Tabular metadata (i.e., attributes in a table) identification and classification is a fundamental problem in large-scale data management of structured corpora, especially for complex tables rich in multi-level hierarchical metadata with nesting. Medical, security, data science research literature, Web tables, contain thousands of such complex tables, but often lack or incorrectly label their complex metadata. In this work, we describe an unsupervised, scalable, contrastive-learning approach for classification of multi-layer, hierarchical metadata in such tables. We compared it to the state of the art (SOTA) as well as the latest Large Language Models (LLMs), such as OpenAI GPT 3.5 and 4 with and without Retrieval Augmented Generation (RAG) on several large-scale heterogeneous datasets. We outperform SOTA and LLMs in classifying horizontal metadata (HMD) of deep levels (3–5) and for all levels (1–3) of vertical metadata (VMD). For HMD levels 1–2, SOTA outperforms us insignificantly, with a delta of ≈1%. LLMs with/without RAG slightly outperform us with deltas of 4–5% in accuracy for HMD level 1, but we significantly outperformed LLMs/LLMs+RAG with delta up to 29% for all other levels 2–5 HMD and up to 87% delta for VMD.

Here, we describe one of the first Web-scale hybrid Knowledge Graph (KG)-Large Language Model (LLM), populated with the latest peer-reviewed medical knowledge on colorectal Cancer. It is currently being evaluated to assist with both medical research and clinical information retrieval tasks at Moffitt Cancer Center and Research Institute, which is one of the top Cancer centers in the U.S. and in the world. Our hybrid is remarkable as it serves the user needs better than just an LLM, KG or a search-engine in isolation. LLMs as is are known to exhibit hallucinations and catastrophic forgetting as well as are trained on outdated corpora. The state of the art KGs, such as PrimeKG, cBioPortal, ChEMBL, NCBI, and other require manual curation, hence are quickly getting stale. CancerKG is unsupervised and is capable of automatically ingesting and organizing the latest medical findings. To alleviate the LLMs shortcomings, the verified KG serves as a Retrieval Augmented Generation (RAG) guardrail. CancerKG exhibits 5 different advanced user interfaces, each tailored to serve different data modalities better and more convenient for the user. We evaluated CancerKG on real user queries and report a high NDCG score on a large-scale corpora of approximately 44K publications.

Data profiling is a "set of statistical data analysis activities to determine properties of a dataset". Historically, it was aimed at data (not meta-data), but at scale, the tables’ meta-data (i.e. title, attribute names, types) becomes abundant, hence its profiling becomes vital, especially in order to understand the contents of large-scale structured datasets.Here we describe and evaluate the algorithms and models behind our scalable Meta-data profiler. It is capable of learning Meta-profiles for a topic of interest in extreme-scale structured datasets, such as WDC [1] or CORD-19 [2] having millions of tables and hundreds of thousands of sources. A 3D Meta-profile visualizes a specific topic (e.g. COVID-19 vaccine side-effects) present in a large-scale structured dataset and simplifies access and comparison for data scientists and end-users.

Accessing large-scale structured datasets such as WDC [31] or CORD-191 is very challenging [ 11, 13, 14, 41, 42]. Even if one topic (e.g. Vaccine Side-Effects) is of interest, the side-effects tables in different papers have hundreds of different schemas, depending on the authors, which significantly complicates both finding and querying them. Here we demonstrate our scalable Meta-data profiler, capable of constructing a standardized interface to a topic of interest in large-scale structured datasets. This interface, called Meta-profile represents a meta-data summary per each topic, representative of the entire dataset. Such profiles can be used as a robust visualization as well as to simplify access to structured data for both data scientists and end users at scale [32, 42]

Accessing large-scale structured datasets such as WDC [21], having millions of tables coming from hundreds of thousands of sources is very challenging [11, 13, 14, 30, 31]. Even if one topic (e.g. Job postings) is of interest, Jobs tables in different sources have hundreds of different schemas, which significantly complicates both finding and querying them.

