One-Class Active Learning

A Julia package for One-Class Active Learning.

This package implements active learning strategies for one-class learning. The package has been developed as part of a benchmark suite for active learning strategies for one-class classification. For more information about this research project, see the OCAL project website, and the companion paper.

Holger Trittenbach, Adrian Englhardt, Klemens Böhm, "An overview and a benchmark of active learning for outlier detection with one-class classifiers", DOI: 10.1016/j.eswa.2020.114372, Expert Systems with Applications, 2021.

Installation

This package requires at least Julia 1.0 (latest tested version is 1.5). This package is not registered yet. Please use the following command to add the package with Pkg3.

using Pkg
Pkg.add("https://github.com/englhardt/OneClassActiveLearning.jl.git")

Overview

One-class classifiers learn to identify if objects belong to a specific class, often used for outlier detection. The package uses the one-class classifiers implemented in SVDD.jl. The package implements a full active-learning cycle. This includes the setup with data loading, splitting and setting up the initial labels. Then it includes several active learning strategies for one-class learning. All the results are stored in .json files.

The scripts in example gives a jump start on how to use this package. To run a minimal example, execute:

> julia --project example/example.jl

The result is then stored in example/example_pool_qs.json.

The following examples are available:

• pool_qs: The default example with a pool based query strategy
• query_synthesis: Use a query synthesis strategy optimized via particle swarm optimization
• batch_qs: Use a TopK batch query strategy

Run any of the above examples by supplying the name, e.g., pool_qs as command line argument:

> julia --project example/example.jl pool_qs

When learning the classifier, the Ipopt solver does not always solve the SVDD optimally. We suggest to use a more sophisticated solver such as Gurobi in experiments.

Data and initial pool init_strategies

The package offers multiple data splitting strategies.

• (Sh) Split holdout: Model fitting and query selection on the training split, and testing on a distinct holdout sample.
• (Sf) Split full: Model fitting, query selection and testing on the full data set.
• (Si) Split inlier: Like Sf, but model fitting on labeled inliers only.
• (Sl) Split labels: Like Sf, but model fitting on labeled observations (inliers and outliers) only.

The package includes multiple strategies to initialize the initial pool before starting the active learning:

• (Pu) Pool unlabeled: All observations are unlabeled.
• (Pp) Pool percentage: Stratified proportion of labels for p percent of the observations.
• (Pn) Pool number: Stratified proportion of labels for a fixed number of observations.
• (Pnin) Pool number inliers: A fixed number of labeled inliers.
• (Pa) Pool attributes: As many labeled inliers as number of attributes.

Active learning strategies

This is a list of the available active learning strategies:

Pool-based Query Strategies

Pool-based query strategies define an informativeness function for a query. They then select the most informative observation from a pool of unlabeled observations.

• Data-based query strategies
  • MinimumMarginPQs and ExpectedMinimumMarginPQs [1]
  • ExpectedMaximumEntropyPQs [1]
  • MinimumLossPQs [2]
• Model-based query strategies
  • HighConfidencePQs [3]
  • DecisionBoundaryPQs
• Hybrid query strategies
  • NeighborhoodBasedPQs [4]
  • BoundaryNeighborCombination [5]
• Baselines
  • RandomPQs
  • RandomOutlierPQs

Batch Query Strategies

The batch query strategies select multiple queries at once in one active learning iteration. Some of them adapt the pool-based query strategies from the previous section.

• Baseline batch strategies
  • RandomBatchQs
  • TopKBatchQs
  • GappedTopKBatchQs
• Filtering batch strategies
  • FilterSimilarBatchQs
  • FilterHierarchicalBatchQs
• Iterative batch strategies
  • IterativeBatchQs
• Partitioning batch strategies
  • ClusterTopKBatchQs
  • EnsembleBatchQs

For more details, we refer to the following publication and the corresponding experiment code here:

Adrian Englhardt, Holger Trittenbach, Dennis Vetter, Klemens Böhm, “Finding the Sweet Spot: Batch Selection for One-Class Active Learning”. In: Proceedings of the 2020 SIAM International Conference on Data Mining (SDM), DOI: 10.1137/1.9781611976236.14, May 7-9, 2020, Cincinnati, Ohio, USA.

Query Synthesis Strategies

Query synthesis is an active learning query scenario where one does not require a pool of unlabeled observations but can query any point in the data space.

• Baselines
  • RandomQss
  • RandomOutlierQss
• Advanced strategies
  • DecisionBoundaryQss
  • NaiveExplorativeMarginQss
  • ExplorativeMarginQss

To synthesize a query with a given informativeness function, one may use particle swarm optimization or any optimizer from BlackBoxOptim.jl.

For more details, we refer to the following publication and the corresponding experiment code here:

Adrian Englhardt, Klemens Böhm, “Exploring the Unknown - Query Synthesis in One-Class Active Learning”. In: Proceedings of the 2020 SIAM International Conference on Data Mining (SDM), DOI: 10.1137/1.9781611976236.17, May 7-9, 2020, Cincinnati, Ohio, USA.

Extending with a new strategy

A new query strategy NewPQs can be implemented with the following steps:

1. First create a new type NewPQs and implement:

qs_score(qs::NewPQs, data::Array{T, 2}, pools::Dict{Symbol, Vector{Int}}) where T <: Real

Here, data is a subset of the full data set which depends on the data splitting strategy. The current labels of the active learning cycle are in pools (:U for a unlabeled, :Lin for labeled a inlier and :Lout for labeled a outlier observation). The argument pools contains a dictionary mapping from the different labels to indices in data. The output of the function must be a score array, where the score at index i belongs to observation i in data. The framework then chooses the observation with the highest score that is still unlabeled.

2. Add a call to the constructor of NewPQs in initialize_qs (src/QueryStrategies/qs_utils.jl).

3. Export NewPQs in src/QueryStrategies/QueryStrategies.jl.

Authors

We welcome contributions and bug reports.

This package is developed and maintained by Holger Trittenbach and Adrian Englhardt.

References

[1] A. Ghasemi, H. R. Rabiee, M. Fadaee, M. T. Manzuri, and M. H. Rohban. Active learning from positive and unlabeled data. In 2011 IEEE 11th International Conference on Data Mining Workshops, pages 244–250, Dec 2011.

[2] A. Ghasemi, M. T. Manzuri, H. R. Rabiee, M. H. Rohban, and S. Haghiri. Active one-class learning by kernel density estimation. In 2011 IEEE International Workshop on Machine Learning for Signal Processing, pages 1–6, Sept 2011.

[3] V. Barnabé-Lortie, C. Bellinger, and N. Japkowicz. Active learning for one-class classification. In 2015 IEEE 14th International Conference on Machine Learning and Applications (ICMLA), pages 390–395, Dec 2015.

[4] N. Görnitz, M. Kloft, K. Rieck, and U. Brefeld. Toward supervised anomaly detection. Journal of Artificial Intelligence Research (JAIR), pages 235–262, Jan. 2013.

[5] L. Yin, H. Wang, and W. Fan. Active learning based support vector data description method for robust novelty detection. Knowledge-Based Systems, pages 40–52, Aug. 2018.