With large amounts of data continuously generated by intelligence devices, efficient analysis of huge data collections to unearth valuable insights has become one of the most elusive challenges for both academia and industry. The key elements to establishing a scalable analyzing framework should involve (1) an intuitive interface to describe the desired outcome, (2) a well-crafted model that integrates all available information sources to derive the optimal outcome and (3) an efficient algorithm that performs the data integration and extraction within a reasonable amount of time. In this dissertation, we address these challenges by proposing (1) a cross-language interface for a succinct expression of recursive queries, (2) a domain specific neural network model that can incorporate information of multiple modalities, and (3) a sample efficient training method that can be used even for extremely-large output-class classifiers.
Our contributions in this thesis are thus threefold: First, for the ubiquitous recursive queries in advanced data analytics, on top of BigDatalog and Apache Spark, we design a succinct and expressive analytics tool encapsulating the functionality and classical algorithms of Datalog, a quintessential logic programming language. We provide the Logical Library (LLib), a Spark MLlib-like high-level API supporting a wide range of recursive algorithms and the Logical DataFrame (LFrame), an extension to Spark DataFrame supporting both relational and logical operations. The LLib and LFrame enable smooth collaborations between logical applications and other Spark libraries and cross-language logical programming in Scala, Java, or Python. Second, we utilize variants of recurrent neural network (RNN) to incorporate some enlightening sequential information overlooked by the conventional works in two different domains including Spoken Language Understanding (SLU) and Internet Embedding (IE). In SLU, we address the problem caused by solely relying on the first best interpretation (hypothesis) of an audio command through a series of new architectures comprising bidirectional LSTM and pooling layers to jointly utilize the other hypotheses' texts or embedding vectors, which are neglected but with valuable information missed by the first best hypothesis. In IE, we propose the DIP, an extension of RNN, to build up the internet coordinate system with the IP address sequences, which are also unnoticed in conventional distance-based internet embedding algorithms but encode structural information of the network. Both DIP and the integration of all hypotheses bring significant performance improvements for the corresponding downstream tasks. Finally, we investigate the training algorithm for multi-class classifiers with a large output-class size, which is common in deep neural networks and typically implemented as a softmax final layer with one output neuron per each class. To avoid expensive computing the intractable normalizing constant of softmax for each training data point, we analyze the well-known negative sampling and improve it to the amplified negative sampling algorithm, which gains much higher performance with lower training cost.