Fact-checked by Grok 2 weeks ago

SRL

Statistical relational learning (SRL) is a subdiscipline of artificial intelligence and machine learning concerned with domain models that exhibit both uncertainty and complex, relational structure.^[1] It addresses learning and inference in relational domains by integrating techniques from statistical learning, probabilistic graphical models, and relational representations such as first-order logic.^[1] SRL emerged in the late 1990s, building on advancements in inductive logic programming from the 1980s and 1990s, and probabilistic graphical models developed in the 1980s and 1990s.^[1] Pioneering work includes probabilistic relational models (introduced in 1999) and Markov logic networks (2000s), enabling scalable handling of uncertainty in structured data like knowledge bases and networks.^[1] The field has applications in areas such as web and knowledge base mining, bioinformatics, and social network analysis, where traditional i.i.d. assumptions fail due to interdependencies.^[1] Research in SRL focuses on challenges like parameter estimation, structure learning, and efficient inference, contributing to advancements in AI systems that reason under uncertainty.^[1]

Overview

Definition and scope

Statistical relational learning (SRL) is an area of machine learning that integrates relational representations, which capture structured data involving entities, attributes, and interrelations, with statistical methods to handle uncertainty and probabilistic dependencies. This combination enables the modeling of complex domains where data exhibits both relational structure and noise, allowing for the representation of joint probability distributions over interconnected objects rather than independent instances.^[2]^[3] The scope of SRL encompasses tasks in domains with inherent relational complexity, such as social networks and bioinformatics, where traditional machine learning approaches fall short due to assumptions of independent data points. It distinguishes itself from pure relational learning, which relies on deterministic logical rules without probabilistic elements, and from conventional statistical learning, which typically operates on flat, propositional data without explicit relational modeling. By focusing on structured prediction and inference over graphs or databases, SRL addresses challenges like incomplete information and variable numbers of entities.^[2]^[3] Key motivations for SRL arise from the need to develop scalable models for large, interconnected datasets that incorporate domain knowledge while managing uncertainty, thereby improving predictions in systems where decisions influence one another, such as collective classification or link prediction. These models facilitate reasoning under partial observability and enable the exploitation of dependencies to enhance accuracy beyond isolated analyses.^[2] SRL traces its early roots to the 1990s, building on advancements in inductive logic programming for learning relational rules and Bayesian networks for probabilistic inference, which together laid the groundwork for unifying logic and statistics in structured settings.^[2]

Historical development

Statistical relational learning (SRL) emerged in the late 1990s as a synthesis of inductive logic programming (ILP), which focused on learning relational rules from data, and probabilistic graphical models for handling uncertainty. ILP techniques, such as the FOIL algorithm developed by J. R. Quinlan in 1990, enabled the induction of logical rules from relational data but lacked mechanisms for probabilistic reasoning.^[4] Concurrently, Judea Pearl's 1988 framework for Bayesian networks provided a foundation for representing and inferring probabilities in complex systems, inspiring extensions to relational domains. Early efforts, including David Poole's Independent Choice Logic in 1993, began bridging logic and probability by allowing probabilistic choices in logical programs. The 2000s marked key milestones in formalizing SRL frameworks. Probabilistic relational models (PRMs), introduced by Nir Friedman, Lise Getoor, Daphne Koller, and Avi Pfeffer in 1999, and further developed including Ben Taskar in the 2002 Journal of Machine Learning Research paper by Getoor et al., extended Bayesian networks to object-oriented relational structures, enabling compact modeling of linked entities.^[5]^[6] Markov logic networks (MLNs), proposed by Matthew Richardson and Pedro Domingos in 2006, unified first-order logic with Markov networks by assigning weights to logical formulas, facilitating probabilistic inference over relational data.^[7] These developments were consolidated in the 2007 edited volume Introduction to Statistical Relational Learning by Lise Getoor and Ben Taskar, which synthesized foundational research and highlighted SRL's potential for real-world applications. Influential researchers such as Lise Getoor, Pedro Domingos, Ben Taskar, and David Poole drove the field's progress, with early work disseminated through conferences like the International Conference on Inductive Logic Programming (ILP, starting 1991) and the Conference on Uncertainty in Artificial Intelligence (UAI, from 1985). In the 2010s, advancements like probabilistic soft logic (PSL), developed by Stephen H. Bach, Bert Huang, and colleagues around 2013, introduced continuous truth values and hinge-loss objectives for scalable relational reasoning. The rise of big data, particularly from social networks and knowledge graphs, accelerated SRL adoption by addressing challenges in collective classification and link prediction. This evolution shifted the paradigm from deterministic, rule-based ILP systems to hybrid probabilistic-relational models capable of managing uncertainty in large-scale, structured datasets.

Core Concepts

Relational structure

In statistical relational learning (SRL), relational structure refers to the organization of data as a collection of entities, each characterized by attributes and connected through relations, forming an interconnected representation that captures dependencies among multiple objects rather than treating them as independent instances. This contrasts sharply with traditional machine learning approaches that assume independent and identically distributed (i.i.d.) data in tabular formats, where each row represents a standalone example without explicit links to others.^[8] Instead, relational structure models real-world domains like social networks or bibliographic databases, where entities such as users or papers interact in complex, non-i.i.d. ways, enabling the representation of multi-relational knowledge. Key elements of relational structure include entities as the core objects (e.g., specific individuals or documents), attributes as their descriptive properties (e.g., a user's age or a paper's topic), and relations as the linkages between entities (e.g., "friend" or "cites").^[8] Relations are often expressed using first-order logic (FOL) predicates, such as friend(X, Y), which denote binary connections between variables representing entities, allowing for flexible, population-level descriptions without enumerating every instance. Multi-relational data, common in domains like social networks with entities including users, posts, and groups connected via relations like "likes" or "follows," exemplify this structure by incorporating multiple types of interactions.^[8] Representationally, these structures draw from directed or undirected graphs—where entities are nodes, attributes are node labels, and relations are edges—or relational databases with schemas defining classes and reference slots (e.g., a "Student" class linking to "Course" via "EnrolledIn"). Challenges in relational structure arise from the open-world assumption, where the number of entities is variable and not all relations are observed, unlike closed-world databases that assume complete knowledge.^[8] For instance, in a citation network, entities are individual papers, attributes include author names and publication year, and the "cites" relation forms directed edges between papers, allowing the model to represent evolving scholarly dependencies without fixed entity counts. This graph-like or logical form supports scalability to large, heterogeneous datasets, such as web graphs with hyperlinks.^[8] The importance of relational structure in SRL lies in its ability to capture cross-instance dependencies, such as homophily in social networks where connected entities share similar attributes, which flat vector representations in standard learning cannot express. By modeling entities and relations explicitly, SRL frameworks enable richer knowledge representations that facilitate tasks like link prediction or entity resolution, laying the groundwork for handling uncertainty in these interconnected domains.^[8]

