MSc. Esteban Vilca

With the release of TABPFN v2.1, you can now fine-tune TabPFN on your own data - unlocking stronger performance for domain-specific tabular tasks. 🧠 TabPFN is a powerful foundation model for tabular data that predicts full posteriors using transformer architectures trained on synthetic tasks. And now, with just a few steps, you can adapt it to your real-world dataset. 📚 A community member, Iliya Valchev, put together a fantastic walkthrough covering: ✅ Running TabPFN in batched mode ✅ Preprocessing & inference-time transformation handling ✅ Fine-tuning the transformer backbone on your dataset 👉 Read the full tutorial here: https://lnkd.in/dCuwNVh2 🔧 Whether you're working with classification or regression problems, there are ready-to-use examples for both: • Fine-tune classifier → https://lnkd.in/dT86vmnq • Fine-tune regressor → https://lnkd.in/d4a5tjtt If you're dealing with niche domains or custom tabular tasks, this update is worth checking out!

1

If you lead a data team and are buried under requests, tool sprawl, and tech debt — Sidecar Data can 10x your data team’s impact Aaron R. and James Webb are building the operating system for data teams, automating the coordination, documentation, and observability work that keeps teams effective and platforms healthy.

5

Double ML is a great contributor to the momentum we’re seeing around causal inference. It’s a flexible modelling technique that lets us estimate causal effects while using machine learning to control for confounding. It also addresses regularization bias and overfitting through orthogonalization and sample splitting. That’s a big improvement over some traditional approaches when working with high-dimensional data. But there’s a risk that comes with the attention for double ML. Because double ML is often packaged as “causal” machine learning, more and more people seem to think the method itself is doing the causal work. But that’s not the case. Double ML helps with the quantification of a causal effect estimate, but the most difficult part comes before that. You are the one who needs to construct a well-defined causal question, taking into account things like treatment variation irrelevance, interference, and so on. You are the one who needs to construct an accurate Causal Graph, considering things like granularity, clustering, aggregation and any cycles that might arise by doing so. Then you are the one that needs to figure out whether a sufficient adjustment set exists to identify your causal quantity of interest. If so, then you’ll probably realize you don’t have all of those variables in your data. You are the one that then needs to figure out how to proxy them as well as possible, in order to control for confounding as much as possible. You are the one that needs to think about whether any selection bias exists in the sample you're using, and if so, how you might use something like inverse probability of selection weighting to try and 'remove' this bias. You are the one that needs to manually perform positivity diagnostics, and perhaps take appropriate actions based on the results. Only then can you use your adjustment set in the double ML procedure to get an effect estimate. Then comes sensitivity analysis. And again, you are the one (often together with domain experts) who has to reason whether unobserved confounding could plausibly overturn your conclusions. All of this is driven by external knowledge. No observational dataset, no algorithm, not even Causal ML, will do that part for you. Causal ML deserves the “causal” label because it is built to respect some of the structure of causal inference, for example by avoiding regularization bias which could systematically draw our estimate towards 0. But that doesn’t mean it automatically makes a model causal. Please don't think that throwing a high-dimensional dataset into double ML will give you a causal answer by default. If you want to get good at causal inference, you need to start from the beginning: identification, the ladder of causation, potential outcomes, and so on. Only when you understand what causal inference from observational data actually requires, can you use causal ML effectively. #CausalInference

46

4 Comments