Projection mining for advanced recommendation systems and data mining

Abstract

A method for projection mining comprises performing a first projection on a first data object of a first type comprising a plurality of data entries and a second data object of a second type comprising a plurality of data entries to create definitions of attributes of the first data object and definitions of attributes of the second data object, performing a second projection of the definitions of the attributes of the first data object and the definitions of the attributes of the second data object into a space of meta-attributes based on semantic relationships among the attributes of the first data object and the second data object, learning relationships between the space of meta-attributes formed by the projections of the first data object and the second data object and a space of meta-attributes relating to new data not included in the first data object and the second data object, and generating at least one new data object of the first or second type based on the new data using the learned relationships.

1 . A method for projection mining comprising:
 performing a first projection on a first data object of a first type comprising a plurality of data entries and a second data object of a second type comprising a plurality of data entries to create definitions of attributes of the first data object and definitions of attributes of the second data object;   performing a second projection of the definitions of the attributes of the first data object and the definitions of the attributes of the second data object into a space of meta-attributes based on semantic relationships among the attributes of the first data object and the second data object;   learning relationships between the space of meta-attributes formed by the projections of the first data object and the second data object and a space of meta-attributes relating to new data not included in the first data object and the second data object; and   generating at least one new data object of the first or second type based on the new data using the learned relationships.   
   
   
       2 . The method of  claim 1 , wherein the relationships are learned by using linear algebra, at least one matrix inversion, a linear algorithm, or generating a data mining model. 
   
   
       3 . The method of  claim 1 , wherein the at least one new data object is generated by using an inverse of the meta-attributes with the new data to map back to objects of the first and second types but containing the new data. 
   
   
       4 . The method of  claim 1 , wherein the first projection performed on the first data object is a different projection than the first projection that is performed on the second data object. 
   
   
       5 . The method of  claim 1 , wherein the first projection is an identity projection. 
   
   
       6 . The method of  claim 1 , wherein the first projection is performed separately on a first data object and a second data object relating to a first data set and on a first data object and a second data object relating to a second data set. 
   
   
       7 . The method of  claim 6 , wherein the first projection creates, for the first data set, definitions of attributes of the first data object comprising a first matrix T including correspondences of first objects and attributes of the first objects, and definitions of attributes of the second data object comprising a second matrix A including correspondences of second objects and attributes of the second objects, and for the second data set, definitions of attributes of the first data object comprising a first matrix T′ including correspondences of first objects and attributes of the first objects, and definitions of attributes of the second data object comprising a second matrix A′ including correspondences of second objects and attributes of the second objects 
   
   
       8 . The method of  claim 7 , wherein the first projection comprises:
 filtering at least some of the fields of the first data object and the second data object to include or exclude certain data or types of data based on filtering criteria;   expanding categorical fields of low dimensionality;   applying text mining to unstructured or high cardinality fields to produce structured document-term matrices; and   integrating the results of the prior steps to form the first matrix T or T′ and the second matrix A or A′.   
   
   
       9 . The method of  claim 7 , wherein the second projection is performed using Principal Components Analysis, Independent Components Analysis, Matrix Decompositions, Vector Quantization, Non-Negative Matrix Factorization, or k-means clustering, self-organizing maps clustering, or other clustering methods that provide a soft-clustering or probabilistic output. 
   
   
       10 . The method of  claim 7 , wherein the second projection is performed using Non-Negative Matrix Factorization comprising:
 factoring the first matrix T and the second matrix A to each form two matrices of lower rank; and   projecting the first matrix T′ and the second matrix A′ into the space of meta-attributes.   
   
   
       11 . The method of  claim 10 , wherein the factoring comprises:
 factoring the first matrix T according to T˜G×M, wherein matrix G includes correspondences of first objects and meta-attributes of the first objects, and matrix M includes correspondences of first objects and attributes of the first objects, and   factoring the second matrix A according to A˜W×H, wherein matrix W includes correspondences of second objects and meta-attributes of the second objects, and matrix H includes correspondences of second objects and attributes of the second objects.   
   
   
       12 . The method of  claim 11 , wherein the projection comprises:
 projecting the first matrix T′ according to G′ T ˜T′×M −1 , wherein matrix G′ T  includes correspondences of first objects and meta-attributes of the first objects, and matrix M −1  is a matrix pseudo-inverse of the matrix M, and   projecting the second matrix A′ according to W′ A ˜A′×H −1 , wherein matrix W′ A  includes correspondences of second objects and meta-attributes of the second objects, and matrix H −1  is a matrix pseudo-inverse of the matrix H.   
   
