BR-MARLENE: Multi-Label Transfer in Streams

• BR-MARLENE is an online method for multi-label transfer learning, decomposing streams into binary relevance sub-classifiers for individual label predictions.
• It employs label-wise drift detection and imbalance-aware resampling to update models and transfer knowledge across source and target streams.
• Empirical results show BR-MARLENE achieves state-of-the-art Macro-G-Mean performance with fast runtime compared to dependency-aware alternatives.

Searching arXiv for BR-MARLENE and the cited paper to ground the article in the current literature. BR-MARLENE is an online method for multi-label transfer learning in non-stationary data streams. The acronym expands to Binary Relevance Multi-Label classification in non-stationary environments with Multi-Source Transfer Learning, where “BR” denotes Binary Relevance (Du et al., 9 Sep 2025). It is designed for settings in which each incoming instance may have multiple binary labels, each label-specific concept may drift over time, and useful predictive structure may be shared not only across source and target streams but also across labels themselves. Within the framework of the source paper, BR-MARLENE is the simpler of two proposed methods, with BRPW-MARLENE extending it by explicitly modelling pairwise label dependencies (Du et al., 9 Sep 2025).

1. Problem setting and conceptual scope

The method is formulated for an online multi-label stream

$$D = \{(\mathbf{x}^{t},\mathbf{y}^{t})\}_{t=0}^\infty$$

with feature vector $\mathbf{x}^{t}\in\mathcal{X}$ and binary relevance vector

$$\mathbf{y}^{t} = \{y_q^t\}_{q=1}^{|\mathcal{L}|}\in \mathcal{Y}=\{0,1\}^{|\mathcal{L}|},$$

where $\mathcal{L}=\{l_q\}_{q=1}^{|\mathcal{L}|}$ is the label set and, for each label $l_q$, $y_q^t=1$ denotes relevance and $0$ otherwise (Du et al., 9 Sep 2025).

A label-specific concept is defined as

$P(y_q\mid \mathbf{x}).$

The stream is non-stationary: concept drift occurs if there exist $t'\neq t''$ such that

$$P^{t'}(y_q\mid \mathbf{x}) \neq P^{t''}(y_q\mid \mathbf{x}).$$

A defining premise of the method is that, in multi-label streams, drift can happen independently for each label, synchronously for some labels, or asynchronously across labels. BR-MARLENE therefore adopts a label-wise drift model rather than assuming one global stream concept change (Du et al., 9 Sep 2025).

The transfer-learning setting is multi-source inductive transfer learning. There are $\mathbf{x}^{t}\in\mathcal{X}$0 streams,

$\mathbf{x}^{t}\in\mathcal{X}$1

where $\mathbf{x}^{t}\in\mathcal{X}$2 are source streams and $\mathbf{x}^{t}\in\mathcal{X}$3 is the target stream. A stream domain is

$\mathbf{x}^{t}\in\mathcal{X}$4

and the formulation allows either same-domain or different-domain streams,

$\mathbf{x}^{t}\in\mathcal{X}$5

while label concepts differ:

$\mathbf{x}^{t}\in\mathcal{X}$6

A central assumption is that, for a target-stream label, all other labels from the target stream and from the source streams are treated as possible sources of transferable knowledge (Du et al., 9 Sep 2025).

2. Core architecture and transfer mechanism

BR-MARLENE applies a Binary Relevance transformation: every multi-label stream is decomposed into $\mathbf{x}^{t}\in\mathcal{X}$7 online binary learning problems, and each resulting learner is a sub-classifier. The ensemble of all such sub-classifiers is denoted $\mathbf{x}^{t}\in\mathcal{X}$8 (Du et al., 9 Sep 2025).

Its defining design choice is that the ensemble is shared across labels, streams, and historical concepts. When predicting a target label $\mathbf{x}^{t}\in\mathcal{X}$9, BR-MARLENE does not restrict itself to classifiers previously trained for that same label. Instead, it allows all sub-classifiers from all labels and streams to contribute, but with target-label-specific weights $$\mathbf{y}^{t} = \{y_q^t\}_{q=1}^{|\mathcal{L}|}\in \mathcal{Y}=\{0,1\}^{|\mathcal{L}|},$$0 (Du et al., 9 Sep 2025).

