Computer scientists have introduced RuleMiner, an open-source, web-based software platform designed to conduct interactive data exploration and predictive modeling through rule-based approaches. Developed to bridge the gap between programmable machine learning libraries and domain specialists who require interpretable models without writing software code, the system wraps the computational capabilities of the RuleKit library within a graphical interface accessible directly through standard web browsers. Unlike conventional machine learning pipelines that present complex internal parameters or opaque predictive outputs, rule-based induction derives structured if-then statements composed of explicit premises and conclusions. These intuitive representations enable practitioners across various scientific and industrial disciplines to understand model logic, inspect training data quality, detect latent dataset biases, and directly integrate their own specialized domain expertise into computational workflows. By providing a hosted graphical deployment alongside freely accessible containerized source code, the developers have established an environment where predictive induction, structural ruleset modification, and immediate statistical validation operate within a single unified workspace.
The system emerges directly from earlier foundational work on the RuleKit library, a well-established sequential covering algorithmic suite created for rule learning. While RuleKit previously demonstrated competitive predictive performance against standard rule-learning techniques and was later expanded into RuleKit2 with optimized computational routines and a basic interface, its primary distribution as a programmable Python library restricted practical adoption primarily to data scientists and software engineers. Furthermore, earlier conference presentations of RuleMiner established the initial feasibility of a visual interface, but the current production release introduces extensive architectural enhancements and broader analytical capabilities. The updated platform is distributed under an AGPL 3.0 open-source license, allowing organizations to deploy local instances tailored to private infrastructure requirements. Beyond its core sequential covering routines, the software introduces a specialized deep-rules algorithm capable of synthesizing complex compound expressions containing conjunctions and alternative logical operators, alongside specialized import pipelines capable of parsing rulesets generated by external algorithms such as MLRules and LORD or external systems following specified tabular formats.
To support responsive user interaction during demanding computational workloads, RuleMiner is built upon a modular microservices architecture distributed across five distinct operational layers encapsulated within independent Docker containers. The presentation layer provides the client interface via an Angular application served by an nginx web server, utilizing the Redux state management library to maintain a single central data store across multiple browser tabs and persistent working sessions. The application layer centers on a portal service implemented with the Django REST Framework, which authenticates incoming client traffic, verifies authorization permissions, manages project entities, interacts with persistent records, and directs computational tasks. By distinguishing between lightweight requests and resource-intensive analytical pipelines, the portal service processes immediate inquiries synchronously while offloading long-running calculations to dedicated background workers, ensuring that browser sessions remain fully responsive and that standard HTTP network connections do not stall while waiting for extensive data induction routines to terminate.
The underlying execution engine is organized within a dedicated computation layer that bifurcates analytical tasks according to operational complexity and execution duration. Lightweight, on-demand requests, such as recalculating rule coverage metrics or dataset performance indicators immediately following manual ruleset edits, are directed through a FastAPI service called rules-evaluation-service via synchronous REST API calls. Conversely, computationally demanding operations, including automated rule induction from large datasets, k-fold cross-validation routines, and comprehensive analytical report generation, are managed by background worker processes built with Celery. These workers communicate with the Django application layer through a RabbitMQ message queue, fetching pending jobs asynchronously as computational capacity becomes available across the host environment. The underlying inductive engine executes sequential covering algorithms via RuleKit, while internal rule structures, indicator scoring, and mathematical evaluations are computed through the companion decision-rules library, allowing processing capacity to scale dynamically by altering worker process counts.
Data storage and operational observability are partitioned across dedicated system tiers to prevent resource contention between metadata transactions and intensive analytical file operations. The data layer separates structured application records from raw analytical datasets by maintaining two independent PostgreSQL databases: a primary instance storing user projects, algorithm hyperparameter configurations, experiment parameters, and generated rulesets, alongside a secondary database designated as db-storage that houses uploaded training and evaluation datasets. Analytical report files generated by Celery workers are archived on an SFTP report-storage server managed through application access credentials. Complementing these storage facilities, the monitoring layer incorporates Prometheus to gather real-time performance metrics across all deployed microservices, Loki to aggregate and index container log outputs, and Grafana to render diagnostic dashboards and issue operational alerts, providing administrators with continuous visibility into backend health, computational bottlenecks, and worker queue stability.
When compared to established visual data mining environments, RuleMiner occupies a distinct functional niche within predictive modeling software. Widely deployed desktop platforms, including Orange, RapidMiner AI Studio, KEEL, KNIME, Weka, and QMAK, offer rule induction as secondary components among broad catalogs of general machine learning algorithms, allowing users to inspect or execute static rulesets without providing tools to alter underlying logic. While KNIME permits manual rule creation, those definitions operate purely as independent filter groups unlinked to induced machine learning models. EasyMiner provides a dedicated web-based rule environment based on association rule algorithms, yet its induction constraints remain limited to rule pattern templates and its scope excludes regression and survival modeling. Among evaluated visual systems, RuleMiner represents the only application that simultaneously supports manual rule definition, user-guided rule induction, and interactive ruleset editing across classification, regression, and survival analysis problems within a single unified deployment.