Uncertainty and probability

In statistical relational learning (SRL), uncertainty arises primarily from incomplete observations, noisy relations among entities, or inherent stochastic processes in the data, which are addressed by defining probability distributions over relational structures and possible worlds. These distributions allow models to quantify the likelihood of unobserved or ambiguous relations, enabling robust reasoning in domains with complex dependencies. For instance, attribute uncertainty pertains to indeterminate values of object properties, reference uncertainty involves ambiguous links between entities, and existence uncertainty deals with unknown entities or relations altogether.^[3]^[9] Key concepts in handling uncertainty include marginal and conditional probabilities adapted to relational settings, where probabilities are computed over interconnected entities rather than isolated variables. Marginal probabilities assess the likelihood of a specific relation or attribute holding true across a domain, while conditional probabilities evaluate it given evidence from related entities, facilitating inference in interdependent structures. Existential uncertainty, such as unknown relations between objects, is managed by probabilistic representations that assign non-zero probabilities to potential existences, avoiding deterministic absences. This approach contrasts with purely logical systems by incorporating graded beliefs over relational configurations.^[3]^[9] At its core, probability in SRL relies on joint distributions defined over populations of entities, capturing collective dependencies in relational data rather than independent samples. These distributions model the entire configuration of attributes and relations as a cohesive probabilistic space, often using weighted logical formulas to specify the form. The choice between closed-world and open-world assumptions significantly impacts these calculations: under the closed-world assumption, unobserved facts are deemed false, simplifying distributions but risking underestimation of uncertainty; in contrast, the open-world assumption permits unknown facts, leading to more flexible but computationally intensive probability assessments over potentially infinite structures.^[10]^[3] A practical example is fraud detection in financial networks, where probabilistic edges represent "suspected links" between accounts with incomplete transaction data, allowing the joint distribution to weigh the likelihood of fraudulent patterns amid noisy or missing relations. This relational probabilistic modeling distinguishes SRL from classical statistics, which typically operates on fixed-dimensional vectors or independent observations; in SRL, probabilities extend over variable-sized or infinite populations of entities and evolving structures, accommodating real-world relational incompleteness.^[1]^[3]

Integration of logic and statistics

In statistical relational learning (SRL), the integration of logic and statistics enables the representation of complex relational structures while accounting for uncertainty, unifying declarative knowledge with probabilistic inference. First-order logic (FOL) provides relational expressiveness by allowing rules to be formulated as compact, interpretable formulas that capture dependencies among entities, such as predicates defining relationships in a domain.^[7] Statistics complements this by incorporating probabilities to model incomplete or noisy data, transforming rigid logical statements into flexible, weighted constructs that reflect degrees of belief.^[11] This fusion allows SRL models to handle structured prediction tasks that neither pure logic nor traditional machine learning can address effectively on their own. Key mechanisms for this integration include lifting logical rules to probabilistic weights and enforcing logical consistency through soft probabilistic constraints. In lifted representations, FOL formulas are assigned real-valued weights that scale their influence across multiple ground instances, enabling efficient computation over relational data without enumerating all possibilities.^[7] Logical constraints, which demand strict satisfaction in classical systems, are relaxed via probabilistic interpretations, where violations incur penalties proportional to their improbability, thus accommodating real-world variability.^[11] For instance, a FOL rule like "friends(x, y) ∧ ¬similar_interests(x, y) ⇒ false" can be weighted to express that friends are likely but not guaranteed to share interests, with the weight quantifying empirical confidence derived from data.^[7] This integration offers significant benefits, including the injection of domain knowledge through background rules that guide learning and inference, thereby reducing data requirements and improving generalization.^[12] It also supports scalable inference by leveraging logical structure to prune search spaces in probabilistic computations, addressing the combinatorial explosion common in relational domains.^[12] Ultimately, it mitigates the brittleness of pure logic, which falters under uncertainty, and the structural limitations of pure statistical methods, which overlook relational dependencies.^[11] Theoretically, the integration rests on combining possible worlds semantics from logic—where interpretations of FOL formulas define possible worlds—with Markov semantics from graphical models, yielding a probability distribution over worlds as a log-linear model parameterized by formula weights.^[12] Hybrid formalisms, such as those implementing weighted logical rules, operationalize this foundation.^[7]

Representation Formalisms

Probabilistic graphical models

Probabilistic graphical models (PGMs) provide a foundational representation formalism in statistical relational learning (SRL) for compactly encoding joint probability distributions over relational data, where nodes represent random variables corresponding to entities or attributes, and edges capture conditional dependencies between them.^[13] This graphical structure leverages conditional independencies to factorize complex distributions, enabling efficient representation of uncertainty in structured domains with multiple interrelated objects. In SRL, PGMs are extended to handle relational aspects through two primary types: directed models, such as Bayesian networks, which model asymmetric or causal dependencies via directed acyclic graphs, and undirected models, such as Markov random fields, which represent symmetric interactions without directional implications.^[13] Relational extensions further adapt these for multi-relational data; for instance, probabilistic relational models (PRMs) build on Bayesian networks by incorporating schema-level classes, attributes, and relations to define probabilistic dependencies across entity types, while relational Markov networks (RMNs) extend Markov random fields using clique templates—defined via relational queries—to capture higher-order interactions in undirected relational settings.^[14]^[15] A key feature of PGMs in SRL is the factorization of the joint distribution, which exploits the graph structure to decompose the probability into local components. For directed PGMs like Bayesian networks, the joint probability over variables X factorizes as

P(X) = \prod_i P(X_i \mid \mathrm{Pa}(X_i)),

where \mathrm{Pa}(X_i) denotes the parents of X_i in the directed acyclic graph; this allows the full distribution to be specified via conditional probability distributions (CPDs) at each node.^[13] Another essential feature is plate notation, which compactly represents repeated substructures for multiple instances of entities, such as drawing a "plate" around a subgraph to indicate replication over a set of objects, thereby avoiding explicit enumeration of variables for each entity in large relational datasets. For example, in a bibliographic network, a PGM might use plate notation to model multiple papers and authors: nodes within a paper plate could include attributes like topic or quality, with edges to parent nodes for publication venue; an author plate might connect via relational edges (e.g., authorship links) to influence topic probabilities, and citation edges could represent dependencies between papers, all factorized according to the directed or undirected structure to capture influences like author expertise on cited work quality.^[14]

