How does This Interaction Affect Me? Interpretable Attribution for Feature Interactions

Authors: Michael Tsang, Sirisha Rambhatla, Yan Liu

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on standard annotation labels indicate our approach provides significantly more interpretable explanations than comparable methods, which is important for analyzing the impact of interactions on predictions. We also provide accompanying visualizations of our approach that give new insights into deep neural networks. We conduct experiments first on Arch Detect in 5.2 then on Arch Attribute in 5.3.
Researcher Affiliation Academia Michael Tsang, Sirisha Rambhatla, Yan Liu Department of Computer Science University of Southern California {tsangm,sirishar,yanliu.cs}@usc.edu
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at: https://github.com/mtsang/archipelago
Open Datasets Yes BERT was fine-tuned on the SST dataset [44], and Res Net152 was pretrained on Image Net [12]. We obtain segment labels from the MS COCO dataset [29] and match them to the label space of Image Net. Recommendation Task: Fig. 6 shows Archipelago s result for this task using a state-of-the-art Auto Int model [45] for adrecommendation. Here, our approach finds a positive interaction between device_id and banner_pos in the Avazu dataset [1].
Dataset Splits No The paper refers to using standard datasets and their test sets (e.g., 'SST test set', 'Image Net for Res Net152') but does not explicitly provide specific training/validation/test dataset split percentages or sample counts to reproduce the data partitioning.
Hardware Specification Yes These experiments were done on a server with 32 Intel Xeon E5-2640 v2 CPUs @ 2.00GHz and 2 Nvidia 1080 Ti GPUs.
Software Dependencies No The paper does not provide specific software dependency details, such as library names with version numbers, required to replicate the experiments.
Experiment Setup No While some setup details like baseline definition and use of Quickshift superpixel segmenter are provided, the paper does not specify concrete hyperparameter values (e.g., learning rate, batch size, number of epochs) or detailed training configurations for the models used.