Interactive Visual Pattern Search on Graph Data via Graph Representation Learning

Abstract

Fig. 1. The visualization interface of GraphQ contains: (1) A query editing panel to specify the subgraph patterns and initiate the search.(2.1) (2.2) Query result panels to display the retrieved results. The graph thumbnails can be displayed in overview and detail modes.(3) A statistics and filtering panel that helps users select a graph to construct example-based query, and visualizes the distribution of the query results in the database. (4) A query option control panel to specify whether fuzzy-pattern search is enabled and whether the node-match should be highlighted. (5) A popup window for pairwise comparison between the query pattern and the returned result.The figure shows a case study on program workflow graph pattern search and the details are described in Section 5.1.

VA + Embeddings STAR: A State‐of‐the‐Art Report on the Use of Embeddings in Visual Analytics

“…Furthermore, the corresponding interface needs to consider providing guidance, expansion hints, and feedback to user‐defined queries. The commonality between TopicSifter , an interactive search space reduction technique [KCD * 19], the interactive visual pattern search technique proposed by Song et al [SDXR22], and VISAtlas , an imaged‐based query system for visualization collections [YHZ22], is about finding ways of creating human‐interpretable embedding vectors. Users can observe object alignments inside the embedding space.…”

“…This field has seen substantial growth after some initial word‐embedding technologies were modified to handle this data type. Common usage scenarios are analysis and comparison of graph topology [CZC*17,PCZ*21, SDXR22], and (just as for word embeddings) comparing the embedding spaces of different algorithms can give important insights to the different algorithms as well as to the underlying network [LNH*18, CZG*22]. Furthermore, TorusTraffic ND [CZIM18] provides an interesting example of embedding network topology onto a Hilbert curve rather than using a numerical vector.…”

“…Filtering can be restricted to documents containing keywords, to keywords and their nearest neighbors, or loosened to include similar documents ranked by some distance measure in the embedding space, giving users multiple levels of constraints with thresholds to customize their search. Embeddings are often integrated into the computational part of a VA pipeline and may not be explicitly mapped to a visual representation [KCD*19, SDXR22]. In other cases where embeddings are explicitly represented, the interactive interface allows for another dimension of user customization.…”

“…We classify the computational approach into three main methods, which can of course be used in combination within the scope of a single paper. The most common, with 88 occurrences, is neu ral network, where different types of ML/DL models are used to obtain the embeddings [ZJQH22,RSL*22, SDXR22]. Even though there are many different variants, some of which are highly customized for the specific tasks, our analysis shows that using the BERT [DCLT19] model or word2vec [MSC*13] (or some of the other versions inspired/derived from it) is a common choice among the surveyed papers.…”

Visual Relationship Detection with Multimodal Fusion and Reasoning

“…For example, Qi et al [ 9 ] proposed a method to caption better sports videos by modeling players’ interactions. Song et al [ 10 ] devised a visual graph network to propagate semantic information to capture relationships.…”