David Arbour

📄 Our paper on comprehensive, low-variance evaluation of slate recommendation strategies using offline logged data was accepted at AAAI-24! Work done during my internship at @AdobeResearch with @darbour26, Georgios Theocharous, and Nikos Vlassis. (arxiv.org/abs/2308.14165)

A couple weeks ago, my, @darbour26, Tung Mai and Anup Rao’s paper on online balancing experimental design appeared in ICML! We show how to reduce variance when treatments are assigned sequentially, e.g. survey experiments or online A/B tests. icml.cc/virtual/2022/s…

Let me be your boss! We're hiring a company expert on moderation on our brand new DS team. Reddit is growing and making big investments in our mod tools. If you're interested in helping us build the next big thing for Reddit, please dm me. boards.greenhouse.io/reddit/jobs/35…

Our new working paper: tell the computer what you know (and don't know!) about a causal question w/ discrete data → automatically get most precise possible answer (bounds, or a point estimate). Joint w/ @guilhermejd1 @nsfinkelstein @dean_c_knox Shpitser.🧵arxiv.org/abs/2109.13471

Another interest talk about doubly-robust confidence sequences: drive.google.com/file/d/1bAh0lX…

📢 Call for abstracts 📢 for the Conference on Digital Experimentation (CODE@MIT) Submit your best work *by Sept 15* on online field experiments, statistical methods, practical & computational challenges, etc. ide.mit.edu/events/2021-co… #mitcode

Tonight at #ICML2021, I will be presenting my work with @purva_pruthi and David Jensen on "How and Why to Use Experimental Data to Evaluate Methods for Causal Inference". Our poster will be in spot A4 in Gather Town Room 7, tonight from 12am - 2am EDT.

🚨New!🚨 You're estimating a population mean from samples observed with varied probabilities. Do you use a Horvitz–Thompson/IPW or a Hájek/self-normalizing estimator? @Stats_samir and I examine an old question due to Trotter & Tukey (1954): why not both?arxiv.org/abs/2106.07695 🧵

@elenadata @NSF Congratulations Elena!

Honored to receive an #NSFCAREER award & psyched to spend the next five years (and beyond) studying relational causal inference and developing algorithms for answering causal questions from network data. Many thanks to @NSF and the anonymous reviewers for the vote of confidence!

Clever work from @iamwillcai & @jugander decomposing peer effects: When we copy others, how much is just based on observing others' choice vs. the rewards they receive from their choices? ojs.aaai.org/index.php/ICWS…

🚨 COOL JOB ALERT 🚨: Use network sci, exploratory analysis, causal inference, surrogacy measurement @Twitter. Help improve measurement, follow-recommendations, & other problems to make twitter more friendly & better for newcomers!

Excited to announce that our tutorial with @darbour26 on "Causal inference from network data" has been accepted as a lecture-style tutorial at #KDD2021. Stay tuned for more details: netcause.github.io. @kdd_news

ALT Highlights - An Equivalence between Private Learning and Online Learning (ALT '21 Tutorial) ift.tt/2QXb6Kj

@iwaudbysmith @mccrinbc These are definitely good references. Should note GPML site has references to packages, but I'd also suggest gpytorch (github.com/cornellius-gp/…) for a nice modern GP library.

@mccrinbc Courtesy of @darbour26: gpss.cc has tutorials, e.g. github.com/gpschool/gpss20 Also, apparently gaussianprocess.org/gpml/chapters/… has code and data.

Calling #BayesTwitter! I'm reading GPs for ML and I'd love to supplement this with some solid Python examples from an open-source Jupyter notebook. Any suggestions? :) @latentjasper @kswersk @ZoubinGhahrama1 @RogerGrosse @andrewgwils

Conformal inference gives rigorous outlier/out-of-distribution detection. We show how to control FDR with conformal p-values -- even though they are dependent, they satisfy the PRDS property! arxiv.org/abs/2104.08279 With E. Candès, @lihua_lei_stat, Y. Romano, and M. Sesia

Check this out 👇 our aim here is to take a closer look at the assumptions in #causalinferene! one of the most exciting aspects of this workshop is bringing in experts from epi, (bio)stats, & philosophy in a major CS conference #ML #icml2021

These are easy to use: take your existing non/semi-parametric estimators & CIs, and replace the root-(1/n) width with these root-(log(n)/n) widths to get a time-uniform coverage guarantee for an infinite time horizon. More extensions, tweaks, etc. coming soon!

Unlike classical confidence intervals, confidence sequences give you the flexibility to make inferences at arbitrary (data-dependent) stopping times (e.g. in a sequential experiment or an observational study where data are collected in an online fashion).

Super excited to share this work with @edwardhkennedy, @darbour26, Aaditya Ramdas, and Ritwik Sinha on confidence sequences for causal effects! arxiv: arxiv.org/pdf/2103.06476… code: github.com/wannabesmith/s…

We're continuing to work hard on methods for experimental design, so if you have any tricky problems in this arena, feel free to reach out! I'd love to help if I can. And more is coming soon :)