Abstract:Summarizing web graphs is challenging due to the heterogeneity of the modeled information and its changes over time. We investigate the use of neural networks for lifelong graph summarization. Assuming we observe the web graph at a certain time, we train the networks to summarize graph vertices. We apply this trained network to summarize the vertices of the changed graph at the next point in time. Subsequently, we continue training and evaluating the network to perform lifelong graph summarization. We use the GNNs Graph-MLP and GraphSAINT, as well as an MLP baseline, to summarize the temporal graphs. We compare $1$-hop and $2$-hop summaries. We investigate the impact of reusing parameters from a previous snapshot by measuring the backward and forward transfer and the forgetting rate of the neural networks. Our extensive experiments on ten weekly snapshots of a web graph with over $100$M edges, sampled in 2012 and 2022, show that all networks predominantly use $1$-hop information to determine the summary, even when performing $2$-hop summarization. Due to the heterogeneity of web graphs, in some snapshots, the $2$-hop summary produces over ten times more vertex summaries than the $1$-hop summary. When using the network trained on the last snapshot from 2012 and applying it to the first snapshot of 2022, we observe a strong drop in accuracy. We attribute this drop over the ten-year time warp to the strongly increased heterogeneity of the web graph in 2022.
Abstract:HyperAggregation is a hypernetwork-based aggregation function for Graph Neural Networks. It uses a hypernetwork to dynamically generate weights in the size of the current neighborhood, which are then used to aggregate this neighborhood. This aggregation with the generated weights is done like an MLP-Mixer channel mixing over variable-sized vertex neighborhoods. We demonstrate HyperAggregation in two models, GraphHyperMixer is a model based on MLP-Mixer while GraphHyperConv is derived from a GCN but with a hypernetwork-based aggregation function. We perform experiments on diverse benchmark datasets for the vertex classification, graph classification, and graph regression tasks. The results show that HyperAggregation can be effectively used for homophilic and heterophilic datasets in both inductive and transductive settings. GraphHyperConv performs better than GraphHyperMixer and is especially strong in the transductive setting. On the heterophilic dataset Roman-Empire it reaches a new state of the art. On the graph-level tasks our models perform in line with similarly sized models. Ablation studies investigate the robustness against various hyperparameter choices. The implementation of HyperAggregation as well code to reproduce all experiments is available under https://github.com/Foisunt/HyperAggregation .
Abstract:Text role classification involves classifying the semantic role of textual elements within scientific charts. For this task, we propose to finetune two pretrained multimodal document layout analysis models, LayoutLMv3 and UDOP, on chart datasets. The transformers utilize the three modalities of text, image, and layout as input. We further investigate whether data augmentation and balancing methods help the performance of the models. The models are evaluated on various chart datasets, and results show that LayoutLMv3 outperforms UDOP in all experiments. LayoutLMv3 achieves the highest F1-macro score of 82.87 on the ICPR22 test dataset, beating the best-performing model from the ICPR22 CHART-Infographics challenge. Moreover, the robustness of the models is tested on a synthetic noisy dataset ICPR22-N. Finally, the generalizability of the models is evaluated on three chart datasets, CHIME-R, DeGruyter, and EconBiz, for which we added labels for the text roles. Findings indicate that even in cases where there is limited training data, transformers can be used with the help of data augmentation and balancing methods. The source code and datasets are available on GitHub under https://github.com/hjkimk/text-role-classification
Abstract:We support scientific writers in determining whether a written sentence is scientific, to which section it belongs, and suggest paraphrasings to improve the sentence. Firstly, we propose a regression model trained on a corpus of scientific sentences extracted from peer-reviewed scientific papers and non-scientific text to assign a score that indicates the scientificness of a sentence. We investigate the effect of equations and citations on this score to test the model for potential biases. Secondly, we create a mapping of section titles to a standard paper layout in AI and machine learning to classify a sentence to its most likely section. We study the impact of context, i.e., surrounding sentences, on the section classification performance. Finally, we propose a paraphraser, which suggests an alternative for a given sentence that includes word substitutions, additions to the sentence, and structural changes to improve the writing style. We train various large language models on sentences extracted from arXiv papers that were peer reviewed and published at A*, A, B, and C ranked conferences. On the scientificness task, all models achieve an MSE smaller than $2\%$. For the section classification, BERT outperforms WideMLP and SciBERT in most cases. We demonstrate that using context enhances the classification of a sentence, achieving up to a $90\%$ F1-score. Although the paraphrasing models make comparatively few alterations, they produce output sentences close to the gold standard. Large fine-tuned models such as T5 Large perform best in experiments considering various measures of difference between input sentence and gold standard. Code is provided under https://github.com/JustinMuecke/SciSen.
Abstract:When training a Neural Network, it is optimized using the available training data with the hope that it generalizes well to new or unseen testing data. At the same absolute value, a flat minimum in the loss landscape is presumed to generalize better than a sharp minimum. Methods for determining flat minima have been mostly researched for independent and identically distributed (i. i. d.) data such as images. Graphs are inherently non-i. i. d. since the vertices are edge-connected. We investigate flat minima methods and combinations of those methods for training graph neural networks (GNNs). We use GCN and GAT as well as extend Graph-MLP to work with more layers and larger graphs. We conduct experiments on small and large citation, co-purchase, and protein datasets with different train-test splits in both the transductive and inductive training procedure. Results show that flat minima methods can improve the performance of GNN models by over 2 points, if the train-test split is randomized. Following Shchur et al., randomized splits are essential for a fair evaluation of GNNs, as other (fixed) splits like 'Planetoid' are biased. Overall, we provide important insights for improving and fairly evaluating flat minima methods on GNNs. We recommend practitioners to always use weight averaging techniques, in particular EWA when using early stopping. While weight averaging techniques are only sometimes the best performing method, they are less sensitive to hyperparameters, need no additional training, and keep the original model unchanged. All source code is available in https://github.com/Foisunt/FMMs-in-GNNs.
Abstract:Privacy preserving deep learning is an emerging field in machine learning that aims to mitigate the privacy risks in the use of deep neural networks. One such risk is training data extraction from language models that have been trained on datasets , which contain personal and privacy sensitive information. In our study, we investigate the extent of named entity memorization in fine-tuned BERT models. We use single-label text classification as representative downstream task and employ three different fine-tuning setups in our experiments, including one with Differentially Privacy (DP). We create a large number of text samples from the fine-tuned BERT models utilizing a custom sequential sampling strategy with two prompting strategies. We search in these samples for named entities and check if they are also present in the fine-tuning datasets. We experiment with two benchmark datasets in the domains of emails and blogs. We show that the application of DP has a huge effect on the text generation capabilities of BERT. Furthermore, we show that a fine-tuned BERT does not generate more named entities entities specific to the fine-tuning dataset than a BERT model that is pre-trained only. This suggests that BERT is unlikely to emit personal or privacy sensitive named entities. Overall, our results are important to understand to what extent BERT-based services are prone to training data extraction attacks.