This github repository (https://github.com/statneutrino/airbnb-newyork-rents) is the final outcome for a project that I undertook at Udacity. I wrote a Machine Learning Pipeline to solve the following problem: a property management company is renting rooms and properties in New York for short periods on various rental platforms.

Brief: To estimate the typical price for a given property based on the price of similar properties. The company receives new data in bulk every week, so the model needs to be retrained with the same cadence, necessitating a reusable pipeline.

This solution is an end-to-end pipeline covering data fetching, validation, segregation, train and validation, test, and release. I have run the model on an initial data sample, re-run it on a new data sample simulating a new data delivery, and it can be reran on new data in production.

This solution implements Bayesian optimisation using Adaptive Experimentation Platform or Ax to search for optimum hyperparameters. This can be found in release v1.0.3

I have used Weights & Biases for tracking for artifacts, experiments and hyperparameters for this project. I have made my W&B project public and can be found here

In order to run the scripts from a linux console, install mlflow and python=3.8. The run the following command

$ mlflow run https://github.com/statneutrino/airbnb-newyork-rents.git -v 1.0.4

This runs the latest release.

If wanting to use hydra-axsweeper to perform Bayesian optimisation for hyperparameter search, you override hydra config using a command like so:

$ mlflow run https://github.com/statneutrino/airbnb-newyork-rents.git \
$ -v 1.0.3 \
$ -P hydra_options="modeling.max_tfidf_features=int(interval(10,30)) modeling.random_forest.max_features=interval(0.1,1.0) -m"

More information can be found on the documentation

There is currently a bug with the current implementation of axsweeper - only the first 5 jobs can be run before an error is raised - I am currently investigating this and will update when ready

What I'd like to do next is implement Bayesian Additive Regression Trees (BART) within an mlflow pipeline. Advantages are:

• Much less parameter optimization required than random forests and gradient-boosting methods
• Provides confidence intervals in addition to just point estimates
• Flexible through use of priors and embedding in bigger models

Personally I think this model's performance could be improved quite easily with new data sources e.g.

• Image data (CNN/deep learning) of homes/rooms features
• Data linkage of latitude and longitude with neighbourhood/spatial economic or tourism data