The functional workflow of the platform spans the complete lifecycle of tabular data exploration, model induction, interactive structural editing, and predictive validation. Users begin by uploading tabular datasets into dedicated project spaces, where the software calculates problem-specific descriptive metrics and allows users to partition raw observations into stratified training and testing splits. During model generation, users can trigger automatic induction or employ guided configurations, defining mandatory, preferred, or forbidden attributes and conditions to steer the search space toward domain-relevant relationships. Users can also manually craft individual rules from scratch or import existing rulesets produced by external systems. Following induction, the analysis interface visualizes the resulting models as interactive graphs, quantifies condition and attribute importance, and assesses predictive coverage across individual training records, enabling researchers to verify which logical conditions govern particular model predictions and conduct detailed semantic and syntactic comparisons across competing rule collections.
A central design feature of RuleMiner is its real-time editing capability, which allows practitioners to interactively reshape rule collections and observe immediate feedback on overall predictive accuracy and quality indicators. Within the graphical interface, users can add novel rule statements, modify numerical condition boundaries, delete specific rules, import individual clauses from external rulesets, assign categorical labels, or temporarily disable rules to test their marginal impact on overall performance. Once modifications are committed, the evaluation service recalculates performance metrics synchronously, showing precisely how structural refinements alter precision, recall, and sample coverage. Once calibrated, models can be applied to generate predictions on unseen test datasets, with the system supplying complete explanatory lists of every covering rule responsible for a given classification or estimate. Furthermore, the platform includes automated reporting modules that compile exploratory analyses, knowledge discovery summaries, and predictive assessments into exportable formats.
The practical utility of the platform was demonstrated through an illustrative classification analysis of the publicly available seismic-bumps dataset, which documents geological sensor readings from underground coal mining shifts. After loading the tabular data and performing a stratified split into training and testing partitions, the researchers executed an automated induction routine guided by an interface questionnaire designed to select analysis preferences without requiring manual algorithmic hyperparameter tuning. The resulting classification model produced a concise ruleset comprising 21 statistically significant rules that collectively achieved complete coverage across the entire training sample. Inspection of attribute importance revealed distinct physical mechanisms governing shift safety: safe shifts, designated as class zero, were predominantly characterized by baseline acoustic signals including gimpuls, nbumps, and nbumps2, whereas hazardous shifts, designated as class one, depended heavily on higher energy indicators, including genergy, senergy, maxenergy, gimpuls, and nbumps2.
Evaluation of the induced seismic model demonstrated balanced predictive generalizability across independent test observations without exhibiting substantial evidence of empirical overfitting. On the testing dataset, the 21-rule model achieved a balanced accuracy of 0.73, mirroring performance indicators observed on the training partition. When confronted with 51 documented hazardous mining shifts within the test partition, the model successfully identified 31 hazardous events while missing 20, incurring a cost of 101 false alarm classifications across 725 safe shifts. Through the interface, investigators could inspect the specific numerical thresholds defining each hazardous alert, alter condition limits to balance detection sensitivity against false alarm rates, and evaluate how individual rule modifications impacted operational risk metrics. The authors noted that full tutorials replicating this classification pipeline, alongside additional instructional walkthroughs covering regression datasets and survival analysis modeling, are documented within the project’s online technical manual.
Prior to formal publication, the software underwent extensive quality assurance, combining automated end-to-end testing suites written with Cypress with practical user evaluation involving approximately 50 master’s students enrolled in the Knowledge Discovery course at the Silesian University of Technology. As of October 30, 2026, the hosted platform recorded 100 registered user accounts, encompassing academic students, platform developers, researchers across the Łukasiewicz Research Network, and external collaborators from independent universities and commercial enterprises. The underlying RuleKit computation engine has achieved widespread community adoption, logging more than 20,000 independent downloads. RuleMiner has already supported specialized clinical research initiatives, including the analysis of pediatric medical records from a clinical study targeting acute lymphoblastic leukemia and cardiovascular datasets from the Department of Electrocardiology at the Upper-Silesian Medical Centre, while incorporating research findings developed across three doctoral dissertations completed at the Silesian University of Technology and the Łukasiewicz AI Institute.
Subject of Research: Web-based interactive rule induction and ruleset editing for classification, regression, and survival analysis
Article Title: RuleMiner: Interactive data exploration using rule-based approach
Article References: Badura, J., Macha, D., Macha, Ł., Wróbel, Ł., & Sikora, M. (2026). RuleMiner: Interactive data exploration using rule-based approach. SoftwareX, 36, Article 103228. https://doi.org/10.1016/j.softx.2026.103228
Image Credits: AI Generated
DOI: 10.1016/j.softx.2026.103228
Keywords: RuleMiner, RuleKit, rule induction, explainable artificial intelligence, interactive machine learning, survival analysis, classification models, regression analysis, knowledge discovery, microservices architecture, decision rules, user driven induction, seismic bumps dataset, model interpretability, open source software, Celery task queue, data quality control
News Source: Blake Davidson. (November 14, 2026). Interactive RuleMiner Platform Brings Rule Induction and Survival Analysis to Web Browsers. Scienmag.



