Roadmap
LukeMathWalker opened this issue · comments
In terms of functionality, the mid-term end goal is to achieve an offering of ML algorithms and pre-processing routines comparable to what is currently available in Python's scikit-learn.
These algorithms can either be:
- re-implemented in Rust;
- re-exported from an existing Rust crate, if available on crates.io with a compatible interface.
In no particular order, focusing on the main gaps:
-
Clustering:
- DBSCAN
- Spectral clustering;
- Hierarchical clustering;
- OPTICS.
-
Preprocessing:
- PCA
- ICA
- Normalisation
- CountVectoriser
- TFIDF
- t-SNE
-
Supervised Learning:
- Linear regression;
- Ridge regression;
- LASSO;
- ElasticNet;
- Support vector machines;
- Nearest Neighbours;
- Gaussian processes; (integrating
friedrich- tracking issue nestordemeure/friedrich#1) - Decision trees;
- Random Forest
- Naive Bayes
- Logistic Regression
- Ensemble Learning
- Least Angle Regression
- PLS
The collection is on purpose loose and non-exhaustive, it will evolve over time - if there is an ML algorithm that you find yourself using often on a day to day, please feel free to contribute it 💯
Hi, I'm eager to help I'll take Linear regression, Lasso and ridge.
Cool! I worked a bit on linear regression a while ago - you can find a very vanilla implementation of it here: https://github.com/rust-ndarray/ndarray-examples/tree/master/linear_regression @Nimpruda
What does Normalization mean, is it like sklearn's StandardScaler or something else?
Exactly @InCogNiTo124.
This is an interesting project and I will work on the PCA implementation
I am the author of the friedrich crate which implements Gaussian Processes.
While it is still a work in progress, it is fully featured and I would be happy to help integrate it into the project if you have directions to do so.
That would be awesome @nestordemeure - I'll have a look at the project and I'll get back to you! Should I open an issue on friedrich's repository when I am ready? Or would you prefer it to be tracked here on the linfa repository?
Both are ok with me.
An issue in friedrich's repository might help avoid overcrowning linfa with issues but do as you prefer.
I'd love to take the Nearest Neighbors implementation
I think this is really great, I just started on a sklearn like implementation of their pipelines, here but more or less for experimentation without anything serious. I'll be sure to keep my eye on issues/goals here and help out where I can. Thanks for the initiative! 👏
Hi there! First off, I don't have any experience in ML, but I read a lot about it (and listen to way too many podcasts on the topic). I'm interested in jumping in. I have quite some experience developing in Rust, and specifically high fidelity simulation tools (cf nyx and hifitime).
I wrote an Ant Colony Optimizer in Rust. ACOs are great for traversing graphs which represent a solution space, a problem which is considered NP hard if I'm not mistaken. Is that something used at all in ML? If so, would it be of interest to this library, or is there a greater interest (for now) to focus on the problems listed in the first post?
Cheers
Hi @ChristopherRabotin I've never heard of ACOs but as it's in relation with graphs you should check if it has any uses with Markov Chains.
So far, I haven't found how both can be used together. The closest I found was finding several papers which use Markov Chains to analyze ACOs.
I would like to take the Naive Bayes one.
I'll take on Gaussian Processes.
I'll put some work towards the text tokenization algorithms (CountVectorizer and TFIDF). I'm also extremely interested in a good SVM implementation in Rust. Whoever is working on that, let me know if you'd like some help or anything.
Please take a look at what is already out there before diving head down into a reimplementation @tyfarnan - I haven't had the time to look at friedrich by @nestordemeure yet (taking a break after the final push to release the blog post and related code 😅) but we should definitely start from there as well as the GP sub-module in rusty-machine.
@tyfarnan, don't hesitate to contact me via an issue on friedrich's repository once @LukeMathWalker has explicited what is expected of code that is integrated into Linfa and how this integration will be done.
I did a quick round up of crates that implement the algorithms listed on the roadmap. Probably missed quite a few too but this can be a good starting point.
It was just a quick search so I don't know how reliavent each crate is but I tried to make a note if the crate was old and unmaintained. Hopefully this can be useful for helping with algorithm design or saving us from having to reimplement something that is already there.
Tracking friedrich<>linfa integration here: nestordemeure/friedrich#1
I have updated the Issue to make sure it's immediately clear who is working on what and what items are still looking for an owner 👍
hey @LukeMathWalker could you add me next to the normalization? I plan to do it by New Year's as I'm still not very experienced with Rust, but I have an idea how to implement it
Done @InCogNiTo124 🙏
Started implementing DBScan in #12.
Also if there are suggestions Gaussian Mixture Models would be cool
Implementation of DBSCAN merged to master - thanks @xd009642 🙏
Hi, really cool project!
I have a question concerning the scope: do you eventually want to have deep learning and reinforcement learning algorithms too? I guess I'm curious to know if adding them is the plan eventually, but you want to start with the easier stuff, or if you think along the line of the scikit dev themselves : here.
