My data set has 13 columns
I used column 1-4 as my X-values which is the baseline for the causality of what i am predicting. Columns 5-12 constitutes the summation of column 13 values which is what I will use as my y values.
column 13 as y values which is the main thing I am predicting.
The problem now is that the model is too awful.
R2 score = -1.2435659415110933e-05
mean_squared_error = 1.8704356137558542e+39
Any suggestion of what I can do?
For better understanding, I recommend you reading these articles listed below:
Article Forecasting methods in energy planning models
Article A review on renewable energy and electricity requirement for...
Article A comprehensive overview on the data driven and large scale ...
you can see this website
https://machinelearningmastery.com/multi-class-classification-tutorial-keras-deep-learning-library/
You must first scatterplots to check whether it is a linear relationship between the variables. If not, try other methods.
Reduce both the independent and dependent variables (i.e your x and y values into group of similar data and recalculate. You may bear in mind not to have too many similar variables as this may cause auto correlation..
Adekunle Ajibode -
If any of those four variables has predictive power for your y-values, you could determine this by graphing each x, one at a time, with y. Note that if the relationship is a curved line, it is still linear regression if you use polynomial regression. Note also that independent variables correlated with each other would mean you have collinearity, which messes with your coefficients and variances, but does not hurt prediction. That is not autocorrelation, which relates estimated residuals. That would be for time series, except for spatial autocorrelation. You did not say anything that indicated that would be an issue to consider.
You could then try one or more multiple regressions. If you research "graphical residual analysis," you could use that to check model fit, including heteroscedasticity. If you compare models on the same scatterplot with the same sample, you can compare fits. If you research "cross-validation," that helps check to see if you have overfit to a particular sample such that the model is not really for the entire population, or subpopulation to which you want it to apply. You could be modeling a random divergence. If you have enough data for two samples, you could do a graphical residual analysis comparing models for one sample on one scatterplot, and for another sample on another scatterplot.
The bottom line for prediction is how closely you predict data that you have, but did not use in your model selection or in estimating coefficients.
Best wishes - Jim
PS - I suggest always starting your regressions with a look at scatterplot graphs. Also, graphical residual analyses are often very helpful.
- Badges
- Science topic