The University of Amsterdam (ILPS) at TREC 2015 Total Recall Track

We describe the participation of the University of Amsterdams ILPS group in the Total Recall track at TREC 2015. Based on the provided Baseline Model Implemention (”BMI”) we set out to provide two more baselines we can compare to in future work. The two methods are bootstrapped by a synthetic document based on the query, use TF/IDF features, and sample with dynamic batch sizes which depend on the percentage of predicted relevant documents. We sample at least 1 percent of the corpus and stop sampling if a batch contains no relevant documents. The methods differ in the classifier used, i.e. Logistic Regression and Random Forest.