Feature vector clustering
One goal of computer services (e.g., email, web pages, blogs, advertisements, etc.) is to provide a user with Kinds (digital representations of everyday things) that may be relevant and interesting to the user. Users and Kinds may be plotted within a multidimensional matrix as feature vectors based upon their respective characteristics. An unsupervised clustering technique may be executed upon the matrix to create a mathematical cluster of feature vectors having similar characteristics. For example, a clothing cluster may comprise a dress Kind, a shoe Kind, a wool Kind, a watch Kind, etc. because the unsupervised clustering technique may determine these Kinds are plotted within the matrix in such a way that they have similar characteristics relating to clothing. The unsupervised clustering technique may also be utilized in determining which Kinds may be relevant to a user given a particular context with which a user is engaged with a computer resource.
1. A method for determining a user has an interest in a Kind comprising:
identifying a dimension comprised within a user feature vector of a user, the dimension comprised within the user feature vector having a user interest probability value above an interest threshold value, the user interest probability value specifying a probability that the user has an interest in a characteristic represented by the dimension;
identifying a Kind feature vector of a Kind based upon the Kind feature vector comprising the dimension, the dimension comprised within the Kind feature vector having a Kind relevancy probability value above a relevancy threshold value, the Kind relevancy probability value specifying a probability that the Kind relates to the characteristic represented by the dimension, the Kind representing a non-digital entity; and
determining that the user has an interest in the Kind based upon the Kind relevancy probability value being above the relevancy threshold value and the user interest probability value being above the interest threshold value.
2. The method of claim 1 , the identifying a Kind feature vector comprising:
plotting the user feature vector and the one or more Kind feature vectors within a matrix;
creating a mathematical cluster comprising at least one Kind feature vector within the matrix using an unsupervised clustering technique; and
identifying the Kind feature vector within the mathematical cluster.
3. The method of claim 1 , comprising:
determining a user context based upon a context with which the user is engaged with a computer resource; and
identifying the dimension comprised within the user feature vector based upon the user context.
4. The method of claim 2 , the creating a mathematical cluster comprising:
computing a metric distance between the user feature vector and the at least one Kind feature vector within the matrix using the unsupervised clustering technique.
5. The method of claim 2 , the unsupervised clustering technique comprising a cosine similarity operation comprising a dot product or an inner product.
6. The method of claim 2 , the unsupervised clustering technique comprising a Laplacian.
7. The method of claim 2 , the unsupervised clustering technique comprising at least one of a Euclidian distance, a Kernel method, an SVM framework, or a principle component analysis.
8. The method of claim 3 , the determining a user context comprising:
determining the user context based upon at least one of user email data, user calendar data, prior user internet search data, prior user interactions with an application, or current user interactions with the application.
9. The method of claim 1 , the Kind representing a tangible entity.
10. The method of claim 1 , comprising:
presenting content associated with the Kind to the user.
11. A system for determining a user has an interest in a Kind comprising:
a relevancy component configured to:
identify a dimension comprised within a user feature vector of a user, the dimension comprised within the user feature vector having a user interest probability value above an interest threshold value, the user interest probability value specifying a probability that the user has an interest in a characteristic represented by the dimension;
identify a Kind feature vector of a Kind based upon the Kind feature vector comprising the dimension, the dimension comprised within the Kind feature vector having a Kind relevancy probability value above a relevancy threshold value, the Kind relevancy probability value specifying a probability that the Kind relates to the characteristic represented by the dimension, the Kind representing a non-digital entity; and
determine that the user has an interest in the Kind based upon the Kind relevancy probability value being above the relevancy threshold value and the user interest probability value being above the interest threshold value, at least some of the relevancy component implemented at least in part via a processing unit.
12. The system of claim 11 , the relevancy component configured to:
plot the user feature vector and one or more Kind feature vectors within a matrix;
create a mathematical cluster comprising at least one Kind feature vector within the matrix using an unsupervised clustering technique; and
identify the Kind feature vector within the mathematical cluster.
13. The system of claim 11 , the relevancy component configured to:
determine a user context based upon a context with which the user is engaged with a computer resource; and
identify the dimension comprised within the user feature vector based upon the user context.
14. The system of claim 12 , the relevancy component configured to:
compute a metric distance between the user feature vector and the at least one Kind feature vector within the matrix using the unsupervised clustering technique.
15. The system of claim 12 , the unsupervised clustering technique comprising a cosine similarity operation comprising a dot product or an inner product.
16. The system of claim 12 , the unsupervised clustering technique comprising a Laplacian.
17. The system of claim 12 , the unsupervised clustering technique comprising at least one of a Euclidian distance, a Kernel method, an SVM framework, or a principle component analysis.
18. The system of claim 13 , the relevancy component configured to:
determine the user context based upon at least one of user email data, user calendar data, prior user internet search data, prior user interactions with an application, or current user interactions with the application.
19. The system of claim 11 , the relevancy component configured to:
present content associated with the Kind to the user.
20. A computer-readable memory comprising processor-executable instructions that when executed perform a method for determining a user has an interest in a Kind comprising:
identifying a dimension comprised within a user feature vector of a user, the dimension comprised within the user feature vector having a user interest probability value above an interest threshold value, the user interest probability value specifying a probability that the user has an interest in a characteristic represented by the dimension; and
identifying a Kind feature vector of a Kind based upon the Kind feature vector comprising the dimension, the dimension comprised within the Kind feature vector having a Kind relevancy probability value above a relevancy threshold value, the Kind relevancy probability value specifying a probability that the Kind relates to the characteristic represented by the dimension, the Kind representing a non-digital entity.