Single-source SimRank query calculates the similarity between a query node and every node in a graph, which traverses the paths starting from the query node for similarity computation. However, the scale of the paths increases exponentially as path length increases, which decreases the computation efficiency. Sampling-based algorithms reduce computational cost by path sampling, but they need to sample sufficient paths to ensure the accuracy, and the performance might be affected by the large scale of paths. In this paper, we propose VecSim for efficient single-source SimRank query by path aggregation. VecSim first aggregates the paths starting from query node with common arrived nodes step by step to obtain the hitting probabilities, and then aggregates the paths starting from the arrived nodes reversely to obtain the first-meeting probabilities in a similar way, in which only several vectors are maintained. The extra-meeting probabilities are excluded from each step, and an efficient sampling-based algorithm is designed, which estimates the extra-meeting probabilities by sampling paths within a specified length. For further speeding up query processing, we propose a threshold-sieved algorithm, which prunes the entries with small values that contribute little to the final similarity scores by setting a threshold. Extensive experiments are done on four small and four large graphs, which demonstrate that VecSim outperforms the competitors in terms of time and space costs on a comparable accuracy. In particular, VecSim achieves an empirical error of 10-4 level in under 0.1 second over all of these graphs.

Link-based similarity computation arises in many real applications, including web search, clustering and recommender system. Lots of similarity measures are devoted recently, but there is one undesirable drawback, called ''path missing'' issue, i.e., the paths between objects are not fully considered for similarity computation. For example, SimRank considers only in-coming paths of equal length from a common ''center'' object, and a large portion of other paths are fully neglected. A comprehensive measure can be modeled by tallying all the possible paths between objects, but a large number of traverses would be required for these paths to fetch the similarities, which might increase the computational difficulty. In this paper, we propose a comprehensive similarity measure, namely RG-SimRank (Random surfer Graph-based SimRank), which resolves the "path missing'' issue with inheriting the philosophy of SimRank. We build a random surfer graph by allowing the surfer to stay at current object, go to other objects against in-links or along out-links. RG-SimRank adopts SimRank to compute similarities in random surfer graph instead of the original network, which has a same form of SimRank and hence inherits the optimization techniques on similarity computation. We prove that RG-SimRank considers all the possible paths of any direction and any length. And it provides a general solution to assess similarities, under which lots of existing similarity measures become its special cases. Other similarity measures besides SimRank can also be enhanced similarly using random surfer graph. Extensive experiments on real datasets demonstrate the performance of the proposed approach.

Prediction on electricity load time series of commercial building and construction sets plays an important role in smart grid applications, building operation and environmental protection, etc. It differs significantly from the traditional aggregated electricity time series prediction problem and aims to provide more accurate prediction for each building respectively. In recent years, deep learning shows significant improvements on time series prediction over classical methods. However, it suffers from the problem of limited data. The problem is more serious in the context of electricity time series prediction of buildings. For instance, new buildings or buildings with newly installed meters have limited historical data, which results in poor model training and forecasts. In this paper we propose a novel generative adversarial learning based prediction method for dealing with this problem. We employ the generative adversarial network framework to obtain adversarially generated data for prediction model training by using long short-term memory units based recurrent neural network and attention mechanism. Extensive experiments on real-life dataset demonstrate the improvements of our proposed method against the traditional methods.

Tag generation aims to find relevant tags for a given entity, which has numerous applications, such as classification, information retrieval and recommender system. Practically, the data of real applications is sparse and lacks sufficient description for entities, which might lead to incomprehensive results. Random walk with restart (RWR) can find the hidden relationship between nodes by utilizing indirect connections. However, the traditional RWR computation is based on the whole structure of the given network, which maintains a matrix for storing all relevances between nodes. And the efficiency problem would be run into as network grows large. In this paper, we propose a top-k tag generation algorithm, namely DRWR, for efficiently generating the tags from entity-term network. The terms are treated as candidate tags, and the most relevant terms are treated as the tags for a given entity. The relevance computation between entity and terms is divided into two stages: off-line stage and on-line stage. In off-line stage, the relevances between terms are computed over the term-term network that is built based on the whole structure of entity-term network. In on-line stage, the relevances between entity and each term are computed based on the relevances between terms. For supporting fast on-line query processing, we develop a pruning algorithm, which skips the operations on relevances between terms smaller than a threshold. Extensive experiments on real datasets demonstrate the efficiency and effectiveness of the proposed approach.

P-Rank is a simple and captivating link-based similarity measure that extends SimRank by exploiting both in- and out-links for similarity computation. However, the existing work of P-Rank computation is expensive in terms of time and space cost and cannot efficiently support similarity computation in large information networks. For tackling this problem, in this paper, we propose an optimization technique for fast P-Rank computation in information networks by adopting the spiritual of partial sums. We write P-Rank equation based on partial sums and further approximate this equation by setting a threshold for ignoring the small similarity scores during iterative similarity computation. An optimized similarity computation algorithm is developed, which reduces the computation cost by skipping the similarity scores smaller than the give threshold during accumulation operations. And the accuracy loss estimation under the threshold is given through extensive mathematical analysis. Extensive experiments demonstrate the effectiveness and efficiency of our proposed approach through comparing with the straightforward P-Rank computation algorithm.

Diversifying query suggestions has emerged recently, by which the recommended queries can be both relevant and diverse. Most existing works diversify suggestions by query log analysis, however, for structured data, not all query logs are available. To this end, this paper studies the problem of suggesting diverse query terms by using topics from Wikipedia. Wikipedia is a successful online encyclopedia, and has high coverage of entities and concepts. We first obtain all relevant topics from Wikipedia, and then map each term to these topics. As the mapping is a nontrivial task, we leverage information from both Wikipedia and structured data to semantically map each term to topics. Finally, we propose a fast algorithm to efficiently generate the suggestions. Extensive evaluations are conducted on a real dataset, and our approach yields promising results.

With the social networks (SNs) becoming ubiquitous and massive, the issue of similarity computation among entities becomes more challenging and draws extensive interests from various research fields. SimRank is a well known similarity measure, however it considers only the meetings between two nodes that walk along equal length paths since the path length increases strictly with the iteration increasing during the similarity computation, besides, it does not differentiate importance for each link. In this paper, we propose a novel structural similarity measure, E-Rank (Entity Rank), towards effectively computing the structural similarity of entities in SNs, based on the intuition that two entities are similar if they can arrive at common entities. E-Rank can be well applied to social networks for measuring similarities of entities. Extensive experiments demonstrate the effectiveness of E-Rank by comparing with the state-of-the-art measures.