🦖 Rax: Composable Learning to Rank using JAX

Rax is a Learning-to-Rank library written in JAX. Rax provides off-the-shelf implementations of ranking losses and metrics to be used with JAX. It provides the following functionality:

• Ranking losses (rax.*_loss): rax.softmax_loss, rax.pairwise_logistic_loss, ...
• Ranking metrics (rax.*_metric): rax.mrr_metric, rax.ndcg_metric, ...
• Transformations (rax.*_t12n): rax.approx_t12n, rax.gumbel_t12n, ...

Ranking

A ranking problem is different from traditional classification/regression problems in that its objective is to optimize for the correctness of the relative order of a list of examples (e.g., documents) for a given context (e.g., a query). Rax provides support for ranking problems within the JAX ecosystem. It can be used in, but is not limited to, the following applications:

• Search: ranking a list of documents with respect to a query.
• Recommendation: ranking a list of items given a user as context.
• Question Answering: finding the best answer from a list of candidates.
• Dialogue System: finding the best response from a list of responses.

Synopsis

In a nutshell, given the scores and labels for a list of items, Rax can compute various ranking losses and metrics:

import jax.numpy as jnp
import rax

scores = jnp.asarray([2.2, -1.3, 5.4])  # output of a model.
labels = jnp.asarray([1., 0., 0.])      # indicates doc 1 is relevant.

rax.ndcg_metric(scores, labels)         # computes a ranking metric.
rax.pairwise_hinge_loss(scores, labels) # computes a ranking loss.

All of the Rax losses and metrics are purely functional and compose well with standard JAX transformations. Additionally, Rax provides ranking-specific transformations so you can build new ranking losses. An example is rax.approx_t12n, which can be used to transform any (non-differentiable) ranking metric into a differentiable loss. For example:

loss_fn = rax.approx_t12n(rax.ndcg_metric)
loss_fn(scores, labels)            # differentiable approx ndcg loss.
jax.grad(loss_fn)(scores, labels)  # computes gradients w.r.t. scores.

Examples

See the examples/ directory for complete examples on how to use Rax.