Most Probable Explanations for Probabilistic Database Queries

Total Page:16

File Type:pdf, Size:1020Kb

Most Probable Explanations for Probabilistic Database Queries Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) Most Probable Explanations for Probabilistic Database Queries Ismail˙ Ilkan˙ Ceylan and Stefan Borgwardt Thomas Lukasiewicz Faculty of Computer Science Department of Computer Science Technische Universitat¨ Dresden, Germany University of Oxford, UK [email protected] [email protected] [email protected] Abstract variables, is one of the key computational problems in PGMs. In its general form, this problem is studied under the name Forming the foundations of large-scale knowledge of maximum a posteriori hypothesis1 (MAP), where the hy- bases, probabilistic databases have been widely pothesis refers to an instantiation of a set of distinguished studied in the literature. In particular, probabilistic variables. The most probable explanation (MPE) is a spe- query evaluation has been investigated intensively cial case of MAP, where the distinguished variables and the as a central inference mechanism. However, des- evidence variables cover all variables in the model. Maximal pite its power, query evaluation alone cannot ex- posterior probability computations have also been lifted to re- tract all the relevant information encompassed in lational probabilistic models such as Markov Logic Networks large-scale knowledge bases. To exploit this po- (MLNs) [Richardson and Domingos, 2006]. tential, we study two inference tasks; namely find- Both MPE and MAP translate to probabilistic databases ing the most probable database and the most prob- in a natural way through the rich structure of queries. The able hypothesis for a given query. As natural coun- most probable database problem (analogous to MPE), first terparts of most probable explanations (MPE) and proposed in [Gribkoff et al., 2014], is the problem of de- maximum a posteriori hypotheses (MAP) in prob- termining the (classical) database with the largest probabil- abilistic graphical models, they can be used in a ity that satisfies a given query. Intuitively, the query defines variety of applications that involve prediction or constraints on the data, and the goal is to find the most prob- diagnosis tasks. We investigate these problems rel- able database that satisfies these constraints. We also in- ative to a variety of query languages, ranging from troduce a more intricate notion, called most probable hypo- conjunctive queries to ontology-mediated queries, thesis, which only asks for partial databases satisfying the and provide a detailed complexity analysis. query (analogous to MAP). The most probable hypothesis contains only tuples that contribute to the satisfaction con- 1 Introduction dition of the query, which allows to more precisely pinpoint Research on building large-scale probabilistic knowledge the most likely explanations of the query. bases (PKBs) has resulted in a number of systems includ- We study the computational complexity of the correspond- MPD MPH ing NELL [Mitchell et al., 2015], Yago [Hoffart et al., ing decision problems, denoted by and , respect- 2013], DeepDive [Shin et al., 2015], and Google’s Know- ively, for a variety of query languages. Our results provide ledge Vault [Dong et al., 2014]. They have been used in a detailed insights about the nature of these problems. We show exist- wide range of areas to automatically build structured know- that the data complexity of both problems is lower for ential queries universal queries MPH ledge bases. Most of them are based on the long tradition and than for . As expected, MPD foundations of probabilistic databases (PDBs) [Imieliski and usually has a higher complexity than . ontology-mediated queries Lipski, 1984; Suciu et al., 2011], which define probability We extend our results towards (OMQs) unions of con- distributions over sets of classical databases. , which enrich the well-known class of junctive queries Probabilistic query evaluation is the key inference task un- with the power of ontological rules based on ± [ et al. derpinning these systems. However, PKBs encompass rich, Datalog (also called existential rules) Baget , 2011; et al. ] ontology- structured knowledge, for which alternative inference mech- Cal`ı , 2013 . This follows the tradition of based data access [ et al. ] anisms beyond probabilistic query evaluation are needed. In- Poggi , 2008 , which allows us to spired by the maximal posterior probability computations in query PDBs in a more advanced manner. For OMQs, our ana- Probabilistic Graphical Models (PGMs) [Pearl, 1988; Koller lysis shows that the computational complexity of these prob- and Friedman, 2009], we investigate the problem of finding lems can change significantly w.r.t. the ontology languages most probable explanations for probabilistic queries to ex- 1There exist other variants of these inference tasks [Kwisthout, ploit the potential of such large databases to their full extent. 