Reproduction Code: Multimodal Relevance Regression

Multimodal Relevance Regression This repository contains the reproduction code and data for the paper: "Towards continuous-valued relevance assessment of social media posts in disaster response: A decay-based regression approach" by David Hanny, Andreas Kramer, Ehsaneddin Jajilian, Sebastian Schmidt, and Bernd Resch. The work was submitted to the AGILE Conference 2026. 💾 Data Our evaluation dataset was collected using the X (formerly Twitter) v1.1 and v2 filtered stream and recent search Application Programming Programming Interfaces (APIs). In accordance with X's developer policies, it can only be shared as a list of post IDs. We therefore provide post IDs, the corresponding relevance labels, and all derived features used for inference. Furthermore, we provide the softmax outputs of the fine-tuned TwHIN-BERT model used for text classification. All data is available in the folder ./data/input: - train_data_public.parquet: Anonymised training data (3,659 posts) - oof_preds_text_train.npy: Out-of-fold softmax predictions on the training data from a fine-tuned TwHIN-BERT text classifier for meta learning - test_data_public.parquet: Anonymised test data (915 posts) - preds_text_test.npy: Softmax predictions of a TwHIN-BERT text classifier, fine-tuned on the full training data (HuggingFace) 🐳 Reproduction Our full experimental pipeline is available as a containerised Docker image.To run the experiments, use:docker compose run --rm --service-ports run-experiments # run experiments as python scriptdocker compose run --rm --service-ports edit-experiments # edit mode as a marimo notebook All results will be stored to ./data/output, and all figures will automatically go to ./data/figures. For editing, a Marimo interface will be available at localhost:8080. Please make sure to use the access tokens from the stdout logs. To create the visualisations, use: docker compose run --rm --service-ports run-visualisations # print and store visualisations as python script docker compose run --rm --service-ports edit-visualisations # edit mode as a marimo notebook