This repository contains the official implementation of our TMLR paper "TabCBM: Concept-based Interpretable Neural Networks for Tabular Data" and its corresponding version in ICML's Workshop on Interpretable Machine Learning for Healthcare (IMLH 2023).

This work was done by Mateo Espinosa Zarlenga, Zohreh Shams, Michael Edward Nelson, Been Kim, and Mateja Jamnik

There has been significant efforts in recent years on designing neural architectures that can explain their predictions using high-level units of information referred to as "concepts". Nevertheless, these methods have thus far never been deployed or designed to be applicable for tabular tasks, leaving crucial domains such as those in healthcare and genomics out of the scope of concept-based interpretable models. In this work, we first provide the a novel definition of what a concept entails in a general tabular domain and then propose Tabular Concept Bottleneck Models (TabCBMs), a family of interpretable self-explaining neural architectures capable of discovering high-level concept explanations for tabular tasks without sacrificing state-of-the-art performance.

Concept-based interpretability addresses the opacity of deep neural networks by constructing an explanation for a model's prediction using high-level units of information referred to as concepts. Research in this area, however, has been mainly focused on image and graph-structured data, leaving high-stakes tasks whose data is tabular out of reach of existing methods. In this paper, we address this gap by introducing the first definition of what a high-level concept may entail in tabular data. We use this definition to propose Tabular Concept Bottleneck Models (TabCBMs), a family of interpretable self-explaining neural architectures capable of learning high-level concept explanations for tabular tasks. As our method produces concept-based explanations both when partial concept supervision or no concept supervision is available at training time, it is adaptable to settings where concept annotations are missing. We evaluate our method in both synthetic and real-world tabular tasks and show that TabCBM outperforms or performs competitively compared to state-of-the-art methods, while providing a high level of interpretability as measured by its ability to discover known high-level concepts. Finally, we show that TabCBM can discover important high-level concepts in synthetic datasets inspired by critical tabular tasks (e.g., single-cell RNAseq) and allows for human-in-the-loop concept interventions in which an expert can identify and correct mispredicted concepts to boost the model's performance.

You can locally install this package by first cloning this repository:

$ git clone https://github.com/mateoespinosa/tabcbm

We provide an automatic mechanism for this installation using Python's setup process with our standalone setup.py. To install our package, therefore, you only need to move into the cloned directory (cd tabcbm) and run:

$ python setup.py install

After running this, you should by able to import our package locally using

import tabcbm

In this repository, we include a standalone TensorFlow implementation of Tabular Concept Bottleneck Models (TabCBMs) which can be easily trained from scratch given a set of samples that may or may not have binary concept annotations.

In order to use our model's implementation, you first need to install all our code's requirements (listed in requirements.txt) or by following the installation instructions above.

After you have installed all dependencies, you should be able to import TabCBM as a standalone keras Model as follows:

from tabcbm.models.tabcbm import TabCBM

#####
# Define your pytorch dataset objects
#####

x_train = ...  # Numpy np.ndarray with shape (batch, features) containing samples
y_train = ...   # Numpy np.ndarray with shape (batch) containing integer labels

#####
# Construct the model's hyperparameters
#####
n_concepts = ...  # Number of concepts we wish to discover in the task

tab_cbm_params = dict(
    features_to_concepts_model=..., # Provide Keras model to be used for the feature to latent code model (e.g., $\phi$)
    concepts_to_labels_model=..., # Provide Keras model to be used for the concept-scores-to-label model (e.g., f)
    loss_fn=..., # An appropiate loss function following tensorflow's loss function APIs
    latent_dims=32, # Size of latent space for concept embeddings
    n_concepts=n_concepts,  # Number of concepts we wish to discover in the task
    n_supervised_concepts=0, # Change to another number if concept supervision is expected (i.e., we have a c_train matrix)

    # Loss hypers
    coherence_reg_weight=..., # Scalar loss weight for the coherence regularizer
    diversity_reg_weight=..., # Scalar loss weight for the diversity regularizer
    feature_selection_reg_weight=..., # Scalar loss weight for the specificity regularizer
    concept_prediction_weight=..., # Scalar loss weight hyper for the concept predictive loss (only relevant if n_supervised_concepts != 0)
)