This means that knowledge transfer is realized through cross-label, cross-stream reuse of sub-classifiers in a shared ensemble with label-specific weighting. The same sub-classifier can be highly weighted for one target label and heavily downweighted for another. The paper explicitly characterizes this as transfer by shared ensemble reuse, rather than by parameter sharing or feature mapping (Du et al., 9 Sep 2025).

Drift adaptation is handled at the label level. For each label $$\mathbf{y}^{t} = \{y_q^t\}_{q=1}^{|\mathcal{L}|}\in \mathcal{Y}=\{0,1\}^{|\mathcal{L}|},$$1, drift is monitored separately using a detector such as DDM-OCI. If drift is detected on label $$\mathbf{y}^{t} = \{y_q^t\}_{q=1}^{|\mathcal{L}|}\in \mathcal{Y}=\{0,1\}^{|\mathcal{L}|},$$2 of stream $$\mathbf{y}^{t} = \{y_q^t\}_{q=1}^{|\mathcal{L}|}\in \mathcal{Y}=\{0,1\}^{|\mathcal{L}|},$$3, BR-MARLENE creates a new sub-classifier for that label and appends it to $$\mathbf{y}^{t} = \{y_q^t\}_{q=1}^{|\mathcal{L}|}\in \mathcal{Y}=\{0,1\}^{|\mathcal{L}|},$$4, while retaining prior sub-classifiers. If drift occurs on a target-stream label, then all sub-classifiers reset their performance indicators for that label, forcing re-estimation of transfer usefulness under the new target concept (Du et al., 9 Sep 2025).

A common misconception is that BR-MARLENE models label dependency in the same sense as classifier chains or pairwise conditional models. It does not. The method transfers across labels only indirectly, through weighted reuse of sub-classifiers. It therefore remains a Binary Relevance method in the sense that it does not explicitly model label dependencies (Du et al., 9 Sep 2025).

3. Imbalance handling, weighting, and prediction rule

The method explicitly addresses class imbalance, which may differ by label and change over time. For each sub-classifier, the numbers of positive and negative examples learned so far are recorded as $$\mathbf{y}^{t} = \{y_q^t\}_{q=1}^{|\mathcal{L}|}\in \mathcal{Y}=\{0,1\}^{|\mathcal{L}|},$$5 and $$\mathbf{y}^{t} = \{y_q^t\}_{q=1}^{|\mathcal{L}|}\in \mathcal{Y}=\{0,1\}^{|\mathcal{L}|},$$6. Given a training sub-example $$\mathbf{y}^{t} = \{y_q^t\}_{q=1}^{|\mathcal{L}|}\in \mathcal{Y}=\{0,1\}^{|\mathcal{L}|},$$7, the number of training repetitions is sampled as

$$\mathbf{y}^{t} = \{y_q^t\}_{q=1}^{|\mathcal{L}|}\in \mathcal{Y}=\{0,1\}^{|\mathcal{L}|},$$8

Minority-class examples are therefore oversampled online by increasing their expected replication count (Du et al., 9 Sep 2025).

For each target label, each sub-classifier maintains confusion counts $$\mathbf{y}^{t} = \{y_q^t\}_{q=1}^{|\mathcal{L}|}\in \mathcal{Y}=\{0,1\}^{|\mathcal{L}|},$$9, $\mathcal{L}=\{l_q\}_{q=1}^{|\mathcal{L}|}$0, $\mathcal{L}=\{l_q\}_{q=1}^{|\mathcal{L}|}$1, and $\mathcal{L}=\{l_q\}_{q=1}^{|\mathcal{L}|}$2. To balance positive and negative contributions when majority and minority status may change over time, BR-MARLENE defines dynamic correction factors $\mathcal{L}=\{l_q\}_{q=1}^{|\mathcal{L}|}$3 and $\mathcal{L}=\{l_q\}_{q=1}^{|\mathcal{L}|}$4 such that