2011], and there are different naming conventions across communit- Computing the maximal posterior probability of a distin- ies; here, we use the terminology from the Bayesian Network (BN) guished set of variables, given evidence about another set of literature [Darwiche, 2009]. 950 Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence (IJCAI-17) under consideration and the complexity-theoretic assump- Vegetarian FriendOf tions employed. Our results provide tight complexity bounds alice 0.7 alice bob 0.7 for all Datalog± languages known in the literature. Detailed bob 0.9 alice chris 0.8 proofs of all results can be found at https://lat.inf. chris 0.6 bob chris 0.1 tu-dresden.de/research/papers.html. Eats Meat 2 Background and Motivation bob spinach 0.7 shrimp 0.7 chris mussels 0.8 mussels 0.9 We recall the basics of classical query languages and data- alice broccoli 0.2 seahorse 0.3 bases and briefly describe the tuple-independent PDB model. 2.1 Queries and Databases Figure 1: The PDB Pv, where each table represents a predicate and each row is a probabilistic tuple. We consider a relational vocabulary consisting of finite sets R of predicates, C of constants, and V of variables.A term is a constant or a variable. An atom is of the form P(s1; : : : ; sn), through which we can ask the probability of vegetarians being where P is an n-ary predicate, and s1; : : : ; sn are terms. A friends with vegetarians. In the given PDB, alice, bob, and tuple is an atom without variables. chris are all vegetarians and friends with each other (with A conjunctive query (CQ) is an existentially quantified for- some probability). The query mula ∃x φ, where φ is a conjunction of atoms, written as a Q′ = ∃y Veg(bob) ∧ FriendOf(bob; y) ∧ Veg(y) comma-separated list. A union of conjunctive queries (UCQ) 1 is a disjunction of CQs. If φ is an arbitrary Boolean combin- is a special case of Q1 which asks whether bob has vegetarian ′ ation of atoms, then we call ∃x φ an existential query (∃Q), friends. Its probability can be computed as P(Q1) = 0:9 ⋅ 0:1 ⋅ and ∀x φ a universal query (∀Q).A first-order query (FOQ) 0:6 = 0:054: ∎ is an arbitrary first-order formula. A query is Boolean if it has This task is known as probabilistic query evaluation, and no free variables. it has been studied extensively in PDBs. In a celebrated res- A database is a finite set of tuples. The central problem ult by [Dalvi and Suciu, 2012], it has been shown that UCQs studied for databases is query evaluation: Finding all answers exhibit a data complexity dichotomy between P and #P for to a query Q over a database D, which are assignments of probabilistic query evaluation. However, while dealing with the free variables in Q to constants such that the resulting large-scale knowledge bases, alternative inference mechan- first-order formula is satisfied in D in the usual sense. In the isms are needed. Consider again our running example, and following, we consider only Boolean queries Q, and focus on suppose that we have observed that Q′ is true and would like the associated decision problem, i.e., deciding whether Q is 1 to learn what best explains this observation w.r.t. the PDB Pv. satisfied in D, denoted as usual by D ⊧ Q. We revisit this example after providing a principled approach 2.2 Probabilistic Databases for dealing with such tasks. The most elementary probabilistic database model is based on the tuple-independence assumption. We adopt this model 3 Most Probable Explanations and refer to [Suciu et al., 2011] for details and alternatives. We investigate the problem of finding most probable explan- A probabilistic database induces a set of classical databases ations for probabilistic database queries under two different (called worlds), each associated with a probability value. semantics. Finding the most probable database is to determ- Formally, a probabilistic database (PDB) P is a finite set ine the world with the largest probability that satisfies a given of probabilistic tuples of the form ⟨t ∶ p⟩ ; where t is a tuple query, as formalized by [Gribkoff et al., 2014]: and p ∈ [0; 1], and, whenever ⟨t ∶ p⟩; ⟨t ∶ q⟩ ∈ P, then p = q.A Definition 2. The most probable database for a query Q over PDB P assigns to every tuple t the probability p if ⟨t ∶ p⟩ ∈ P, a PDB P is and otherwise the probability 0. Since we usually consider arg max P(D); a single, fixed PDB, we denote this probability assignment D⊧Q simply by P. We concentrate on the well-known possible where D ranges over all worlds induced by P. worlds semantics. Under the tuple-independence assumption, P induces the following unique probability distribution over Intuitively, a PDB defines a probability distribution over classical databases D: exponentially many classical databases, and the most prob- able database is the element in this collection that has the P(D) ∶= M P(t) M(1 − P(t)): highest probability while still satisfying the query.
