At least some of the users and items can be grouped into groups according to the correlations. Input on one or more of the groups can be received. A second model can be generated based on the input on the one or more of the groups.
1. A method comprising:
generating, by a processor, a first user-item preference model by decomposing a rating matrix comprising users, items, and ratings of the items provided by the users into a product of two matrices U and V, wherein the items, and the items include a plurality of components which are factors related to the items that caused the users to provide the ratings of the items, wherein the matrix U includes correlations between the users in the rating matrix and the factors related to the items and the matrix V includes correlations between the items in the rating matrix and the factors related to the items;
grouping, by the processor, the users and items correlated to dominant factors in the matrices U and V into groups, wherein the dominant factors are factors whose correlations with the users and items in the matrices U and V are greater than a predetermined threshold representing a separation between users who prefer the items and users who do not prefer the items;
receiving an input that revises the groups of the users and items;
generating, by the processor, a second user-item preference model of the product of the matrices U and V based on the revised groups of the users and items after receipt of the input; and
rendering for display, on a user interface, recommended items based on the second user-item preference model.
2. The method of claim 1 , wherein the input comprises approving or disapproving that one or more of the users or items belongs to a particular group.
3. The method of claim 1 , wherein the input comprises moving one or more of the users or items from one group to another group.
4. The method of claim 1 , wherein the input comprises indicating that one or more users or items are similar to one or more other users or items.
5. The method of claim 1 , wherein the input comprises merging or splitting one or more of the groups.
6. The method of claim 1 , wherein the input comprises deleting one or more of the users or items such that the one or more of the users or items are not used to generate the second user-item preference model.
7. The method of claim 1 , wherein generating the first user-item preference model by decomposing the rating matrix into a product of two matrices U and V comprises:
generating the first user-item preference model by decomposing the rating matrix into a product of two matrices U and V that minimize an objective function.
8. The method of claim 7 , wherein generating the second user-item preference model comprises:
adding a mathematical constraint reflecting the input to the objective function; and
generating the second user-item preference model by decomposing the rating matrix into a product of two new matrices U and V that minimize the objective function with the added mathematical constraint.
9. The method of claim 1 , further comprising providing a first accuracy score for the first user-item preference model and a second accuracy score for the second user-item preference model, the first and second accuracy scores representing how closely the respective first and second user-item preference models' ratings match the ratings from the rating matrix.
10. The method of claim 1 , wherein the input is received from a domain expert.
11. The method of claim 1 , wherein the input is received from a user not represented in the rating matrix.
12. A system comprising:
an interface;
a processor; and
a memory storing instructions that when executed cause the processor to:
store a rating matrix comprising users, items, and ratings of the items provided by the users, wherein the items include a plurality of components which are factors related to the items that caused the users to provide the ratings of the items;
generate a first user-item preference model by decomposing the rating matrix into a product of two matrices U and V that minimize an objective function, wherein the matrix U includes correlations between the users in the rating matrix and the factors related to the items and the matrix V includes correlations between the items in the rating matrix and the factors related to the items;
group the users and items correlated to dominant factors in the matrices U and V into groups, wherein the dominant factors are factors whose correlations with the users and items in the matrices U and V are greater than a predetermined threshold representing a separation between users who prefer the items and users who do not prefer the items;
receive, by the interface, an input that revises the groups of the users and items;
generate a second user-item preference model of the product of the matrices U and V based on the revised groups of the users and items; and
render for display, on a user device, recommended items based on the second user-item preference model.
13. The system of claim 12 , wherein to generate the second user-item preference model, the instructions are executable to cause the processor to: add a mathematical constraint reflecting the input to the objective function and generate the second user-item preference model by decomposing the rating matrix into a product of two new matrices U and V that minimize the objective function with the added mathematical constraint.
14. The system of claim 12 , wherein the instructions are executable to cause the processor to calculate a first accuracy score for the first user-item preference model and a second accuracy score for the second user-item preference model, the first and second accuracy scores representing how closely the respective first and second user-item preference models' ratings match the ratings from the rating matrix.
15. A non-transitory computer-readable storage medium storing instructions for execution by a processor, the instructions when executed causing the processor to:
access a rating matrix comprising users, items, and ratings of the items provided by the users, wherein the items include components which are factors related to the items that caused the users to provide the ratings of the items;
generate a first user-item preference model based on decomposing the rating matrix into a product of two matrices U and V that minimize an objective function, wherein the matrix U includes correlations between the users in the rating matrix and the factors related to the items and the matrix V includes correlations between the items in the rating matrix and the factors related to the items;
group the users and items correlated to dominant factors in the matrices U and V into groups, wherein the dominant factors are factors whose correlations with the users and items in the matrices U and V are greater than a predetermined threshold representing a separation between users who prefer the items and users who do not prefer the items;
receive an input that revises the groups of the users and items;
add a constraint to the objective function reflecting the input;
generate a second user-item preference model based on decomposing the rating matrix into a product of two new matrices U and V that minimize the objective function with the added constraint; and
render for display, on a user interface, recommended items based on the second user-item preference model.
16. The non-transitory computer-readable storage medium of claim 15 , wherein the input comprises moving a user or item from one group to another group, or deleting a user or item from one of the groups.
17. The non-transitory computer-readable storage medium of claim 15 , wherein the input comprises merging two of the groups or splitting one of the groups.
18. The system of claim 12 , wherein the input comprises moving a user or item from one group to another group, or deleting a user or item from one of the groups.
19. The system of claim 12 , wherein the input comprises merging two of the groups or splitting one of the groups.