Tamara Kolda, Distinguished Member of Technical Staff, Sandia National Laboratories @ MLconf SF Part
summary
Describes using CP (CANDECOMP/PARAFAC) decompositions with a Poisson tensor factorization objective (KL divergence) for count data, enforcing nonnegative factors and optional weights; optimization uses alternating updates, multiplicative rules or Newton/quasi-Newton and handles sparse tensors. Demonstrates on an Enron email example: 8,500 messages, 100 users over 28 months, components reveal a legal-department cluster and a pre-SEC-investigation activity spike and role/gender patterns. Recommends Tensor Toolbox or Factorization Machine, subset-first analysis, and notes scalability is doable.
AI-written summary · about
chapters (7)
- 0:00Poisson Distribution better than Gaussian for Count Data
- 0:17New Approach: Poisson Tensor Factorization (PTF)
- 1:38Interpreting PTF: Balls & Urns
- 2:23Solving the Poisson Regression DE Problem
- 2:57PTF for Time-Evolving Social Network
- 3:22Enron Email Data Component
- 4:48More on Computing Tensor Factorizations
related talks
10mDale Smith, Data Scientist, Nexidia
01h 11mTensor Decompositions: Uniqueness and Smoothed Analysis
58mTensor Decomposition, Sparse Representations and Applications
52mAverage-Case Overcomplete Tensor Decomposition
56mTensor Decompositions for Estimating Latent Variable Models
01h 03mRecent Progress in High-Dimensional Learning
48mOrthogonal Tensor Decomposition
01h 00mSmoothed Analysis of Tensor Decompositions