Does traditional predictive models in statistics, such as logistics regression and ARIMA model, necessarily perform bad than nowadays' fancy computational methods such as neural networks, random forest tree, deep learning, etc?Don't you think, in the real applied world, the fancy named machine learning techniques are overly advocated nowadays?
Hi, All of the methods have advantages and disadvantages. Attached is a paper that looks at a comparison involving logistic regression with variable selection. Hope this helps. Best, David Booth
Hi. I have experimented with exponential smoothing methods and ANN for time series forecasting with less data. Performance of both the algorithms is similar.
It depends on the problem. If the boundary between categories is linear and the predictive relationships are linear, the regression can perform just as well. When the problem is less well-behaved, a more flexible model may be necessary.
ANN kind have learning methods for building models. ANN is employed generally when the relationship is not clearly definable by simple mathematical equation. Where as the regression kind of statistical methods bring out relationship based on the test data set. Generally statistical and process control data are solved thru regression.
Both methods are not directly comparable.
As Michael pointed out it depends on the problem at hand. Deep Learners perform really well for image and natural language processing problems. They are well suited for medical imaging, e. g. early detection of fibrosis or Alzheimer's. But they need a carnivorous amount of training data. Also they have the problem of fading signal if there are many layers. It is often good practice to start training a simple learner such as Bayesian model as benchmark, then use more sophisticated methods such as Logistic Regression which need much more "fine tuning". May be this is of interest:
Conference Paper Multimedia data mining using deep learning
One way of looking at it is the tradeoff between bias and variance. Nonlinear models like ANN may have low bias but in the same time high variance. Since both sources of errors determine the overall model accuracy a complex model may in effect (depending on the data) be less accurate than simpler models.
- Badges
- Science topic