Logical and relational representations

Logical representations in statistical relational learning (SRL) employ formalisms from mathematical logic to declaratively model relational structures and knowledge about entities and their interactions. These representations use predicates to denote relations, variables to refer to entities, and quantifiers to express generalizations over domains, allowing compact encoding of complex dependencies. For example, the formula \forall x \forall y \, (\text{friend}(x,y) \to \text{similar}(x,y)) states that for all individuals x and y, if x and y are friends, then x and y are similar.^[16] Key types of logical representations include propositional logic, first-order logic (FOL), and description logics. Propositional logic handles simple cases by treating relations as ground atoms—propositions without variables—suitable for finite, non-relational domains where each possible fact is explicitly represented. First-order logic extends propositional logic to relational settings by incorporating variables, predicates of arbitrary arity, functions, and quantifiers (\forall and \exists), enabling succinct descriptions of knowledge applicable to potentially infinite populations.^[16] Description logics, a decidable fragment of FOL, focus on terminological knowledge for ontologies, using concepts (unary predicates), roles (binary predicates), and constructors like intersections and existential restrictions to define class hierarchies and properties.^[17] Relational aspects of these representations involve grounding abstract logical formulas to concrete data instances. Herbrand interpretations provide a semantics for this by considering the Herbrand universe—all terms constructible from constants and function symbols—and the Herbrand base—all ground atoms formed by applying predicates to these terms; an interpretation then assigns truth values to these ground atoms to evaluate formula satisfaction.^[12] Clausal forms, such as conjunctive normal form or definite clause programs (e.g., Horn clauses of the form A \leftarrow B_1, \dots, B_n), normalize FOL knowledge for efficient processing, supporting automated reasoning via resolution and facilitating learning algorithms by restricting hypothesis spaces to clause-based structures.^[16] A representative example appears in knowledge bases, where relational facts are encoded as triples in the form (subject, predicate, object), corresponding directly to logical atoms such as \text{knows}(\text{John}, \text{Mary}), which ground predicates over named entities to assert specific relations. These representations are inherently deterministic, treating relations as true or false without inherent uncertainty, though SRL incorporates probabilistic extensions to handle noisy or incomplete data.^[16]

Hybrid formalisms

Hybrid formalisms in statistical relational learning combine elements of first-order logic with probabilistic reasoning, enabling the representation of relational structures alongside uncertainty through weighted logical formulas that act as soft constraints rather than hard rules.^[7] These approaches extend pure logical representations by associating probabilities or weights with formulas in first-order logic (FOL), allowing models to capture both relational dependencies and statistical regularities in complex domains.^[18] A prominent example is Markov logic networks (MLNs), which unify FOL with Markov networks by treating logical formulas as cliques in an undirected graphical model.^[7] In an MLN, each FOL formula F_i is assigned a real-valued weight w_i, reflecting the strength of the constraint it imposes; higher weights indicate formulas that are more likely to hold true. The probability of a possible world x (a Herbrand interpretation assigning truth values to all ground atoms) is given by

P(x) = \frac{1}{Z} \exp\left( \sum_i w_i n_i(x) \right),

where n_i(x) is the number of true groundings of formula F_i in x, and Z is the normalization constant (partition function) summing the exponentials over all possible worlds.^[7] This formulation derives from the Markov network framework, where each ground formula corresponds to a feature function that contributes additively to the log-probability; grounding the FOL formulas generates the network structure, with weights learned from data to balance logical consistency and probabilistic fit.^[7] Another key formalism is probabilistic soft logic (PSL), which relaxes the binary truth values of traditional logic using continuous relaxations based on linear inequalities and the Łukasiewicz t-norm.^[19] In PSL, rules are expressed as weighted implications between linear combinations of truth values (ranging from 0 to 1), enabling scalable inference over large relational domains by formulating objectives as convex optimization problems rather than exact probabilistic enumeration.^[20] Other hybrid approaches include BLOG, a probabilistic programming language that incorporates first-order logical constructs to define generative models over worlds with unknown numbers of objects and relations,^[21] and ProbLog, which extends Prolog with probabilistic facts and clauses to compute probabilities of logical queries in relational settings.^[18] These hybrid formalisms offer expressive power for handling complex relational queries that blend deterministic rules with probabilistic softening, while achieving scalability through techniques like lazy grounding, where only relevant portions of the logical formulas are instantiated during inference.^[7]

Learning and Inference Tasks

Parameter estimation

Parameter learning in statistical relational learning (SRL) involves inferring the quantitative parameters of a model—such as conditional probability distributions in directed graphical models or weights on logical rules in hybrid formalisms—from observed relational data, assuming the underlying structure is fixed. This process enables the model to capture probabilistic dependencies while leveraging the predefined relational schema, distinguishing it from structure learning tasks.^[2]^[7] The standard objective for parameter estimation is maximum likelihood estimation (MLE), which identifies parameters \theta that maximize the log-likelihood of the data:

\arg\max_\theta \sum \log P(\text{data} \mid \theta)

Exact MLE is often intractable due to the exponential size of relational groundings, leading to approximations like pseudo-likelihood, which factors the joint distribution into local conditionals, or variational methods that bound the likelihood. Supervised parameter learning uses fully labeled data to directly optimize these objectives, while unsupervised approaches employ the expectation-maximization (EM) algorithm: the E-step computes expected values over latent variables given current parameters, and the M-step updates \theta to maximize the resulting expected log-likelihood. For further handling of intractability, maximum a posteriori (MAP) estimation incorporates priors (e.g., Gaussian on weights) to regularize the optimization.^[2]^[22] In Markov logic networks (MLNs), parameters are real-valued weights on weighted first-order logic formulas, and learning proceeds via gradient-based optimization techniques such as limited-memory BFGS (L-BFGS) applied to the pseudo-likelihood or Monte Carlo approximations of the true likelihood. For undirected relational Markov fields, such as relational Markov networks, discriminative training methods like the voted perceptron approximate MLE by iteratively updating weights based on the difference between observed and predicted feature expectations under the current model.^[7]^[2] A representative example is estimating the probability of friendship links in a social network modeled as a probabilistic graphical model, where observed edges and node attributes (e.g., shared interests) inform MLE via pseudo-likelihood to account for collective dependencies among potential links.^[2]

