Work presented at CPRML 2015 for model selection using efficiency of conformal predictors. Work done in collaboration with Ritvik Jaiswal and Dr. Vineeth Balasubramanian from IIT Hyderabad
2. PROPOSED APPROACH
• Efficiency - means of selecting model parameters in classifiers
• Narrow conformal prediction regions desirable
• Model selection posed as an optimisation problem
• Objective to optimise the value of the S-criterion of efficiency
• k-Nearest Neighbour (k-NN) classifier used for validating
approach
3. PROPOSED METHODOLOGY
• k-Nearest Neighbour (k-NN) classifier used for validating idea
• Objective: Minimise S-criterion of efficiency
• The k which minimises the objective function in general gives high accuracy and
efficiency.
• S-criterion of efficiency is defined as:
• where are the p-values defined as follows:
4. PROPOSED METHODOLOGY
• Smaller values preferable for S-criterion
• Ensures smaller size for the prediction set →
• Intuition: For an incoming test point, we want most of the training points to have a
higher conformity score than test points
• Small values for the expression are desirable
• and are conformity scores for test and training points respectively
5. PROPOSED METHODOLOGY
• This gives a proxy for the S-criterion which is to be minimized
where n and m are the number of test and training points respectively.
• The conformity score for an incoming point for the k-NN classifier is defined as:
6. PROPOSED METHODOLOGY
• This leads us to the objective function:
where n, m and k are the number of test points, training points and the
number of nearest neighbours respectively.
8. EMPIRICAL STUDY → USPS DATASET
k vs OBJECTIVE FUNCTION
ρ = 0.8589
k vs ACCURACY
ρ = -0.9271
9. EMPIRICAL STUDY → USPS DATASET
(EFFICIENCY RESULTS)
k vs OBJECTIVE FUNCTION k vs PREDICTION SET SIZE
(80% CONFIDENCE)
10. EMPIRICAL STUDY → STANDARD WAVEFORM DATASET
k vs OBJECTIVE FUNCTION
ρ = -0.6501
k vs ACCURACY
ρ = 0.7989
11. EMPIRICAL STUDY → STANDARD WAVEFORM DATASET
(EFFICIENCY RESULTS)
k vs OBJECTIVE FUNCTION k vs PREDICTION SET SIZE
(80% CONFIDENCE)
12. CONCLUSIONS AND FUTURE/ONGOING WORK
• While validity is guaranteed, efficiency varies with classifier parameters
• Proposed approach shows promise – a baby step
– k vs Objective function (Validation Set) and k vs Accuracy (Test Set) are
negatively correlated
– As value of the objective function decreases, efficiency increases (expectedly)
• Future/Ongoing work
– What would other measures of efficiency lead to?
– Can we frame this as a convex/submodular/other objective function with
guaranteed performance bounds?