An important class of database queries in machine learning and data science workloads is the opaque filter query: a query with a selection predicate that is implemented with a UDF, with semantics that are unknown to the query optimizer. Some typical examples would include a CNN-style trained image classifier, or a textual sentiment classifier. Because the optimizer does not know the predicate's semantics, it cannot employ standard optimizations, yielding long query times. We propose voodoo indexing, a two-phase method for optimizing opaque filter queries. Before any query arrives, the method builds a hierarchical "query-independent" index of the database contents, which groups together similar objects. At query-time, the method builds a map of how much each group satisfies the predicate, while also exploiting the map to accelerate execution. Unlike past methods, voodoo indexing does not require insight into predicate semantics, works on any data type, and does not require in-query model training. We describe both standalone and SparkSQL-specific implementations, plus experiments on both image and text data, on more than 100 distinct opaque predicates. We show voodoo indexing can yield up to an 88% improvement over standard scan behavior, and a 79% improvement over the previous best method adapted from research literature.

Querying the content of images and video requires expensive content extraction methods. Modern extraction techniques are based on deep convolutional neural networks (CNNs) and can classify objects within images with astounding accuracy. Unfortunately, these methods are slow: processing a single image can take about 10 milliseconds on modern GPU-based hardware. As massive video libraries become ubiquitous, running a content-based query over millions of video frames is prohibitive. One promising approach to reduce the runtime cost of queries of visual content is to use a hierarchical model, such as a cascade, where simple cases are handled by an inexpensive classifier. Prior work has sought to design cascades that optimize the computational cost of inference by, for example, using smaller CNNs. However, we observe that there are critical factors besides the inference time that dramatically impact the overall query time. Notably, by treating the physical representation of the input image as part of our query optimization-that is, by including image transformations such as resolution scaling or color-depth reduction within the cascade-we can optimize data handling costs and enable drastically more efficient classifier cascades. In this paper, we propose TAHOMA, which generates and evaluates many potential classifier cascades that jointly optimize the CNN architecture and input data representation. Our experiments on a subset of ImageNet show that TAHOMA's input transformations speed up cascades by up to 35 times. We also find up to a 98x speedup over the ResNet50 classifier with no loss in accuracy and a 280x speedup if some accuracy is sacrificed.

Data transformation is a critical first step in modern data analysis: before any analysis can be done, data from a variety of sources must be wrangled into a uniform format that is amenable to the intended analysis and analytical software package. This data transformation task is tedious, time-consuming, and often requires programming skills beyond the expertise of data analysts. In this paper, we develop a technique to synthesize data transformation programs by example, reducing this burden by allowing the analyst to describe the transformation with a small input-output example pair, without being concerned with the transformation steps required to get there. We implemented our technique in a system, FOOFAH, that efficiently searches the space of possible data transformation operations to generate a program that will perform the desired transformation. We experimentally show that data transformation programs can be created quickly with FOOFAH for a wide variety of cases, with 60% less user effort than the well-known WRANGLER system.

Advancements in new data analysis and visualization technologies have resulted in wide applicability of data-driven decision making. However, raw data from various sources must be wrangled into a suitable form before they are processed by the downstream data tools. People traditionally write data transformation programs to automate this process, and such work is cumbersome and tedious.

The application of machine learning to large datasets has become a vital component of many important and sophisticated software systems built today. Such trained systems are often based on supervised learning tasks that require features, signals extracted from the data that distill complicated raw data objects into a small number of salient values. A trained system's success depends substantially on the quality of its features. Unfortunately, feature engineering—the process of writing code that takes raw data objects as input and outputs feature vectors suitable for a machine learning algorithm—is a tedious, time-consuming experience. Because “big data” inputs are so diverse, feature engineering is often a trial-and-error process requiring many small, iterative code changes. Because the inputs are so large, each code change can involve a time-consuming data processing task (over each page in a Web crawl, for example). We introduce Zombie, a data-centric system that accelerates feature engineering through intelligent input selection, optimizing the “inner loop” of the feature engineering process. Our system yields feature evaluation speedups of up to 8× in some cases and reduces engineer wait times from 8 to 5 hours in others.