Structure discovery

Structure discovery in statistical relational learning (SRL) involves identifying unknown relations, predicates, or graph topologies from relational data to construct the underlying model structure. Unlike parameter estimation, which optimizes weights within a fixed structure, structure discovery builds the relational skeleton, such as logical rules or dependency graphs, enabling the model to capture complex interdependencies among entities. This process is essential for domains with heterogeneous objects and links, like social networks or knowledge bases, where the topology is not predefined.^[1] Key methods for structure discovery include subset selection techniques for deriving logical rules, often employing beam search within inductive logic programming (ILP) frameworks to efficiently explore the space of possible clauses while limiting the search to promising candidates. Score-and-search approaches systematically evaluate candidate structures by computing a score that balances data fit and model simplicity, such as using the Bayesian Information Criterion (BIC) to add or remove graph edges in relational Markov networks. For more intricate topologies, evolutionary algorithms like genetic algorithms evolve populations of rule sets or graphs, applying selection, crossover, and mutation to navigate the vast hypothesis space and produce compact, high-performing structures in probabilistic logic programs.^[23]^[24] These methods face significant challenges, including the combinatorial explosion of potential structures, which grows exponentially with the number of predicates and entities, rendering exhaustive search infeasible. Handling noisy or incomplete data requires regularization, such as priors on clause length or complexity penalties, to prevent overfitting and ensure generalizable models.^[25] A representative example is learning citation patterns from academic publication data, where structure discovery might identify rules like "if two papers share a co-author, then the likelihood of mutual citation increases," derived from datasets such as CiteSeer or high-energy physics corpora using probabilistic relational trees.^[26] Structure scoring formalizes this trade-off mathematically:

\text{Score}(G) = \log \text{Likelihood}(\text{data} \mid G, \theta) - \text{Complexity}(G)

Common complexity metrics include the Akaike Information Criterion (AIC), BIC, or Minimum Description Length (MDL), which penalize overly complex graphs to favor parsimonious explanations.^[23]

Inference techniques

Inference in statistical relational learning (SRL) involves computing probabilities over relational queries given evidence, such as the marginal probability P(Q \mid E), where Q is a relational query and E is observed evidence in a probabilistic relational model.^[1] This process addresses the challenges of uncertainty and relational structure by deriving posterior distributions over atoms, predicates, or entire subgraphs in domains with complex dependencies.^[7] Exact inference methods in SRL, such as variable elimination or junction tree algorithms, are applicable to small grounded networks but become intractable due to the exponential growth in the number of ground atoms.^[7] Variable elimination operates by systematically summing out variables to compute marginals, while junction trees exploit cliques in the moralized graph for exact marginalization, both grounded to propositional Markov networks in SRL models like Markov logic networks (MLNs).^[27] These techniques provide precise results but are limited to domains with low treewidth after grounding.^[1] Approximate inference dominates practical SRL applications, employing methods like Markov chain Monte Carlo (MCMC) sampling, loopy belief propagation (BP), and mean-field variational inference to handle large-scale relational data.^[7] MCMC, particularly Gibbs sampling or the MC-SAT algorithm, generates samples from the posterior by walking the state space of ground atoms, efficiently incorporating logical constraints through satisfiability solvers.^[28] Loopy BP approximates marginals by iteratively passing messages between nodes in the factor graph, converging to fixed points even on loopy structures, while mean-field methods optimize factorized distributions to bound the log-likelihood.^[29] Relational aspects of SRL inference require techniques to mitigate the grounding explosion, where full instantiation of logical formulas yields exponentially many atoms. Grounding reduction employs lazy grounding, which instantiates only relevant portions of the network during inference, avoiding exhaustive enumeration by dynamically expanding based on evidence and queries.^[30] Lifted inference advances this further by performing computations at the logical level, grouping isomorphic atoms into representatives and operating on counts rather than individuals.^[29] For instance, lifted BP extends propositional BP by aggregating messages over equivalence classes of atoms, scaling to populations of thousands.^[29] In lifted inference, the number of groundings of a formula F with free variables is computed without enumeration using multiplicity terms derived from domain sizes and constants. For a formula F(x_1, \dots, x_k) where variables range over domains D_1, \dots, D_k with evidence fixing some constants, the count is:

\#\text{groundings}(F) = \prod_{i=1}^k (|D_i| - c_i)

where c_i is the number of constants fixed for variable x_i by evidence, ensuring symmetry preservation.^[27] An illustrative application is inferring hidden links in a knowledge graph represented as an MLN, where MCMC sampling, such as MC-SAT, explores configurations satisfying weighted logical rules to estimate link probabilities, enabling scalable predictions in large graphs like those in citation networks.^[28]

Applications and Examples

Web and knowledge bases

Statistical relational learning (SRL) addresses the incompleteness and uncertainty inherent in large-scale web knowledge bases, such as Freebase and DBpedia, by modeling relational data through probabilistic frameworks that incorporate logical rules and statistical inference.^[31] These models represent entities and their interconnections using relations like "locatedIn" (e.g., connecting cities to countries) or "worksFor" (e.g., linking people to organizations), enabling the handling of noisy, crowdsourced extractions from sources like Wikipedia.^[31] By combining first-order logic with probabilistic distributions, SRL infers missing facts while accounting for dependencies across the graph structure.^[31] Key applications of SRL in web and knowledge bases include entity linking, which aligns textual mentions to canonical entities in the graph; knowledge completion, which predicts absent relational triples; and ontology matching, which identifies semantic correspondences between schemas of different knowledge bases.^[31] In entity linking, SRL leverages collective inference to resolve ambiguities by considering relational context, improving accuracy over independent classification.^[32] Knowledge completion uses embedding techniques to forecast links, such as predicting that a person "worksFor" an organization based on co-occurrence patterns.^[31] Ontology matching employs SRL to align concepts across bases like DBpedia and Freebase, capturing hierarchical and relational constraints.^[33] Notable examples demonstrate SRL's practical deployment. Google's Knowledge Vault project integrates SRL-inspired methods, such as embedding-based models and path-ranking algorithms, to fuse web extractions into a probabilistic knowledge base with over 271 million high-confidence triples, expanding beyond Freebase's coverage by 33%.^[34] In DBpedia, Markov Logic Networks (MLNs) enhance quality by inferring missing types and relations, achieving up to 94% F-measure in type completion tasks through weighted logical rules and scalable inference.^[35] For link prediction in RDF graphs, Probabilistic Soft Logic (PSL) applies hinge-loss Markov random fields to model soft constraints, enabling efficient prediction of new edges in relational domains like social or citation networks represented as RDF triples.^[20] Techniques central to these applications include scalable grounding, which approximates inference over billions of ground atoms in large graphs, and embedding methods integrated with probabilistic logic, such as those in Knowledge Vault that combine neural embeddings with fusion models for uncertainty handling.^[31] PSL facilitates this scalability by relaxing logical rules into continuous probabilities, supporting distributed optimization for web-scale data.^[20] The impact of SRL in web and knowledge bases lies in enhancing search accuracy, such as through better entity disambiguation in query processing, and managing uncertainty in crowdsourced or extracted data, leading to more reliable knowledge graphs for applications like semantic search.^[34] These advancements have enabled the construction of robust, probabilistic repositories that support real-world inference at scale.^[31]