Contextual embeddings, such as ELMo and BERT [21], [49], assign each word a representation based on its context [44]. This research builds on an observation that the context for structured data can be encoded very differently from the traditional sentence and text-based context. This means that embeddings for the structured data can be constructed on different principles, and if properly optimized, they can be used to improve performance ML and AI-based tasks related to the structured data. Here we present several new types of tabular embeddings taking into consideration structure of columns, rows, and presence of metadata. We demonstrated that properly optimized embeddings in combination with ML and DL models show significant improvement on such important tasks as tabular column and tuple recognition.

Anecdotal evidence suggests that Web-search engines, together with the Knowledge Graphs and Bases, such as YAGO [46], DBPedia [13], Freebase [16], Google Knowledge Graph [52] provide rapid access to most structured information on the Web. However, taking a closer look reveals a so called "knowledge gap" [18] that is largely in the dark. For example, a person searching for a relevant job opening has to spend at least 3 hours per week for several months [2] just searching job postings on numerous online job-search engines and the employer websites. The reason why this seemingly simple task cannot be completed by typing in a few keyword queries into a search-engine and getting all relevant results in seconds instead of hours is because access to structured data on the Web is still rudimentary. While searching for a job we have many parameters in mind, not just the job title, but also, usually location, salary range, remote work option, given a recent shift to hybrid work places, and many others. Ideally, we would like to write a SQL-style query, selecting all job postings satisfying our requirements, but it is currently impossible, because job postings (and all other) Web tables are structured in many different ways and scattered all over the Web. There is neither a Web-scale generalizable algorithm nor a system to locate and normalize all relevant tables in a category of interest from millions of sources.Here we describe and evaluate on a corpus having hundreds of millions of Web tables [39], a new scalable iterative training data generation algorithm, producing high quality training data required to train Deep- and Machine-learning models, capable of generalizing to Web scale. The models, trained on such en-riched training data efficiently deal with Web scale heterogeneity compared to poor generalization performance of models, trained without enrichment [20], [25], [38]. Such models are instrumental in bridging the knowledge gap for structured data on the Web.

Correctly identifying the semantic label of a table column (e.g. artist ) or the domain of a tuple (e.g. Song) is crucial for data science tasks, such as schema matching, data cleaning and discovery. Existing data preparation and integration systems are known to make mistakes that need to be corrected by humans, which is labor-intensive and expensive especially at scale [19], [20], [25], [9], [8]. The accuracy is also known to suffer on dirty data. In this paper we define and evaluate tabular embeddings to help increase the accuracy.Embeddings, is a well-known dimensionality reduction technique, usually applied to represent (one-dimensional) text documents/sentences as vectors for further analytics [28], [7], [23]. Using embeddings lowers dimensionality, as well as the models trained with embeddings often exhibit higher accuracy, compared to the same trained without such, because the embeddings vectors store the context information. Our 2-dimensional embeddings are more suitable for tables rather than text. We justify their efficiency on fundamental tasks on tables such as classifying columns and tuples.We perform an extensive experimental evaluation and compare the Neural Network trained with tabular embeddings against the same trained without such embeddings. We report significant accuracy gains in tuple classification when using the Neural Network trained with our tabular embeddings - up to 17.6% delta in F-measure for Songs compared with the same without our embeddings. For training and evaluating all our embeddings and models, we use a large-scale WebTables dataset having ≈15 million tables coming from ≈ 248K English Web sources [12].

Schema matching, historically, is a sub-area of Data Integration, responsible for matching relational or semi-structured schemas to facilitate further data integration process. In a standard schema matching scenario, with two schemas, a semi-supervised matching algorithm would generate pairwise table and attribute matches. Having a correct mapping between two schemas enables many data integration scenarios, such as schema integration, data translation, schema evolution, mediated/global schema, reverse-engineering, and others [18], [20], [23], [29], [30].For Web scale datasets with millions of tables from hundreds of thousands of sources, for example WEBTABLES [11], schema matching in its classical format becomes computationally infeasible due to its quadratic complexity in the number of schemas. We make a step forward, by noticing, that such brute-force matching is no longer feasible, and is also not needed at scale. Instead, a scalable solution would be to match only the semantically relevant tables, which are much less numerous.WebLens, a scalable data integration system, first, trains Deep Learning models to find and match semantically similar tables, then derives mediated schemas for these subsets to enable uniform access to all relevant data. In this paper, we focus on a high-level description of the entire process and give an example of query processing. For all experiments in the paper, we use a large-scale structured dataset having more than 15 million of relational Web tables in English coming from more than 248 thousand of Web sources.

