Stratified splits constraint in XGBoost - xgboost

I'm using a large dataset spanning many years to cross-validate hyperparameters for an XGBoost model. This data can look different in different years, so to reduce generalization error I would like to disallow the model from making any splits that are imbalanced with respect to years, i.e. don't let it split on year. For example, adding a constraint that all splits must contain at least n samples from each year, or adding a penalty on how far the ratio of each year's data in the split differs from 1/2. I don't have the timestamp as a feature but there are other features that would allow it to do effectively this. I don't see anything in the documentation that covers this use-case, but I was wondering if there might be some trick (eg. with monotonicity constraints) that could work.

Related

A huge number of discrete features

I'm developing a regression model. But I ran into a problem when preparing the data. 17 out of 20 signs are categorical, and there are a lot of categories in each of them. Using one-hot-encoding, my data table is transformed into a 10000x6000 table. How should I prepare this type of data?
I used PCA, trying to reduce the dimension, but even 70% of the variance is in 2500 features. That's why I joined.
Unfortunately, I can't attach the dataset, as it is confidential
How do I prepare the data to achieve the best results in the learning process?
Can the data be mapped more accurately in a non-linear manner? If so, you might want to try using an autoencoder for dimensionality reduction.
One thing to note about PCA is that it computes an orthogonal projection of the data into linear space. This means that it only gives a linear mapping of the data. Autoencoders, on the other hand, can give you a non-linear mapping, and so is able to represent a greater amount of variance in the data in fewer dimensions. Just be sure to use non-linear activation functions in your autoencoder architecture.
It really depends on exactly what you are trying to do. Getting a covariance matrix (and also PCA decomp.) will give you great insight about which classes tend to come together (and this requires one-hot encoded categories), but training a model off of that might be problematic.
In general, it really depends on the model you want to use.
One option would be a random forest. They can definitely be used for regression, though they need to be trained specifically for that. SKLearn has a class just for this:
https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html
The benifits of random forest is that it is great for tabular data (as is the case here), and can easily be trained using numerical values for class features, meaning your data vector can only be of dimension 20!
Decision tree models (such as random forest) are being shown to out-preform deep-learning in many cases, and this may be one of them.
TLDR; If you use a random forest, it can take learn even with numerical values for categories, and you can avoid creating incredibly large vectors for data.

How can I study the properties of outliers in high-dimensional data?

I have a bundle of high-dimensional data and the instances are labeled as outliers or not. I am looking to get some insights around where these outliers reside within the data. I seek to answer questions like:
Are the outliers spread far apart from each other? Or are they clustered together?
Are the outliers lying 'in-between' clusters of good data? Or are they on the 'edge' boundaries of the data?
If outliers are clustered together, how do these cluster densities compare with clusters of good data?
'Where' are the outliers?
What kind of techniques will let me find these insights? If the data was 2 or 3-dimensional, I can easily plot the data and just look at it. But I can't do it high-dimensional data.
Analyzing the Statistical Properties of Outliers
First of all, if you can choose to focus on specific features. For
example, if you know a featues is subject to high variation, you can
draw a box plot. You can also draw a 2D graph if you want to focus on
2 features. THis shows how much the labelled outliers vary.
Next, there's a metric called a Z-score, which basically says how
many standard devations a point varies compared to the mean. The
Z-score is signed, meaning if a point is below the mean, the Z-score
will be negative. This can be used to analyze all the features of the
dataset. You can find the threshold value in your labelled dataset for which all the points above that threshold are labelled outliers
Lastly, we can find the interquartile range and similarly filter
based on it. The IQR is simply the difference between the 75
percentile and 25 percentile. You can also use this similarly to Z-score.
Using these techniques, we can analyze some of the statistical properties of the outliers.
If you also want to analyze the clusters, you can adapt the DBSCAN algorithm to your problem. This algorithm clusters data based on densities, so it will be easy to apply the techniques to outliers.

Algorithm - finding the order of HMM from observations

I am given a data that consists of N sequences of variable lengths of hidden variables and their corresponding observed variables (i.e., I have both the hidden variables and the observed variables for each sequence).
Is there a way to find the order K of the "best" HMM model for this data, without exhaustive search? (justified heuristics are also legitimate).
I think there may be a confusion about the word "order":
A first-order HMM is an HMM which transition matrix depends only on the previous state. A 2nd-order HMM is an HMM which transition matrix depends only on the 2 previous states, and so on. As the order increases, the theory gets "thicker" (i.e., the equations) and very few implementations of such complex models are implemented in mainstream libraries.
A search on your favorite browser with the keywords "second-order HMM" will bring you to meaningful readings about these models.
If by order you mean the number of states, and with the assumptions that you use single distributions assigned to each state (i.e., you do not use HMMs with mixtures of distributions) then, indeed the only hyperparameter you need to tune is the number of states.
You can estimate the optimal number of states using criteria such as the Bayesian Information Criterion, the Akaike Information Criterion, or the Minimum Message Length Criterion which are based on model's likelihood computations. Usually, the use of these criteria necessitates training multiple models in order to be able to compute some meaningful likelihood results to compare.
If you just want to get a blur idea of a good K value that may not be optimal, a k-means clustering combined with the percentage of variance explained can do the trick: if X clusters explain more than, let say, 90% of the variance of the observations in your training set then, going with an X-state HMM is a good start. The 3 first criteria are interesting because they include a penalty term that goes with the number of parameters of the model and can therefore prevent some overfitting.
These criteria can also be applied when one uses mixture-based HMMs, in which case there are more hyperparameters to tune (i.e., the number of states and the number of component of the mixture models).

Multiple trained models vs Multple features and one model

I'm trying to build a regression based M/L model using tensorflow.
I am trying to estimate an object's ETA based on the following:
distance from target
distance from target (X component)
distance from target (Y component)
speed
The object travels on specific journeys. This could be represented as from A->B or from A->C or from D->F (POINT 1 -> POINT 2). There are 500 specific journeys (between a set of points).
These journeys aren't completely straight lines, and every journey is different (ie. the shape of the route taken).
I have two ways of getting around this problem:
I can have 500 different models with 4 features and one label(the training ETA data).
I can have 1 model with 5 features and one label.
My dilemma is that if I use option 1, that's added complexity, but will be more accurate as every model will be specific to each journey.
If I use option 2, the model will be pretty simple, but I don't know if it would work properly. The new feature that I would add are originCode+ destinationCode. Unfortunately these are not quantifiable in order to make any numerical sense or pattern - they're just text that define the journey (journey A->B, and the feature would be 'AB').
Is there some way that I can use one model, and categorize the features so that one feature is just a 'grouping' feature (in order separate the training data with respect to the journey.
In ML, I believe that option 2 is generally the better option. We prefer general models rather than tailoring many models to specific tasks, as that gets dangerously close to hardcoding, which is what we're trying to get away from by using ML!
I think that, depending on the training data you have available, and the model size, a one-hot vector could be used to describe the starting/end points for the model. Eg, say we have 5 points (ABCDE), and we are going from position B to position C, this could be represented by the vector:
0100000100
as in, the first five values correspond to the origin spot whereas the second five are the destination. It is also possible to combine these if you want to reduce your input feature space to:
01100
There are other things to consider, as Scott has said in the comments:
How much data do you have? Maybe the feature space will be too big this way, I can't be sure. If you have enough data, then the model will intuitively learn the general distances (not actually, but intrinsically in the data) between datapoints.
If you have enough data, you might even be able to accurately predict between two points you don't have data for!
If it does come down to not having enough data, then finding representative features of the journey will come into use, ie. length of journey, shape of the journey, elevation travelled etc. Also a metric for distance travelled from the origin could be useful.
Best of luck!
I would be inclined to lean toward individual models. This is because, for a given position along a given route and a constant speed, the ETA is a deterministic function of time. If one moves monotonically closer to the target along the route, it is also a deterministic function of distance to target. Thus, there is no information to transfer from one route to the next, i.e. "lumping" their parameters offers no a priori benefit. This is assuming, of course, that you have several "trips" worth of data along each route (i.e. (distance, speed) collected once per minute, or some such). If you have only, say, one datum per route then lumping the parameters is a must. However, in such a low-data scenario, I believe that including a dummy variable for "which route" would ultimately be fruitless, since that would introduce a number of parameters that rivals the size of your dataset.
As a side note, NEITHER of the models you describe could handle new routes. I would be inclined to build an individual model per route, data quantity permitting, and a single model neglecting the route identity entirely just for handling new routes, until sufficient data is available to build a model for that route.

Select important features then impute or first impute then select important features?

I have a dataset with lots of features (mostly categorical features(Yes/No)) and lots of missing values.
One of the techniques for dimensionality reduction is to generate a large and carefully constructed set of trees against a target attribute and then use each attribute’s usage statistics to find the most informative subset of features. That is basically we can generate a large set of very shallow trees, with each tree being trained on a small fraction of the total number of attributes. If an attribute is often selected as best split, it is most likely an informative feature to retain.
I am also using an imputer to fill the missing values.
My doubt is what should be the order to the above two. Which of the above two (dimensionality reduction and imputation) to do first and why?
From mathematical perspective you should always avoid data imputation (in the sense - use it only if you have to). In other words - if you have a method which can work with missing values - use it (if you do not - you are left with data imputation).
Data imputation is nearly always heavily biased, it has been shown so many times, I believe that I even read paper about it which is ~20 years old. In general - in order to do a statistically sound data imputation you need to fit a very good generative model. Just imputing "most common", mean value etc. makes assumptions about the data of similar strength to the Naive Bayes.