Doubly Robust Bayesian Inference for Non-Stationary Streaming Data with $\beta$-Divergences
Authors: Jeremias Knoblauch, Jack E. Jewson, Theodoros Damoulas
NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Reducing False Discovery Rates of CPS from over 90% to 0% on real world data, this offers the state of the art. ... Lastly, Section 5 showcases the substantial gains in performance of robust BOCPD when compared to its standard version on real world data in terms of both predictive error and CP detection. ... Figure 4 shows that robust BOCPD deals with outliers on-line. |
| Researcher Affiliation | Academia | Jeremias Knoblauch The Alan Turing Institute Department of Statistics University of Warwick Coventry, CV4 7AL j.knoblauch@warwick.ac.uk Jack Jewson Department of Statistics University of Warwick Coventry, CV4 7AL j.e.jewson@warwick.ac.uk Theodoros Damoulas The Alan Turing Institute Department of Computer Science & Department of Statistics University of Warwick Coventry, CV4 7AL t.damoulas@warwick.ac.uk |
| Pseudocode | Yes | Stochastic Variance Reduced Gradient (SVRG) inference for BOCPD Input at time 0: Window & batch sizes W, B , b ; frequency m, prior θ0, #steps K, step size η s.t. W > B > b ; and denotes sampling without replacement for next observation yt at time t do for retained run-lengths r R(t) do if τr = 0 then if r < W then θr θ r Full Opt (ELBO(yt r:t)); τr m else if r W then θ r θr; τr Geom (B /(B + b )) B min(B , r) ganchor r 1 i I ELBO(θ r, yt i), where I Unif{0, . . . , min(r, W)}, |I| = B for j = 1, 2, . . . , K do b min(b , r) and e I Unif{0, . . . , min(r, W)} and |e I| = b gold r 1 i e I ELBO(θ r, yt i), gnew r 1 i e I ELBO(θr, yt i) θr θr + η gnew r gold r + ganchor r ; τr τr 1 r r + 1 for all r R(t); R(t) R(t) {0} |
| Open Source Code | Yes | Software and simulation code is available as part of a reproducibility award at https://github.com/alan-turing-institute/rbocpdms/. |
| Open Datasets | Yes | The well-log data set was first studied in Ruanaidh et al. [39] and has become a benchmark data set for univariate CP detection. ... we also analyze Nitrogen Oxide (NOX) levels across 29 stations in London using spatially structured Bayesian Vector Autoregressions [see 25]. |
| Dataset Splits | No | The paper mentions using specific datasets for experiments but does not provide details on training, validation, or test splits (e.g., percentages, sample counts, or cross-validation setup). |
| Hardware Specification | Yes | The robust version has more computational overhead than standard BOCPD, but still needs less than 0.5 seconds per observation using a 3.1 GHZ Intel i7 and 16GB RAM. |
| Software Dependencies | No | The paper mentions 'Python scipy’s L-BFSG-B optimization routine' but does not specify a version number for scipy or any other key software dependencies. |
| Experiment Setup | Yes | for which we initialize βp = 0.05 and βp = 0.005 for d = 1 and d = 29, respectively. ... In our experiments L is a bounded absolute loss. |