Inducing Probabilistic Relational Rules from Probabilistic Examples
Authors: Luc De Raedt, Anton Dries, Ingo Thon, Guy Van den Broeck, Mathias Verbeke
IJCAI 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We answer two questions experimentally. |
| Researcher Affiliation | Collaboration | KU Leuven, Department of Computer Science Celestijnenlaan 200A, BE-3001 Heverlee, Belgium. Now at Siemens AG, Otto-Hahn-Ring 6, GE-81739 Munich. Now at Sirris, A. Reyerslaan 80, BE-1030 Brussels |
| Pseudocode | Yes | Algorithm 1 The Prob FOIL+ learning algorithm. |
| Open Source Code | Yes | Prob FOIL+ and the datasets used in this paper in Prob FOIL+ format can be downloaded from https://dtai.cs.kuleuven.be/software/probfoil/. |
| Open Datasets | Yes | We use BNGenerator to randomly generate a Bayesian network structure. The generated network has 45 nodes, 70 edges, a maximal degree of 6 and an induced width of 5. ... we extracted the facts for all predicates related to the sports domain from iteration 850 of the NELL knowledge base. |
| Dataset Splits | Yes | For each of these, we trained Prob FOIL, Prob FOIL+ and standard regression learners from the Weka suite on 500 training examples. The learned models are evaluated on 500 test examples... We used 3-fold cross-validation. To create the folds, for each target predicate, the facts were randomly split into 3 parts. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, memory, or cloud instance types used for running experiments. It only generally states that experiments were performed. |
| Software Dependencies | No | The paper mentions using the "Weka suite" and "Prob Log2 system" but does not specify their version numbers (e.g., Weka 3.9, Prob Log2 vX.Y.Z) which are crucial for reproducibility. |
| Experiment Setup | Yes | For all predicates, the m-estimate s m value was set to 1 and the beam width to 5. The value of p for rule significance was set to 0.99. Furthermore, to avoid a bias towards the majority class, the examples are balanced, i.e., a part of the negative examples is removed. ...we also tested all settings with a high m-value (1000), and a rule significance p of 0.9 (parameter setting B). |