Semi-supervised Embedding in Attributed Networks with Outliers

Abstract

In this paper, we propose a novel framework, called Semi-supervised Embedding in Attributed Networks with Outliers (SEANO), to learn a low-dimensional vector representation that systematically captures the topological proximity, attribute affinity and label similarity of vertices in a partially labeled attributed network (PLAN). Our method is designed to work in both transductive and inductive settings while explicitly alleviating noise effects from outliers. Experimental results on various datasets drawn from the web, text and image domains demonstrate the advantages of SEANO over the state-of-the-art methods in semi-supervised classification under transductive as well as inductive settings. We also show that a subset of parameters in SEANO are interpretable as outlier scores and can significantly outperform baseline methods when applied for detecting network outliers. Finally, we present the use of SEANO in a challenging real-world setting -flood mapping of satellite images and show that it is able to outperform modern remote sensing algorithms for this task. on the output outlier scores from SEANO and demonstrate its advantages over baseline methods specializing in network outlier detection. We finally conduct a case study of flood mapping to visually show the power of SEANO when applied to flood mapping. Related WorkNetwork Embedding: Network embedding strategies have gained increasing importance in recent years. Early ideas include IsoMap [35] and Locally Linear Embedding (LLE) [29], which exploited the manifold structure of vector data to compute low-dimensional embeddings. More recently, due to the emergence of naturally arising network data, other network embedding methods have been proposed [34,27,9]. In addition to learning embeddings for homogeneous networks, several researchers have proposed ideas for embedding attributed networks [36,13,12,26,37,15,10]. While they incorporate the attributes and/or label information into the embeddings, most are inherently transductive and cannot generate embeddings for vertices unseen during training. The two exceptions are Planetoid [37] and GraphSAGE [10] for inductive learning. However, Planetoid [37] is specialized for semi-supervised classification and the output network embedding, as a byproduct, does not capture all the information (as can be seen from their model architecture). Therefore, its embeddings might not be generalized to other applications such as visualization and clustering. GraphSAGE [10], on the other hand, only works on unsupervised learning or fully supervised learning setting and cannot be directly applied in a semi-supervised manner. Finally, none of the existing work on network embedding explicitly accounts for the impact of outliers. We summarize the differences between the proposed SEANO model with some of these recent efforts in Table 1. Method Attributes Labels Semisupervised Inductive Address Outliers DeepWalk [27] Node2Vec [9] TADW [36] LANE [13] TriDNR [26] Planetoid [37] GCN [15] GraphSAGE [10] SEANO Table 1: A comparison of SEANO w...

Micro- and macro-level churn analysis of large-scale mobile games

“…where w u * v * and w u v are the vectors of weights for edges (u * , v * ) and (u , v ) in the softmax layer, respectively. The denominator is computed by negative sampling [20]. Although the embedding function is learned by training a context inference task, it is still considered as "unsupervised" because the contexts are calculated by sampling on the attributed graph, which is independent of any supervised learning task [21].…”

A Semi-Supervised and Inductive Embedding Model for Churn Prediction of Large-Scale Mobile Games

“…where w u * v * and w u v are the vectors of weights for edges (u * , v * ) and (u , v ) in the softmax layer, respectively. The denominator is computed by negative sampling [14]. Although the embedding function is learned by training a context inference task, it is still considered as "unsupervised" because the contexts are calculated by sampling on the attributed graph, which is independent of any supervised learning task [15].…”

“…Temporal loss refers to the loss related to graph dynamics. Different from existing works [14]- [16], the proposed embedding model takes into account graph dynamics. Understanding the temporal dynamics of attributed bipartite graphs is crucial to precisely model the churn behavior.…”

“…Several very recent works have proposed the idea of inductive graph embedding [14]- [16], inspired by which we propose our inductive edge embedding model for churn prediction. Our model improves these works in several major ways.…”

ImVerde: Vertex-Diminished Random Walk for Learning Imbalanced Network Representation

“…Most existing work on network embedding either works in the unsupervised fashion, i.e., not considering the label information at all, or is implicitly assuming a balanced setting, i.e., not suitable for the imbalanced setting. More specifically, unsupervised embedding approaches [5], [13], [30], [36] aimed to learn the node representations which preserve the graph topological structure in the feature space; and semi-supervised methods [1], [2], [32] implicitly assumed that the labeled set consisted of roughly equal number of examples from each class. However, in the presence of imbalanced labeled set, these methods will inevitably suffer from imbalanced node-context pairs, resulting in node representations not amiable to the following classification task.…”