Recommended publications
  • Probabilistic Databases
    Series ISSN: 2153-5418 SUCIU • OLTEANU •RÉ •KOCH M SYNTHESIS LECTURES ON DATA MANAGEMENT &C Morgan & Claypool Publishers Series Editor: M. Tamer Özsu, University of Waterloo Probabilistic Databases Probabilistic Databases Dan Suciu, University of Washington, Dan Olteanu, University of Oxford Christopher Ré,University of Wisconsin-Madison and Christoph Koch, EPFL Probabilistic databases are databases where the value of some attributes or the presence of some records are uncertain and known only with some probability. Applications in many areas such as information extraction, RFID and scientific data management, data cleaning, data integration, and financial risk DATABASES PROBABILISTIC assessment produce large volumes of uncertain data, which are best modeled and processed by a probabilistic database. This book presents the state of the art in representation formalisms and query processing techniques for probabilistic data. It starts by discussing the basic principles for representing large probabilistic databases, by decomposing them into tuple-independent tables, block-independent-disjoint tables, or U-databases. Then it discusses two classes of techniques for query evaluation on probabilistic databases. In extensional query evaluation, the entire probabilistic inference can be pushed into the database engine and, therefore, processed as effectively as the evaluation of standard SQL queries. The relational queries that can be evaluated this way are called safe queries. In intensional query evaluation, the probabilistic Dan Suciu inference is performed over a propositional formula called lineage expression: every relational query can be evaluated this way, but the data complexity dramatically depends on the query being evaluated, and Dan Olteanu can be #P-hard. The book also discusses some advanced topics in probabilistic data management such as top-kquery processing, sequential probabilistic databases, indexing and materialized views, and Monte Carlo databases.
    [Show full text]
  • Adaptive Schema Databases ∗
    Adaptive Schema Databases ∗ William Spothb, Bahareh Sadat Arabi, Eric S. Chano, Dieter Gawlicko, Adel Ghoneimyo, Boris Glavici, Beda Hammerschmidto, Oliver Kennedyb, Seokki Leei, Zhen Hua Liuo, Xing Niui, Ying Yangb b: University at Buffalo i: Illinois Inst. Tech. o: Oracle {wmspoth|okennedy|yyang25}@buffalo.edu {barab|slee195|xniu7}@hawk.iit.edu [email protected] {eric.s.chan|dieter.gawlick|adel.ghoneimy|beda.hammerschmidt|zhen.liu}@oracle.com ABSTRACT in unstructured or semi-structured form, then an ETL (i.e., The rigid schemas of classical relational databases help users Extract, Transform, and Load) process needs to be designed in specifying queries and inform the storage organization of to translate the input data into relational form. Thus, clas- data. However, the advantages of schemas come at a high sical relational systems require a lot of upfront investment. upfront cost through schema and ETL process design. In This makes them unattractive when upfront costs cannot this work, we propose a new paradigm where the database be amortized, such as in workloads with rapidly evolving system takes a more active role in schema development and data or where individual elements of a schema are queried data integration. We refer to this approach as adaptive infrequently. Furthermore, in settings like data exploration, schema databases (ASDs). An ASD ingests semi-structured schema design simply takes too long to be practical. or unstructured data directly using a pluggable combina- Schema-on-query is an alternative approach popularized tion of extraction and data integration techniques. Over by NoSQL and Big Data systems that avoids the upfront time it discovers and adapts schemas for the ingested data investment in schema design by performing data extraction using information provided by data integration and infor- and integration at query-time.