Either way, I'll be glad to help spread the rust gospel. Right know I'm going through the Reinforcement Learning book, and I will implement some of the algorithms; if that's in the scope of linfa, I'll be glad to try adding them to it. If not, I plan to read through Understanding Machine Learning afterwards, and thus will eventually reach some of the algorithms in the roadmap. Then I will help by implementing them. :)
From previous discussions deep learning etc is out of scope for the same reasons as it is for sci-kit. @LukeMathWalker might have more to say about it or reinforcement learning 😄
Ok, thanks.
I would consider both of them to be out of scope for this project - it's already incredibly broad as it is right now 😅
I'd love to see something spawn up for reinforcement learning, especially gym environments!
Can you also include Non-Negative Matrix Factorization (NMF) in the list for pre-processing steps. Its a standard algorithm in NLP/audio enhancement and decomposes a matrix into the product of two positive valued matrices. (https://en.wikipedia.org/wiki/Non-negative_matrix_factorization)
One of the nice properties is that there is a simple incremental algorithm for solving the the problem, with simple modification for sparsity constraints.
Please take a look at the PR rust-ndarray/ndarray-linalg#184 which adds TruncatedEig and TruncatedSvd routines to the library. Both are based on the LOBPCG algorithm and allow an iterative approach to eigenvalue and singular value decomposition. This is used in PCA, manifold learning (e.g. spectral clustering) and discriminant analysis and is therefore useful here too. The algorithm also supports sparse problems, because the operator is defined in a matrix free way. (the matrix A is provided as a closure in the function call to LOBPCG)
Can you also add me to the spectral clustering task? Will try to implement classical Multidimensional Scaling. The t-SNE technique is interesting too, but requires more time because it is based on a custom optimization problem. Furthermore gaussian embedding is another interesting technique, I used recently, but requires at-least SGD for a single layer NN. See papers here:
Hi there!
I've been working on an implementation of decision trees here. It's still a WIP and needs documentation but it's a start at least. Once it's a bit more polished I can look at random forests also.
Hello, if possible, I would like to assume the implementation of ICA algorithm. My implementation will take this paper as a guide.
@VirtualSpaceman That would be great--any pull requests are very welcome
ordinary linear regression was added in #20, thanks to @Nimpruda and @paulkoerbitz
linear decision trees were added in #18, kudos to @mossbanay
fast Independent Component Analysis was added in #47, kudos to @VasanthakumarV
I'm gonna start working on OPTICS at some point soon 👀
Hello, I am working on an implementation of the Approximated DBSCAN algorithm here and I was wondering if that is something that could be interesting for this project. Right now it has all the basic functionalities implemented and I would happily make any changes to make it fit here
@Sauro98 I believe linfa currently an implementation of the vanilla DBSCAN algorithm here, but an approximate version would be a great addition! It would be great if you could open a pull request following that same style under the linfa-clustering sub-crate (i.e. using the appx_dbscan vs. dbscan), in a way that maintains some closeness between the API for the existing algorithm and your own ( Dbscan::predict(&hyperparams, &dataset); -> AppxDbscan::predict(&hyperparams, &dataset);), which I believe usually accepts data in the form of an ndarray &Array2<f64> structure variant. I'd also think it would be really interesting to see benchmarks comparing performance between the two!
Hello, I've started to work on a port of sklearn Gaussian Mixture Model (mentioned by @xd009642 above). I would be happy to contribute to the linfa-clustering sub-crate. Btw, thanks for the linfa initiative which is really promising.
Gaussian naïve Bayes was added in #51, kudos to @VasanthakumarV
Fast K-medoids clustering (PAM, FasterPAM) implementations: https://crates.io/crates/kmedoids
Markov Chains was mentioned above, but I'd really like to see Hidden Markov Models. I think it'd fit better under the "supervised learning" set of algorithms, even though it has unsupervised applications as well.
I made a toy project with https://github.com/paulkernfeld/hmmm a while back and it seemed solid enough.
There are no (published) API docs, but the code itself is quite small, and very well documented.
I'd like to implement Random Forrest.
Nearest neighbours merged in #120, thanks to @YuhanLiin
hi all i'd like to help implement too. what's the best way to pick up a task?
hi all i'd like to help implement too. what's the best way to pick up a task?
not difficult, just mention your interest here and I will add you to the list once you've submitted the initial draft :)
Is there any interest for linfa supporting model selection algorithms such as grid search or hyperparameter tuning?
@Clara322 I personally think that would be a good candidate for a new linfa crate if you want to open an issue for it specifically then there can be some discussion on the specifics of what the design will look like and the steps to implement it 👍
Could it be added "causal inference" like https://github.com/microsoft/dowhy library or would be out of scope for linfa?
Hi everyone, I've implemented the semi-supervised learning algorithm called dynamic label propagation using Rust. I'm getting accuracy score up to 98% for one of the datasets I've been using. I don't think this algorithm is very well known, but could it be added to the Linfa library?
