In week 1, we have had a look at the 2019 NYC taxi dataset, after some preprocessing we realised that it was not suitable for our Federate Learning approach so we moved towards the 2013 dataset.
From week 2 to week 6 we have work on the dataset:
In week 7 we started using Tensorflow Keras to get some skill, I have tested out a simple recurrent sequence-to-sequence model which takes as input a sequence of locations and try to predict the next location for each input location.
On week 8 we learned how to use feature columns to embed and encode categorial features in some useful representation and how to define custom training cycles for our Keras model.
After some tutorials in week 9 and 10, in week 11 we started working with Keras and TFF to build a federated model for our taxi dataset. We have developed 2 centralized models, one using the keras simple Sequential model and another using the Keras functional API, and we tested them out as centralized models.
Then we have taken the second model and federated it doing experiments in week 12. The experimental results gave us some partial performance comparison between the centralized and the federated model which was able to achieve little higher scores (we would need to go deeper with these experiments to confirm their results).
In week 13 we developed an algorithm to generate synthetic mobility data in order to test out the Paul thesis conclusions relative to demographic groups models. The following week we have generated some easy to use code templates to use Tensorboard to track the experiments and training-evaluation pipelines to test the models against validation and test sets. In this week we have also used our synthetic dataset to test the Paul conclusions with 3 different models confirming his thesis, a group model seems not to improve performance significantly and suffer from lack of data.