    [Show full text]
  • Finding Interesting Itemsets Using a Probabilistic Model for Binary Databases
    Finding interesting itemsets using a probabilistic model for binary databases Tijl De Bie University of Bristol, Department of Engineering Mathematics Queen’s Building, University Walk, Bristol, BS8 1TR, UK [email protected] ABSTRACT elegance and e±ciency of the algorithms to search for fre- A good formalization of interestingness of a pattern should quent patterns. Unfortunately, the frequency of a pattern satisfy two criteria: it should conform well to intuition, and is only loosely related to interestingness. The output of fre- it should be computationally tractable to use. The focus quent pattern mining methods is usually an immense bag has long been on the latter, with the development of frequent of patterns that are not necessarily interesting, and often pattern mining methods. However, it is now recognized that highly redundant with each other. This has hampered the more appropriate measures than frequency are required. uptake of FPM and FIM in data mining practice. In this paper we report results in this direction for item- Recent research has shifted focus to the search for more set mining in binary databases. In particular, we introduce useful formalizations of interestingness that match practical a probabilistic model that can be ¯tted e±ciently to any needs more closely, while still being amenable to e±cient binary database, and that has a compact and explicit repre- algorithms. As FIM is arguably the simplest special case of sentation. We then show how this model enables the formal- frequent pattern mining, it is not surprising that most of the ization of an intuitive and tractable interestingness measure recent work has focussed on itemset patterns, see e.g.
    [Show full text]
  • Open-World Probabilistic Databases
    Open-World Probabilistic Databases Ismail˙ Ilkan˙ Ceylan and Adnan Darwiche and Guy Van den Broeck Computer Science Department University of California, Los Angeles [email protected],fdarwiche,[email protected] Abstract techniques, from part-of-speech tagging until relation ex- traction (Mintz et al. 2009). For knowledge-base comple- Large-scale probabilistic knowledge bases are becoming in- tion – the task of deriving new facts from existing knowl- creasingly important in academia and industry alike. They edge – statistical relational learning algorithms make use are constantly extended with new data, powered by modern information extraction tools that associate probabilities with of embeddings (Bordes et al. 2011; Socher et al. 2013) database tuples. In this paper, we revisit the semantics under- or probabilistic rules (Wang, Mazaitis, and Cohen 2013; lying such systems. In particular, the closed-world assump- De Raedt et al. 2015). In both settings, the output is a pre- tion of probabilistic databases, that facts not in the database dicted fact with its probability. It is therefore required to de- have probability zero, clearly conflicts with their everyday fine probabilistic semantics for knowledge bases. The classi- use. To address this discrepancy, we propose an open-world cal and most-basic model is that of tuple-independent prob- probabilistic database semantics, which relaxes the probabili- abilistic databases (PDBs) (Suciu et al. 2011), which in- ties of open facts to intervals. While still assuming a finite do- deed underlies many of these systems (Dong et al. 2014; main, this semantics can provide meaningful answers when Shin et al. 2015).
    [Show full text]
  • A Compositional Query Algebra for Second-Order Logic and Uncertain Databases
    A Compositional Query Algebra for Second-Order Logic and Uncertain Databases Christoph Koch Department of Computer Science Cornell University, Ithaca, NY, USA [email protected] ABSTRACT guages for querying possible worlds. Indeed, second-order World-set algebra is a variable-free query language for uncer- quantifiers are the essence of what-if reasoning in databases. tain databases. It constitutes the core of the query language World-set algebra seems to be a strong candidate for a core implemented in MayBMS, an uncertain database system. algebra for forming query plans and optimizing and execut- This paper shows that world-set algebra captures exactly ing them in uncertain database management systems. second-order logic over finite structures, or equivalently, the It was left open in previous work whether world-set alge- polynomial hierarchy. The proofs also imply that world- bra is closed under composition, or whether definitions are set algebra is closed under composition, a previously open adding to the expressive power of the language. Compo- problem. sitionality is a desirable and rather commonplace property of query algebras, but in the case of WSA it seems rather unlikely to hold. The reason for this is that the algebra con- 1. INTRODUCTION tains an uncertainty-introduction operation that on the level Developing suitable query languages for uncertain data- of possible worlds is nondeterministic. First materializing a bases is a substantial research challenge that is only cur- view and subsequently using it multiple times in the query is rently starting to get addressed. Conceptually, an uncertain semantically quite different from composing the query with database is a finite set of possible worlds, each one a re- the view and thus obtaining several copies of the view def- lational database.
