Materialisation-Based Reasoning in DatalogMTL with Bounded Intervals

Authors: Przemysław A. Wałęga, Michał Zawidzki, Dingmin Wang, Bernardo Cuenca Grau

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

Reproducibility Variable Result LLM Response
Research Type Experimental We have implemented our algorithm and conducted two experiments. The first experiment compares our implementation with Me Teo R (Wang et al. 2022), which we take as the baseline;... The second experiment tests the scalability of our implementation on datasets of increasing size. All experiments were conducted on a Dell Power Edge R730 server with 512 GB of RAM and two Intel Xeon E5-2640 2.6 GHz processors running Fedora 33, kernel version 5.8.17. We used as benchmarks a temporal extension of the Lehigh University Benchmark (LUBM) (Wang et al. 2022) and the i Temporal bechmark generator (Bellomarini, Nissl, and Sallinger 2022).
Researcher Affiliation Academia Przemysław A. Wał ega, Michał Zawidzki, Dingmin Wang, Bernardo Cuenca Grau Department of Computer Science, University of Oxford, UK {przemyslaw.walega, michal.zawidzki, dingmin.wang, bernardo.cuenca.grau}@cs.ox.ac.uk
Pseudocode Yes Algorithm 1: Reasoning via Periods Detection
Open Source Code Yes Our implementation can be accessed through the following link: https://github.com/wdimmy/Datalog MTLPeriodicity.
Open Datasets Yes We used as benchmarks a temporal extension of the Lehigh University Benchmark (LUBM) (Wang et al. 2022) and the i Temporal bechmark generator (Bellomarini, Nissl, and Sallinger 2022).
Dataset Splits No For the LUBM benchmark, Figure 2 (left), we considered a dataset D with 5 million facts and query facts F@t, where F is a fixed relational atom (about the predicate Full Professor) and t { 500, 400, . . . , 500}.
Hardware Specification Yes All experiments were conducted on a Dell Power Edge R730 server with 512 GB of RAM and two Intel Xeon E5-2640 2.6 GHz processors running Fedora 33, kernel version 5.8.17.
Software Dependencies No The paper specifies the operating system (Fedora 33) and kernel version (5.8.17), but does not provide specific versions for programming languages, libraries, or other application-level software components used for the implementation.
Experiment Setup Yes For the LUBM benchmark, Figure 2 (left), we considered a dataset D with 5 million facts and query facts F@t, where F is a fixed relational atom (about the predicate Full Professor) and t { 500, 400, . . . , 500}; we observed the same type of behaviour as for LUBM (see Figure 2, right). In the second experiment, we analysed how saturation times increase with the size of the input dataset. We considered the same programs as in the first experiment, together with sequences of datasets of increasing size. For the LUBM benchmark, we generated two sequences, each containing 6 datasets. In the first sequence all the intervals of temporal facts are contained within the range [0, 50] and we increase the size of datasets by introducing atoms with new constants. In the second sequence, the number of constants is the same but the number of facts increases; facts occupy increasing ranges of time, namely, [0, 5 10i], for i {1, . . . , 6}. In both sequences, the datasets have 103, 104, 105, 106, 107, and 108 facts, respectively.