Biological and chemical domains

Statistical relational learning (SRL) excels in modeling complex biological networks, such as those connecting genes, proteins, and diseases, by incorporating relational uncertainty and handling incomplete or noisy data inherent in bioinformatics datasets.^[36] These models represent entities as nodes in relational structures and use probabilistic frameworks to capture dependencies, enabling predictions that account for interconnectedness rather than isolated attributes.^[36] In cheminformatics, SRL similarly addresses chemical networks by modeling compound interactions and properties within large repositories like PubChem. Key applications in biological domains include pathway prediction, where SRL infers gene or protein interactions from co-expression patterns and other relational evidence, aiding in the reconstruction of signaling or metabolic pathways.^[36] Another prominent use is collective classification for gene function annotation, which leverages protein-protein interaction networks and sequence similarities to propagate functional labels across related genes, improving accuracy over independent predictions. In chemical domains, SRL supports drug discovery through relational modeling of compound-target interactions, facilitating the identification of potential therapeutics from bioactivity data. Notable examples include the application of Markov Logic Networks (MLNs) in multi-level protein-protein interaction prediction, where SRL frameworks integrate relational data from multiple biological scales to forecast interactions with improved precision over traditional methods. For protein subcellular localization, relational approaches using probabilistic graphical models have enhanced predictions by incorporating network-based evidence, such as protein associations, achieving higher accuracy in eukaryotic species. In cheminformatics, relational learning models applied to PubChem data enable chemical compound similarity assessment, using graph-based representations to cluster molecules by structural and functional relations for virtual screening in drug discovery. Techniques central to these domains involve multi-relational learning, which combines heterogeneous data sources like sequence alignments, interaction graphs, and expression profiles to model biological entities comprehensively.^[36] To address sparsity common in biological and chemical datasets—where interactions are often underrepresented—SRL employs methods like Relational Markov Networks and Inductive Logic Programming to infer missing links through probabilistic inference.^[36] Semantic-based regularization, a SRL variant, further refines collective classification by incorporating relational priors for tasks like protein function prediction. The impact of SRL in these areas is significant, as it accelerates drug target identification by prioritizing candidates through relational pathway analysis and interaction forecasting.^[36] Additionally, it enables integration of multi-omics data, such as genomics and proteomics, to uncover holistic disease mechanisms and support precision medicine initiatives.^[36] Statistical Relational Learning (SRL) provides a powerful framework for modeling dynamic relations in social graphs, such as those involving users, posts, and follows, by incorporating probabilistic models to account for uncertainty in user behaviors and interactions. Unlike traditional graph analysis, SRL integrates relational structure with statistical inference, enabling the representation of complex dependencies like homophily—where connected nodes share similar attributes—and influence propagation across networks. This approach captures evolving social dynamics, such as how information spreads through ties, while handling incomplete or noisy data inherent in platforms like social media.^[7]^[37] Key applications of SRL in social networks include link prediction in friendship networks, where models infer potential connections based on relational patterns; rumor detection through relational inference, which identifies anomalous propagation paths; and personalized recommendation systems that leverage user-item interactions modeled as probabilistic relations. For instance, link prediction uses techniques like Markov Logic Networks (MLNs) to weigh evidence from multiple relational features, improving accuracy over non-relational methods in sparse networks. Rumor detection benefits from collective inference to assess credibility across interconnected posts and users, while recommendations employ relational classifiers to suggest links or content by propagating preferences through the graph. These applications enhance understanding of social behaviors by jointly modeling entities and their relations.^[38]^[39] Notable examples illustrate SRL's efficacy: Twitter user modeling with MLNs has been applied to track topic propagation, where logical rules encode relational dependencies between users, tweets, and topics, enabling probabilistic predictions of influence spread. Similarly, collective classification on the Cora citation network dataset demonstrates how SRL improves node labeling by considering citation links and co-author relations, achieving higher accuracy than independent classifiers through iterative inference over the graph. These cases highlight SRL's ability to exploit network structure for tasks involving interdependent predictions.^[39]^[40] Techniques in SRL for social networks often involve temporal extensions to handle evolving structures, such as dynamic Bayesian networks or time-augmented MLNs that model link formation over snapshots of the graph. Multi-modal integration combines text from posts with graph edges, using hybrid models like probabilistic soft logic to fuse content features with relational data for richer representations. These methods address the temporal and heterogeneous nature of social data, allowing for scalable inference on large, changing networks.^[41]^[42] The impact of SRL in social and network analysis is significant, enhancing fraud detection by identifying suspicious cliques or anomalous relational patterns in financial or online networks, and supporting targeted advertising through models of influence and homophily that predict user engagement. By quantifying relational uncertainties, SRL enables more robust interventions, such as early detection of fraudulent behaviors or optimized marketing campaigns based on propagation dynamics.^[43]^[7]

Challenges and Future Directions

Scalability issues

One of the primary computational challenges in statistical relational learning (SRL) arises from the exponential growth in the number of groundings required for relational models. For a domain with n entities and a predicate of arity k, the number of possible ground atoms can reach O(n^k), leading to a combinatorial explosion that severely limits scalability when applying first-order logic rules to large datasets.^[44] This issue is exacerbated in relational data, where the joint distribution over states grows exponentially in both the number of attributes and objects, resulting in O(2^{nm}) possible states for m objects and n binary attributes per object.^[3] Exact inference in such grounded models, particularly in frameworks like Markov logic networks (MLNs), is #P-complete, rendering it intractable for domains beyond small scales.^[7] To address these bottlenecks, researchers have developed approximate inference methods, such as Markov chain Monte Carlo (MCMC) sampling and approximate counting techniques, which reduce the need for full grounding by estimating probabilities over subsets of the model.^[7]^[45] Lifted inference approaches further mitigate scalability by operating on groups of symmetric variables rather than individual groundings, enabling polynomial-time computations in the domain size for certain models.^[46] Hardware accelerations, including GPU-based parallel processing for rule coverage and grounding operations, have also been integrated into relational learning systems, achieving up to 8x speedups on datasets involving thousands of examples compared to CPU implementations.^[47] Scalability is often measured by time and space complexity, where full grounding incurs exponential space requirements (O(2^{nm})) and inference times that grow superlinearly with domain size, while lifted methods achieve near-linear scaling in practice for benchmarks like protein interaction networks with millions of random variables.^[3]^[46] For instance, systems like Tuffy scale MLN inference to large knowledge bases with billions of triples by leveraging database optimizations, completing tasks like entity resolution in minutes that crash in-memory approaches.^[48] However, even advanced MLN implementations face limitations on web-scale data, where sequential processing and memory overflows prevent handling datasets exceeding hundreds of millions of facts without approximations.^[48] In comparison to non-relational machine learning methods like deep learning, SRL models exhibit poorer scalability for joint relational reasoning due to their explicit handling of dependencies, though SRL provides more accurate uncertainty estimates.^[49] Historically, SRL inference evolved from exact grounding methods dominant in the 2000s, such as first-order variable elimination introduced in 2003, which still suffered from propositional explosion, to lifted approximations in the 2010s and 2020s, including knowledge compilation techniques like first-order negation normal form circuits that support tractable model counting over thousands of objects.^[46] These advancements, building on early symmetries exploitation, have enabled SRL applications on datasets previously infeasible, shifting focus from brute-force computation to symmetry-aware algorithms.^[46]

