In this paper, we study the following problem: given a knowledge graph (KG) and a set of input vertices (representing concepts or entities) and edge labels, we aim to find the smallest connected subgraphs containing all of the inputs. This problem plays a key role in KG-based search engines and natural language question answering systems, and it is a natural extension of the Steiner tree problem, which is known to be NP-hard. We present RECON, a system for finding approximate answers. RECON aims at achieving high accuracy with instantaneous response (i.e., sub-second/millisecond delay) over KGs with hundreds of millions edges without resorting to expensive computational resources. Furthermore, when no answer exists due to disconnection between concepts and entities, RECON refines the input to a semantically similar one based on the ontology, and attempts to find answers with respect to the refined input. We conduct a comprehensive experimental evaluation of RECON. In particular we compare it with five existing approaches for finding approximate Steiner trees. Our experiments on four large real and synthetic KGs show that RECON significantly outperforms its competitors and incurs a much smaller memory footprint.

Document clustering requires a deep understanding of the complex structure of longtext; in particular, the intra-sentential (local) and inter-sentential features (global).Existing representation learning models do not fully capture these features.To address this, we present a novel graph-based representation for document clustering that builds a graph autoencoder (GAE) on a Keyword Correlation Graph.The graph is constructed with topical keywords as nodes and multiple local and global features as edges.A GAE is employed to aggregate the two sets of features by learning a latent representation which can jointly reconstruct them.Clustering is then performed on the learned representations, using vector dimensions as features for inducing document classes.Extensive experiments on two datasets show that the features learned by our approach can achieve better clustering performance than other existing features, including term frequency-inverse document frequency and average embedding.

Relation Extraction is a way of obtaining the semantic relationship between entities in text. The state-of-the-art methods use linguistic tools to build a graph for the text in which the entities appear and then a Graph Convolutional Network (GCN) is employed to encode the pre-built graphs. Although their performance is promising, the reliance on linguistic tools results in a non end-to-end process. In this work, we propose a novel model, the Self-determined Graph Convolutional Network (SGCN), which determines a weighted graph using a self-attention mechanism, rather using any linguistic tool. Then, the self-determined graph is encoded using a GCN. We test our model on the TACRED dataset and achieve the state-of-the-art result. Our experiments show that SGCN outperforms the traditional GCN, which uses dependency parsing tools to build the graph.

Text classification requires a deep understanding of the linguistic features in text; in particular, the intra-sentential (local) and inter-sentential features (global). Models that operate on word sequences have been successfully used to capture the local features, yet they are not effective in capturing the global features in long-text. We investigate graph-level extensions to such models and propose a novel architecture for combining alternative text features. It uses an attention mechanism to dynamically decide how much information to use from a sequence- or graph-level component. We evaluated different architectures on a range of text classification datasets, and graph-level extensions were found to improve performance on most benchmarks. In addition, the attention-based architecture, as adaptively-learned from the data, outperforms the generic and fixed-value concatenation ones.

Efficiently answering reachability queries on a directed graph is a fundamental problem and many solutions - theoretical and practical - have been proposed. A common strategy to make reachability query processing efficient, accurate and scalable is to precompute indexes on the graph. However this often becomes impractical, particularly when dealing with large graphs that are highly dynamic or when queries have additional constraints known only at the time of querying. In the former case, indexes become stale very quickly and keeping them up to date at the same speed as changes to the graph is untenable. For the latter setting, currently proposed indexes are often quite bulky and are highly customized to handle only a small class of constraints. In this paper, we propose a first practical attempt to address these issues by abandoning the traditional indexing approach altogether and operating directly on the graph as it evolves. Our approach, called ARROW, uses random walks to efficiently approximate reachability between vertices, building on ideas that have been prevalent in the theory community but ignored by practitioners. Not only is ARROW well suited for highly dynamic settings - as it is index-free, but it can also be easily adapted to handle many different forms of ad-hoc constraints while being competitive with custom-made index structures. In this paper, we show that ARROW, despite its simplicity, is near-accurate and scales to graphs with tens of millions of vertices and hundreds of millions of edges. We present extensive empirical evidence to illustrate these advantages.

In this paper, we address the problem of sampling from a set and reconstructing a set stored as a Bloom filter. To the best of our knowledge our work is the first to address this question. We introduce a novel hierarchical data structure called BloomSampleTree that helps us design efficient algorithms to extract an almost uniform sample from the set stored in a Bloom filter and also allows us to reconstruct the set efficiently. In the case where the hash functions used in the Bloom filter implementation are partially invertible, in the sense that it is easy to calculate the set of elements that map to a particular hash value, we propose a second, more space-efficient method called HashInvert for the reconstruction. We study the properties of these two methods both analytically as well as experimentally. We provide bounds on run times for both methods and sample quality for the BloomSampleTree based algorithm, and show through an extensive experimental evaluation that our methods are efficient and effective.

In this paper, we address the problem of sampling from a set and reconstructing a set stored as a Bloom filter. To the best of our knowledge our work is the first to address this question. We introduce a novel hierarchical data structure called BloomSampleTree that helps us design efficient algorithms to extract an almost uniform sample from the set stored in a Bloom filter and also allows us to reconstruct the set efficiently. In the case where the hash functions used in the Bloom filter implementation are partially invertible, in the sense that it is easy to calculate the set of elements that map to a particular hash value, we propose a second, more space-efficient method called HashInvert for the reconstruction. We study the properties of these two methods both analytically as well as experimentally. We provide bounds on run times for both methods and sample quality for the BloomSampleTree based algorithm, and show through an extensive experimental evaluation that our methods are efficient and effective.

We describe a dynamic graph generator with overlapping communities that is capable of simulating community scale events while at the same time maintaining crucial graph properties. Such a benchmark generator is useful to measure and compare the responsiveness and efficiency of dynamic community detection algorithms. Since the generator allows the user to tune multiple parameters, it can also be used to test the robustness of a community detection algorithm across a spectrum of inputs. In an experimental evaluation, we demonstrate the generator's performance and show that graph properties are indeed maintained over time. Further, we show that standard community detection algorithms are able to find the generated community structure. To the best of our knowledge, this is the first time that all of the above have been combined into one benchmark generator, and this work constitutes an important building block for the development of efficient and reliable dynamic, overlapping community detection algorithms.

We consider the problem of scheduling jobs that require multiple resources such as memory, bandwidth and processors. For each job, the input specifies start time, finish time and profit; the input also specifies the job's requirement for each resource. Each resource has a fixed capacity (called bandwidth). A feasible solution is a subset of jobs such that for any timeslot and any resource, the total requirement of the jobs active at the timeslot does not exceed the capacity of the resource. The goal is to maximize the profit of the jobs selected. We present an approximation algorithm with provable guarantees and effective heuristics for this problem. The algorithm has an approximation ratio of O(r), where r is the number of resources. We present an experimental evaluation of our algorithms that exhibit their effectiveness.