Similarity join, which can find similar objects (e.g., products, names, addresses) across different sources, is powerful in dealing with variety in big data, especially web data. Threshold-driven similarity join, which has been extensively studied in the past, assumes that a user is able to specify a similarity threshold, and then focuses on how to efficiently return the object pairs whose similarities pass the threshold. We argue that the assumption about a well set similarity threshold may not be valid for two reasons. The optimal thresholds for different similarity join tasks may vary a lot. Moreover, the end-to-end time spent on similarity join is likely to be dominated by a back-and-forth threshold-tuning process.

In this paper we consider the problem of mining frequently occurring interesting phrases in large document collections in an ad-hoc fashion. Ad-hoc refers to the ability to perform such analyses over text corpora that can be an arbitrary subset of a global set of documents. Most of the times the identification of these ad-hoc document collections is driven by a user or application defined query with the aim of gathering statistics describing the sub-collection, as a starting point for further data analysis tasks. Our approach to mine the top-k most interesting phrases consists of a novel indexing technique, called Sequence Pattern Indexing (SeqPattIndex), that benefits from the observation that phrases often overlap sequentially. We devise a forest based index for phrases and an further improved version with additional redundancy elimination power. The actual top-k phrase mining algorithm operating on these indices is a combination of a simple merge join and inspired by the pattern-growth framework from the data mining community, making use of early termination and search space pruning technologies that enhance the runtime performance. Overall, our approach has on average a lower index space consumption as well as a lower runtime for the top-k phrase mining task, as we demonstrate in the experimental evaluation using real-world data.

As a typical data mining research topic, sequential pattern mining has been studied extensively for the past decade. Recently, mining various sequential patterns incrementally over stream data has raised great interest. Due to the challenges of mining stream data, many difficulties not so obvious in static data mining have to be reconsidered carefully. In this paper, we propose a novel algorithm which stores only frequent closed prefixes in its enumeration tree structure, used for mining and maintaining patterns in the current sliding window, to solve the frequent closed sequential pattern mining problem efficiently over stream data. Some effective search space pruning and pattern closure checking strategies have been also devised to accelerate the algorithm. Experimental results show that our algorithm outperforms other state-of-the-art algorithm significantly in both running time and memory use.

In this demonstration, we propose an interactive query completion system on structural data like DBLP, called SEQUEL. It is novel in several aspects: with patterns mined on the structural data using newly devised algorithm, SEQUEL offers high-utility completions composed with not only words but also phrases, and requires no explicit indications of corresponding columns. Instead of using query logs exploited previously for unstructured data, more effective completions are provided based on patterns mined directly from the records. Moreover, an effective index structure helps SEQUEL respond fast at millisecond level for each keystroke.

Classification is one of the most essential tasks in data mining. Unlike other methods, associative classification tries to find all the frequent patterns existing in the input categorical data satisfying a user-specified minimum support and/or other discrimination measures like minimum confidence or information-gain. Those patterns are used later either as rules for rule-based classifier or training features for support vector machine (SVM) classifier, after a feature selection procedure which usually tries to cover as many as the input instances with the most discriminative patterns in various manners. Several algorithms have also been proposed to mine the most discriminative patterns directly without costly feature selection. Previous empirical results show that associative classification could provide better classification accuracy over many datasets.

Mining generator patterns has raised great research interest in recent years. The main purpose of mining itemset generators is that they can form equivalence classes together with closed itemsets, and can be used to generate simple classification rules according to the MDL principle. In this paper, we devise an efficient algorithm called StreamGen to mine frequent itemset generators over a stream sliding window. We adopt a novel enumeration tree structure to help keep the information of mined generators and the border between generators and non-generators, and propose some optimization techniques to speed up the mining process. We further extend the algorithm to directly mine a set of high quality classification rules over stream sliding windows while keeping high performance. The extensive performance study shows that our algorithm outperforms other state-of-the-art algorithms which perform similar tasks in terms of both runtime and memory usage efficiency, and has high utility in terms of classification.

Sequential pattern mining has raised great interest in data mining research field in recent years. However, to our best knowledge, no existing work studies the problem of frequent sequence generator mining. In this paper we present a novel algorithm, FEAT (abbr. Frequent sEquence generATor miner), to perform this task. Ex-perimental results show that FEAT is more efficient than traditional sequential pattern mining algorithms but generates more concise re-sult set, and is very effective for classifying Web product reviews.