   
       13 . The method of  claim 12 , wherein learning relationships comprises:
 creating a matrix S comprising correspondences between first objects and second objects for the first dataset; and   creating a matrix Z, according to Z=G T ×S×W, comprising correspondences between meta-attributes of the first objects and meta-attributes of the second objects for the first dataset.   
   
   
       14 . The method of  claim 12 , wherein generating at least one new data object comprises:
 creating a matrix S′, according to S′=(G T   T ) −1 ×Z×(W A ) −1 , comprising correspondences between the first objects and the second objects for the second dataset.   
   
   
       15 . The method of  claim 1 , wherein the relationships are learned by generating a data mining model and generating at least one new data object comprises generating recommendations using the data mining model. 
   
   
       16 . The method of  claim 15 , wherein the recommendations are generated by:
 generating a set of scoring vectors using the data mining model;   ranking the generated set of scoring vectors; and   selecting at least a portion of the generated set of scoring vectors as the recommendations.   
   
   
       17 . The method of  claim 16 , wherein the recommendations are further generated by:
 ranking the generated set of scoring vectors by comparing dot products of the vectors or using another comparison function; and   ordering the scoring vectors by sorting, filtering, or selecting vectors by class.   
   
   
       18 . A method for automatically generating a conference schedule for an attendee of a conference comprising:
 performing a first projection on data relating to sessions of at least one conference and comprising a plurality of session data entries and on data relating to attendees at the at least one conference and comprising a plurality of data entries to create definitions of attributes of the sessions and definitions of attributes of the attendees;   performing a second projection of the definitions of attributes of the sessions and definitions of attributes of the attendees into a space of meta-attributes based on semantic relationships among the attributes of the sessions and the attendees;   learning relationships between the space of meta-attributes formed by the projections of the sessions and the attendees and a space of meta-attributes relating to new data relating to at least one new conference and including new data relating to a plurality of new sessions not included in the data relating to sessions and a plurality of new attendees not included in the data relating to attendees;   generating a ranking of matches between new sessions and new attendees using the learned relationships; and   generating a conference schedule of an attendee of the new conference using the ranking of matches between new sessions and new attendees.   
   
   
       19 . The method of  claim 18 , wherein the conference schedule is generated by:
 assigning sessions in the conference schedule of the attendee based on each highest ranked unassigned session for the attendee until the conference schedule of the attendee is full or partially full; and   skipping or assigning as backup sessions a lower ranked unassigned session occurring at the same time as an assigned session.   
   
   
       20 . The method of  claim 18 , further comprising:
 assigning session based on spatial proximity of sessions so as to satisfy distance or time constraints between sessions.   
   
   
       21 . A system for projection mining comprising:
 a processor operable to execute computer program instructions;   a memory operable to store computer program instructions executable by the processor; and   computer program instructions stored in the memory and executable to perform the steps of:   performing a first projection on a first data object of a first type comprising a plurality of data entries and a second data object of a second type comprising a plurality of data entries to create definitions of attributes of the first data object and definitions of attributes of the second data object;   performing a second projection of the definitions of the attributes of the first data object and the definitions of the attributes of the second data object into a space of meta-attributes based on semantic relationships among the attributes of the first data object and the second data object;   learning relationships between the space of meta-attributes formed by the projections of the first data object and the second data object and a space of meta-attributes relating to new data not included in the first data object and the second data object; and   generating at least one new data object of the first or second type based on the new data using the learned relationships.   
   
   
       22 . The system of  claim 21 , wherein the relationships are learned by using linear algebra, at least one matrix inversion, a linear algorithm, or generating a data mining model. 
   
   
       23 . The system of  claim 21 , wherein the at least one new data object is generated by using an inverse of the meta-attributes with the new data to map back to objects of the first and second types but containing the new data. 
   
   
       24 . The system of  claim 21 , wherein the first projection performed on the first data object is a different projection than the first projection that is performed on the second data object. 
   
   
       25 . The system of  claim 21 , wherein the first projection is performed separately on a first data object and a second data object relating to a first data set and on a first data object and a second data object relating to a second data set.