Improvement Discussion

[miqbalrp]

From Nadhira (feel free to add more):

• Feature Engineering: (need domain knowledge, other features that can be added)
• Correlation between columns + VIF (reduce redundancy and improve modeling capabilities)
• Data exploration (to make the core hypothesis)
• Weighted Average and Information Value (importance from each variable)
• Using GridSearchCV (sklearn)

[miqbalrp]

From Iqbal

• Need to consider more models such as SVM, etc
• Project structure (need advice from Mr. Dudung)
• Optimization on running the simulation, maybe we can remove the visual?

[miqbalrp]

Considering using another simulation, ie: brazil nut effect on binary segregation. Simulation: Click here

Advantage:

• More practical in industry
• More straight-forward objective: predict in what condition/in what frequency the system will segregate
• Propose the initial condition is when the system is fully mixed

Disadvantage:

• So far no data generated, need to change the code for generating data automatically
• At a glance, the system tends to stay at the initial condition

[miqbalrp]

After a look at some basic visualization from the data, I think the distribution is very random based on the labels.

[miqbalrp]

After a look at some basic visualization from the data, I think the distribution is very random based on the labels.

I think we need to remove some features to minimize the random. I suggest to fix the kn and gn.

[dudung]

Let's give a try to make a hyperplane with z, where x ≡ f^2, y ≡ A --> z = c xy, with c = 4 π^2 / g, since Γ = 4 π^2 f^2 A / g

[miqbalrp]

Here is the hyperplane that shows the relationship between frequency, amplitude, and Gamma. I have included the code in the visualization notebook. Hope this is the correct way to plot