февраля 10, 2020 —
Posted by Mahdi Milani Fard, Software Engineer, Google Research
Most ML practitioners have encountered the typical scenario where the training data looks very different from the run-time queries on which the model is evaluated. As a result, flexible ML solutions such as DNNs or forests that rely solely on the training dataset often act unexpectedly and even wildly in parts of the input space not …
Most ML practitioners have encountered the typical scenario where the training data looks very different from the run-time queries on which the model is evaluated. As a result, flexible ML solutions such as DNNs or forests that rely solely on the training dataset often act unexpectedly and even wildly in parts of the input space not covered by the training and validation datasets. This behaviour is especially problematic in cases where important policy or fairness constraints can be violated.
Unconstrained models can behave unexpectedly where there is little training data coverage. Here, DNN and GBT predictions are far from the ground truth on the testing data.
Even though common forms of regularization can result in more sensible extrapolation, standard regularizers cannot guarantee reasonable model behaviour across the entire input space, especially with high-dimensional inputs. Switching to simpler models with more controlled and predictable behaviour can come at a severe cost to the model accuracy.
TF Lattice makes it possible to keep using flexible models, but provides several options to inject domain knowledge into the learning process through semantically meaningful common-sense or policy-driven shape constraints. For example, you can specify that the model output should be monotonically increasing with respect to a given input. These extra pieces of domain knowledge can help the model learn beyond just the training dataset and makes it behave in a manner controlled and expected by the user.
TensorFlow Lattice Library
TensorFlow Lattice is a library for training constrained and interpretable lattice based models. A lattice is an interpolated look-up table that can approximate arbitrary input-output relationships in your data.
The simple example above is a lattice function over 2 input features and with 4 parameters, which are the function's values at the corners of the input space; the rest of the function is interpolated from these parameters. You can use higher dimensional lattices and a finer grained grid of parameters to get a more flexible function. The library implements the lattice with the tfl.layers.Lattice Keras layer.
TensorFlow Lattice also provides piecewise linear functions (with tfl.layers.PWLCalibration Keras layer) to calibrate and normalize the input features to the range accepted by the lattice: 0 to 1 in the example lattice above.
For categorical features, TensorFlow Lattice provides categorical calibration (with tfl.layers.CategoricalCalibration Keras layer) with similar output bounding to feed into a lattice. Combining the calibrators and the lattice above, we can get a calibrated lattice model.
There are several forms of constraints you can impose on TensorFlow Lattice layers to inject your knowledge of the problem domain into the training process:
Monotonicity: You can specify that the output should only increase/decrease with respect to an input. In our example, you may want to specify that increased distance to a coffee shop should only decrease the predicted user preference.
Convexity/Concavity: You can specify that the function shape can be convex or concave. Mixed with monotonicity, this can force the function to represent diminishing returns with respect to a given feature.
Unimodality: You can specify that the function should have a unique peak or unique valley. This lets you represent functions that are expected to have a sweet spot with respect to a feature.
Pairwise trust: This constraint suggests that one input feature semantically reflects trust in another feature. For example, a higher number of reviews makes you more confident in the average star rating of a restaurant. The model will be more sensitive with respect to the star rating (i.e. will have a larger slope with respect to the rating) when the number of reviews is higher.
Pairwise dominance: This constraint suggests that the model should treat one feature as more important than another feature. This is done by making sure the slope of the function is larger with respect to the dominant feature.
In addition to shape constraints, TensorFlow lattice provides a number of regularizers to control the flexibility and smoothness of the function on a per-feature basis. These include Laplacian Regularizer (for a flatter function), Hessian Regularizer (for a more linear calibration function), Wrinkle Regularizer (for a smoother calibration function) and Torsion Regularizer (for a more co-linear lattice function).