    [Show full text]
  • Answering Queries from Statistics and Probabilistic Views∗
    Answering Queries from Statistics and Probabilistic Views∗ Nilesh Dalvi Dan Suciu University of Washington, Seattle, WA, USA 1 Introduction paper we will assume the Local As View (LAV) data integration paradigm [17, 16], which consists of defin- Systems integrating dozens of databases, in the scien- ing a global mediated schema R¯, then expressing each tific domain or in a large corporation, need to cope local source i as a view vi(R¯) over the global schema. with a wide variety of imprecisions, such as: different Users are allowed to ask queries over the global, me- representations of the same object in different sources; diated schema, q(R¯), however the data is given as in- imperfect and noisy schema alignments; contradictory stances J1; : : : ; Jm of the local data sources. In our information across sources; constraint violations; or in- model all instances are probabilistic, both the local sufficient evidence to answer a given query. If standard instances and the global instance. Statistics are given query semantics were applied to such data, all but the explicitly over the global schema R¯, and the probabil- most trivial queries will return an empty answer. ities are given explicitly over the local sources, hence We believe that probabilistic databases are the right over the views. We make the Open World Assumption paradigm to model all types of imprecisions in a uni- throughout the paper. form and principled way. A probabilistic database is a probability distribution on all instances [5, 4, 15, 1.1 Example: Using Statistics 12, 11, 8]. Their early motivation was to model im- precisions at the tuple level: tuples are not known Suppose we integrate two sources, one show- with certainty to belong to the database, or represent ing which employee works for what depart- noisy measurements, etc.
    [Show full text]
  • Uva-DARE (Digital Academic Repository)
    UvA-DARE (Digital Academic Repository) On entity resolution in probabilistic data Ayat, S.N. Publication date 2014 Document Version Final published version Link to publication Citation for published version (APA): Ayat, S. N. (2014). On entity resolution in probabilistic data. General rights It is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), other than for strictly personal, individual use, unless the work is under an open content license (like Creative Commons). Disclaimer/Complaints regulations If you believe that digital publication of certain material infringes any of your rights or (privacy) interests, please let the Library know, stating your reasons. In case of a legitimate complaint, the Library will make the material inaccessible and/or remove it from the website. Please Ask the Library: https://uba.uva.nl/en/contact, or a letter to: Library of the University of Amsterdam, Secretariat, Singel 425, 1012 WP Amsterdam, The Netherlands. You will be contacted as soon as possible. UvA-DARE is a service provided by the library of the University of Amsterdam (https://dare.uva.nl) Download date:01 Oct 2021 On Entity Resolution in Probabilistic Data Naser Ayat On Entity Resolution in Probabilistic Data Naser Ayat On Entity Resolution in Probabilistic Data Naser Ayat On Entity Resolution in Probabilistic Data Academisch Proefschrift ter verkrijging van de graad van doctor aan de Universiteit van Amsterdam op gezag van de Rector Magnificus prof.dr. D.C. van den Boom ten overstaan van een door het college voor promoties ingestelde commissie, in het openbaar te verdedigen in de Agnietenkapel op dinsdag 30 september 2014, te 12.00 uur door Seyed Naser Ayat geboren te Esfahan, Iran Promotors: prof.
