Accelerating the Global Aggregation of Local Explanations

Authors: Alon Mor, Yonatan Belinkov, Benny Kimelfeld

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Third, we design and conduct an experimental study that shows the effectiveness of our solutions.
Researcher Affiliation Academia Technion Israel Institute of Technology, Haifa, Israel almr16@campus.technion.ac.il, belinkov@technion.ac.il, bennyk@cs.technion.ac.il
Pseudocode No The paper describes algorithms and optimizations in prose but does not provide pseudocode or a clearly labeled algorithm block.
Open Source Code Yes The code and data are available at https://github.com/alonm16/anchor.
Open Datasets Yes We used the following tasks from Kaggle. URL references to the tasks can be found in the long version (Mor, Belinkov, and Kimelfeld 2023).
Dataset Splits Yes Coronavirus tweets (sentiment). The dataset consists of 16,000, 4200, and 12,500 documents (training, validation, test).
Hardware Specification Yes All experiments were conducted on a machine with 96 of Intel Xeon Gold 6336Y 2.40GHz CPUs with 24 cores, 512GB RAM, 8 of 50GB Nvidia A40 GPUs running Ubuntu 20.04 LTS.
Software Dependencies Yes The algorithms were programmed in Python 3.10 with the libraries CUDA 11.6, Py Torch 2.0, and Numpy 1.23.
Experiment Setup Yes We use k = 20. In each task, the documents are organized into three collections: training, validation, and test. ... During runtime, we reduce τ for a word to reflect its past occurrences as an anchor. Consequently, we generate fewer documents to decide whether the word is an anchor. Instead of τ, we use τ(w, c) = τ ω Gpr(w, c)/Nw where Nw is the number of occurrences of w in the dataset, and ω is a hyperparameter. We use ω = 0.4 in our experiments... We observed that decreasing this number (to just 50 in our experiments) reduced the execution cost without harming the quality... We also filtered out rare words, where we defined a word to be rare if it occurs fewer times than some threshold (5 in our experiments).