A Bayesian Approach To Model Overlapping Objects Available As Distance Data
summary
Introduces POCD, a Bayesian model that clusters objects directly from pairwise distance matrices allowing overlapping cluster membership via an IBP prior. Likelihood is an augmented Wishart that is rotation- and translation-invariant, so embeddings are unnecessary. In experiments on protein contact maps for 10 HIV-1 protease inhibitor drugs POCD returns three blocks, isolating one drug with low cross-resistance. Inference uses Metropolis-within-Gibbs MCMC, runtime about O(n^3) with efforts toward parallel and amortized acceleration.
AI-written summary · about
chapters (8)
- 0:00Intro
- 1:46What is clustering
- 3:52The probabilistic model
- 6:47Bayesian substring model
- 7:40Prior over Z
- 9:31Monte Carlo
- 13:49Conclusion
- 14:51Data
related talks
48mLocally Private Bayesian Inference for Count Models
01h 06mNonparametric Bayesian Methods: Models, Algorithms, and Applications I
43mPufferfish Privacy Mechanisms for Correlated Data
24mStochastic Block Models with R: Statistically rigorous clustering
01h 04mNonparametric Bayesian Methods: Models, Algorithms, and Applications II
37mPower to the Points: Local Certificates for Clusterings
25mLucky Talk: From Coding to Clustering
34mTopological Data Analysis