Could it be added "causal inference" like https://github.com/microsoft/dowhy library or would be out of scope for linfa?
there are many interesting pattern which linfa can learn from but we would need first to support graphical models
Hi everyone, I've implemented the semi-supervised learning algorithm called dynamic label propagation using Rust. I'm getting accuracy score up to 98% for one of the datasets I've been using. I don't think this algorithm is very well known, but could it be added to the Linfa library?
cool, sure! Once you have a working prototype, submit a PR and I will review the integration. We have to see how to add support for incomplete datasets though
@bytesnake I'm playing with it, creating graph and identification support. If one day i feel it can be ready i'll submit a PR. https://github.com/vaijira/linfa/tree/causal/algorithms/linfa-causal-inference
Infrastructure Goals
Aside from just adding new algorithms, there are also some infrastructure tasks that will significantly improve the ergonomics and performance of Linfa. They are listed here in descending order of importance, in my opinion:
- #220: Python bindings allows
linfato target the same userbase asscikit-learn, broadening the reach of the project. Having Python benchmarks also allows a fair performance comparison betweenlinfaandscikit-learn. - #228: This task potentially allows removing BLAS dependencies from
linfacompletely, significantly increasing code quality. It also has the side-effect of increasing benchmark coverage. - #103: More benchmarks allow us to make performance optimizations with confidence.
- #161: Allows users to have more visibility into the internals of longer-running algorithms, similar to other mainstream ML libraries.
Can I work on adding Linear Discriminant Analysis to linfa? Here is a link to the Sklearn analog
I've been working on some features like: Categorical Encoding, MAPE and random forest. How can I contribute?
Can I work on adding Linear Discriminant Analysis to linfa? Here is a link to the Sklearn analog
Does LDA output the dimensionally-reduced data at all? If so it should go into linfa-reduction
I've been working on some features like: Categorical Encoding, MAPE and random forest. How can I contribute?
Random forests are covered by this PR.
Categorical encoding would go into linfa-preprocessing. I'm pretty sure we don't have it but just check to make sure.
MAPE is a simple function that would go into linfa/src/metrics_regression.rs
Can I work on adding Linear Discriminant Analysis to linfa? Here is a link to the Sklearn analog
Does LDA output the dimensionally-reduced data at all? If so it should go into
linfa-reduction
It can perform dimensionality-reduction (transform). It can also just be used to predict classes (predict). The parentheses hold the method analog in Sklearn. Is there a preference for which should be implemented? Also, I am still getting familiar with Rust so it may take a few weeks to get done.
Preferably implement both if possible.
Are there plans to implement ridge regression in the linear sub-package? Looking for models to contribute.
Ridge regression should already be in linfa-elasticnet
What about imputation, similar to scikit imput?
We don't have that. That can go in linfa-preprocessing
Hi!
Can I take up Random Forests?
Also can we look to implement xgboost and adaboost?
#229 implements bootstrap aggregation, which is a generalization of random forests, so you could work on that.
xgboost and adaboost seem to both be ensemble algorithms that are not necessarily tied to decision trees (correct me if I'm wrong), so we should probably put them in a new algorithm crate called linfa-emsemble or something. Bootstrap aggregation should probably go in there as well.
I'd like to contribute quantile regression
In terms of functionality, the mid-term end goal is to achieve an offering of ML algorithms and pre-processing routines comparable to what is currently available in Python's
scikit-learn.These algorithms can either be:
* re-implemented in Rust; * re-exported from an existing Rust crate, if available on [crates.io](crates.io) with a compatible interface.In no particular order, focusing on the main gaps:
* Clustering: * [x] DBSCAN * [x] Spectral clustering; * [x] Hierarchical clustering; * [x] OPTICS. * Preprocessing: * [x] PCA * [x] ICA * [x] Normalisation * [x] CountVectoriser * [x] TFIDF * [x] t-SNE * Supervised Learning: * [x] Linear regression; * [x] Ridge regression; * [x] LASSO; * [x] ElasticNet; * [x] Support vector machines; * [x] Nearest Neighbours; * [ ] Gaussian processes; (integrating `friedrich` - tracking issue [Integrating friedrich into linfa nestordemeure/friedrich#1](https://github.com/nestordemeure/friedrich/issues/1)) * [x] Decision trees; * [ ] Random Forest * [x] Naive Bayes * [x] Logistic Regression * [ ] Ensemble Learning * [ ] Least Angle Regression * [x] PLSThe collection is on purpose loose and non-exhaustive, it will evolve over time - if there is an ML algorithm that you find yourself using often on a day to day, please feel free to contribute it 💯
Id love to take on Random Forest! I have previously implemented it simplistically in Go, but I'd love to make it happen in Rust. This is my first open source contribution - let me know how I can make it happen :)
I'd also would like to help this.
I'm interested in the least angle regression (lars). It seems that PR #115 was trying to implement it but it has paused for 3 years. So I guess it's basically abolished. I'm going to pick it up.
I am interested in random forests.
@MarekJReid @giorgiozoppi did either of you take a chance at random forests?
i look into. At school we did this week. For python binding maturin is perfect. @zenconnor should i look inside scitkit-learn? I was looking at scikit learn implementation, as soon I can i provide a class diagram of that.