July 7, 2017 •Brian Wong
Last week, I attended the Data Intelligence Conference hosted by Capital One in McLean, Virginia. The three-day conference focused on a community of practitioners in the field of machine learning and artificial intelligence modeling. The talks at the conference centered on predictive modeling methods, real-life case studies involving machine learning and predictive analytics, data governance and security issues, and other data cleaning and visualization methods.
There were certainly a wide-range of topics and expertise levels throughout the talks, and I was able to pick up a few ideas on how to approach different real-world problems. Two talks were of particular interest to me: one on differentially private algorithms for modeling and another on methods in outlier detection.
Differentially private algorithms are methods used to scramble the input data into a model while keeping strong model performance. In areas like health care and other industries with personally identifiable information, this can be vital to protecting the private information of data subjects. We were shown that it is certainly possible to back out private input data based on the results of the dataset and the model that is being used.
Outlier and anomaly detection is a common problem in predictive modeling as erroneous outliers can have a strong effect on the results of a model. I learned of a variety of new methods for anomaly detection, such as the t-Distributed Stochastic Neighbor Embedding (t-SNE) method, and ways to visualize data sets with the goal of finding outliers. Besides these newer methods, it was a good refresher to see more standard methods, such as scatterplots, histograms, and summary statistics, being used as a first resort when inspect data.
While this is the first year of this conference being held, it was run very smoothly, and I hope to see what next year’s conference looks like!