Machine learning (ML) training data is often scattered across disparate collections of datasets, called data silos. This fragmentation poses a major challenge for data-intensive ML applications: integrating and transforming data residing in different sources demand a lot of manual work and computational resources. With data privacy and security constraints, data often cannot leave the premises of data silos, hence model training should proceed in a decentralized manner. In this work, we present a vision of how to bridge the traditional data integration (DI) techniques with the requirements of modern machine learning. We explore the possibilities of utilizing metadata obtained from data integration processes for improving the effectiveness and efficiency of ML models. Towards this direction, we analyze two common use cases over data silos, feature augmentation and federated learning. Bringing data integration and machine learning together, we highlight new research opportunities from the aspects of systems, representations, factorized learning and federated learning.

Data lakes are becoming increasingly prevalent for Big Data management and data analytics. In contrast to traditional 'schema-on-write' approaches such as data warehouses, data lakes are repositories storing raw data in its original formats and providing a common access interface. Despite the strong interest raised from both academia and industry, there is a large body of ambiguity regarding the definition, functions and available technologies for data lakes. A complete, coherent picture of data lake challenges and solutions is still missing. This survey reviews the development, architectures, and systems of data lakes. We provide a comprehensive overview of research questions for designing and building data lakes. We classify the existing approaches and systems based on their provided functions for data lakes, which makes this survey a useful technical reference for designing, implementing and deploying data lakes. We hope that the thorough comparison of existing solutions and the discussion of open research challenges in this survey will motivate the future development of data lake research and practice.

Data science workflows often require extracting, preparing and integrating data from multiple data sources. This is a cumbersome and slow process: most of the times, data scientists prepare data in a data processing system or a data lake, and export it as a table, in order for it to be consumed by a Machine Learning (ML) algorithm. Recent advances in the area of factorized ML, allow us to push down certain linear algebra (LA) operators, executing them closer to the data sources. With this work, we revisit classic data integration (DI) systems and see how these fit into modern data lakes that are meant to support LA as a first-class citizen.

Schema matching is the process of capturing correspondence between attributes of different datasets and it is one of the most important prerequisite steps for analyzing heterogeneous data collections. State-of-the-art schema matching algorithms that use simple schema- or instance-based similarity measures struggle with finding matches beyond the trivial cases. Semantics-based algorithms require the use of domain-specific knowledge encoded in a knowledge graph or an ontology. As a result, schema matching still remains a largely manual process, which is performed by few domain experts. In this paper we present the Relational Embeddings MAtcher, or rema, for short. rema is a novel schema matching approach which captures semantic similarity of attributes using relational embeddings: a technique which embeds database rows, columns and schema information into multidimensional vectors that can reveal semantic similarity. This paper aims at communicating our latest findings, and at demonstrating rema's potential with a preliminary experimental evaluation.

In modern enterprises, both operational and organizational data is typically spread across multiple heterogeneous systems, databases and file systems. Recognizing the value of their data assets, companies and institutions construct data lakes, storing disparate datasets from dierent departments and systems. However, for those datasets to become useful, they need to be cleaned and integrated. Data can be well documented, structured and encoded in dierent schemata, but also unstructured with implicit, human-understandable semantics. Due to the sheer scale of the data itself but also the multitude of representations and schemata, data integration techniques need to scale without relying heavily on human labor. Existing integration approaches fail to address hidden semantics without human input or some form of ontology, making large scale integration a daunting task. The goal of my doctoral work is to devise scalable data integration methods, employing modern machine learning to exploit semantics and facilitate discovery of novel relationship types. In order to capture semantics with minimal human intervention, we propose a new approach which we call Data as a Language (DaaL). By leveraging embeddings from the Natural Language Processing (NLP) literature, DaaL aims at extracting semantics from structured and semi-structured data, allowing the exploration of relevance and similarity among dierent data sources. This paper discusses existing data integration mechanisms and elaborates on how NLP techniques can be used in data integration, alongside challenges and research directions.