    [Show full text]
  • BAYESSTORE: Managing Large, Uncertain Data Repositories with Probabilistic Graphical Models
    BAYESSTORE: Managing Large, Uncertain Data Repositories with Probabilistic Graphical Models Daisy Zhe Wang∗ , Eirinaios Michelakis∗ , Minos Garofalakisy∗ , and Joseph M. Hellerstein∗ ∗ Univ. of California, Berkeley EECS and y Yahoo! Research ABSTRACT Of course, the fundamental mathematical tools for managing un- certainty come from probability and statistics. In recent years, these Several real-world applications need to effectively manage and reason about tools have been aggressively imported into the computational do- large amounts of data that are inherently uncertain. For instance, perva- main under the rubric of Statistical Machine Learning (SML). Of sive computing applications must constantly reason about volumes of noisy special note here is the widespread use of Graphical Modeling tech- sensory readings for a variety of reasons, including motion prediction and niques, including the many variants of Bayesian Networks (BNs) human behavior modeling. Such probabilistic data analyses require so- and Markov Random Fields (MRFs) [13]. These techniques can phisticated machine-learning tools that can effectively model the complex provide robust statistical models that capture complex correlation spatio/temporal correlation patterns present in uncertain sensory data. Un- patterns among variables, while, at the same time, addressing some fortunately, to date, most existing approaches to probabilistic database sys- computational efficiency and scalability issues as well. Graphical tems have relied on somewhat simplistic models of uncertainty that can be models have been applied with great success in applications as di- easily mapped onto existing relational architectures: Probabilistic informa- verse as signal processing, information retrieval, sensornets and tion is typically associated with individual data tuples, with only limited pervasive computing, robotics, natural language processing, and or no support for effectively capturing and reasoning about complex data computer vision.
    [Show full text]
  • Brief Tutorial on Probabilistic Databases
    Brief Tutorial on Probabilistic Databases Dan Suciu University of Washington Simons 2016 1 About This Talk • Probabilistic databases – Tuple-independent – Query evaluation • Statistical relational models – Representation, learning, inference in FO – Reasoning/learning = lifted inference • Sources: – Book 2011 [S.,Olteanu,Re,Koch] – Upcoming F&T survey [van Den Broek,S] Simons 2016 2 Background: Relational databases Smoker Friend X Y X Z Database D = Alice 2009 Alice Bob Alice 2010 Alice Carol Bob 2009 Bob Carol Carol 2010 Carol Bob Query: Q(z) = ∃x (Smoker(x,’2009’)∧Friend(x,z)) Z Constraint: Q(D) = Bob Q = ∀x (Smoker(x,’2010’)⇒Friend(x,’Bob’)) Carol Q(D) = true Probabilistic Database Probabilistic database D: X y P a1 b1 p1 a1 b2 p2 a2 b2 p3 Possible worlds semantics: Σworld PD(world) = 1 X y P (Q) = Σ P (world) X Xy y D world |= Q D a1 b1 X y a1 a1b2 b1 X y a1 b2 a1 b1 X y p1p2p3 a2 a2b2 b2 a2 b2 X y a2 b2 a1 b2 a1 b1 X y (1-p1)p2p3 a1 b2 (1-p1)(1-p2)(1-p3) Outline • Model Counting • Small Dichotomy Theorem • Dichotomy Theorem • Query Compilation • Conclusions, Open Problems Simons 2016 5 Model Counting • Given propositional Boolean formula F, compute the number of models #F Example: X1 X2 X3 F F = (X1∨X2) ∧ (X2∨X3) ∧ (X3∨X1) 0 0 0 0 0 0 1 0 #F = 4 0 1 0 0 0 1 1 1 1 0 0 0 1 0 1 1 1 1 0 1 [Valiant’79] #P-hard, even for 2CNF 1 1 1 1 Probability of a Formula • Each variable X has a probability p(X); • P(F) = probability that F=true, when each X is set to true independently Example: X1 X2 X3 F F = (X1∨X2) ∧ (X2∨X3) ∧ (X3∨X1) 0 0 0 0 0 0 1 0 P(F) = (1-p1)*p2*p3 + 0 1 0 0 p1*(1-p2)*p3 + p1*p2*(1-p3) + 0 1 1 1 p1*p2*p3 1 0 0 0 1 0 1 1 n If p(X) = ½ for all X, then P(F) = #F / 2 1 1 0 1 1 1 1 1 Algorithms for Model Counting [Gomes, Sabharwal, Selman’2009] Based on full search DPLL: • Shannon expansion.