#####
# Perform its self-supervised pre-training first
#####

ss_tabcbm = TabCBM(
    self_supervised_mode=True,
    **tab_cbm_params,
)
ss_tabcbm.compile(optimizer=tf.keras.optimizers.Adam(1e-3))
ss_tabcbm._compute_self_supervised_loss(x_test[:2, :])
ss_tabcbm.set_weights(ss_tabcbm.get_weights())

ss_tabcbm.fit(
    x=x_train,
    y=y_train,
    validation_split=0.2,
    epochs=100,
    batch_size=256
)

#####
# And now do its supervised training stage
#####

tabcbm = TabCBM(
    self_supervised_mode=False,
    concept_generators=ss_tabcbm.concept_generators,
    prior_masks=ss_tabcbm.feature_probabilities,
    **tab_cbm_params,
)
tabcbm.compile(optimizer=optimizer_gen())
tabcbm._compute_supervised_loss(
    x_test[:2, :],
    y_test[:2],
    c_true=(
        c_train_real[:2, :]
        if c_train_real is not None else None
    ),
)
tabcbm.fit(
    x=x_train,
    y=y_train,
    validation_split=0.2,
    epochs=100,
    batch_size=256
)

For a step-by-step example showing how to generate a dataset and configure a TabCBM for training on your own custom dataset, please see our Synth-Nonlin example notebook.

Further documentation on this model's parameters, as well as test scripts showing examples of usage, will be incoorporated with the aforementioned refactor that is currently going on. Until then, feel free to raise an issue or reach out if you want specific details on a parameter of TabCBM.

Our TabCBM module takes the following initialization key arguments:

• features_to_concepts_model (tf.keras.Model): A tensorflow model mapping input features with shape $(B, ...)$ to a set of latent codes with shape $(B, m)$. This is the latent code encoder $\phi$ used in Figure 1 of our paper.
• concepts_to_labels_model (tf.keras.Model): A tensorflow model mapping a set of concept scores in $[0, 1]$ with shape $(B, k')$ to a set of output probabilities for each of the task classes (i.e., a tensor with shape $(B, L)$). This is the label predictor $f$ used in Figure 1 of our paper.
• latent_dims (int): The dimensionality $m$ of the latent code and the concept embeddings.
• n_concepts (int): Number of total concepts to use for this model. This number must include both supervised concepts (if any) and unsupervised/discovered concepts.
• masking_values (np.ndarray or None): The values to use when masking each of the input features. This should be an array with $n$ elments in it, one for each input feauture. If None, as it is defaulted, we use an array with all zeros (i.e., we will mask all samples using zero masks).
• features_to_embeddings_model (tf.keras.Model or None): An optional tensorflow model used to preprocess the input features before passing them to the feature_to_concepts_model. This argument can be used to incoorporate learnable embedding-based models when working with categorical features where we would like to learn embeddings for each of the categories in each discrete feature. If not provided (i.e., set to None) then we assume no input preprocessing is needed. If provided, then it is expected that the effective input shape of features_to_concepts_model is the effective output shape of features_to_embeddings_model.
• cov_mat (np.ndarray or None): Empirical $(n \times n)$ covariance matrix for the input training features. This covariance is used for learning correlated gate maskings that take into account cross-feature correlations to avoid leakage when a feature is masked (as in SEFS). If not provided (i.e., set to None) then we will assume that all features are independent of each other (i.e., the covariance matrix is the identity matrix).

The different components of TabCBM's loss can be configured through the following arguments:

• loss_fn (Callable[tf.Tensor, tf.Tensor]): A loss function to use for the downstream task. This is a differientable TF function that takes the true labels y_true and the predicted labels y_pred for a batch of B inputs, and returns a vector of size (B) describing the loss for each sample. Defaults to TF's unreduced categorical cross entropy.
• coherence_reg_weight (float): Weight for the coherence regulariser (called $\lambda_\text{co}$ in the paper). Defaults to 0.1.
• diversity_reg_weight (float): Weight for the diversity regulariser (called $\lambda_\text{div}$ in the paper). Defaults to 5.
• feature_selection_reg_weight (float): Weight for the specificity regulariser (called $\lambda_\text{spect}$ in the paper). Defaults to 5.
• top_k (int): Number of k-nearest neighbors to use when computing the coherency loss. This argument is important to fine tune and must be less than the batch size. Defaults to 32.