Example: Ranking Restaurants
This example is from our end-to-end shape constraint tutorial that covers many of the above mentioned constraints with canned estimators. Imagine a scenario where we want to determine whether or not users will click on a restaurant search result. The task is to predict the clickthrough rate (CTR) given input features:
average rating: a numeric feature in the range 1 to 5
number of reviews: a numeric feature in range 0 to 200
sollar rating: a categorical feature with values “$” to “$$$$” represented as 0 to 3 and missing value represented as -1
We have these as the domain knowledge to confine and control our model’s behavior:
Output is monotonically increasing in average rating
Output is monotonically increasing in number of reviews, but with diminishing returns
The model should trust the average rating more when there are more reviews
Users typically prefer “$$” restaurants to “$” restaurants
We can construct a calibrated lattice model using the Keras layers provided by the library:
model = tf.keras.models.Sequential()
model.add(
tfl.layers.ParallelCombination([
# Feature: average rating
tfl.layers.PWLCalibration(
# Input keypoints for the piecewise linear function
input_keypoints=np.linspace(1., 5., num=20),
# Output is monotonically increasing in this feature
monotonicity='increasing',
# This layer is feeding into a lattice with 2 vertices
output_min=0.0,
output_max=1.0),
# Feature: number of reviews
tfl.layers.PWLCalibration(
input_keypoints=np.linspace(0., 200., num=20),
# Output is monotonically increasing in this feature
monotonicity='increasing',
# There is diminishing returns on the number of reviews
convexity='concave',
# Regularizers defined as a tuple ('name', l1, l2)
kernel_regularizer=('wrinkle', 0.0, 1.0),
# This layer is feeding into a lattice with 3 vertices
output_min=0.0,
output_max=2.0),
# Feature: dollar rating
tfl.layers.CategoricalCalibration(
# 4 rating categories + 1 missing category
num_buckets=5,
default_input_value=-1,
# Partial monotonicity: calib(0) <= calib(1)
monotonicities=[(0, 1)],
# This layer is feeding into a lattice with 2 vertices
output_min=0.0,
output_max=1.0),
]))
model.add(
tfl.layers.Lattice(
# A 2x3x2 grid lattice
lattice_size=[2, 3, 2],
# Output is monotonic in all inputs
monotonicities=['increasing', 'increasing', 'increasing']
# Trust: more responsive to input 0 if input 1 increases
edgeworth_trusts=(0, 1, 'positive')))
model.compile(...)
The resulting trained model satisfies all the specified constraints, and the added regularization makes the function smooth:
The above model can also be constructed using canned estimators provided by the library. Check out in our shape constraints tutorial for more details in an end-to-end colab describing the effect of each of the described constraints. TF Lattice Keras layers can also be used in combination with other Keras layers to construct partially constrained or regularized models. For example, lattice or PWL calibration layers can be used at the last layer of deeper networks that include embeddings or other Keras layers. For further information check out the Tensorflow Lattice website. There are many guides and tutorials available to get you started: shape constraints, canned estimators, custom estimators, and Keras layers. Also check out our presentation at the TF Dev Summit:
Feedback
We are looking forward to hearing your thoughts and comments on the library. For bugs or issues, please reach out to us on Github.
Acknowledgements
This release was made possible with contributions from Oleksandr Mangylov, Mahdi Milani Fard, Taman Narayan, Yichen Zhou, Nobu Morioka, William Bakst, Harikrishna Narasimhan, Andrew Cotter and Maya Gupta.
Publications
For further details on the models and algorithms used within the library, check out our publications on lattice models:
Shape Constraints for Set Functions, Andrew Cotter, Maya Gupta, H. Jiang, Erez Louidor, Jim Muller, Taman Narayan, Serena Wang, Tao Zhu. International Conference on Machine Learning (ICML), 2019
Monotonic Calibrated Interpolated Look-Up Tables, Maya Gupta, Andrew Cotter, Jan Pfeifer, Konstantin Voevodski, Kevin Canini, Alexander Mangylov, Wojciech Moczydlowski, Alexander van Esbroeck, Journal of Machine Learning Research (JMLR), 2016
Lattice Regression, Eric Garcia, Maya Gupta, Advances in Neural Information Processing Systems (NeurIPS), 2009
Next post
TensorFlow Core·
TensorFlow Lattice: Flexible, controlled and interpretable ML
февраля 10, 2020
—
Posted by Mahdi Milani Fard, Software Engineer, Google Research
Most ML practitioners have encountered the typical scenario where the training data looks very different from the run-time queries on which the model is evaluated. As a result, flexible ML solutions such as DNNs or forests that rely solely on the training dataset often act unexpectedly and even wildly in parts of the input space not …