    [Show full text]
  • Efficient In-Database Analytics with Graphical Models
    Efficient In-Database Analytics with Graphical Models Daisy Zhe Wang, Yang Chen, Christan Grant and Kun Li fdaisyw,yang,cgrant,[email protected]fl.edu University of Florida, Department of Computer and Information Science and Engineering Abstract Due to recent application push, there is high demand in industry to extend database systems to perform efficient and scalable in-database analytics based on probabilistic graphical models (PGMs). We discuss issues in supporting in-database PGM methods and present techniques to achieve a deep integration of the PGM methods into the relational data model as well as the query processing and optimization engine. This is an active research area and the techniques discussed are being further developed and evaluated. 1 Introduction Graphical models, also known as probabilistic graphical models (PGMs), are a class of expressive and widely adopted machine learning models that compactly represent the joint probability distribution over a large number of interdependent random variables. They are used extensively in large-scale data analytics, including infor- mation extraction (IE), knowledge base population (KBP), speech recognition and computer vision. In the past decade, many of these applications have witnessed dramatic increase in data volume. As a result, PGM-based data analytics need to scale to terabytes of data or billions of data items. While new data-parallel and graph- parallel processing platforms such as Hadoop, Spark and GraphLab [1, 36, 3] are widely used to support scalable PGM methods, such solutions are not effective for applications with data residing in databases or with need to support online PGM-based data analytics queries. More specifically, large volumes of valuable data are likely to pour into database systems for many years to come due to the maturity of the commercial database systems with transaction support and ACID properties and due to the legal auditing and data privacy requirements in many organizations.
    [Show full text]
  • Indexing Correlated Probabilistic Databases
    Indexing Correlated Probabilistic Databases Bhargav Kanagal Amol Deshpande [email protected] [email protected] Dept. of Computer Science, University of Maryland, College Park MD 20742 ABSTRACT 1. INTRODUCTION With large amounts of correlated probabilistic data being Large amounts of correlated probabilistic data are be- generated in a wide range of application domains includ- ing generated in a variety of application domains includ- ing sensor networks, information extraction, event detection ing sensor networks [14, 21], information extraction [20, 17], etc., effectively managing and querying them has become data integration [1, 15], activity recognition [28], and RFID an important research direction. While there is an exhaus- stream analysis [29, 26]. The correlations exhibited by these tive body of literature on querying independent probabilistic data sources can range from simple mutual exclusion depen- data, supporting efficient queries over large-scale, correlated dencies, to complex dependencies typically captured using databases remains a challenge. In this paper, we develop Bayesian or Markov networks [30] or through constraints on efficient data structures and indexes for supporting infer- the possible values that the tuple attributes can take [19, 12]. ence and decision support queries over such databases. Our Although several approaches have been proposed for rep- proposed hierarchical data structure is suitable both for in- resenting complex correlations and for querying over corre- memory and disk-resident databases. We represent the cor- lated databases, the rapidly increasing scale of such databases relations in the probabilistic database using a junction tree has raised unique challenges that have not been addressed over the tuple-existence or attribute-value random variables, before.
    [Show full text]
  • Analyzing Uncertain Tabular Data
    Analyzing Uncertain Tabular Data Oliver Kennedy and Boris Glavic Abstract It is common practice to spend considerable time refining source data to address issues of data quality before beginning any data analysis. For ex- ample, an analyst might impute missing values or detect and fuse duplicate records representing the same real-world entity. However, there are many sit- uations where there are multiple possible candidate resolutions for a data quality issue, but there is not sufficient evidence for determining which of the resolutions is the most appropriate. In this case, the only way forward is to make assumptions to restrict the space of solutions and/or to heuristically choose a resolution based on characteristics that are deemed predictive of \good" resolutions. Although it is important for the analyst to understand the impact of these assumptions and heuristic choices on her results, evalu- ating this impact can be highly non-trivial and time consuming. For several decades now, the fields of probabilistic, incomplete, and fuzzy databases have developed strategies for analyzing the impact of uncertainty on the outcome of analyses. This general family of uncertainty-aware databases aims to model ambiguity in the results of analyses expressed in standard languages like SQL, SparQL, R, or Spark. An uncertainty-aware database uses descriptions of potential errors and ambiguities in source data to derive a correspond- ing description of potential errors or ambiguities in the result of an analysis accessing this source data. Depending on technique, these descriptions of uncertainty may be either quantitative (bounds, probabilities), or qualita- tive (certain outcomes, unknown values, explanations of uncertainty).
    [Show full text]