Show HN: Mljar Automated Machine Learning for Tabular Data (Explanation,AutoDoc) | Hacker News Reader

Show HN: Mljar Automated Machine Learning for Tabular Data (Explanation,AutoDoc)

github.com

33 pointspplonski868 comments

The creator here. I'm working on AutoML since 2016. I think that latest release (0.7.15) of MLJAR AutoML is amazing. It has ton of fantastic features that I always want to have in AutoML:

- Operates in three modes: Explain, Perform, Compete.

- `Explain` is for data exploratory and checking the default performance (without HP tuning). It has Automatic Exploratory Data Analysis.

- `Perform` is for building production-ready models (HP tuning + ensembling).

- `Compete` is for solving ML competitions in limited time amount (HP tuning + ensembling + stacking).

- All ML experiments have automatic documentation which creates Markdown reports ready to commit to the repo (https://github.com/mljar/mljar-examples/tree/master/Income_c...).

- The package produces extensive explanations: decision tree visualization, feature importance, SHAP explanations, advanced metrics values.

- It has advanced feature engineering, like: Golden Features, Features Selection, Time and Text Transformations, Categoricals handling with target, label, or one-hot encodings.

This is really cool! I'm curious to see how this performs against other AutoML systems like autogluon, autokeras, and AutoML Tables. Do you have any model quality benchmarks on standard research tabular datasets (UCI, OpenML, PMLB) or Kaggle competitions?

I'm also curious how does it compare! The package will be included in the newest comparison done by OpenML people https://github.com/openml/automlbenchmark

I have some old comparison of closed-source old system https://mljar.com/blog/automl-comparison/ but the open source version is much better and have more features.

I will certainly be taking a good look at this! My pandemic project was building a system that processes raw text documents into different sets of features (BOW, sentiment, etc) and I ended up developing a generic library for easily processing a database table into the appropriate sort of numpy array. So it has started to veer into AI flavored automl... hopefully I could contribute something to your project, but it seems like you've got quite a lot there already.

This really cool. The examples in github doesn't seem to show any defining of feature types (categorical, binary, continuous, etc). Is there automatic detection of feature type?

Thank you! Yes, there is an automatic feature detection. Based on selected algorithm and feature type the preprocessing method is selected.

Automatic documentation seems to be an advatage. But what about explainabity?

There is explanation of data done with Automatic Exploratory Data Analysis.

There is explanation of ML models and predictions:

- learning curves from training

- decision tree visualization

- linear model coeffs printout

- features importance (permutation based)

- SHAP features importance

- SHAP dependence plots

- SHAP decision plots

Reply on news.ycombinator.com