Amir Abolhassani

• PhD
• Senior Data Scientist at General Motors Company

What is the right way of interpreting the relation between Y & X in the following graphs? I have calculated the Pearson and Spearman correlation. However, they are significant, but the correlation coefficient is too small (There is a high number of observations around zero on the X axis, I'm wondering if I need to use the dataset as it is or do something specific prior calculating the correlation coefficient between Y & X). Thank you.

I have a conceptual question related to fuzzy logic method. I have a fuzzy environment and planning to use a fuzzy logic method to define/develop a productivity measure. However, I’m confused about the relation between the sample size (with predefined range for each input/feature which will limit the inputs range) in the data and fuzzy logic method parameters such as membership function etc. To me it seems there is no relationship, if you have one observation or 1000, you can define the Fuzzy classification of input and output variables/levels (with predefined range which the range is not coming from the data) and the functions, no matter how many observations are available. In that case, the surface viewer is something that is identified by the rules only and not available data (no matter if there is one observation or 1000 ones) and the hypothetical values of a use case that you are identifying.

Am I right or I'm missing something? Thank you.

P.S. I have attached a surface viewer example.

Hello, There is a dataset with several KPIs which are varying between (0,1). What is the best analytical approach to split the data and define a line in two dimensional (or define a plane in multi-dimensional space) based on data behavior and practical assumptions/considerations (there is some recommended ranges for each KPIs etc.)?

For instance in the attached screenshot, I want to flag the individuals/observations in A_e area for more investigation. I want to be able to apply the proposed approach in multi-dimensional space with several KPIs as well. Any thoughts would be appreciated.

Imagine that I have a regression model equation as,

Y= B₀ + B₁X₁ + B₂X₂ + B₃X3 + B4X₄ + e

Where,

X1 is a variable with continuous value

X2 is a categorical variable at two levels such as gender

X3 is a variable with integer value

X4 is another categorical variable at two levels such as if a person is experienced or inexperience

I can solve it by using an OLS regression or using an unbalanced panel data. In the OLS regression, I can compare male group verses female group or experience group versus inexperience group which seems like I'm treating them as fixed effects in an unbalanced panel data. In that case, both of them are the same and I have to get the same results, am I right? Is there any advantages using OLS regression versus unbalanced panel data? Can I have a panel data without considering the time impact on the dependent variable? Thank you.

In one hand, there are limited number of data-points and so many features for the analysis, let’s say 50 data-points and 60 features. On the other hand, the unusual data-points are more important than the usual-points and you want to know the underline facts of happening those unusual data-points. How do you approach the analysis to tackle these two issues at the same time?

Any help would be appreciated.

I'm interested in knowing the limitations of different open-source solvers such as PuLP, GLPK, and Symphony comparing to commercial optimization solvers such as CPLEX or Gurobi.

Any information would be appreciated.

I would appreciate if you could define what different steps are necessary to develop a policy for determining the frequency of a PMs on a specific equipment. For instance, we have a pump that it has three different PMs and we are not gathering any sensors' information such as current, vibration, etc. How can we define the frequency of these PMs for this pump that might have various failure modes? Thank you.

There are 11 predictors, combination of continuous and categorical variables. The response is a continuous variable and there isn’t any multicollinearity between variables, all VIFs are lower than 2.9. Multiple linear regression, ridge, lasso, and weighted regression methods (the weights are calculated through a robust method) are used. In contrary to other regression methods, the sign of a particular predictor coefficients, a binary variable, is flipping when weighted regression method is used.

I appreciate for any feedback and thought about the flip sign of the binary variable.