%0 Conference Paper %1 buttner2023joining %A Büttner, Maik %A Beyer, Christian %A Spiliopoulou, Myra %B Discovery Science %D 2023 %E Bifet, Albert %E Lorena, Ana Carolina %E Ribeiro, Rita P. %E Gama, João %E Abreu, Pedro H. %I Springer Nature Switzerland %K active-feature-acquisition data-streams imputation myown stream-mining %P 308-322 %R 10.1007/978-3-031-45275-8_21 %T Joining Imputation and Active Feature Acquisition for Cost Saving on Data Streams with Missing Features %X Replacing missing features in data streams is an important task in order to enable many machine learning algorithms that require feature-complete instances for training and prediction. Two popular methods for dealing with missing features are imputation and active feature acquisition, where in the former missing values are approximated, whereas in the latter, missing features are provided by an expert for a cost and within a limited budget. In this work, we present a hybridized approach, where we employ an active feature acquisition method in the first stage to pick candidate features on which we would require a costly expert and then check in a second stage how well we could impute these candidate features. If the imputation is expected to be of a certain quality, we skip the purchase and impute instead. We provide a framework for such a scenario and used it to run extensive experiments. Our results on 6 data sets show that our proposed method can achieve a similar classification performance while spending 1\% to 27\% less budget. %@ 978-3-031-45275-8

@inproceedings{buttner2023joining, abstract = {Replacing missing features in data streams is an important task in order to enable many machine learning algorithms that require feature-complete instances for training and prediction. Two popular methods for dealing with missing features are imputation and active feature acquisition, where in the former missing values are approximated, whereas in the latter, missing features are provided by an expert for a cost and within a limited budget. In this work, we present a hybridized approach, where we employ an active feature acquisition method in the first stage to pick candidate features on which we would require a costly expert and then check in a second stage how well we could impute these candidate features. If the imputation is expected to be of a certain quality, we skip the purchase and impute instead. We provide a framework for such a scenario and used it to run extensive experiments. Our results on 6 data sets show that our proposed method can achieve a similar classification performance while spending 1{\%} to 27{\%} less budget.}, added-at = {2023-12-06T14:37:59.000+0100}, author = {Büttner, Maik and Beyer, Christian and Spiliopoulou, Myra}, biburl = {https://www.bibsonomy.org/bibtex/244730f1915dcdeca65a6ff396896fd61/kmd-ovgu}, booktitle = {Discovery Science}, doi = {10.1007/978-3-031-45275-8_21}, editor = {Bifet, Albert and Lorena, Ana Carolina and Ribeiro, Rita P. and Gama, João and Abreu, Pedro H.}, interhash = {9ab58e65861556c90bb6f5d2bcf30bf6}, intrahash = {44730f1915dcdeca65a6ff396896fd61}, isbn = {978-3-031-45275-8}, keywords = {active-feature-acquisition data-streams imputation myown stream-mining}, pages = {308-322}, publisher = {Springer Nature Switzerland}, timestamp = {2023-12-11T11:20:03.000+0100}, title = {Joining Imputation and Active Feature Acquisition for Cost Saving on Data Streams with Missing Features}, year = 2023 }