previous article
PAW has grown on three parallel tracks in the meantime, and become correspondingly rich in content. Most of my favourites on this first day belong to the longer (more precisely: one-hour) lecture format for deep dives, which gets right to the point and into the details. More time, more input.
Sven Crone's deep dive revolved around the problems of time series: How to deal with them in a contemporary and competent manner whilst holding a lecture which is not boring even for a second. In contrast to many other time-series specialists, Sven did not portray classical statistical time-series methods (ARIMA & co.) as the only correct way to solve problems of time series. Instead, he showed how to meaningfully use neural networks for time-series forecasting. A use of neural networks here is not hype-driven vogue, but a consequence of the fact that neural networks in the case of such problems perform a lot better than many of the other robust workhorses such as random forest. Of interest here: Deep nets typically offer little added value for time-series problems. Flat architectures provide similar accuracy at significantly lower complexity. Deep neural networks can be used meaningfully by switching to the meta-level: Based on time-series data, deep learning can be used to decide which algorithm works best for these data.
Dr. John Snow saved human lives with data visualization. That was during the cholera epidemic in London, when he was able to use a map of death cases in a particularly stricken district to find the cause of the proliferation and render it harmless. With this example, Alfred Inselberg began his lecture on interactive data visualization using parallel coordinates. Frankly, I'm still not sure how useful visualisations with parallel coordinates are for my work - whenever I've had a look at them so far, I've never seen anything. I learnt that this might be because I was looking at static plots. If it works, then only with interactive plots additionally combined with other graphics (scatter plots, for example) which are updated simultaneously and consistently.
Phil Winters dealt with a (former?) trend: Bots. Not exactly my favourite topic, to be honest: Too much hype, too little practical benefit so far. But I figured that if Phil were dealing with it, he would do a fine job: He has a talent for bringing hype down to earth and testing its practical benefits with plenty of common sense. The bot "Emil" created specially for the presentation served consistently as an example of a pleasingly detailed depiction, coupled with a convincing plea for greater integration of user feedback into predictive models (active learning).
Last but not least, there was a topic which I, frankly, had initially suspected of being perhaps not entirely serious. I had to revise my prejudice: Jonathan Mall delivered an exciting, well-founded and extremely well-presented introduction to the psychologically clever use of word embeddings for marketing purposes. For me, this was one of the most stimulating lectures because it combined two favourite themes: Word embeddings and psychology. He showed how word embeddings can be used to quantify country- and language-specific term associations, find appropriate associations and modify inappropriate ones.