Interpretability and evaluation

Evaluating the performance of statistical relational learning (SRL) models requires metrics that account for both predictive accuracy and the handling of relational structures, such as dependencies among entities. Common metrics include accuracy for classification tasks, where models achieve reported rates like 87.2% for venue prediction and 93.1% for link prediction on benchmark datasets such as Cora.^[50] For ranking-based tasks like link prediction, the area under the receiver operating characteristic curve (AUC-ROC) is widely used, with examples showing values around 0.92 on databases like UW-CSE.^[51] Generative models are often assessed via log-likelihood, which measures the fit to observed data and guides parameter estimation. Relational-specific metrics, such as precision@K, evaluate the quality of top-K predictions in link prediction scenarios, emphasizing the relevance of predicted relations over exhaustive recall. Interpretability in SRL focuses on methods that reveal the underlying relational logic, making "black-box" predictions more transparent. In Markov logic networks (MLNs), rule extraction involves identifying top-weighted logical formulas, where weights represent confidence levels and can be interpreted as log-odds ratios for relational patterns. This approach allows users to inspect learned clauses, such as those encoding entity dependencies, directly from the model. Visualization techniques further enhance interpretability by rendering relational graphs, highlighting node connections and edge probabilities to illustrate how inferences propagate across structures. These methods contrast with non-relational baselines like graph neural networks (GNNs), where interpretability often relies on post-hoc explanations rather than inherent logical representations.^[52] Challenges in SRL evaluation stem from the lack of standardized benchmarks, leading to inconsistent comparisons across diverse relational datasets and tasks. Relational data often introduces biases, such as overrepresentation of certain entity types, which can skew model generalization and fairness assessments. Evaluation methods include adapted cross-validation in relational settings, where splits are performed at the entity level to prevent information leakage from shared relations, ensuring robust out-of-sample performance estimates. Ablation studies dissect the contributions of structure discovery versus parameter estimation, revealing, for instance, that relational structure learning significantly boosts accuracy over parameter-only models in tasks like entity resolution. A key gap in current SRL coverage is the need for causal interpretability, which has emerged post-2020 to address limitations in correlational predictions. Frameworks like Causal Relational Learning (CARL) enable declarative modeling of causal effects in relational domains, supporting counterfactual reasoning over graphs to explain intervention outcomes. This is particularly vital for applications requiring "what-if" analyses, where traditional SRL metrics fall short in distinguishing causation from association.^[53]

Emerging trends

One prominent emerging trend in statistical relational learning (SRL) is the integration of neuro-symbolic AI, which combines SRL's probabilistic logical frameworks with neural networks to enhance reasoning and learning capabilities. This approach addresses limitations in pure neural methods by incorporating symbolic structure for better interpretability and generalization. A key example is DeepProbLog, introduced in the late 2010s, which extends probabilistic logic programming with neural predicates to support hybrid inference and learning over relational data.^[54] DeepProbLog has been applied to tasks like program induction and visual question answering, demonstrating improved performance on benchmarks involving uncertainty and relational structure compared to traditional SRL or neural methods alone.^[55] Recent advances as of 2024 continue to emphasize neurosymbolic methods, including surveys on transitions from SRL to broader neurosymbolic AI.^[56] Another trend involves federated learning adaptations for SRL to preserve privacy in distributed relational datasets, such as those in healthcare or social networks. By aligning basis representations and applying weight penalties, federated SRL models enable collaborative training without sharing raw data, mitigating privacy risks while maintaining model accuracy. This is particularly relevant for scenarios where relational dependencies span multiple entities across decentralized sources.^[57] SRL is expanding into new domains, including robotics for relational planning under uncertainty. In multi-object manipulation tasks, SRL models learn affordance relations—such as grasping dependencies between objects—using probabilistic rules to handle noisy sensor data and partial observability. For instance, relational affordance models for two-arm robots integrate logical predicates with statistical inference to predict feasible actions in dynamic environments, outperforming non-relational baselines in simulation and real-world trials.^[58] Similarly, in climate modeling, spatio-temporal extensions of SRL, such as relational random forests, capture dependencies in weather patterns like severe storms, enabling probabilistic forecasts that account for geographic and temporal relations in large-scale environmental data.^[59] Advancements in automated machine learning (AutoML) for SRL focus on streamlining hyperparameter tuning and model selection in relational settings. Tools like AutoSmart automate preprocessing and optimization for temporal relational data, reducing manual effort while achieving competitive performance on link prediction tasks. Quantum-inspired inference methods further address scalability challenges in SRL by embedding relational knowledge into quantum-like structures for efficient probabilistic reasoning over large graphs. These approaches leverage superposition-inspired representations to approximate lifted inference, speeding up computations on knowledge bases with millions of relations.^[60]^[61] Future directions emphasize ethical AI in SRL, particularly mitigating bias in social models that infer relations from networked data. Frameworks like probabilistic soft logic (PSL) extend SRL to define relational fairness constraints, quantifying group-level biases in link prediction and ensuring equitable outcomes across demographics. Standardization of benchmarks is also gaining traction, with suites like RelBench (introduced in 2024) providing unified evaluation protocols for graph-based relational learning tasks, facilitating comparisons across models and promoting reproducible research.^[62]^[63] Recent works post-2020 highlight SRL's integration with graph transformers, which combine attention mechanisms with relational probabilistic models for enhanced representation learning on knowledge graphs. This synergy improves tasks like complex query answering by pre-training transformers on masked relational patterns, achieving state-of-the-art results on datasets like WebQuestionsSP. Open-source tools such as Tuffy continue to support these advancements by enabling scalable inference in Markov logic networks (MLNs), a foundational SRL paradigm, through database-backed optimization for real-world applications.^[64]^[65]

