Minimax Classification under Concept Drift with Multidimensional Adaptation and Performance Guarantees
Authors: Verónica Álvarez, Santiago Mazuelas, Jose A Lozano
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on multiple benchmark datasets show the classification improvement of AMRCs compared to the stateof-the-art and the reliability of the presented performance guarantees. |
| Researcher Affiliation | Academia | 1BCAM-Basque Center for Applied Mathematics, Bilbao, Spain 2IKERBASQUE-Basque Foundation for Science 3Intelligent Systems Group, University of the Basque Country UPV/EHU, San Sebasti an, Spain. Correspondence to: Ver onica Alvarez <valvarez@bcamath.org>, Santiago Mazuelas <smazuelas@bcamath.org>, Jose A. Lozano <jlozano@bcamath.org>. |
| Pseudocode | Yes | Algorithm 1 Learning AMRCs Input: (xt 1, yt 1), bτ t 1, λt 1, and µt 1 Output: bτ t, λt, µt, and R(Ut) Update bτ t and λt using (xt 1, yt 1), bτ t 1, and λt 1 (see Alg. 3) Update µt and obtain R(Ut) solving (7) using bτ t, λt, and µt 1 (see Alg. 4) Algorithm 2 Prediction with AMRCs Input: xt and µt Output: ˆyt for AMRC ht or for deterministic AMRC hd t cx P y Y Φ(xt, y)Tµt ϕ(µt) + if cx = 0 then for y Y do ht(y|xt) 1/|Y| else for y Y do ht(y|xt) Φ(xt, y)Tµt ϕ(µt) + /cx Draw ˆyt from distribution ht(y|xt) or obtain ˆyt from arg maxy Y ht(y|xt) |
| Open Source Code | Yes | The implementation of the proposed AMRCs is publicly available in Python and Matlab languages.1 1https://github.com/Machine Learning BCAM/ AMRC-for-concept-drift-ICML-2022 |
| Open Datasets | Yes | AMRCs are compared with the state-of-the-art using 12 datasets that have been often used as benchmarks for supervised classification under concept drift (Pavlidis et al., 2011; Kumagai & Iwata, 2016; Lu et al., 2016; Nguyen et al., 2017; Webb et al., 2018): Weather , Elec2 , Airlines , German , Chess , Usenet1 , Usenet2 , Email Spam , Credit card , Smart grid sta-bility , Shuttle , and Poker . The last 2 datasets are multi-class problems and the rest are binary (see further details in the supplementary materials). The benchmark datasets can be obtained from UCI repository and from the Massive On-line Analysis library (Bifet et al., 2010). |
| Dataset Splits | Yes | The results of the 9 methods that use kernels utilize a scaling parameter calculated with a twostage five-fold cross validation. Specifically, at the first stage, the values for the scaling factor are selected from 2i for i = { 6, 3, 0, 3, 6}. At the second stage, if γ0 = 2i where i is the best parameter obtained at the first stage, then the values for the scaling parameters are selected from γ02i, i = { 2, 1, 0, 1, 2}. The final value is γ = γ02i where i is the best parameter obtained at second stage. |
| Hardware Specification | No | No specific hardware details (like GPU/CPU models, memory, or cloud instances) used for running experiments are explicitly mentioned in the paper. |
| Software Dependencies | No | The paper mentions that the implementation is in "Python and Matlab languages" but does not provide specific version numbers for these languages or any other software libraries or dependencies. |
| Experiment Setup | Yes | Specifically, for the synthetic dataset we use the linear map Ψ(x) = x and for the benchmark datasets we use RFFs with D = 200 as given by (2). In addition, we use the recursive approach presented in Akhlaghi et al., 2017 to obtain the variances Qt,i and r2 t,i of noise processes wt,i and vt,i, in (13); we obtain the probability of the labels using (12) with W = 200; and the ASM in (19) is implemented with N = 100 and K = 2000. |