Theoretical Description¶
Two methods for multi-label uncertainty-quantification have been implemented in MAPIE so far : Risk-Controlling Prediction Sets (RCPS) [1] and Conformal Risk Control (CRC) [2]. The difference between these methods is the way the conformity scores are computed.
For a multi-label classification problem in a standard independent and identically distributed (i.i.d) case, our training data has an unknown distribution .
For any risk level between 0 and 1, the methods implemented in MAPIE allow the user to construct a prediction set for a new observation with a guarantee on the recall. RCPS and CRC give two slightly different guarantees:
RCPS:
CRC:
1. Risk-Controlling Prediction Sets¶
1.1. General settings¶
Let’s first give the settings and the notations of the method:
Let be a set-valued function (a tolerance region) that maps a feature vector to a set-valued prediction. This function is built from the model which was previously fitted on the training data. It is indexed by a one-dimensional parameter which is taking values in such that:
Let be a loss function on a prediction set with the following nesting property:
Let be the risk associated to a set-valued predictor:
The goal of the method is to compute an Upper Confidence Bound (UCB) of and then to find as follows:
The figure bellow explains this procedure:
Following those settings, the RCPS method gives the following guarantee on the recall:
1.2. Bounds calculation¶
In this section, we will consider only bounded losses (as for now, only the loss is implemented. We will show three different Upper Calibration Bounds (UCB) (Hoeffding, Bernstein and Waudby-Smith–Ramdas) of based on the empirical risk which is defined as follows:
1.2.1. Hoeffding Bound¶
Suppose the loss is bounded above by one, then we have by the Hoeffding inequality that:
Which implies the following UCB:
1.2.2. Bernstein Bound¶
Contrary to the Hoeffding bound, which can sometimes be too simple, the Bernstein UCB is taking into account the variance and gives smaller prediction set size:
Where:
1.2.3. Waudby-Smith–Ramdas¶
This last UCB is the one recommended by the authors of [1] to use when using a bounded loss as this is the one which gives the smallest prediction sets size while having the same risk guarantees. This UCB is defined as follows:
Let and
Further let:
Then:
2. Conformal Risk Control¶
The goal of this method is to control any monotone and bounded loss. The result of this method can be expressed as follows:
Where
In the case of multi-label classification,
To find the optimal value of , the following algorithm is applied:
With :
3. References¶
[1] Lihua Lei Jitendra Malik Stephen Bates, Anastasios Angelopoulos and Michael I. Jordan. Distribution-free, risk-controlling prediction sets. CoRR, abs/2101.02703, 2021. URL https://arxiv.org/abs/2101.02703.39
[2] Angelopoulos, Anastasios N., Stephen, Bates, Adam, Fisch, Lihua, Lei, and Tal, Schuster. “Conformal Risk Control.” (2022).