Why Multiple Linear Regression performs much better than Neural Network?

Question

I am working on a problem with eight inputs and one output with the following sample training data. https://i.stack.imgur.com/F3iyB.png

I used two methods to try to solve this using neural network (NN) and multi-linear regression (MLR). and the prediction results using MLR massively outperforms NN. I tried almost everything to the NN such as grid search to find the optimal "settings", and a lot of trial and error but still couldn't get the model to perform.

(blue lines are predictions and orange lines are the targets), a perfect model would fit the two lines perfectly.

Results using MLR:

https://i.stack.imgur.com/UDl5i.png

Results Using NN:

https://i.stack.imgur.com/oYm0O.png

I'm not really looking for an answer to fix the NN, but more of an explanation of why this might happen.

Answer 1

Look into overfitting or bias/variance tradeoff. Generally, more complex models require a lot more data to train on. If you have too few data points then the model will memorize the data set rather than learn the pattern in there