New answers tagged

0

Following my comment above, here is a reproducible answer: library(CARBayesST) ################################################# #### Run the model on simulated data on a lattice ################################################# #### set up the regular lattice x.easting <- 1:10 x.northing <- 1:10 Grid <- ...


1

Using PCA to project the data to a lower dimension can yield worse results. This is as true for neural networks as it is for any other model. This is because the projection to a lower dimension is not aware of the outcome. All it does is retain the PCs with the largest variance; if the informative features PCs with lowest variance, then PCA will make the ...


5

You don't exactly know $\omega$ but you have some idea, a distribution based on the previous data you've seen, which is described by $p(\omega|X,Y)$. If you had a constant $\omega_0$, the posterior predictive distribution would be $p(y^*|x^*,\omega_0)$, but the integral is basically an expected value (i.e. a weighted average) over all possible $\omega$. By ...


0

I think you want to first use R's predict function to first generate Prediction intervals or Confidence intervals around the independent values. Read this link for more info. R prediction info The next step is to use those intervals to generate possible values with a random number generator scaled appropriately for those values.


0

A simple method is bootstrapping. Estimate regression, get the residuals. Get an empirical cdf of x. Then use inverse of it to generate new x set. Plug new x into the regression model and add bootstrapped residuals That was if you think x causes y. If there’s no causality then it’s easier. Get eCDFs of x and y. Then estimate correlation of eCDF outputs ...


1

While somewhat unlikely this phenomenon can indeed happen. The results from local explainer by LIME can disagree (on occasion substantially) with the results of the global model. Probably it is worth considering different kernel widths as well as checking the goodness-of-fit of the LIME explainer too. More details: LIME is training a model in the "...


0

This would be awesome! In fact it would be even better, if the firm would know how the weights would have been generated (for example if the weights might be biased due to ridge regression/lasso). This knowledge could be added in various ML models as some kind of "pre-knowledge" for example in a Baysian Framework. Or just be used as a guidline in model ...


0

The simulation is indeed often used. For example, in alpha-zero style MCTS+RL algorithms, the game tree is rolled out according to the known rules at test time. (You may argue the rules of chess are too simple to qualify as a "simulation" of any real environment of course). And a recent line of research is to make physics engines whose dynamics are ...


1

If I've properly understood what you've written, then yes. The only caveat here is that the data preprocessing is not based on the outcome. For example, many people select features by computing correlation with the outcome. Were you to do something like this, your validation procedure would look slightly different. Yes. The different modes of data ...


0

You have an enormous amount of data, but no immediately visible structure. You could try spline transformations for your percentage cover to account for possible nonlinearities, but it doesn't really look like this is very promising. Whatever you do, do a sensitivity analysis by removing the three very high data points and checking whether your results hold ...


1

Your response variable is ordinal with three levels, so try ordinal logistic regression, two models, one with each of the competing predictors, and compare them. You can for instance evaluate each of the models with cross-validation and see which is best. Or fit one model with both predictors?


0

Relatedly for regression with multi-dependent variables, per a source: ...the default metric used is the relative error, which is the mean squared error of predictions divided by the mean squared error of a default hypothesis always predicting the mean. which provides a general frame leading to a metric. Note, this relates to one of your mentioned ...


1

Don't use the hit rate as a quality measure for interval predictions. (Or if you do, do not be surprised if your winning algorithm predicts an interval of $(0,300)$ for all instances and gets a hit rate of 100%.) Your quality measure needs to balance coverage and length of the prediction intervals: yes, we want high coverage, but we also want short ...


1

If you addressed the multicollinearity issue by mean centering the data, then no issue with short-range prediction. With longer horizons, however, the question is do the historical means, to which you are adding the expected change per the regression model, remain valid? A lesser issue if your interest is in percent change forecast and your variables are ...


3

Like you quoted, multicollinearlity presents a much bigger problem for inference than prediction. Prediction is mostly fine under multicollinearity as long as the variables stay collinear (although if you get close to perfect collinearity, you can get some precision issues with regards to things like matrix inversion). But there is the issue that we don't ...


0

You mentioned an ANOVA as well as prediction of an outcome. So you can run a linear regression with the predictors being an indicator for the binary variable and dummy variable for three of the four categories of the ordinal variable. You'd only need 3 dummy variables since you need to drop one as the base case. (In case you don't know, a dummy variable is ...


2

Don't use thresholds in assessing models. (See here for some of the problems with them.) Instead, find out which of your models yield well-calibrated and sharp probabilistic predictions, using proper scoring rules. We have a scoring-rules tag. (Consider combining models, this often improves predictions.) Once you have a well-performing model, consider using ...


0

All KPIs in classification can be important. (See the Wikipedia article on sensitivity and specificity for the entire zoo of common KPIs.) Which one is more important depends on the costs of correct and wrong classifications. Sometimes false positives are more costly than false negatives, or vice versa, and the relationships between these costs (and those of ...


0

I can offer a real world answer, but it is anecdotal. From a long life sampling way too many doctors, what doctors actually weigh very heavily is what they are seeing, to the point they fail to diagnose situations that do not fit the population. To make an example, doctors often see sebaceous (fatty) cysts under the skin, so tend to diagnose all ...


2

MAE=10 implies that, on average, the forecast's distance from the true value is 10 (e.g true value is 200 and forecast is 190 or true value is 200 and forecast is 210 would be a distance of 10). MAPE=10 implies that, on average, the forecast's distance from the true value is 10% of the true value (e.g true value is 100 and forecast is 90 or true value is 100 ...


0

Hard to know exactly without knowing what kind of model you're using, but it would be worth repeating the model without age to see what you get. It could be that you're missing important correlations by adjusting for age. Collider bias is one potential issue here. I would report univariable correlations and a multivariable model in this case.


0

Both Kendall Tau and Spearman R rank are all kind of correlation indicators, each has it's own assumptions and focus. But none of the focuses is about "prediction power". Prediction power is influenced by not only correlations, but also assumptions implied in your model. From this perspective, your second approach, which in essence is a model selection ...


1

This is a regression problem, so why are you using accuracy as a metric? It makes sense that your predictions will not exactly match the label value, as the output is continuous. So try MSE, RMSE, or some other continuous metric. Your loss function is converging, so there might not be a problem with the model structure at all.


Top 50 recent answers are included