Node classification in networks is a common graph mining task. In this paper, we examine how separating identity (a node’s attribute) and preference (the kind of identities to which a node prefers to link) is useful for node classification in social networks. Building upon recent work by Chin et al. (2019), where the separation of identity and preference is accomplished through a technique called łdecoupled smoothingž, we show how models that characterize both identity and preference are able to capture the underlying structure in a network, leading to improved performance in node classification tasks. Specifically, we use probabilistic soft logic (PSL) [2], a flexible and declarative statistical reasoning framework, to model identity and preference. We compare our approach with the original de-coupled smoothing method and other node classification methods implemented in PSL, and show that our approach outperforms the state-of-the-art decoupled smoothing method as well as the other node classification methods across several evaluation metrics on a real-world Facebook dataset [24].
Identifying biological entities such as genes and proteins from scientific documents is crucial for further downstream tasks such as question answering and information retrieval. This task is challenging because the same surface text can refer either to a gene or a protein based on the context. Traditional approaches such as Huang et al. [12] consider the words present in the surrounding text to infer the context. However, they fail to consider the semantics of these words which are better represented by contextual word embeddings such as BERT [6]. Deep learning based approaches, on the other hand, fail to make use of the relational structure of scientific documents. We introduce a novel probabilistic approach that jointly classifies all entity references using a class of undirected graphical models called hinge-loss Markov random fields [1]. Our approach can combine relational information with embeddingbased word semantics. Further, our approach can be easily extended to incorporate new sources of information. Our initial evaluation on the JNLPBA shared task corpus [4] shows that our joint classification approach outperforms both traditional machine learning approaches and semantic models based on word embeddings by up to 7.5% on F1 score. ACM Reference Format: Alexander Miller, Naum Markenzon, Varun Embar, and Lise Getoor. 2020. Collective Bio-Entity Recognition in Scientific Documents using Hinge-Loss Markov Random Fields. In Proceedings of International Workshop on Mining and Learning with Graphs (MLG ’20). ACM, New York, NY, USA, 6 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn
Community detection in graphs, data clustering, and local pattern mining are three mature fields of data mining and machine learning. In recent years, attributed subgraph mining is emerging as a new powerful data mining task in the intersection of these areas. Given a graph and a set of attributes for each vertex, attributed subgraph mining aims to find cohesive subgraphs for which (a subset of) the attribute values has exceptional values in some sense. While research on this task can borrow from the three abovementioned fields, the principled integration of graph and attribute data poses two challenges: the definition of a pattern language that is intuitive and lends itself to efficient search strategies, and the formalization of the interestingness of such patterns. We propose an integrated solution to both of these challenges. The proposed pattern language improves upon prior work in being both highly flexible and intuitive. We show how an effective and principled algorithm can enumerate patterns of this language. The proposed approach for quantifying interestingness of patterns of this language is rooted in information theory, and is able to account for prior knowledge on the data. Prior work typically quantifies interestingness based on the cohesion of the subgraph and for the exceptionality of its attributes separately, combining these in a parameterized trade-off. Instead, in our proposal this trade-off is implicitly handled in a principled, parameter-free manner. Extensive empirical results confirm the proposed pattern syntax is intuitive, and the interestingness measure aligns well with actual subjective interestingness.
Graph models are widely used to analyse diffusion processes embedded in social contacts and to develop applications. A range of graph models are available to replicate the underlying social structures and dynamics realistically. However, most of the current graph models can only consider concurrent interactions among individuals in the co-located interaction networks. However, they do not account for indirect interactions that can transmit spreading items to individuals who visit the same locations at different times but within a certain time limit. The diffusion phenomena occurring through direct and indirect interactions is called same place different time (SPDT) diffusion. This paper introduces a model to synthesize co-located interaction graphs capturing both direct interactions, where individuals meet at a location, and indirect interactions, where individuals visit the same location at different times within a set timeframe. We analyze 60 million location updates made by 2 million users from a social networking application to characterize the graph properties, including the space-time correlations and its time evolving characteristics, such as bursty or ongoing behaviors. The generated synthetic graph reproduces diffusion dynamics of a realistic contact graph, and reduces the prediction error by up to 82% when compare to other contact graph models demonstrating its potential for forecasting epidemic spread.
