Training examples: 135 (Test samples: 15)
Features: 4 (length and width of sepals and petals)
Classes: 3 (Setosa, Versicolor, and Virginica)

There are substantial overlap with the Versicolor and Virginica classes.
(Click the red Forward arrow to advance to the next Step).

Training examples: 1697 (Test samples: 100)
Features: 64 (8x8 image of a digit)
Classes: 10 (All digits [0..9])

The 8x8 representation of each digit. The poor resolution is a challenge for this dataset.
(Click the red Forward arrow to advance to the next Step).

The formal objective is: Given the Iris dataset with 135 training examples $B = (T_1, T_2, ..., T_{135})$, 1-NN underlying algorithm, and 15 test samples, we need to predict the label for these samples.

Note that Conformal Prediction can work with any algorithm, and will still produce valid predictions.

The formal objective is: Given the Iris dataset with 140 training examples $B = (T_1, T_2, ..., T_{140})$, SVM underlying algorithm, and 10 test samples, we need to predict the label for these samples.

Note that Conformal Prediction can work with any algorithm, and will still produce valid predictions.

The formal objective is: Given the Digits dataset with 1697 training examples $B = (T_1, T_2, ..., T_{1697})$, 1-NN underlying algorithm, and 100 test samples, we need to predict the label for these samples.

Note that Conformal Prediction can work with any algorithm, and will still produce valid predictions.

The formal objective is: Given the Digits dataset with 1697 training examples $B = (T_1, T_2, ..., T_{1697})$, SVM underlying algorithm, and 100 test samples, we need to predict the label for these samples.

Note that Conformal Prediction can work with any algorithm, and will still produce valid predictions.

With 1-Nearest Neighbour, one way to calculate $\alpha_i$ is using the Euclidean distance of the object value from the training example $i$ to the nearest training example that also has the same label $y$, as follows. $$\alpha_i = min\{|x_j - x_i| : 1 \leq j \leq N+1 \phantom{2}\& \phantom{2} j \ne i \phantom{2}\& \phantom{2} y_i = y_i\}$$ Each test sample $S$ will be given a potential label (i.e. Setosa, Versicolor, or Virginica). Then, the sample is put into the training set, as a new example $T_{N+1}$. Our new dataset is: $B = (T_1, T_2, ..., T_N, T_{N+1})$
For each potential label, we calculate $\alpha$ for all training examples, including the new sample.

With 1-Nearest Neighbour, one way to calculate $\alpha_i$ is using the Euclidean distance of the object value from the training example $i$ to the nearest training example that also has the same label $y$, as follows. $$\alpha_i = min\{|x_j - x_i| : 1 \leq j \leq N+1 \phantom{2}\& \phantom{2} j \ne i \phantom{2}\& \phantom{2} y_i = y_i\}$$ Each test sample $S$ will be given a potential label (i.e. $[0,9]$). Then, the sample is put into the training set, as a new example $T_{N+1}$. Our new dataset is: $B = (T_1, T_2, ..., T_N, T_{N+1})$
For each potential label, we calculate $\alpha$ for all training examples, including the new sample.