$\mathcal{L}=\{l_q\}_{q=1}^{|\mathcal{L}|}$5

yielding

$\mathcal{L}=\{l_q\}_{q=1}^{|\mathcal{L}|}$6

These are used to compute corrected predictive values

$\mathcal{L}=\{l_q\}_{q=1}^{|\mathcal{L}|}$7

and

$\mathcal{L}=\{l_q\}_{q=1}^{|\mathcal{L}|}$8

If a sub-classifier outputs probabilities $\mathcal{L}=\{l_q\}_{q=1}^{|\mathcal{L}|}$9 and $l_q$0, they are calibrated as

$l_q$1

and

$l_q$2

The sub-classifier weight is then

$l_q$3

where $l_q$4 and $l_q$5 are cumulative correct and wrong scores updated online (Du et al., 9 Sep 2025).

The paper notes that some weighting equations in the text suffer formatting corruption, but also states that the intended meaning is that a sub-classifier receives more credit when it is correct on difficult or minority-like examples, especially when other ensemble members are wrong. This suggests that the weighting rule is intended as both a transfer-selection mechanism and an imbalance-sensitive competence estimator (Du et al., 9 Sep 2025).

For target label $l_q$6, BR-MARLENE aggregates all sub-classifiers with that label’s specific weights:

$l_q$7

and predicts

$l_q$8

There is therefore no separate decision threshold such as $l_q$9; prediction is by comparing the two class scores (Du et al., 9 Sep 2025).

4. Online workflow, memory model, and complexity

The algorithmic pipeline is defined as follows. On receiving a new example $y_q^t=1$0 from stream $y_q^t=1$1, the method initializes new stream-specific sub-classifiers if the stream has not been seen before; performs drift detection for each label; appends a new sub-classifier when drift is detected; trains each label-specific model using the Poisson resampling scheme; and, if the example came from the target stream, updates all sub-classifiers’ weights using the weighting scheme (Du et al., 9 Sep 2025).

This produces an asymmetric training regime. On source streams, BR-MARLENE updates sub-classifiers and drift detectors but does not update target-label weights. On the target stream, it updates target sub-classifiers, drift detectors, and all transfer weights. During prediction on the target stream, all sub-classifiers vote for every target label using a distinct weight vector for each label (Du et al., 9 Sep 2025).

The method retains previous sub-classifiers after drift rather than replacing them. It therefore has memory of old concepts, but the ensemble size grows over time. The source paper emphasizes that BR-MARLENE is designed without method-specific hyperparameters; only the base classifier and drift detector must be chosen. In the reported experiments, the base learner is Hoeffding Trees and the drift detector is DDM-OCI (Du et al., 9 Sep 2025).

The target-stream learning complexity is given as

$y_q^t=1$2

while for a source stream $y_q^t=1$3 it is

$y_q^t=1$4

Prediction complexity is

$y_q^t=1$5

The paper therefore characterizes BR-MARLENE as scaling linearly in the number of target labels and in the current ensemble size, and as substantially more scalable than the pairwise extension BRPW-MARLENE (Du et al., 9 Sep 2025).

5. Empirical evaluation and reported performance

BR-MARLENE was evaluated on seven real-world multi-label datasets: Slashdot, Ohsumed, Reuters, Yeast, 20NG, TMC2007, and IMDB. These datasets vary in number of labels from 14 to 103 and generally have low label imbalance rates. Synthetic datasets were also used to study source versus no-source transfer, similar versus non-similar sources, and different drift patterns. Each synthetic dataset has 2 numeric features, 5 binary labels, drift combinations SS, IS, II, IA, AA, AS, target sizes $y_q^t=1$6, and source streams with either similar or non-similar distributions (Du et al., 9 Sep 2025).

The comparison set consists of 11 state-of-the-art multi-label stream methods from MOA/MEKA: EBR, ECC, EPS, ERT, EaBR, EaCC, EaPS, GOBR, GOCC, GOPS, GORT. Evaluation uses sliding-window prequential testing with window size $y_q^t=1$7 of $y_q^t=1$8, and the paper proposes three imbalance-aware metrics: Macro-G-Mean, Micro-G-Mean, and Label-set-based G-Mean (LS-G-Mean) (Du et al., 9 Sep 2025).

