scikit-learn: machine learning in Python. Please feel free to ask specific questions about scikit-learn. Please try to keep the discussion focused on scikit-learn usage and immediately related open source projects from the Python ecosystem.
I have two questions about HistGradientBoostingClassifier. a) When using early stopping do you end up with the "best model" according to the validation loss or the most recent one after it stops?
b) Is the validation set chosen by HistGradientBoostingClassifier chosen at random and is it the same set for every iteration of the training?
maybe these should be asked on github as an issue?
a) there's no notion of best model. early stopping stops the training process if the score hasn't improved by more than
b) yes and yes
tol in the last n_iter_no_change iterations. The score can be the loss or an arbitrary scorer and it can be computed on the training set or on the validation setb) yes and yes
I'm not sure what you mean by best model. There's no notion of best model, only one model is built. If you mean "model with the lowest training loss" that's basically the model at the last iteration, under the assumption that the training loss is always supposed to decrease (unless your learning rate becomes too high). If you mean "model with the lowest training loss that doesn't make the validation loss go up", that's what early stopping is supposed to give you (and it's preferable to the former)
It is common in early stopping for the final valudation loss to be higher than the loss a few epochs before. How long you wait to see if the loss will start going down again is sometimes called "patience" . I think that's what pytorch lightning calls it
I'm trying to use a Lasso in a machine learning context, and I want to keep updating it with each test run I do
partial_fit is doing an update
hrm... it seems like all of the ones with that method only are for classification, not for regressed output.
I will use that in the project my team is doing. Thanks for helping
I wrote a short blog post that might be of interest to the community - deploying Scikit-Learn models to Kuberentes using Bodywork (an open source deployment tool that I have developed).
I would like to ask how to join two pereprocessors I did saved in two separate files. I have one file with a model and another with prerpocessor (doing average and filling NaN boxes) then another file with a model (same estimator) and forth file is preprocessor for second model? I would like to merge these four files into two (one joint model and one joint estimator).
3 replies
I have transformer1.file, model1.file, transformer2.file and model2.file (same estimator in model1 and model2). I would like to have tranformer_composite.file and model_composite.file.
Hi everyone, I'm a graduate student at Cornell and I had a paper (https://dl.acm.org/doi/abs/10.1145/3429445) published a while ago in correcting the bias of feature importance in tree-based methods. Impurity-based feature importances can be misleading for high cardinality features (many unique values), which is already noted in the docstring of featureimportances in RandomForest. I just opened a new pull request #20058 to implement a new feature importance measurement based on out-of-bag samples, which is guaranteed to remove this bias. I think this feature is going to be useful for scikit-learn users. Any comments or suggestions will be helpful!
Thanks! I have a list of particles with their caracteristics (position, direction, energy and stadistic weight). This variables are correlated. I want to fit that curves and later sample new particles. I want to know how scikit-learn keep the correlation. I'm sorry if it is a beginner question. And thanks again.
Well you need to define what are your feature variables and the target variable. So for instance you could try to predict the position from all the other variables. Correlations would be taken into account depending on the model. So for instance if your model is linear the target would be a linear combination of the features. If you do have a known analytical relation between your variables it might be easier and more reliable to use scipy.optimize or scipy.odr to find the coefficients you would like to learn though.
Thanks for your answer, I forgot to mention that I fit my data with KDE and I sample new particles from this model. Is this model keeping the correlation between the different variables? Or it assumes the variables are independent from each other? Thanks again!
2 replies
Hi everyone, I'm trying to create a streamlit app but face the following message when I run
streamlit run <file.py>: Make this Notebook Trusted to load map: File -> Trust Notebook. I Googled this issue, and even after making Chrome my default browser, nothing changes. Please help.
Hello! in sklearn.decomposition.PCA, how do I tell it which column represents the label?
For example, I have a dataframe with the following columns:
feature_0 feature_1 feature_2 label
How do I tell PCA that label is the dependent variable?
Hello
Is the
Thanks
Is the
accuracy_at_k (https://scikit-learn.org/stable/modules/generated/sklearn.metrics.top_k_accuracy_score.html#sklearn.metrics.top_k_accuracy_score) an implementation of the hit ratio at k(https://www.researchgate.net/publication/344486356_Hit_ratio_An_Evaluation_Metric_for_Hashtag_Recommendation)Thanks
Hi folks, I am a master's student in CS and I have a question for you. I am working on a multi-class text classification problem, and I am using scikit-learn to implement my solution. I want to predict for a paragraph x if x belongs to one out of seven categories of information. I already implemented my solution using your library, but I am not confident if the steps I am following are correct or not, or if I am missing something. Could you please take a look at the image below and give your opinion? If this is not the right place for this kind of question, please let me know. Thank you in advance for your contribution! 
4 replies
Hi all! We're working on a generic implementation of a discrete time survival model for random forests. Similar to this and this. Basically, the idea is to split on hazard curves which are a bit like the class probabilities of regular classification random forests but then stratified per duration since inception of an observation. We want to use scikit-learn for a base. Is anyone here familiar with the random forest code? Also tips for a good PR are very welcome.
It really depends on the kind of data that you have. If you have a corpus of documents LDA would be one way to get cluster/topics https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.LatentDirichletAllocation.html
You could also try pre-trained embeddings like word2vec and the likes
why do they have to be of the same length?