Here we present machine learning-based approach to automatic classify live and dead bacteria that can be used for rapid search for optimal antibiotics in case of bacterial infections. The patients must be promptly administered a most efficient medication because all delays significantly increase morbidity and mortality. We engineered a new technology allowing us to efficiently and rapidly capture bacterial cells from different biological samples and proceed with a rapid antibiotic susceptibility testing thereby bypassing the need to culture the bacterium. We developed a new machine learning and microscopy-based approach for rapid assessment of bacterial viability following tests with antibiotics. Also, we created a labeled dataset with ~100 images of live and dead bacteria stained with DAPI (DNA; blue) and FM4-64 (membrane; red) either treated with an antibiotic or untreated. We analyzed wild type (WT) and ampicillin-resistant (ampR) E. coli, WT and ampR S. aureus, and B. subtilis. For antibiotic susceptibility testing we used ampicillin, chloramphenicol and erythromycin. We extracted information about red and blue channels from the images and tried two machine learning classifiers for rapid assessment of viability of the bacteria. The classifiers Random Forest and J48 Decision Tree demonstrated precision 90.7% and 96%, recall 94.4% and 100%, and F-measure 92.5% and 95.2%, correspondingly, on 10-fold cross-validation.

Data profiling is a "set of statistical data analysis activities and processes to determine properties of a given dataset". Historically,most of the data profiling tasks were aimed at data. At scale, when a dataset has millions of tables, their meta-data (i.e. titles, attribute names and types) becomes abundant similar to data instances, and its profiling starts playing a vital role.

Anecdotal evidence suggests the Variety of Big data is one of the most challenging problems in Computer Science research today [1]. First, Big data arrives from a myriad of data sources, hence its shape and flavor differ. Second, hundreds of different Big data management systems support different APIs, storage/indexing schemes, and expose data to the users through their data model lens, each specific to their own system. All of these offer a significant impediment for Big data users who just want an easy to use interface to all relevant data regardless of its shape, format, size, and a back-end system used to store it. Naturally, these differences also complicate development of any analytical algorithms on top of large-scale, heterogeneous datasets. Here we describe HYBRID.POLY- a consolidated in-memory polystore engine [2], designed to support heterogeneous large-scale data and interactively process complex analytical work-loads. We execute and evaluate several popular analytical work-loads including Data Fusion, Machine Learning, and Music search at scale.

As data analytics has become an important application for modern data management systems, a new category of data management system has appeared recently: the scalable linear algebra system. In this paper, we argue that a parallel or distributed database system is actually an excellent platform upon which to build such functionality. Most relational systems already have support for cost-based optimization-which is vital to scaling linear algebra computations-and it is well-known how to make relational systems scale. We show that by making just a few changes to a parallel/distributed relational database system, such a system can be a competitive platform for scalable linear algebra. Taken together, our results should at least raise the possibility that brand new systems designed from the ground up to support scalable linear algebra are not absolutely necessary, and that such systems could instead be built on top of existing relational technology. Our results also suggest that if scalable linear algebra is to be added to a modern dataflow platform such as Spark, they should be added on top of the system's more structured (relational) data abstractions, rather than being constructed directly on top of the system's raw dataflow operators.

Here we discuss our initial efforts towards unsupervised clustering of a large-scale Web tables dataset. We improve our previous approach of weakly-supervised clustering, where an operator would provide a few descriptive keywords to generate an entity-identifying classifier, which is applied to the corpora to form a cohesive entity-centric cluster [1]. Here, we make a next step towards fully unsupervised algorithm by automatically generating these descriptive keywords. These keywords then can be used to generate high-precision training data and train a classifier to form a cluster. Here, we describe and evaluate this new unsupervised keyword generation algorithm and apply it to a large-scale Web tables corpus to form initial small high-precision clusters.

