Boosting with Tempered Exponential Measures

Authors: Richard Nock, Ehsan Amid, Manfred Warmuth

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We have performed experiments on a testbed of 10 UCI domains, whose details are given in APP. (Section A3). Experiments were carried out using a 10-fold stratified cross-validation procedure. To compare t-ADABOOST with ADABOOST, we ran t-ADABOOST with a first range of values of t P t0.0, 0.2, 0.4, 0.6, 0.8, 0.9u.
Researcher Affiliation Industry Richard Nock Google Research richardnock@google.com Ehsan Amid Google Deep Mind eamid@google.com Manfred K. Warmuth Google Research manfred@google.com
Pseudocode Yes Algorithm 1 t-ADABOOSTpt, S, Jq Input: t P r0, 1s, training sample S, #iterations J; Output: classifiers HJ, Hp1{1 tq J (see (9)); Step 1 : initialize tempered weights: q1 p1{mt q 1 p P mq; Step 2 : for j 1, 2, ..., J Step 2.1 : get weak classifier hj Ð weak_learnerpqj, Sq; Step 2.2 : choose weight update coefficient µj P R; Step 2.3 : @i P rms, for uji . yihjpxiq, update the tempered weights as qpj 1qi qji bt exptp µjujiq Ztj , where Ztj qj bt exptp µjujq 1{t . (8) Step 2.4 : choose leveraging coefficient αj P R;
Open Source Code No The paper does not provide a statement or link indicating that the source code for their methodology is publicly available.
Open Datasets Yes We have performed experiments on a testbed of 10 UCI domains, whose details are given in APP. (Section A3). Table A3: Public domains considered in our experiments (m total number of examples, d total number of example s features, including class), ordered in increasing m ˆ d (see text).
Dataset Splits Yes Experiments were carried out using a 10-fold stratified cross-validation procedure.
Hardware Specification No The paper does not provide any specific details about the hardware used for running the experiments (e.g., CPU, GPU models, memory).
Software Dependencies No The paper states that the implementation is "programmed in Java" but does not specify any version numbers for Java or other software libraries/dependencies.
Experiment Setup Yes All boosting models were trained for a number of J 20 decision trees (The appendix provides experiments on training bigger sets). Each decision tree is induced using the tempered loss with the corresponding value of t (see Theorem 4) following the classical top-down template, which consists in growing the current heaviest leaf in the tree and picking the best split for the leaf chosen. We implemented t-ADABOOST exactly as in Section 5, including computing leveraging coefficients as suggested. Thus, we do not scale models. More details are provided in the appendix. In our implementation of decision-tree induction, when the number of possible splits exceeds a fixed number S (currently, 2 000), we pick the best split in a subset of S splits picked at random.