Rivu

Overview

Rivu is a Rust reimplementation of incremental learning ideas popularized by the Massive Online Analysis (MOA) framework. It focuses on prequential evaluation (test-then-train) for streaming classification with learners such as Naive Bayes and Hoeffding Trees, while providing an interactive command line wizard and real-time console output.

Features

• Prequential evaluation runner – Interleaves prediction and training while honoring optional limits on processed instances and wall-clock time. Periodically samples performance metrics and RAM-hours usage so you can track drift and resource consumption during execution.
• Interactive CLI wizard – Guides you through picking a task, configuring a stream, evaluator, and learner. Each prompt includes contextual help, default values, and validation (including .arff path checks) to keep configuration friction low.
• Streaming data sources – Supports .arff file streams and synthetic generators for SEA, Agrawal, and Asset Negotiation concepts. Generators expose knobs for seeds, class balancing, noise, drift, and concept duration.
• Incremental learners – Ships with a classic Naive Bayes classifier and a configurable Hoeffding Tree (VFDT) that lets you choose the numeric estimator, split criterion, and leaf prediction strategy.
• Online metrics – Basic classification evaluator emits accuracy, Cohen's kappa, optional precision/recall/F1 aggregates, and per-class statistics. Snapshots feed the live console renderer to display throughput, accuracy, kappa variants, elapsed time, and RAM-hours.

Getting Started

Prerequisites

• Rust 1.76 or newer (Edition 2024) with Cargo.

Download a release

Grab the latest build from the project's GitHub Releases page. On macOS you can install the signed package and then launch the CLI with rivu. For Linux and Windows download the corresponding archive and run the rivu executable directly; these builds are not signed, so follow your platform's guidelines for running unsigned binaries.

Launch the wizard

cargo run

Select the prequential evaluation task and answer the wizard prompts for stream, evaluator, and learner. The runner prints a header describing the session and refreshes a live status line with metrics, throughput, and progress bars.

Run a task non-interactively

If you already know the configuration you want, the rivu run subcommand lets you provide everything up front without stepping through the wizard. The flags mirror the wizard prompts and accept additional --*-param key=value overrides for nested settings.

For example, the following command mirrors MOA's EvaluatePrequential invocation shown above, training a Hoeffding Tree on the covtypeNorm.arff stream while reporting basic classification metrics:

rivu run \
  --task evaluate-prequential \
  --learner hoeffding-tree \
  --stream arff-file \
  --evaluator basic-classification \
  --stream-param path=/Users/rafaelvenetikides/Developer/rivu/data/covtypeNorm.arff

The split_criterion in the Hoeffding Tree defaults to Gini (matching MOA's -s GiniSplitCriterion flag), so no extra learner override is required. If a value contains spaces, wrap it in quotes so the shell passes it through as a single argument. Any unspecified flags fall back to the same defaults used by the interactive wizard.

Run the test suite

cargo test

The tests cover the prequential evaluator's guards, curve updates, UI schema helpers, and utility modules that support the CLI and evaluation pipeline.

Sample Data

Example .arff files are available under data/ (airlines, covtypeNorm, and giveMeLoanKaggle). Use the "Arff File Stream" option in the wizard and supply one of these paths along with the zero-based class index to get started quickly.

Project Structure

src/
├── classifiers/        # Naive Bayes and Hoeffding Tree implementations
├── core/               # Shared instance headers and type utilities
├── evaluation/         # Online metrics, snapshots, and evaluators
├── streams/            # ARFF reader and synthetic stream generators
├── tasks/              # Prequential evaluator orchestration
├── ui/                 # CLI wizard, prompt drivers, and schema builders
├── utils/              # Math, parsing, and system helpers
├── testing/            # Test doubles for learners, streams, evaluators
└── main.rs             # CLI entry point with live console renderer

Development Workflow

1. Format the codebase with cargo fmt.
2. Run lint checks via cargo clippy.
3. Execute the automated tests with cargo test before opening a pull request.

To add a new learner, stream, or evaluator:

• Implement the component under src/classifiers, src/streams, or src/evaluation.
• Expose it in the CLI by extending the corresponding enums in src/ui/types/choices/.
• Update the builders in src/ui/types/build/ so the wizard can construct the new option from user selections.

License

Rivu is distributed under the AGPL-3.0 license.