Anecdotal evidence suggests that Variety is one of the most challenging problems in Big data research [1]. Different data providers use different data models and formats to represent their data, which causes significant impediment to data scientists, whose goal is to make sense of all relevant data regardless of the source. Hybrid.Poly [2], [3] is the analytical polystore data management system designed to make all data accessible to the analyst, oblivious of the source differences.In this paper, we focus on the in-depth analysis and performance evaluation of the Linear Algebra extensions added to the Hybrid.Poly language.

It is difficult to access data generated by different data sources due to the representation and format differences. ETL, KETL, Jedox, Apatar are some examples of data translation and fusion packages that can be used to resolve representation differences of data coming from different data sources have been favored. However, most tools require significant manual effort to map the data sources. Structural mismatch of data between objects with the same semantics reduces the accessibility of data. Here we discuss our initial efforts toward a scalable unsu-pervised system and algorithms to generate Unified Famous Objects (UFO) - the self-learning “intelligent” data structures that help automate data fusion at scale [Gubanov et al., 2009], [Gubanov et al., 2011]. UFO is a data structure encapsulating different representations of the same data object (e.g. Songs), hence capable of automatically recognizing and mapping such object in different data sources, and significantly reducing manual effort during data integration process. We evaluate our algorithms on a large-scale Web tables corpus having ≈ 64 million of tuples.

Hybrid.poly is an in-memory polystore data management system, able to ingest various kinds of data and run complex analytical workloads on the ingested data [Gubanov, 2017], [Podkorytov et al., 2017]. Hybrid.JSON, a part of [Gubanov, 2017] focuses on ingesting and querying JSON documents in the polystore.

Recent advances in image recognition algorithms are making it possible to query databases for static images which look like a desired target image. This technology is expanding to image recognition within video files as image recognition algorithms advance. Blending the ideas of static image query with video image recognition leads to the subject of our present work. Hybrid.media is our work in progress to ingest video files into an Hybrid In-Memory Polystore Database [Gubanov, 2017], [Podkorytov et al., 2017] and add query support for finding those video files. This work will provide the basis for future work to query videos containing specific items based on content in a video.

Internet mail spam is a problem for most organizations and individuals. Spam can be classified into two categories: fraud and commercial. The fraud category includes phishing, scams, malware, counterfeit products and any other criminal activities. The commercial category includes promotional messages and newsletters that we do not want to receive, being sent illegally from legitimate organizations. Fraud can be seen as being a high threat with high volume while commercial spam is the opposite. Similar to mail, there are spam Web tables that do not have any useful content. Here we describe our machine-learning classifier for efficient and effective Web tables spam filtering that was tested on a large-scale Web tables corpus of ≈ 36 million tables.

The big data era brought us petabytes of data together with the challenges of storing and efficiently accessing largescale datasets. However, it unexpectedly surprised everyone with an enormous variety of data sources and types, and corresponding different data models. Dealing with a variety of those data models turned out to be a “hard nut to crack” for almost all existing data management engines. Data integration, a mature field addressing problems of accessing and fusing data residing in more than one datasource, over the years came up with feasible semi-automatic solutions, most of which efficiently handle a handful of data sources represented in one or two different data models (e.g. relational and semistructured) [Haas et al., 2005], [Gubanov et al., 2008]. While this is significant progress, most of the solutions do not easily scale up, since they usually require some sort of human assistance, infeasible at scale. Unified Famous Object (UFO) [Gubanov et al., 2009], [Bellahsene et al., 2011], [Gubanov et al., 2011a] was one of the first attempts to crack data fusion at scale by introducing a new abstraction called UFO that could incrementally learn different representations of data objects in different sources, and, over time, train itself to recognize and map them with high accuracy without supervision. Having trained many such UFOs, the system would get more and more powerful, as it learned to recognize and access data objects across many sources without supervision. While UFO was definitely progress towards scaling up data fusion, it was not a “silver bullet”, because it was built mostly for relational data. Hence, there is a need for a new large-scale data integration system that could handle many types of data at scale. This paper sketches its architecture, and envisions potential research challenges. First, a few key principles that such a system should follow are described, then follows the discussion on architecture and research avenues.