References

[1]
[PDF] Becoming a Self-Regulated Learner: An Overview
Recent research shows that self-regulatory processes are teachable and can lead to increases in students' motivation and achievement (Schunk. & Zimmerman, 1998) ...
[2]
A Review of Self-regulated Learning: Six Models and Four ...
Apr 28, 2017 · Self-regulated learning (SRL) is a core conceptual framework to understand the cognitive, motivational, and emotional aspects of learning.
[3]
A Review of Self-regulated Learning: Six Models and Four ...
Self-regulated learning (SRL) is a core conceptual framework to understand the cognitive, motivational, and emotional aspects of learning. SRL has made a major ...
[4]
[PDF] Self-Regulated Learning - LINCS
Instead, self-regulation is a self-directive process and set of behaviors whereby learners trans- form their mental abilities into skills (Zimmerman,. Bonnor, & ...
[5]
Supporting students' self-regulated learning in online learning using ...
Jun 26, 2023 · SRL is defined as “the process whereby students activate and sustain cognition, behaviors, and affects, which are systematically oriented toward ...Background · Methods · Results
[6]
Harnessing Self-Regulated Learning to Empower Underperforming ...
Self-regulated learning (SRL) represents a critical educational framework through which learners proactively govern their learning processes using self- ...
[7]
Development of self-regulated learning: a longitudinal study on ...
Dec 10, 2021 · Self-regulated learning (SRL) encompasses the strategies and behaviours that allow students to transform cognitive abilities into task-specific ...Introduction · Methods · Discussion<|control11|><|separator|>
[8]
[PDF] Introduction to Statistical Relational Learning
Page 1. EDITED BY LISE GETOOR AND BEN TASKAR. STATISTICAL RELATIONAL LEARNING. INTRODUCTION ... Statistical Relational Learning and its Connections to Other ...
[9]
[PDF] A Survey on Statistical Relational Learning - UBC Computer Science
Statistical Relational Learning(SRL) is a new branch of machine learning that tries to model a joint distribution over relational data[9]. SRL is a combination ...
[10]
Learning logical definitions from relations
FOIL, a system that constructs definitions. This section introduces a learning algorithm called FOIL whose ancestry can be traced to both AO and ID3. The ...
[11]
[PDF] Learning Probabilistic Models of Link Structure
Popescul et al. (2002) use an approach that uses a relational learner to guide in the construction of features to be used by a (statistical) propositional ...
[12]
[PDF] Markov Logic Networks - CSE Home
Abstract. We propose a simple approach to combining first-order logic and probabilistic graphical models in a single representation. A Markov logic network ...
[13]
[PDF] Statistical Relational Learning - cs.wisc.edu
• structure learning: can use ordinary relational learning methods ... • The dependency structure is defined by associating with each attribute X.A ...
[14]
[PDF] Markov Logic: A Unifying Framework for Statistical Relational Learning
The framework should facilitate the extension to. SRL of techniques from statistical learning, induc- tive logic programming, probabilistic inference and.
[15]
http://www.robotics.stanford.edu/~koller/Papers/Taskar+al:SRL07.pdf
[16]
Introduction to Statistical Relational Learning - MIT Press Direct
Lise Getoor is Assistant Professor in the Department of Computer Science at the University of Maryland. ... Ben Taskar is Assistant Professor in the Computer and ...
[17]
[PDF] Probabilistic Inductive Logic Programming - People | MIT CSAIL
Kersting and L. De Raedt). 5 Conclusions. In this paper, we have presented three settings for probabilistic inductive logic programming ...
[18]
[PDF] Markov Logic: A Unifying Framework for Statistical Relational Learning
Abstract. Interest in statistical relational learning (SRL) has grown rapidly in recent years. Several key SRL tasks have been identified, and a large ...Missing: survey | Show results with:survey<|control11|><|separator|>
[19]
None
Summary of each segment:
[20]
[PDF] Learning Probabilistic Relational Models - IJCAI
Prospective assessment of ai technologies for fraud detection: A case study. In AAAI Workshop on Al Approaches to Fraud Detection and Risk Management, 1997.
[21]
None
### Summary of Relational Markov Networks
[22]
Logical and Relational Learning - SpringerLink
Book Title: Logical and Relational Learning · Editors: Luc De Raedt · Series Title: Cognitive Technologies · Publisher: Springer Berlin, Heidelberg · eBook Packages ...
[23]
[PDF] Log-Linear Description Logics - IJCAI
Log-linear description logics are a family of prob- abilistic logics integrating various concepts and methods from the areas of knowledge representa- tion and ...
[24]
[PDF] ProbLog: A Probabilistic Prolog and Its Application in Link Discovery
We introduce ProbLog, a probabilistic extension of. Prolog. A ProbLog program defines a distribution over logic programs by specifying for each clause.
[25]
Hinge-Loss Markov Random Fields and Probabilistic Soft Logic - arXiv
May 17, 2015 · In this paper, we introduce two new formalisms for modeling structured data, and show that they can both capture rich structure and scale to big data.
[26]
[PDF] Hinge-Loss Markov Random Fields and Probabilistic Soft Logic
PSL provides a natural interface to represent hinge-loss potential templates using two types of rules: logical rules and arithmetic rules. Logical rules are ...
[27]
[PDF] BLOG: Probabilistic Models with Unknown Objects - People @EECS
This paper introduces and illustrates BLOG, a formal lan- guage for defining probability models over worlds with unknown objects and identity uncertainty.
[28]
[PDF] 1 Bayesian Logic Programming: Theory and Tool - People | MIT CSAIL
Learning in Graphical Models. MIT Press, 1998. K. Kersting and L. De Raedt. Adaptive Bayesian Logic Programs. In C. Rouveirol and M.
[29]
[PDF] Learning the Structure of Markov Logic Networks
Richardson and Domingos first learned the structure of an MLN using CLAUDIEN, and then learned maximum pseudo-likelihood weights for it us- ing L-BFGS. In the ...
[30]
Evolutionary learning of probabilistic logic programs - ScienceDirect
Aug 3, 2025 · In this paper, we propose a genetic algorithm for solving the structure learning task. We built it on top of two existing parameter learning ...
[31]
[PDF] Structure Learning for Statistical Relational Models
This work will be extended to analyze hypothesis testing techniques used in learning directed and undirected joint models, and to develop accurate structure.
[32]
[PDF] Exploiting Relational Structure to Understand Publication Patterns in ...
We focus our analyses on four related areas: understanding and identify- ing patterns of citations, examining publication patterns at the author level, ...
[33]
[PDF] Lifted Probabilistic Inference with Counting Formulas
We will use gr (L) to denote the set of ground substitutions that map all the variables in L to constants. We will also consider restricted sets of groundings ...
[34]
[PDF] Sound and Efficient Inference with Probabilistic and Deterministic ...
In this paper we propose MC-SAT, an infer- ence algorithm that combines ideas from MCMC and satis- fiability. MC-SAT is based on Markov logic, which defines.
[35]
[PDF] Lifted First-Order Belief Propagation - CSE Home
Unfortunately, variable elimination has exponential cost in the treewidth of the graph, making it in- feasible for most real-world applications. Scalable approx ...
