Big Data Science – LMU München

The task of peer-reviewing is central to the area of scientific publications. But it is a cumbersome process, due to the growing review workload and newer fields. Argument mining has enabled understanding not only the positions adopted by the people but also the reason why they do so. In the domain of the scientific discourse, specifically in peer-reviewing, this can be potentially used to drive the process.
In our work, we defined the objective to design and develop a reviewing expert system that can assist the editors, meta-reviewers, and reviewers. We demonstrated that the decision process in peer-reviewing is driven by arguments and automatic argument identification can be used to facilitate the process. Within the scope of our project, we defined, structured, and annotated our own scientific-reviews dataset and used it to train our model to extract relevant arguments and then make predictions on the acceptability of the paper. Our machine-learning model achieved near-human performance and showed that arguments are paramount for the publication decision. The process remains interpretable since the extracted arguments can be highlighted in a review without detaching them from their context.

Ressources: Code, Dataset, AAAI-2021 Paper
Team: Siddharth Bhargava, Ruoxia Qi, Yao Zhang, Lukas Dennert, Sophia Selle, Yang Mao
Betreuung: Michael Fromm
Leitung des Innovationslabors: Prof. Dr. Thomas Seidl, Prof. Dr. Bernd Bischl
Infrastruktur: Prof. Dr. Dieter Kranzlmüller
LMU Innovation Lab: https://innolab.ifi.lmu.de/

Climate Change alters the atmospheric circulation over Europe and increases the risk of heavy precipitation. ‚Tief Mitteleuropa‘ and ‚Trog Mitteleuropa‘ are two atmospheric circulation patterns that are associated with heavy precipitation over Central Europe.

Thus the research question of how climate change influences their occurrence is of high relevance. However, the spatio-temporal data structure and the imbalanced classes demand for sophisticated modelling architectures in order to detect these circulation patterns in climate models. This project introduces and compares deep learning algorithms that are able to tackle the problem of detecting the two atmospheric circulation classes. A ResNet18 and a Convolutional LSTM are set up and fitted to the data.

Overall, both model types are generally able to detect the atmospheric circulation patterns and, moreover, carry immense potential in the way they can be set up and fitted to the data. These promising results open the gate for further research in the future.

The model is embedded in an intuitive pipeline that includes modules for preprocessing, modelling and training in pytorch. This pipline is available as a pip-installable python package. The package can be downloaded from Github. In consequence, users are able to train their own weather models by one single command or exchange preprocessing, modelling and training steps in the pipeline by their own modules.

This summer, a short paper was published and presented at the EnviroInfo Conference 2021 in Berlin:
Funk, H., Becker, C., Hofheinz, A., Xi, G., Zhang, Y., Pfisterer, F., Weigert, M., & Mittermeier, M. (forthcoming). Towards an automated classification of Hess & Brezowsky's atmospheric circulation patterns Tief and Trog Mitteleuropa using Deep Learning Methods. In Environmental Informatics: A bogeyman or saviour to achieve the UN Sustainable Development Goals?. Shaker.

Team: Carolin Becker, Henri Funk, Andreas Hofheinz, Guoren Xi, Yao Zhang
Betreuung: Florian Pfisterer, Christoph Molnar, Magdalena Mittermeier, Maximilian Weigert
Leitung des Innovationslabors: Prof. Dr. Thomas Seidl, Prof. Dr. Bernd Bischl
LMU Innovation Lab: https://innolab.ifi.lmu.de/