Endor outperformed nearly 1,000 teams on a Kaggle benchmark
Yaniv Altshuler · Rethinking Predictive Analytics // Yaniv Altshuler, Endor (FirstMark's Data Driven) · Apr 6, 2017 · at 15:08
Yaniv Altshuler, CTO at Endor, details Endor's prediction accuracy results on a Kaggle competition involving 300 million point-of-sale transactions.
0:00 / 0:20exact quote · 20.4s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“Almost a thousand teams participated. And they were given three months, and you can see that the team who won the first place after us, they submitted almost 200 models, so they actually worked pretty hard, and as you can see here, our accuracy is significantly better than all of them.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Yaniv Altshuler
Altshuler: Deep learning is inadequate for predicting dynamic human behavior
“Deep learning requires a lot of data, relatively speaking, and therefore, by definition, cannot detect recent changes in the data, emerging patterns, ah, and therefore, it's, ah, really, really, ah, inadequate to predict human behavior that is, ah, actually do…”
Altshuler: Mathematical invariances guaranteed in all structured human data
“There is a set of mathematical invariances that Are guaranteed to happen in any structured human data.”
Altshuler: Modern machine learning technologies were developed roughly 50 years ago
“All the technologies that we have today were developed around 50 years ago, plus minus a decade. Even deep learning that is so trendy today, it's basically very large neural networks. Concepts from the fifties matured around the seventies.”
Altshuler: Endor can run predictive analytics on one week of data
“In addition, because we use social physics as a prior to the data, we can analyze data from very short periods of time. We don't need to analyze five years worth of data. In some cases, even one week is enough”
Altshuler: Domain and ML expertise can be replaced with computation
“Can we replace domain expertise and machine learning expertise with computation? Our answer, yes we can, and it's not even so computationally intense, but it would require us to take a very, very different approach about data.”
Altshuler: Internal data science teams cannot be replaced by automated platforms
“I actually don't think this is the case because I think that nothing and no one can replace internal teams of data scientists and of analytics people because they are the only ones who are familiar with data and their business, and therefore they are the only …”