[36]
[PDF] Speeding Up Inference in Markov Logic Networks by Preprocessing ...
Poon, P. Domingos, and M.Sumner. A gen- eral method for reducing the complexity of relational inference and its application to MCMC. In ...
[37]
[PDF] A Review of Relational Machine Learning for Knowledge Graphs
Apr 7, 2015 · In the following sections we discuss how statistical relational learning can be applied to knowledge graphs. We will assume that all the ...
[38]
[PDF] Markov Logic Networks for Knowledge Base Completion
We learn the weights of a Markov logic network using maximum likelihood estimation on this knowledge base and then use the learned. Markov logic network to ...Missing: Google | Show results with:Google
[39]
Anatomy Ontology Matching Using Markov Logic Networks - PMC
Ontology matching is a kind of solutions to find semantic correspondences between entities of different ontologies. Markov logic networks which unify ...
[40]
A web-scale approach to probabilistic knowledge fusion
Knowledge Vault is a web-scale probabilistic knowledge base combining web content extractions with prior knowledge, using machine learning to fuse information.
[41]
[PDF] ENHANCING DBPEDIA QUALITY USING MARKOV LOGIC ...
Jun 30, 2018 · Statistical relational learning techniques have not been used previously to infer missing types in knowledge bases whether DBpedia or other ...
[42]
https://www.ll.mit.edu/sites/default/files/publication/doc/social-network-analysis-content-graphs-campbell-ja-22727.pdf
[43]
(PDF) Social networks and statistical relational learning: a survey
Aug 6, 2025 · This combination facilitates probabilistic inference and statistical analysis on top of graphical representations of relational knowledge.
[44]
[PDF] Statistical Relational Learning for Link Prediction | Semantic ...
Statistical Relational Learning for Link Prediction ... Chapter 1 LINK PREDICTION IN SOCIAL NETWORKS Link Prediction ... A Survey of Link Prediction in Social ...
[45]
Probabilistic Inference on Twitter Data to Discover Suspicious Users ...
SocialKB is based on Markov Logic Networks (MLNs), a popular representation in statistical relational learning. It learns a knowledge base (KB) on the ...
[46]
[PDF] Collective Classification in Network Data - Tina Eliassi-Rad
The Cora dataset contains a number of Machine Learning papers divided into one of 7 classes while the CiteSeer dataset has 6 class labels. For both datasets, we ...Missing: SRL | Show results with:SRL
[47]
Representations and Ensemble Methods for Dynamic Relational ...
Nov 22, 2011 · We propose a framework for discovering temporal representations of relational data to increase the accuracy of statistical relational learning ...
[48]
[PDF] Social network analysis with content and graphs
Lincoln Laboratory used statistical relational learning algorithms to predict leadership roles of individuals in a group on the basis of patterns of activity, ...
[49]
(PDF) Using relational knowledge discovery to prevent securities fraud
Using statistical relational learning algorithms, we developed models that rank brokers with respect to the probability that they would commit a serious ...
[50]
[PDF] Grounding Methods for Neural-Symbolic AI - IJCAI
This leads to a combinatorial explosion in the num- ber of ground formulas to consider and, therefore, strongly limits their scalability.
[51]
[PDF] Scalable Training of Markov Logic Networks using Approximate ...
In this paper, we propose principled weight learning algo- rithms for Markov logic networks that can easily scale to much larger datasets and application ...
[52]
[PDF] Lifted Inference and Learning in Statistical Relational Models
The techniques presented in this thesis are evaluated empirically on statistical relational models of thousands of interacting objects and millions of random.
[53]
(PDF) Relational Learning with GPUs: Accelerating Rule Coverage
Aug 7, 2025 · In this work we present the design of a relational learning system, that takes advantage of graphics processing units (GPUs) to perform the most ...
[54]
[PDF] Tuffy: Scaling up Statistical Inference in Markov Logic Networks ...
ABSTRACT. Over the past few years, Markov Logic Networks (MLNs) have emerged as a powerful AI framework that combines statistical and logical reasoning.
[55]
A comparison of statistical relational learning and graph neural ...
Jun 17, 2021 · Statistical relational learning (SRL) and graph neural networks (GNNs) are two powerful approaches for learning and inference over graphs.
[56]
https://www.sciencedirect.com/science/article/pii/S0004370223002084
[57]
Causal Relational Learning | Proceedings of the 2020 ACM ...
May 31, 2020 · In this paper, we present a formal framework for causal inference from such relational data. We propose a declarative language called CARL for capturing causal ...
[58]
[PDF] DeepProbLog: Neural Probabilistic Logic Programming - arXiv
Dec 12, 2018 · DeepProbLog Inference Inference in DeepProbLog works exactly as described above, except ... Introduction to statistical relational learning. MIT ...
[59]
Neural probabilistic logic programming in DeepProbLog
We now recall the basics of probabilistic logic programming using ProbLog (see De Raedt and Kimmig [6] for more details), and then introduce our new language ...
[60]
Improving Federated Relational Data Modeling via Basis Alignment ...
Nov 23, 2020 · tional data over the federated networks in the later section. 2.1 Federated Learning. Federated ... Statistical Relational Learning. MIT Press.
[61]
[PDF] Learning Relational Affordance Models for Two-Arm Robots
uncertainty. In ... For both these needs, statistical relational learning (SRL) [1], [2], ... probabilistic relational planning rules,” in ICAPS, 2004, pp.
[62]
[PDF] using spatiotemporal relational random forests to improve our ...
Keywords: spatiotemporal data mining, statistical relational learning, severe weather, random ... Spatio-temporal multi- dimensional relational framework trees.Missing: climate | Show results with:climate
[63]
[PDF] AutoSmart: An Efficient and Automatic Machine Learning framework ...
Sep 9, 2021 · In this work, we propose a general AutoML framework that sup- ports temporal relational data, including automatic data prepro- cessing, ...
[64]
[PDF] Quantum Embedding of Knowledge for Reasoning - NIPS papers
Statistical Relational Learning (SRL) methods are the most widely used techniques to generate distributional representations of the symbolic Knowledge Bases ...
[65]
Fairness in Relational Domains - ACM Digital Library
Furthermore, we extend an existing statistical relational learning framework, probabilistic soft logic (PSL), to incorporate our definition of relational ...
[66]
[PDF] Graph Representation Learning on Relational Databases - RelBench
Nov 27, 2023 · The core idea is to view relational tables as a heterogeneous graph, with a node for each row in each table, and edges specified by primary- ...
[67]
Graph Neural Networks: Taxonomy, Advances, and Trends
Jan 10, 2022 · ... Statistical Relational Learning (SRL). Specifically, the SRL usually models ... Text generation from knowledge graphs with graph transformers. In The ...
[68]
Tuffy: A Scalable Markov Logic Inference Engine
Tuffy is an open-source Markov Logic Network inference engine, and part of Felix. Check out our new demos built with Tuffy/Felix! Markov Logic Networks ...

SRL