Wednesday, January 14, 2015

Paris Machine Learning Meetup #5 Season 2, Time Series and FinTech, Adversarial Algos

Paris Machine Learning Meetup #5 Season 2 is tonight and it will be was a joint meetup with the Paris Financial Engineering meetup. Our current meetup membership has risen to 1700+ members making it the largest Machine Learning Meetup outside the U.S. (and top 6 worldwide).

We have some growth issues, this is why we are trying a new location today. Our sponsor and host is the Maltem Consulting Group.

Let us note that we also have about 600 professionals listed in the LinkedIn Paris Machine Learning group (this is where the jobs should be posted). The meetup hashtag is #MLParis (all the previous speakers and sponsors on Twitter, are on this list). The video below will restransmit the meetup. It starts at 7:00PM Paris time. All the presentations are already available below and will be shortly on the meetup archives:
  Warning: It is very likely that all the talks will be in French. However, all the slides are listed below and are in English.
The program focused on times series and specifically on adversarial algorithms i.e algorithms that have to discover time changing patterns when the other side also have similar probing capabilities. We had the example of Botnet detections using a mix of unsupervised learning and metric evaluation in one presentation and the decomposition of financial series using a new metric in order to different correlated movements between series and those linked to specific distributions. We also had the presentation of a simple API that can easily provide low latency financial data as well as an historically remarkable invariant trend over the years in financial data.

Au programme, un focus sur l'apport du machine learning appliqué aux séries temporelles et à la finance quantitative:

    In this talk, we present a novel non-parametric approach for clustering high-dimensional Markov processes by splitting apart dependency and distribution information without losing any. This approach is able to recover any generative-random-walk-model ground-truth unlike a straightforward application of clustering techniques to time series. 
    We have applied the method to a large financial market dataset of  time series, the CDS market prices of more than 600 companies over the last 8 years through the global financial crisis. Results are compared both with the basic workflow clusters and with an expert's classification. We illustrate their clustering dissimilarities using the HCMapper, a bespoke data visualisation designed to highlight clustering mismatches.
    Some experiment code and internal research are released at on an ongoing basis. 
    Join the CompressiveSensing subreddit or the Google+ Community and post there !
    Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.

    No comments: