Modelling tabular data with Google’s TabNet

Released in 2019, Google Research’s TabNet is claimed in a preprint manuscript to outperform existing methods on tabular data. How does it work and how can one try it?

Tabular data probably make up the majority of business data today. Think of things like retail transactions, click stream data, temperature and pressure sensors in factories, KYC information… the variety is endless.

In another post, I introduced CatBoost, one of my favorite methods for building prediction models on tabular data, and its neural network counterpart, NODE. But around the same time as the NODE manuscript came out, Google Research released a manuscript taking a totally different approach to tabular data modelling with neural networks. Whereas NODE mimics decision tree ensembles, Google’s proposed TabNet tries to build a new kind of architecture suitable for tabular data.

The paper describing the method is called TabNet: Attentive Interpretable Tabular Learning, which nicely summarizes what the authors are trying to do. The “Net” part tells us that it is a type of neural network, the “Attentive” part implies it is using an attention mechanism, it aims to be interpretable, and it is used for machine learning on tabular data.

How does it work?

TabNet uses a kind of soft feature selection to focus on just the features that are important for the example at hand. This is accomplished through a sequential multi-step decision mechanism. That is, the input information is processed top-down in several steps. As the manuscript puts it, “The idea of top-down attention in sequential form is inspired from its applications in processing visual and language data such as for visual question answering (Hudson & Manning, 2018) or in reinforcement learning (Mott et al., 2019) while searching for a small subset of relevant information in high dimensional input.”

The building blocks for performing this sequential attention are called transformer blocks even though they are a bit different from the transformers used in popular NLP models such as BERT. The soft feature selection is accomplished by using the sparsemax function.

The first figure from the paper, reproduced below, sketches how information is aggregated to form a prediction.

One nice property of TabNet is that it does not require feature preprocessing (in contrast to e.g. NODE). Another one is that it has interpretability built in “for free” in that the most relevant features are selected for each example. This means that you don’t have to apply an external explanation module such as shap or LIME.

It is not so easy to wrap one’s head around what is happening inside this architecture when reading the paper, but luckily there is published code which clarifies things a bit and shows that it is not as complicated as you might think.

How can I use it?

 

The original code and modifications

As already mentioned, the code is available, and the authors show how to use it together with the forest covertype dataset. To facilitate this, they have provided three dataset-specific files: one file that downloads and prepares the data (download_prepare_covertype.py), another one that defines the appropriate Tensorflow Feature Columns and a CSV reader input function (data_helper_covertype.py), and the file that contains the training loop (experiment_covertype.py).

The repo README states:

To modify the experiment to other tabular datasets:

– Substitute the train.csv, val.csv, and test.csv files under “data/” directory,

– Modify the data_helper function with the numerical and categorical features of the new dataset,

– Reoptimize the TabNet hyperparameters for the new dataset.

After having gone through this process a couple of times with other datasets, I decided to write my own wrapper code to streamline the process. This code, which I must stress is a totally unofficial fork, is on GitHub.

In terms of the README points above:

• Rather than making new train.csv, val.csv and test.csv files for each dataset, I preferred to read the entire dataset and do the splitting in-memory (as long as it is feasible, of course), so I wrote a new input function for Pandas in my code.
• It can take a bit of work to modify the data_helper.py file, at least initially when you aren’t quite sure what it does and how the feature columns should be defined (this was certainly the case with me). There are also many parameters which need to be changed but which are in the main training loop file rather than the data helper file. In view of this, I also tried to generalize and streamline this process in my code.
• I added some quick-and-dirty code for doing hyperparameter optimization, but so far only for classification.
• It is also worth mentioning that the example code from the authors only shows how to do classification, not regression, so that extra code also has to be written by the user. I have added regression functionality with a simple mean squared error loss.

Using the command-line interface

Execute a command like:

python train_tabnet.py \
  --csv-path data/adult.csv \
  --target-name "<=50K" \
  --categorical-features workclass,education,marital.status,\
occupation,relationship,race,sex,native.country\
  --feature_dim 16 \
  --output_dim 16 \
  --batch-size 4096 \
  --virtual-batch-size 128 \
  --batch-momentum 0.98 \
  --gamma 1.5 \
  --n_steps 5 \
  --decay-every 2500 \
  --lambda-sparsity 0.0001 \
  --max-steps 7700

The mandatory parameters are — -csv-path(pointing to the location of the CSV file),--target-name(the name of the column with the prediction target) and--categorical-featues (a comma-separated list of the features that should be treated as categorical). The rest of the input parameters are hyperparameters that need to be optimized for each specific problem. The values shown above, though, are taken directly from the TabNet manuscript, so they have already been optimized for the Adult Census dataset by the authors.

By default, the training process will write information to the tflog subfolder of the location where you execute the script. You can point tensorboard at this folder to look at training and validation stats:

tensorboard --logdir tflog

and point your web browser to localhost:6006.

If you don’t have a GPU…

… you could try this Colaboratory notebook. Note that if you want to look at the Tensorboard logs, your best bet is probably to create a Google Storage bucket and have the script write the logs there. This is accomplished by using the tb-log-locationparameter. E.g. if your bucket’s name were camembert-skyscrape, you could add--tb-log-location gs://camembert-skyscraperto the invocation of the script. (Note, though, that you have to set the permissions for the storage bucket correctly. This can be a bit of a hassle.)

Then you can point tensorboard, from your own local computer, to that bucket:

tensorboard --logdir gs://camembert-skyscraper

Hyperparameter optimization

There is also a quick-and-dirty script for doing hyperparameter optimization in the repo (opt_tabnet.py). Again, an example is shown in the Colaboratory notebook. The script only works for classification so far, and it is worth noting that some training parameters are still hard-coded although they shouldn’t really be (for example, the patience parameter for early stopping [how many steps do you continue while the best validation accuracy does not improve].)

The parameters that are varied in the optimization script are N_steps, feature_dim, batch-momentum, gamma, lambda-sparsity. (output_dim is set to be equal to feature_dim, as suggested in the optimization tips just below.)

The paper has the following tips on hyperparameter optimization:

Most datasets yield the best results for N_steps ∈ [3, 10]. Typically, larger datasets and more complex tasks require a larger N_steps. A very high value of N_steps may suffer from overfitting and yield poor generalization.

Adjustment of the values of Nd [feature_dim] and Na [output_dim] is the most efficient way of obtaining a trade-off between performance and complexity. Nd = Na is a reasonable choice for most datasets. A very high value of Nd and Na may suffer from overfitting and yield poor generalization.

An optimal choice of γ can have a major role on the overall performance. Typically a larger N_steps value favors for a larger γ.

A large batch size is beneficial for performance — if the memory constraints permit, as large as 1–10 % of the total training dataset size is suggested. The virtual batch size is typically much smaller than the batch size.

Initially large learning rate is important, which should be gradually decayed until convergence.

Results

I’ve tried TabNet via this command line interface for several datasets, including the Adult Census dataset that I used in the post about NODE and CatBoost for reasons that can be found in that post. Conveniently, this dataset had also been used in the TabNet manuscript, and the authors present the best parameter settings they found there. With repeated runs using those setting, I noticed that the best validation error (and test error) tends to be at around 86%, similar to CatBoost without hyperparameter tuning. The authors report a test set performance of 85.7% in the manuscript. When I did hyperparameter optimization with hyperopt, I unsurprisingly reached a similar performance around 86%, albeit with a different parameter setting.

For other datasets such as the Poker Hand dataset, TabNet is claimed to beat other methods by a considerable margin. I have not yet devoted much time to that, but everyone is of course invited to try TabNet with hyperparameter optimization on various datasets for themselves!

Conclusions

TabNet is an interesting architecture that seems promising for tabular data analysis. It operates directly on raw data and uses a sequential attention mechanism to perform explicit feature selection for each example. This property also gives it a sort of built-in interpretability.

I have tried to make TabNet slightly easier to work with by writing some wrapper code around it. The next step is to compare it to other methods across a wide range of datasets.

Please try it on your own datasets and/or send pull requests and help me improve the interface if you are interested!

 

Modelling tabular data with CatBoost and NODE

CatBoost from Yandex, a Russian online search company, is fast and easy to use, but recently researchers from the same company released a new neural network based package, NODE, that they claim outperforms CatBoost and all other gradient boosting methods. Can this be true? Let’s find out how to use both CatBoost and NODE!

Who is this blog post for?

Although I wrote this blog post for anyone who is interested in machine learning and in particular tabular data, it is helpful if you are familiar with Python and the scikit-learn library if you want to follow along with the code. If you aren’t, hopefully you will find the theoretical and conceptual parts interesting anyway!

CatBoost introduction

CatBoost is my go-to package for modelling tabular data. It is an implementation of gradient boosted decision trees with a few tweaks that make it slightly different from e.g. xgboost or LightGBM. It works for both classification and regression problems.

Some nice things about CatBoost:

• It handles categorical features (get it?) out of the box, so you don’t need to worry about how to encode them.
• It typically requires very little parameter tuning.
• It avoids certain subtle types of data leakage that other methods may suffer from. 
• It is fast, and can be run on GPU if you want it to go even faster.

These factors make CatBoost, for me, a no-brainer as the first thing to reach for when I need to analyze a new tabular dataset.

Technical details of CatBoost

Skip this section if you just want to use CatBoost!

On a more technical level, there are some interesting things about how CatBoost is implemented. I highly recommend the paper Catboost: unbiased boosting with categorical features if you are interested in the details. I just want to highlight two things.

1. In the paper, the authors show that standard gradient boosting algorithms are affected by subtle types of data leakage which result from the way that the models are iteratively fitted. In a similar manner, the most effective ways to encode categorical features numerically (like target encoding) are prone to data leakage and overfitting. To avoid this leakage, CatBoost introduces an artificial timeline according to which the training examples arrive, so that only “previously seen” examples can be used when calculating statistics.
2. CatBoost actually doesn’t use regular decision trees, but oblivious decision trees. These are trees where, at each level of the tree, the same feature and the same splitting criterion is used everywhere! This sounds weird, but has some nice properties. Let’s look at what is meant by this.

Regular decision tree. Any feature or split point can be present at each level.

Oblivious decision tree. Each level has the same splits.

Left: Regular decision tree. Any feature or split point can be present at each level. Right: Oblivious decision tree. Each level has the same splits.

In a normal decision tree, feature to split on and the cutoff value both depend on what path you have taken so far in the tree. This makes sense, because we can use the information we already have to decide the most informative next question (like in the “20 questions” game). With oblivious decision trees, the history doesn’t matter; we pose the same question no matter what. The trees are called “oblivious” because they keep “forgetting” what has happened before. 

Why is this useful? One nice property of oblivious decision trees is that an example can be classified or scored really quickly – it is always the same N binary questions that are posed (where N is the depth of the tree). This can easily be done in parallel for many examples. That is one reason why CatBoost is fast. Another thing to keep in mind is that we are dealing with a tree ensemble here. As a stand-alone algorithm, the oblivious decision tree might not work so well, but the idea of tree ensembles is that a coalition of weak learners often works well because errors and biases are “washed out”. Normally, the weak learner is a standard decision tree, and here it is something even weaker, namely the oblivious decision tree. The CatBoost authors argue that this particular weak base learner works well for generalization.

Installing CatBoost

Although installing CatBoost should be a simple matter of typing

pip install catboost

I’ve sometimes encountered problems with that when on a Mac. On Linux systems such as the Ubuntu system I am typing on now, or on Google Colaboratory, it should “just work”. If you keep having problems installing it, consider using a Docker image, e.g.

docker pull yandex/tutorial-catboost-clickhouse
docker run -it yandex/tutorial-catboost-clickhouse

Using CatBoost on a dataset

Link to Colab notebook with code

Let’s have a look at how to use CatBoost on a tabular dataset. We start by downloading a lightly preprocessed version of the Adult/Census Income  dataset which is, in the following, assumed to be located in datasets/adult.csv. I chose this dataset because it has a mix of categorical and numerical features, a nice manageable size in the tens of thousands of examples and not too many features. It is often used to exemplify algorithms, for instance in Google’s What-If Tool and many other places.  

The adult census dataset has the columns ‘age’, ‘workclass’, ‘education’, ‘education-num’, ‘marital-status’, ‘occupation’, ‘relationship’, ‘race’, ‘sex’, ‘capital-gain’, ‘capital-loss’, ‘hours-per-week’, ‘native-country’, and ‘<=50K‘. The task is to predict the value of the last column, ‘<=50K’, which indicates if the person in question earns 50,000 USD or less per year (the dataset is from 1994). We regard the following features as categorical rather than numerical: ‘workclass’, ‘education’, ‘marital-status’, ‘occupation’, ‘relationship’, ‘race’, ‘sex’, ‘native-country’.

The code is pretty similar to scikit-learn except for the Pool datatype that CatBoost uses to bundle feature and target values for a dataset while keeping them conceptually separate. (I have to admit I don’t really know why Pool is there – I just use it, and it seems to work fine.)

The code is available on Colab, but I will copy it here for reference. CatBoost needs to know which features are categorical and will then handle them automatically. In this code snippet, I also use 5-fold (stratified) cross-validation to estimate the prediction accuracy.

from catboost import CatBoostClassifier, Pool
from hyperopt import fmin, hp, tpe
import pandas as pd
from sklearn.model_selection import StratifiedKFold

df = pd.read_csv("https://docs.google.com/uc?" + 
                 "id=10eFO2rVlsQBUffn0b7UCAp28n0mkLCy7&" + 
                 "export=download")
labels = df.pop('<=50K')

categorical_names = ['workclass', 'education', 'marital-status',
                     'occupation', 'relationship', 'race',
                     'sex', 'native-country']  
categoricals = [df.columns.get_loc(i) for i in categorical_names]

nfolds = 5
skf = StratifiedKFold(n_splits=nfolds, shuffle=True)
acc = []

for train_index, test_index in skf.split(df, labels):
  X_train, X_test = df.iloc[train_index].copy(), \
                    df.iloc[test_index].copy()
  y_train, y_test = labels.iloc[train_index], \
                    labels.iloc[test_index]
  train_pool = Pool(X_train, y_train, cat_features = categoricals)
  test_pool = Pool(X_test, y_test, cat_features = categoricals)
  model = CatBoostClassifier(iterations=100,
                             depth=8,
                             learning_rate=1,
                             loss_function='MultiClass') 
  model.fit(train_pool)
  predictions = model.predict(test_pool)
  accuracy = sum(predictions.squeeze() == y_test) / len(predictions)
  acc.append(accuracy)

mean_acc = sum(acc) / nfolds
print(f'Mean accuracy based on {nfolds} folds: {mean_acc:.3f}')
print(acc)

What we tend to get from running this (CatBoost without hyperparameter optimization) is a mean accuracy between 85% and 86%. In my last run, I got about 85.7%.

If we want to try to optimize the hyperparameters, we can use hyperopt (if you don’t have it, install it with pip install hyperopt). In order to use it, you need to define a function that hyperopt tries to minimize. We will just try to optimize the accuracy here. Perhaps it would be better to optimize e.g. log loss, but that is left as an exercise to the reader 😉 

The main parameters to optimize are probably the number of iterations, the learning rate, and the tree depth. There are also many other parameters related to over-fitting, for instance early stopping rounds and so on. Feel free to explore on your own!

# Optimize between 10 and 1000 iterations and depth between 2 and 12

search_space = {'iterations': hp.quniform('iterations', 10, 1000, 10),
                'depth': hp.quniform('depth', 2, 12, 1),
                'lr': hp.uniform('lr', 0.01, 1)
               }

def opt_fn(search_space):

  nfolds = 5
  skf = StratifiedKFold(n_splits=nfolds, shuffle=True)
  acc = []

  for train_index, test_index in skf.split(df, labels):
    X_train, X_test = df.iloc[train_index].copy(), \
                      df.iloc[test_index].copy()
    y_train, y_test = labels.iloc[train_index], \
                      labels.iloc[test_index]
    train_pool = Pool(X_train, y_train, cat_features = categoricals)
    test_pool = Pool(X_test, y_test, cat_features = categoricals)

    model = CatBoostClassifier(iterations=search_space['iterations'],
                             depth=search_space['depth'],
                             learning_rate=search_space['lr'],
                             loss_function='MultiClass',
                             od_type='Iter')

    model.fit(train_pool, logging_level='Silent')
    predictions = model.predict(test_pool)
    accuracy = sum(predictions.squeeze() == y_test) / len(predictions)
    acc.append(accuracy)

  mean_acc = sum(acc) / nfolds
  return -1*mean_acc

best = fmin(fn=opt_fn, 
            space=search_space, 
            algo=tpe.suggest, 
            max_evals=100)

When I last ran this code, it took over 5 hours but resulted in a mean accuracy of 87.3%, which is on par with the best results I got when trying the Auger.ai AutoML platform.

Sanity check: logistic regression

At this point we should ask ourselves if these fancy new-fangled methods are really needed. How would a good old logistic regression perform out of the box and after hyperparameter optimization?

I’ll omit reproducing the code here for brevity’s sake, but it is available in the same Colab notebook as before. One detail with the logistic regression implementation is that it doesn’t handle categorical variables out of the box like CatBoost does, so I decided to code them using target encoding, specifically leave-one-out target encoding, which is the approach taken in NODE and a fairly close though not identical analogue of what happens in CatBoost.

Long story short, untuned logistic regression with this type of encoding yields around 80% accuracy, and around 81% (80.7% in my latest run) after hyperparameter tuning. Here, an interestin alternative is to try automated preprocessing libraries such as vtreat and Automunge, but I will save those for an upcoming blog post!

Taking stock

What do we have so far, before trying NODE?

• Logistic regression, untuned: 80.0%
• Logistic regression, tuned: 80.7%
• CatBoost, untuned: 85.7%
• CatBoost, tuned: 87.2%

 

NODE: Neural Oblivious Decision Ensembles

A recent manuscript from Yandex researchers describes an interesting neural network version of CatBoost, or at least a neural network take on oblivious decision tree ensembles (see the technical section above if you want to remind yourself what “oblivious” means here.) This architecture, called NODE, can be used for either classification or regression.

One of the claims from the abstract reads: “With an extensive experimental comparison to the leading GBDT packages on a large number of tabular datasets, we demonstrate the advantage of the proposed NODE architecture, which outperforms the competitors on most of the tasks.” This naturally piqued my interest. Could this tool be better than CatBoost?

How does NODE work?

You should go to the paper for the full story, but some relevant details are:

• The entmax activation function is used as a soft version of a split in a regular decision tree. As the paper puts it, “The entmax is capable to produce sparse probability distributions, where the majority of probabilities are exactly equal to 0. In this work, we argue that entmax is also an appropriate inductive bias in our model, which allows differentiable split decision construction in the internal tree nodes. Intuitively, entmax can learn splitting decisions based on a small subset of data features (up to one, as in classical decision trees), avoiding undesired influence from others.” The entmax functions allows a neural network to mimic a decision tree-type system while keeping the model differentiable (weights can be updated based on the gradients).
• The authors present a new type of layer, a “node layer”, which you can use in a neural network (their implementation is in PyTorch). A node layer represents a tree ensemble.
• Several node layers can be stacked, yielding a hierarchical model where the input is fed through one tree ensemble at a time. Successive concatenation of input representations can be used to give a model which is reminiscent of the popular DenseNet model for image processing, just specialized in tabular data.
• The parameters of a NODE model are:
  • Learning rate (always 0.001 in the paper)
  • The number of node layers (k)
  • The number of trees in each layer (m)
  • The depth of the trees in each layer (d)

 

How is NODE related to tree ensembles?

To get a feeling for how the analogy between this neural network architecture and decision tree ensembles looks, Figure 1 is reproduced here.

How should the parameters be chosen?

There is not much guidance in the manuscript; the authors suggest using hyperparameter optimization. They do mention that they optimize over the following space:

• num layers: {2, 4, 8} 
• total tree count: {1024, 2048} 
• tree depth: {6, 8} 
• tree output dim: {2, 3}

In my code, I don’t do grid search but rather let hyperopt sample values within certain ranges. The way I thought about it (which could be wrong) is that each layer represents a tree ensemble (a single instance of CatBoost, let’s say). For each layer that you add, you may add some representation power, but you also make the model much heavier to train and potentially risk overfitting. The total tree count seems roughly analogous to the number of trees in CatBoost/xgboost/random forests, and has the same tradeoffs: with many trees, you can express more complicated functions, but the model will take much longer to train and risk overfitting. The tree depth, again, has the same type of tradeoff. As for the output dimensionality, frankly, I don’t quite understand why it is a parameter. Reading the paper, it seems it should be equal to one for regression and equal to the number of classes for classification.

How does one use NODE?

The authors have made code available on GitHub. They do not provide a command-line interface but rather suggest that users run their models in the provided Jupyter notebooks. One classification example and one regression example is provided in those notebooks.

The repo README page also strongly suggests using a GPU to train NODE models. (This is a factor in favor of CatBoost.) 

I have prepared a Colaboratory notebook with some example code on how to run classification on NODE and how to optimize hyperparameters with hyperopt. 

Please move to the Colaboratory notebook right now to keep following along! 

Here I will just highlight some parts of the code.

General problems adapting the code

The problems I encountered when adapting the authors’ code were mainly related to data types. It’s important that the input datasets (X_train and X_val) are arrays (numpy or torch) in float32 format; not float64 or a mix of float and int. The labels need to be encoded as long (int64) for classification, and float32 for regression. (You can see this handled in the cell titled “Load, split and preprocess the data”.)

Other problems were related to memory. The models can quickly blow up the GPU memory, especially with the large batch sizes used in the authors’ example notebooks. I solved this simply by using the maximum batch size I could get away with on my laptop (and later, on Colab).

In general, though, it was not that hard to get the code to work. The documentation was a bit sparse, but sufficient.

 

Categorical variable handling

Unlike CatBoost, NODE does not support categorical variables, so you have to prepare those yourself into a numerical format. We do it for the Adult Census dataset in the same way the NODE authors do it, using LeaveOneOutEncoder from the category_encoders library. Here we just use a regular train/test split instead of 5-fold CV out of convenience, as it takes a long time to train NODE (especially with hyperparameter optimization).

from category_encoders import LeaveOneOutEncoder
import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split

df = pd.read_csv('https://docs.google.com/uc' + 
                 '?id=10eFO2rVlsQBUffn0b7UCAp28n0mkLCy7&' + 
                 'export=download')
labels = df.pop('<=50K')
X_train, X_val, y_train, y_val = train_test_split(df,
                                                  labels,
                                                  test_size=0.2)

class_to_int = {c: i for i, c in enumerate(y_train.unique())}                                                                                                               
y_train_int = [class_to_int[v] for v in y_train]                                                                                                                            
y_val_int = [class_to_int[v] for v in y_val] 

cat_features = ['workclass', 'education', 'marital-status',
                'occupation', 'relationship', 'race', 'sex',
                'native-country']
  
cat_encoder = LeaveOneOutEncoder()
cat_encoder.fit(X_train[cat_features], y_train_int)
X_train[cat_features] = cat_encoder.transform(X_train[cat_features])
X_val[cat_features] = cat_encoder.transform(X_val[cat_features])

# Node is going to want to have the values as float32 at some points
X_train = X_train.values.astype('float32')
X_val = X_val.values.astype('float32')
y_train = np.array(y_train_int)
y_val = np.array(y_val_int)

Now we have a fully numeric dataset. 

Model definition and training loop

The rest of the code is essentially the same as in the authors’ repo (except for the hyperopt part). They created a Pytorch layer called DenseBlock, which implements the NODE architecture. A class called Trainer holds information about the experiment, and there is a straightforward training loop that keeps track of the best metrics seen so far and plots updated loss curves.

Results & conclusions

With some minimal trial and error, I was able to find a model with around 86% validation accuracy. After hyperparameter optimization with hyperopt (which was supposed to run overnight on a GPU in Colab, but in fact timed out after about 40 iterations), the best performance was 87.2%. In other runs I have achieved 87.4%. In other words, NODE did outperform CatBoost, albeit slightly, after hyperopt tuning.

However, accuracy is not everything. It is not convenient to have to do costly optimization for every dataset. 

Pros of NODE vs CatBoost:

• It seems that slightly better results can be obtained (based on the NODE paper and this test; I will be sure to try many other datasets!)

Pros of CatBoost vs NODE:

• Much faster
• Less need of hyperparameter optimization
• Runs fine without GPU
• Has support for categorical variables

Which one would I use for my next projects? Probably CatBoost will still be my go-to tool, but I will keep NODE in mind and maybe try it just in case…

It’s also important to realize that performance is dataset-dependent and that the Adult Census Income dataset is not representative of all scenarios. Perhaps more importantly, the preprocessing of categorical features is likely rather important in NODE. I’ll return to the subject of preprocessing in a future post!