Deep Learning in Business Analytics: A Clash of Expectations and Reality

by belter • 4 years ago

19 10 4 years ago

belter• 4 years ago

PDF: https://arxiv.org/ftp/arxiv/papers/2205/2205.09337.pdf

"...This paper explains why DL - despite its popularity - has difficulties speeding up its adoption within business analytics. It is shown - by a mixture of content analysis and empirical study - that the adoption of deep learning is not only affected by computational complexity, lacking big data architecture, lack of transparency (black-box), and skill shortage, but also by the fact that DL does not outperform traditional ML models in the case of structured datasets with fixed-length feature vectors. Deep learning should be regarded as a powerful addition to the existing body of ML models instead of a one size fits all solution..."

stanfordkid• 4 years ago

I'd be curious to understand why this is the case. It seems that businesses generally have highly network style data -- think of your average system for a "widget" business... it ranges from orders to suppliers to manufacturing processes, advertising, all the way to orders from customers. All stored in different systems with disparate join keys.

Most of the hard part of BI I have seen is joining these pieces of information together that are otherwise uncorrelated -- e.g which clicks corresponded to conversion is a tracking problem, then correlating them with actual business actions (e.g running a marketing campaign or improving a process)

It's more about data joining and data aggregation than "ML"

mountainriver3• 4 years ago

This is an incredibly shallow paper. GNNs are mostly used on structured data today and are a part of the Deep Learning family, as are transformers.

The authors seem to be unaware of that and are pegging deep learning as a very narrow implementation.

lumost• 4 years ago

Traditional methods typically outperform DNNs on fixed numeric features. It’s very difficult to beat a GBDT in these scenarios, and the GBDT doesn’t have many hyper parameters and trains without specialized hardware.

If you add in ingratitudes data like text the situation changes.

mountainriver3• 4 years ago

This https://arxiv.org/pdf/2106.11959.pdf would say differently. It shows there isn't a goto architecture when compared to modern DL

civilized• 4 years ago

If you're saying that neural nets should be a tool of first resort on tabular data, well, I've been watching for years and the theme has been extremely consistent: neural nets don't bring much to the table in this case, compared to existing ML methods. XGBoost is still the Kaggler tool of choice on tabular data, for example.

mountainriver3• 4 years ago

Not saying they should be the first tool of choice, saying that the study doesn't compare GB to more modern DL techniques like transformers. This is a good analysis https://arxiv.org/pdf/2106.11959.pdf and finds there isn't a go to architecture

rwilson4• 4 years ago

And here I am mostly doing t-tests...

crispyambulance• 4 years ago

Keep doing them!

Most "business analytics" or "business intelligence" is bullshit work. They justify enormous expense and labor yet spend very little on the quality of the data that they rely on.

"Garbage in ==> Garbage out" is something that needs to be emphasized more when folks try to apply fancy-pants techniques to business "intelligence".

g42gregory• 4 years ago

Very few businesses have the Google/Meta level of data collection. In the low-data environments, SVMs (as an example), more often than not, outperform corresponding DL models.

Deep Learning in Business Analytics: A Clash of Expectations and Reality

Related Stories