Interactive Matrix Factorization
Sameer Singh, Carlos Guestrin

Citation
Sameer Singh, Carlos Guestrin. "Interactive Matrix Factorization". Talk or presentation, 28, October, 2014.

Abstract
Matrix factorization is one of the most commonly applied machine learning formulations that encapsulates a diverse set of applications such as recommendations systems, natural language processing, information extraction, and bioinformatics. For practical deployment of matrix factorization (or in fact, any machine learning system), model designer has to be able to predict the performance on unseen data, understand the types of errors that can be made, and predictably fix such issues. Existing approaches focus on informative selection of instances to provide to human annotators that are used to improve the system (active learning). Unfortunately, this is not a viable solution for matrix factorization since (1) a large number of labels are required to be effective, and (2) expects the designer to know, for e.g., whether a given user would like a specific movie. There is a need for interactivity in machine learning that allows model designers to explore the model, and concisely encode domain information in an intuitive manner in order to improve performance. In this project, we propose interactive matrix factorization. Instead of restricting supervision to labels on single cells, we use logic-based constraints on the whole matrix. These constraints are often intuitive to specify (as rules), capture supervision concisely (quantifiers over sets of rows/columns), and powerful enough to capture complex domain knowledge. For example, even though the model designer may not know the rating for each user and movie, it is quite straightforward to find groups of movies that should have similar ratings over all the users. We will design novel machine learning algorithms that are capable of extracting such rules from a given model (for exploration), and incorporating user-provided constraints into the model (in a probabilistic manner, using the framework of posterior regularization). Due to the large data sets, we will explore the use of distributed stochastic gradient descent. Enabling powerful, constraint-based supervision using distributed machine learning applications will facilitate deployment of accurate, interactive machine learning systems on massive-sized NLP and recommendation system applications.

Electronic downloads

Internal. This publication has been marked by the author for TerraSwarm-only distribution, so electronic downloads are not available without logging in.

Citation formats

• HTML

  Sameer Singh, Carlos Guestrin. <a
  href="http://www.terraswarm.org/pubs/457.html"
  ><i>Interactive Matrix
  Factorization</i></a>, Talk or presentation, 
  28, October, 2014.

• Plain text

  Sameer Singh, Carlos Guestrin. "Interactive Matrix
  Factorization". Talk or presentation,  28, October,
  2014.

• BibTeX

  @presentation{SinghGuestrin14_InteractiveMatrixFactorization,
      author = {Sameer Singh and Carlos Guestrin},
      title = {Interactive Matrix Factorization},
      day = {28},
      month = {October},
      year = {2014},
      abstract = {Matrix factorization is one of the most commonly
                applied machine learning formulations that
                encapsulates a diverse set of applications such as
                recommendations systems, natural language
                processing, information extraction, and
                bioinformatics. For practical deployment of matrix
                factorization (or in fact, any machine learning
                system), model designer has to be able to predict
                the performance on unseen data, understand the
                types of errors that can be made, and predictably
                fix such issues. Existing approaches focus on
                informative selection of instances to provide to
                human annotators that are used to improve the
                system (active learning). Unfortunately, this is
                not a viable solution for matrix factorization
                since (1) a large number of labels are required to
                be effective, and (2) expects the designer to
                know, for e.g., whether a given user would like a
                specific movie. There is a need for interactivity
                in machine learning that allows model designers to
                explore the model, and concisely encode domain
                information in an intuitive manner in order to
                improve performance. In this project, we propose
                interactive matrix factorization. Instead of
                restricting supervision to labels on single cells,
                we use logic-based constraints on the whole
                matrix. These constraints are often intuitive to
                specify (as rules), capture supervision concisely
                (quantifiers over sets of rows/columns), and
                powerful enough to capture complex domain
                knowledge. For example, even though the model
                designer may not know the rating for each user and
                movie, it is quite straightforward to find groups
                of movies that should have similar ratings over
                all the users. We will design novel machine
                learning algorithms that are capable of extracting
                such rules from a given model (for exploration),
                and incorporating user-provided constraints into
                the model (in a probabilistic manner, using the
                framework of posterior regularization). Due to the
                large data sets, we will explore the use of
                distributed stochastic gradient descent. Enabling
                powerful, constraint-based supervision using
                distributed machine learning applications will
                facilitate deployment of accurate, interactive
                machine learning systems on massive-sized NLP and
                recommendation system applications. },
      URL = {http://terraswarm.org/pubs/457.html}
  }
  

Posted by Sameer Singh on 10 Nov 2014.
Groups: services