Classification rules based on distribution functions of functional depth (Q2423182)

Given a data cloud \(\{x\}\), the data depth \(D(x)\) measures how central or deep an observation \(x\) is with respect to \(\{x\}\), i.e. \(D(x)\) has large values for central \(x\) and small values for extreme \(x\). In the article, the notion of data depth is considered as applied to functional objects, e.g., continuous curves. Let \(C(I)\) be a set of continuous functions on a compact interval \(I\) and let \(X=X(t)\) \((t\in I)\) be a random function drawn from \(C(I)\). Given some depth function \(D(X)\), define the distribution function \[ F_D(r)=\Pr(D(X) \le r).\] One can see that large values \(F_D(D(x))\) imply that \(x\) is a central observation and small values \(F_D(D(x))\) imply that \(x\) is an extreme observation from \(\{X\}\). Let observations \(\{X\}\) belong to \(J \ge 2\) classes with prior probabilities \(p_j\) and depth functions \(D_j(X)\), \(j=1,\dots,J\). A classification rule prescribes to assign \(z\) to \(\{X\}_j\) if \(F_{D_j}(D_j(z))=\max_i F_{D_i}(D_i(z))\). The probability of the misclassification is equal to \[\sum_{j=1}^J p_j \Pr\{\text{erroneous classification of } z \in \{X\}_j\}.\] In the article, the learning procedure is proposed if depth functions are a priori unknown and only sample is available, its asymptotic optimality is proved. Simulation results are presented and compared with the results from existing methods.