As data analytics has become an important application for modern data management systems, a new category of data management system has appeared recently: the scalable linear algebra system. In this paper, we argue that a parallel or distributed database system is actually an excellent platform upon which to build such functionality. Most relational systems already have support for cost-based optimization-which is vital to scaling linear algebra computations-and it is well-known how to make relational systems scale. We show that by making just a few changes to a parallel/ distributed relational database system, such a system can be a competitive platform for scalable linear algebra. Taken together, our results should at least raise the possibility that brand new systems designed from the ground up to support scalable linear algebra are not absolutely necessary, and that such systems could instead be built on top of existing relational technology. Our results also suggest that if scalable linear algebra is to be added to a modern dataflow platform such as Spark, they should be added on top of the system's more structured (relational) data abstractions, rather than being constructed directly on top of the system's raw dataflow operators.

Keyword-search engines (e.g. Web-search) usually can be outperformed by a specialized system optimized for a spe- cic domain, type of data, or queries (8, 2, 12, 5, 11, 9). For example, Halevy et. al. in (13) demonstrate how a special- ized Google Fusion Tables spatial search can outperform the general-purpose Google Web-search on bike trails search in San Francisco Bay Area. At the same time, Web content providers usually exhibit a specic focus for their postings. For example, information at http://www.csail.mit.edu is devoted to Computer Science research and education, Han- nah Montana is mostly tweeting about music, and the same is true for most sources. This paper describes the work in progress on a new Type- aware Web-search system that uses topical focus of informa- tion sources to process a large class of queries better than a regular Web search-engine. It leverages semantic proles similar to (10, 6, 7) and a new Type-aware Locality-Sensiti Hashing (TLSH) scheme to accomplish it.

Web-search engines usually can be outperformed by spe-cialized systems optimized for a specific domain or type of data. Halevy et al in [1] demonstrate a use case for a spe-cialized spatial search of Google Fusion Tables, whereby the

Large-scale text data research has recently started to regain momentum [1]-[10], because of the wealth of up to date information communicated in unstructured format. For example, new information in online media (e.g. Web blogs, Twitter, Facebook, news feeds, etc) becomes instantly available and is refreshed regularly, has very broad coverage and other valuable properties unusual for other data sources and formats. Therefore, many enterprises and individuals are interested in integrating and using unstructured text in addition to their structured data.

Relevance of search-results is a key factor for any search engine. In order to return and rank the Web-pages that are most relevant to the query, contemporary search engines use complex ranking functions that depend on hundreds of features. For example, presence or absence of the query keywords on the page, their proximity, frequencies, HTML markup are just a few to name. Additional features might include fonts, tags, hyperlinks, metadata, and parts of the Web-page description. All this information is used by the search-engine to rank HTML Web pages returned to the user, but is unfortunately absent in free text that has no HTML markup, tags, hyperlinks, and any other metadata, except implicit natural language structure.

Currently, WWW , large enterprises, and desktop users suffer from an inability to efficiently access and manage differently structured data. The same data objects (e.g. Product) stored by different databases, repositories, distributed web storage systems, etc are named, referenced, and combined internally into schemas or data structures differently. This leads to structural mismatch of data that often consists of the same semantic objects (e.g. EBay and Yahoo! online auction offers).

Model management is a high-level programming language designed to efficiently manipulate schemas and mappings. It is comprised of robust operators that combined in short programs can solve complex metadata-oriented problems in a compact way. For instance, countless enterprise data integration scenarios can be easily expressed in this high-level language thus saving hundreds of development man-hours. Here we present the first model management engine that has reverse-engineering support for data integration, which is one of the most pressing metadata-oriented problems. It merges two schemas based on the mappings between them and allows user to correct the result keeping all the mappings in sync automatically. For user it is much more convenient than determining which mappings to correct in order to get desired result. In addition, the engine supports restructuring merging which is important when the sources are structured differently and cannot be mapped directly. While making schema merging fully automatic is not yet possible, our work simplifies and automates this process to make it practical in complex data integration scenarios.