The strongest real-world evidence reported for BR-MARLENE is on Macro-G-Mean, where it achieves the best average Friedman rank and significantly higher Macro-G-Mean than all other approaches across datasets (Du et al., 9 Sep 2025).

Dataset	BR-MARLENE Macro-G-Mean
Slashdot	0.357
Ohsumed	0.571
Reuters	0.415
Yeast	0.539
20NG	0.699
TMC2007	0.752
IMDB	0.411

Its average Friedman rank for Macro-G-Mean is 1.025. For Micro-G-Mean, BR-MARLENE is best on Ohsumed, Reuters, 20NG, and TMC2007, but not on Slashdot, Yeast, and IMDB; the paper attributes this to the weighting scheme giving less emphasis to majority classes. Even so, its average Friedman rank remains best at 2.699. For LS-G-Mean, BR-MARLENE again attains the best average Friedman rank, 2.692 (Du et al., 9 Sep 2025).

The runtime evidence positions BR-MARLENE as one of the fastest methods. It is reported as the fastest on Slashdot, Ohsumed, Yeast, and 20NG, second fastest on TMC2007, and third fastest on Reuters and IMDB (Du et al., 9 Sep 2025). This supports the paper’s characterization of BR-MARLENE as the more practical method when the label count is moderate to large.

On synthetic data, the similar-sources BR-MARLENE configuration is reported as the strongest BR-only configuration. The paper also defines the source sub-classifier weight ratio

$y_q^t=1$9

and the average source weight ratio

$0$0

It reports that $0$1 remains high over time on both a synthetic AA dataset and on Yeast, which the authors interpret as evidence that both source-to-target and inter-label transfer are genuinely being used (Du et al., 9 Sep 2025).

6. Relation to BRPW-MARLENE, strengths, and limitations

BRPW-MARLENE is explicitly presented as an extension of BR-MARLENE. Beyond BR classifiers, it introduces pairwise classifiers for ordered label pairs and combines direct BR estimates with dependency terms of the form

$0$2

BR-MARLENE, by contrast, remains the efficient base method and does not explicitly capture pairwise co-occurrence or conditional dependence between labels (Du et al., 9 Sep 2025).

The paper’s Yeast comparison illustrates this tradeoff. On Yeast, BR-MARLENE reports 0.539 Macro-G-Mean, 0.644 Micro-G-Mean, 0.606 LS-G-Mean, and 1562 ms runtime. BRPW-MARLENE improves these to 0.567, 0.663, 0.636, and 82604 ms, respectively, and is stated to be nearly 53 times slower (Du et al., 9 Sep 2025). A plausible implication is that BR-MARLENE occupies the methodologically important middle ground between standard Binary Relevance baselines and far more expensive dependency-aware transfer models.

According to the paper, BR-MARLENE’s principal strengths are that it is the first explicit multi-label transfer method for non-stationary streams, that it handles label-wise concept drift, transfers across both labels and streams, uses strong imbalance-aware weighting, operates online and incrementally, and is designed without method-specific hyperparameters (Du et al., 9 Sep 2025).

Its limitations are equally clear in the source description. It inherits the BR independence assumption, it may suffer from growing ensemble size as old models are retained, its weighting mechanism may show possible sensitivity to noise, and transfer quality depends on the quality of the online weight estimates. The paper also notes that some implementation details are under-specified in the text because of formatting corruption, including exact notation in several weighting equations and precise handling of probabilistic outputs (Du et al., 9 Sep 2025).

The overall characterization that emerges is precise. BR-MARLENE is not novel because it uses Binary Relevance; its novelty lies in embedding Binary Relevance into a transfer-learning, drift-aware, imbalance-aware online ensemble in which useful models—whether trained on the same label, another label, a source stream, or an older concept—can be selectively reused through target-label-specific weights (Du et al., 9 Sep 2025).