Multi-Label Classification Neural Networks with Hard Logical Constraints
Authors: Eleonora Giunchiglia, Thomas Lukasiewicz
JAIR 2021 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct an extensive experimental analysis showing the superior performance of both C-HMCNN(h) and CCN(h) when compared to state-of-the-art models in both the HMC and the general MC setting with hard logical constraints. |
| Researcher Affiliation | Academia | Eleonora Giunchiglia EMAIL Department of Computer Science, University of Oxford, UK Thomas Lukasiewicz EMAIL Department of Computer Science, University of Oxford, UK |
| Pseudocode | No | The paper describes methods and mathematical formulations but does not contain any explicitly labeled "Pseudocode" or "Algorithm" blocks. |
| Open Source Code | Yes | Link: https://github.com/EGiunchiglia/C-HMCNN/ Link: https://github.com/EGiunchiglia/CCN/ |
| Open Datasets | Yes | We tested CCN(h) on 20 real-world datasets commonly used to compare HMC systems (see, e.g., Bi & Kwok, 2011; Nakano et al., 2019; Vens et al., 2008; Wehrmann et al., 2018): 16 are functional genomics datasets (Clare, 2003), 2 contain medical images (Dimitrovski et al., 2008), 1 contains images of microalgae (Dimitrovski et al., 2012), and 1 is a text categorization dataset (Klimt & Yang, 2004).8 The characteristics of these datasets are summarized in Table 1. ... Links: https://dtai.cs.kuleuven.be/clus/hmcdatasets and http://kt.ijs.si/Dragi Kocev/Ph D/resources ... Being the first paper on LCMC problems, we created 16 real-world LCMC datasets, each obtained by enriching a popular and publicly available MC dataset with constraints extracted using the apriori algorithm (Agrawal & Srikant, 1994). |
| Dataset Splits | Yes | The datasets consisted of 5000 (50/50 train/test split) data points sampled from a uniform distribution over [0, 1]2. ... The datasets consisted of 5000 (50/50 train/test split) data points sampled from a uniform distribution over [0, 1]2. ... Table 1: Summary of the 20 real-world datasets. Number of features (D), number of classes (n), and number of data points for each dataset split. ... Table 5: Summary of the real-world MC datasets. For each dataset, we report from left to right: (i) name, (ii) number of features (D), (iii) number of classes (L), (iv-vi) number of data points for each split |
| Hardware Specification | Yes | All experiments were run on an Nvidia Titan Xp with 12 GB memory. ... For this experiment, we used an Nvidia Titan Xp with 12 GB memory as GPU and an Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz as CPU. |
| Software Dependencies | No | The paper mentions software like Adam optimization, PyTorch, and scikit-multilearn, but does not provide specific version numbers for these dependencies. |
| Experiment Setup | Yes | f and g were trained with binary cross-entropy loss using Adam optimization (Kingma & Ba, 2015) for 20k epochs with learning rate 10 2 (β1 = 0.9, β2 = 0.999). ... In all experiments, the loss was minimized using Adam optimizer with weight decay 10 5, and patience 20 (β1 = 0.9, β2 = 0.999). The dropout rate was set to 70% and the batch size to 4. ... In all experiments, the loss was minimized using Adam optimizer with batch size equal to 4, learning rate equal to 10 4, and patience 20 (β1 = 0.9, β2 = 0.999). Since some datasets have very few data points, we set the dropout rate equal to 80% and the weight decay equal to 10 4. |