MLJ is a machine learning framework for Julia aiming to provide a convenient way to use and combine a multitude of tools and models available in the Julia ML/Stats ecosystem.
MLJ is released under the MIT license and sponsored by the Alan Turing Institute.
Extend the categorical variable encoding of MLJ.
Difficulty. Moderate. Duration. 350 hours
MLJ provides basic one-hot encoding of categorical variables but no sophisticated encoding techniques. One-hot encoding is rather limited, in particular when a categorical has a very large number of classes. Many other techniques exists, and this project aims to make some of these available to the MLJ user.
Mentors. Anthony Blaom (best contact: direct message on Julia slack)
Julia language fluency is essential.
Git-workflow familiarity is strongly preferred.
Experience with machine learning and data science workflows.
Familiarity with MLJ's API a plus.
In this project you will survey popular existing methods for one-hot encoding categorical variables. In collaboration with the mentor, you will make a plan for integrating some of these techniques into MLJ. You will begin work on the plan, initially focusing on simple methods, providing MLJ interfaces to existing julia packages, or new implementations where needed. If the project advances well, you will implement more advanced techniques, such as entity embedding via MLJFlux.jl (MLJ's neural network interface).
Existing encoding in MLJ: OneHotEncoder; ContinuousEncoder; UnivariateContinuousTimeEncoder
StatsModels.jl encoders
MLJ feature request
Guo and Berkhahn [(2016]](https://arxiv.org/abs/1604.06737) "Entity Embeddings of Categorical Variables"
Implement survival analysis models for use in the MLJ machine learning platform.
Difficulty. Moderate - hard. Duration. 350 hours
Survival/time-to-event analysis is an important field of Statistics concerned with understanding the distribution of events over time. Survival analysis presents a unique challenge as we are also interested in events that do not take place, which we refer to as 'censoring'. Survival analysis methods are important in many real-world settings, such as health care (disease prognosis), finance and economics (risk of default), commercial ventures (customer churn), engineering (component lifetime), and many more. This project aims to implement models for performing survivor analysis with the MLJ machine learning framework.
mlr3proba is currently the most complete survival analysis interface, let's get SurvivalAnalysisA.jl to the same standard - but learning from mistakes along the way.
Mentors. Sebastian Vollmer, Anthony Blaom,
Julia language fluency is essential.
Git-workflow familiarity is strongly preferred.
Some experience with survival analysis.
Familiarity with MLJ's API a plus.
A passing familiarity with machine learning goals and workflow is
preferred.
You will work towards creating a survival analysis package with a range of metrics, capable of making distribution predictions for classical and ML models. You will bake in competing risks in early, as well as prediction transformations, and include both left and interval censoring. You will code up basic models (Cox PH and AFT), as well as one ML model as a proof of concept (probably decision tree is simplest or Coxnet).
Specifically, you will:
Familiarize yourself with the training and evaluation machine
learning models in MLJ.
For SurvivalAnalysis.jl, implement the MLJ model interface.
Consider Explainability of SurvivalAnalysis through SurvSHAP(t)
Develop a proof of concept for newer advanced survival analysis
models not currently implemented in Julia.
Mateusz Krzyziński et al., SurvSHAP(t): Time-Dependent Explanations of Machine Learning Survival Models, Knowledge-Based Systems 262 (February 2023): 110234
Kvamme, H., Borgan, Ø., & Scheel, I. (2019). Time-to-event prediction with neural networks and Cox regression. Journal of Machine Learning Research, 20(129), 1–30.
Lee, C., Zame, W. R., Yoon, J., & van der Schaar, M. (2018). Deephit: A deep learning approach to survival analysis with competing risks. In Thirty-Second AAAI Conference on Artificial Intelligence.
Katzman, J. L., Shaham, U., Cloninger, A., Bates, J., Jiang, T., & Kluger, Y. (2018). DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Medical Research Methodology, 18(1), 24.
Gensheimer, M. F., & Narasimhan, B. (2019). A scalable discrete-time survival model for neural networks.](https://peerj.com/articles/6257/) PeerJ, 7, e6257.
Bayesian methods and probabilistic supervised learning provide uncertainty quantification. This project aims increasing integration to combine Bayesian and non-Bayesian methods using Turing.
Difficulty. Difficult. Duration. 350 hours.
As an initial step reproduce SOSSMLJ in Turing. The bulk of the project is to implement methods that combine multiple predictive distributions.
Interface between Turing and MLJ
Comparisons of ensembling, stacking of predictive distribution
reproducible benchmarks across various settings.
Mentors: Hong Ge Sebastian Vollmer
Help data scientists using MLJ track and share their machine learning experiments using MLflow. The emphasis iin this phase of the project is:
support asynchronous workflows, as appear in parallelized model tuning
support live logging while training iterative models, such as neural networks
Difficulty. Moderate. Duration. 350 hours.
MLflow is an open source platform for the machine learning life cycle. It allows the data scientist to upload experiment metadata and outputs to the platform for reproducing and sharing purposes. MLJ already allows users to report basic model performance evaluation to an MLflow service and this project seeks to greatly enhance this integration.
Julia language fluency essential
Understanding of asynchronous programming principles
Git-workflow familiarity strongly preferred.
General familiarity with data science workflows
You will familiarize yourself with MLJ, MLflow and MLflowClient.jl client APIs.
You will familiarize yourself with the MLJFlow.jl package providing MLJ <–> MLflow integration
Implement changes needed to allow correct asynchronous logging of model performance evaluations
Extend logging to (parallelized) model tuning (MLJ's TunedModel wrapper)
Extend logging to controlled training of iterative models (MLJ's IteratedModel wrapper)
MLflow website.
Mentors. Anthony Blaom
Diagnose and exploit opportunities for speeding up common MLJ workflows.
Difficulty. Moderate. Duration. 350 hours.
In addition to investigating a number of known performance bottlenecks, you will have some free reign in this to identify opportunities to speed up common MLJ workflows, as well as making better use of memory resources.
Julia language fluency essential.
Experience with multi-threading and multi-processor computing essential, preferably in Julia.
Git-workflow familiarity strongly preferred.
Familiarity with machine learning goals and workflow preferred
In this project you will:
familiarize yourself with the training, evaluation and tuning of machine learning models in MLJ
benchmark and profile common workflows to identify opportunities for further code optimizations, with a focus on the most popular models
work to address problems identified
roll out new data front-end for iterative models, to avoid unnecessary copying of data
experiment with adding multi-processor parallelism to the current learning networks scheduler
implement some of these optimizations
MLJ Roadmap. See, in particular "Scalability" section.
Data front end for MLJ models.
Mentors. Anthony Blaom, Okon Samuel.