If some ground-truth concepts labels are provided during training, then this can be indicated through the following arguments:

• n_supervised_concepts (int): Number of concepts that will be provided supervision for. If non-zero, then we expect, for each sample, to be provided with a vector of n_supervised_concepts with binary concept annotations. The value of n_supervised_concepts should be less than n_concepts. Defaults to 0 (i.e., no ground-truth concepts provided).
• concept_prediction_weight (int): When provided with ground-truth concepts during training, this value specifies the weight of the concept prediction loss used during training for the supervised concepts. Defaults to 0 (i.e., no ground-truth concepts provided).

A quick note that if ground truth concepts are provided during training, then the first n_supervised_concepts concept scores will correspond to the provided concepts in the same order they are given in the training concept label vector. Moreover, our TabCBM implementation supports partial concept annotations (i.e., some samples may have some concepts annotatated and some may not). This can be done by setting unknown concept labels as NaNs in the corresponding training samples.

Our TabCBM's class also provides arguments to aid with the end-to-end incorporation of the Self-supervised pipeline as part of its training (as shown in the example above). These arguments are:

• self_supervised_mode (bool): Whether or not this model's mask generator modules have been pretrained. If True, then it will use the SEFS pre-text self-supervised task to pretrain these modules when one calls the .fit(...) function. Otherwise, if False, it assumes mask generators have already been pre-trained and .fit(...) will proceed directly to end-to-end training of the entire TabCBM model. See TabCBM notebook example to see how to use this parameter for pretraining. Defaults to False.
• g_model (tf.keras.Model or None): Model to be used for reconstructing the input features from the learnt latent code during self-supervised pretraining. If not provided (i.e., set to None) then it defaults to a simple 3-layer ReLU MLP model with a hidden layer with 500 activation in it. Notice that this model is only relevant during self-supervised pretraining but it is irrelevant during end-to-end TabCBM training and any subsequent inferences.
• gate_estimator_weight: (float): Weight to be used self-supervised mask generator pretraining for the regularizer penalizing the model for not correctly predicting the mask applied to the sample. Defaults to 1.
• include_bn (bool): Whether or not we include a learnable batch normalization layer that preprocesses the input features before any concept embeddings/scores. Defaults to False.
• rec_model_units (List[int]): The size of the layers for the MLP used for the the reconstruction model during the self-supervised pre-training stagte. Defaults to [64].

Finally, we provide some control over the architecture used for the concept generators via the following arguments:

• concept_generator_units (List[int]): The size of the layers for the MLP used for the concept generator models (i.e., $\rho^{(i)}$ models.). Defaults to [64].
• concept_generators (list[tf.keras.Model]): A list of n_concepts TF model to be used as concept generators $\rho$. If not provided (i.e., set to None), then will instantiate each concept generator using a ReLU MLP with layer sizes concept_generator_units.
• prior_masks (np.ndarray or None): Initial values for TabCBM's masks logit probabilities. If not provided (i.e., set to None), then we will randomly initialize every mask's logit probability to a value uniformly at random from $[-1, 1]$.

To reproduce the experiments discussed in our paper, please use our run_experiments.py script in experiments/ after installing the package as indicated above. You should then be able to run all experiments by running this script with the appropiate config from the experiments/configs/ directory. For example, to run our experiments on the Synth-Linear dataset (see our paper), you can execute the following command:

$ python experiments/run_experiments.py dot -c experiments/configs/linear_tab_synth_config.yaml -o results/synth_linear

This should generate a summary of all the results after execution has terminated and dump all results/trained models/logs into the given output directory (results/synth_linear/ in this case).