Social media nowcasting, the process of estimating real-world phenomena from social media data, has grown in popularity over the last several years as an alternative to traditional data collection methods like phone surveys. Unfortunately, current nowcasting methods depend on pre-existing, traditionally collected survey data as an aid to sift through the huge number of signals that can be derived from social media. This dependence severely limits the applicability of current nowcasting techniques. If we could remove this need for conventional data, social media signals could describe a much wider range of target phenomena. We have built a nowcasting querying system that estimates real-world phenomena without requiring any conventional data, relying instead upon an interactive exploration with users. Specifically, our system exploits a user-provided multi-part query consisting of semantic and signal components. The user can explore in real time the tradeoff between these two components to find the most relevant social media signals to estimate the target phenomenon. Our demonstration system lets users search for signals within a large Twitter corpus using a dynamic web-based interface. Also, users can share results with the general public, review and comment on others' shared results, and clone these results as starting points for further exploration and querying.

Nowcasting is the practice of using social media data to quantify ongoing real-world phenomena. It has been used by researchers to measure flu activity, unemployment behavior, and more. However, the typical nowcasting workflow requires either slow and tedious manual searching of relevant social media messages or automated statistical approaches that are prone to spurious and low-quality results. In this paper, we propose a method for declaratively specifying a nowcasting model; this method involves processing a user query over a very large social media database, which can take hours. Due to the human-in-the-loop nature of constructing nowcasting models, slow runtimes place an extreme burden on the user. Thus we also propose a novel set of query optimization techniques, which allow users to quickly construct nowcasting models over very large datasets. Further, we propose a novel query quality alarm that helps users estimate phenomena even when historical ground truth data is not available. These contributions allow us to build a declarative nowcasting data management system, R accoon DB, which yields high-quality results in interactive time. We evaluate R accoon DB using 40 billion tweets collected over five years. We show that our automated system saves work over traditional manual approaches while improving result quality---57% more accurate in our user study---and that its query optimizations yield a 424x speedup, allowing it to process queries 123x faster than a 300-core Spark cluster, using only 10% of the computational resources.

The application of machine learning to large datasets has become a core component of many important and exciting software systems being built today. The extreme value in these trained systems is tempered, however, by the difficulty of constructing them. As shown by the experience of Google, Netflix, IBM, and many others, a critical problem in building trained systems is that of feature engineering. High-quality machine learning features are crucial for the system's performance but are difficult and time-consuming for engineers to develop. Data-centric developer tools that improve the productivity of feature engineers will thus likely have a large impact on an important area of work. We have built a demonstration integrated development environment for feature engineers. It accelerates one particular step in the feature engineering development cycle: evaluating the effectiveness of novel feature code. In particular, it uses an index and runtime execution planner to process raw data objects ( e.g. , Web pages) in order of descending likelihood that the data object will be relevant to the user's feature code. This demonstration IDE allows the user to write arbitrary feature code, evaluate its impact on learner quality, and observe exactly how much faster our technique performs compared to a baseline system.

Stacked sparse denoising auto-encoders (SSDAs) have recently been shown to be successful at removing noise from corrupted images. However, like most denoising techniques, the SSDA is not robust to variation in noise types beyond what it has seen during training. We present the multi-column stacked sparse denoising autoencoder, a novel technique of combining multiple SSDAs into a multi-column SSDA (MC-SSDA) by combining the outputs of each SSDA. We eliminate the need to determine the type of noise, let alone its statistics, at test time. We show that good denoising performance can be achieved with a single system on a variety of different noise types, including ones not seen in the training set. Additionally, we experimentally demonstrate the efficacy of MC-SSDA denoising by achieving MNIST digit error rates on denoised images at close to that of the uncorrupted images.

No abstract available.

No abstract available.