With the proliferation of large datasets, sampling has become pervasive in data analysis. Sampling has numerous benefits—from reducing the computation time and cost to increasing the scope of interactive analysis. A popular task in data science, well-suited toward sampling, is the computation of fast-but-approximate aggregations over sampled data. Aggregation is a foundational block of data analysis, with data cube being its primary construct. We observe that such aggregation queries are typically issued in an ad-hoc, interactive setting. In contrast to one-off queries, a typical query session consists of a series of quick queries, interspersed with the user inspecting the results and formulating the next query. The similarity between session queries opens up opportunities for reusing computation of not just query results, but also error estimates. Error estimates need to be provided alongside sampled results for the results to be meaningful. We propose Sesame , a rewrite and caching framework that accelerates the entire interactive session of aggregation queries over sampled data. We focus on two unique and computationally expensive aspects of this use case: query speculation in the presence of sampling, and error computation, and provide novel strategies for result and error reuse. We demonstrate that our approach outperforms conventional sampled aggregation techniques by at least an order of magnitude, without modifying the underlying database.

Supporting sampling in the presence of joins is an important problem in data analysis, but is inherently challenging due to the need to avoid correlation between output tuples. Current solutions provide either correlated or non-correlated samples. Sampling might not always be feasible in the non-correlated sampling-based approaches -- the sample size or intermediate data size might be exceedingly large. On the other hand, a correlated sample may not be representative of the join. This paper presents a unified strategy towards join sampling, while considering sample correlation every step of the way. We provide two key contributions. First, in the case where a correlated sample is acceptable, we provide techniques, for all join types, to sample base relations so that their join is as random as possible. Second, in the case where a correlated sample is not acceptable, we provide enhancements to the state-of-the-art algorithms to reduce their execution time and intermediate data size.

Modern computing devices and user interfaces have necessitated highly interactive querying. Some of these interfaces issue a large number of dynamically changing and continuous queries to the backend. In others, users expect to inspect results during the query formulation process, in order to guide or help them towards specifying a full-fledged query. Thus, users end up issuing a fast-changing workload to the underlying database. In such situations, the user's query intent can be thought of as being in flux. In this paper, we show that the traditional query execution engines are not well-suited for this new class of highly interactive workloads. We propose a novel model to interpret the variability of likely queries in a workload. We implemented a cyclic scan-based approach to process queries from such workloads in an efficient and practical manner while reducing the overall system load. We evaluate and compare our methods with traditional systems and demonstrate the scalability of our approach, enabling thousands of queries to run simultaneously within interactive response times given low memory and CPU requirements.

Interactive ad-hoc analytics over large datasets has become an increasingly popular use case. We detail the challenges encountered when building a distributed system that allows the interactive exploration of a data cube. We introduce DICE, a distributed system that uses a novel session-oriented model for data cube exploration, designed to provide the user with interactive sub-second latencies for specified accuracy levels. A novel framework is provided that combines three concepts: faceted exploration of data cubes, speculative execution of queries and query execution over subsets of data. We discuss design considerations, implementation details and optimizations of our system. Experiments demonstrate that DICE provides a sub-second interactive cube exploration experience at the billion-tuple scale that is at least 33% faster than current approaches.

The interactive exploration of data cubes has become a popular application, especially over large datasets. In this paper, we present DICE , a combination of a novel frontend query interface and distributed aggregation backend that enables interactive cube exploration. DICE provides a convenient, practical alternative to the typical offline cube materialization strategy by allowing the user to explore facets of the data cube, trading off accuracy for interactive response-times, by sampling the data. We consider the time spent by the user perusing the results of their current query as an opportunity to execute and cache the most likely followup queries. The frontend presents a novel intuitive interface that allows for sampling-aware aggregations, and encourages interaction via our proposed faceted model. The design of our backend is tailored towards the low-latency user interaction at the frontend, and vice-versa. We discuss the synergistic design behind both the frontend user experience and the backend architecture of DICE ; and, present a demonstration that allows the user to fluidly interact with billion-tuple datasets within sub-second interactive response times.

This paper presents a multi-modal emotion recognition framework that is capable of estimating the human emotional state through analyzing and fusing a number of non-invasive external cues. The proposed framework consists of a set of data analysis, feature extraction and emotion recognition modules for processing heterogeneous sensory data (e.g., visual appearance and speech) and a novel probabilistic information fusion model to accurately estimate the human emotional state. Experimental results demonstrate that the proposed emotion recognition framework can automatically and robustly recognize human emotional states. Our results also proof that by fusing complementary information such as facial expression analysis and voice intonation analysis results, the emotion recognition performance can be boosted and outperform each individual modal analysis. The proposed emotion recognition framework can be integrated into existing Intelligent Tutoring Systems (ITSs) for improving the effectiveness of the learning systems by providing feedbacks to the ITSs.