4 ms·
This reflects something I say often as a data scientist: the best data science implementations augment humans. Domain understanding is key. Integration of the
by tomrod 5y ago
This reflects something I say often as a data scientist: the best data science implementations augment humans.
Domain understanding is key. Integration of the domain into ML models are a key part of generalizability -- far beyond train/validate/test, which are subject to data collection biases and temporal trends.
Zillow had a high, general equilibrium impact on a large market. We saw microcosms of this behavior in the "good old days" of Amazon book-selling when competitors would mark up a books price to one penny more than another, resulting in outrageously priced books valued far below the offering price. No behavioral governors were in place, review was inadequate, and they become part of Data Science lore we tell ourselves as our single threaded scripts running Python payloads warm our homes and melt our s'mores whilst our GPU weeps.
Data Science as a field is distinct from ML in that it uses the tools of ML to augment and enhance digital services. Despite the exuberance of boot camps to market prediction and neural nets as the end-all-be-all solution for hard business problems, value in data science adoption isn't solely in prediction, but in validation. Causal modeling and inference are key in this space, but also simply quantitative understanding of statistics and algorithms. As an example, you need someone in the room who understands multi-armed bandit sometimes. But, more importantly, you'll always need someone in the room who understands how multi-armed bandit would improve a business' value proposition.