Multi-Attribution Learning (MAL)

• Multi-Attribution Learning is a paradigm that integrates multiple, heterogeneous attribution signals to capture interdependencies and improve predictive modeling.
• It enhances tasks such as visual attribute prediction and conversion rate estimation by jointly optimizing primary and auxiliary objectives.
• MAL frameworks employ techniques like image-conditioned masked language modeling and multi-task optimization to mitigate negative transfer and boost calibration.

Searching arXiv for recent and foundational papers on Multi-Attribution Learning to ground the article. arXiv search query: "multi-attribution learning arXiv Label2Label MAC conversion rate prediction" Multi-Attribution Learning (MAL) denotes a family of joint-learning formulations in which prediction depends on multiple attribution signals rather than a single isolated target. In visual recognition, MAL assigns each image a set of semantic attributes and seeks to model their correlations; in conversion rate (CVR) prediction, MAL learns from labels generated by multiple attribution mechanisms such as First-Click, Last-Click, Linear, and Data-Driven Attribution while serving one primary target; in personalized attribute learning, MAL jointly models global consensus, attribute-specific structure, and user-specific opinion diversity (Li et al., 2022, Wu et al., 2 Mar 2026, Yang et al., 2017). This suggests that MAL is not a single fixed task, but a broader paradigm for exploiting structured, complementary, and sometimes heterogeneous attribution signals.

1. Problem scope and formal definitions

In one standard visual formulation, MAL is the problem of assigning to each image $x \in X$ a set of attributes $y \subset A$ drawn from an attribute universe $A=\{a_1,\dots,a_M\}$. For binary attributes, the data are written as

$$X=\{x_i\}_{i=1}^N,\qquad Y=\{y_i\}_{i=1}^N,\qquad y_i\in\{0,1\}^M,$$

where $y_{ij}=1$ indicates that image $x_i$ possesses attribute $a_j$. A conventional formulation learns $M$ independent sigmoid classifiers

$l_j(x)=\sigma(w_j^\top f(x)+b_j)$

and minimizes

$$L_{BCE} = -\sum_{i=1}^N \sum_{j=1}^M \left[y_{ij}\log l_j(x_i)+(1-y_{ij})\log(1-l_j(x_i))\right].$$

This is the baseline from which later MAL formulations depart (Li et al., 2022).

In CVR prediction, the observed unit is a click-through sequence, or conversion path,

$y \subset A$0

where $y \subset A$1 encodes user, item, context and behavior-sequence features at the $y \subset A$2-th click. The supervision is defined by a collection of attribution mechanisms $y \subset A$3. Under mechanism $y \subset A$4, each click has a non-negative attribution weight that is normalized to a binary label, one mechanism is selected as the primary target, and the remaining mechanisms serve as auxiliary attribution signals. The total objective is a weighted sum of the primary and auxiliary losses:

$y \subset A$5

This formulation makes the multiplicity of attribution mechanisms explicit and treats MAL as a multi-objective prediction problem (Wu et al., 2 Mar 2026).

A third formulation appears in personalized visual attribute learning. There, for attribute $y \subset A$6 and user $y \subset A$7, labels are observed as $y \subset A$8, and each personalized linear predictor

$y \subset A$9

is decomposed as

$A=\{a_1,\dots,a_M\}$0

where $A=\{a_1,\dots,a_M\}$1 is a global common cognition factor, $A=\{a_1,\dots,a_M\}$2 is an attribute-specific bias factor, and $A=\{a_1,\dots,a_M\}$3 is a user-specific bias factor (Yang et al., 2017).

Setting	Observed unit	Supervision structure
Visual MAL	Image $A=\{a_1,\dots,a_M\}$4	Set of binary or multi-valued attributes
CVR MAL	Click-through sequence	Labels under multiple attribution mechanisms
Personalized MAL	User-attribute sample block	Joint global, attribute-specific, and user-specific labels

2. Why independent or single-view formulations are insufficient

A central motivation for MAL is that naive decompositions discard structure that is present in the labels themselves. In the visual case, independent binary classifiers assume

$A=\{a_1,\dots,a_M\}$5

thereby ignoring sample-specific co-dependencies. Label2Label makes this limitation explicit and argues that instance-wise attribute relations matter because the model should be able to infer missing attributes from the remaining attribute hints together with image context. The paper gives examples such as “brown hair + no beard ⇒ female” and “blond hair + heavy makeup ⇒ attractive,” emphasizing that such couplings vary per image rather than being fixed global rules (Li et al., 2022).

In personalized opinion modeling, two previously separate traditions are identified: methods that learn user-specific labels separately for different attributes, and methods that learn crowd-sourced global labels jointly for multiple attributes. The cited work argues that both categories ignore the joint effect of personal diversity with respect to the global consensus and the intrinsic correlation among multiple attributes. MAL is introduced precisely to unify these two factors in a common-to-special manner (Yang et al., 2017).

In multi-touch attribution for marketing, rule-based strategies and earlier data-driven algorithmic strategies are described as failing to address channel interaction, time dependency, and user characteristics. DNAMTA responds by learning from whole customer journeys and using attention to derive touchpoint-level attribution weights from a conversion predictor. This earlier sequence-modeling line is not identical to later CVR MAL, but it addresses the same structural deficiency: a single attribution rule or a flat count-based model cannot represent the dynamic interaction effects between media channels (Li et al., 2018).

These formulations converge on a common methodological point: MAL is introduced when the supervision has relational structure that is lost under per-label independence, per-user isolation, or single-mechanism optimization. This suggests that MAL is best viewed as a response to negative transfer from oversimplified factorization and to information loss from collapsing multiple attribution views into one.

3. Label-as-word modeling and image-conditioned reconstruction

Label2Label reframes visual MAL as an image-conditioned masked language modeling problem. Each attribute label is treated as a “word,” and a sample’s full attribute vector is verbalized as an unordered but meaningful “sentence.” Concretely, each attribute $A=\{a_1,\dots,a_M\}$6 and its two possible values $A=\{a_1,\dots,a_M\}$7 become tokens in a vocabulary $A=\{a_1,\dots,a_M\}$8 of size $A=\{a_1,\dots,a_M\}$9, where the extra token is [MASK]. For image $$X=\{x_i\}_{i=1}^N,\qquad Y=\{y_i\}_{i=1}^N,\qquad y_i\in\{0,1\}^M,$$0 with label vector $$X=\{x_i\}_{i=1}^N,\qquad Y=\{y_i\}_{i=1}^N,\qquad y_i\in\{0,1\}^M,$$1, the sentence is

$$X=\{x_i\}_{i=1}^N,\qquad Y=\{y_i\}_{i=1}^N,\qquad y_i\in\{0,1\}^M,$$2

optionally bracketed by [CLS] and [SEP]. Positional embeddings are omitted on the token axis because attribute labels are unordered, and the reported empirical result is that this permutation-free design works best (Li et al., 2022).

The image-conditioned masked LLM (IC-MLM) couples token reconstruction with image features. An image encoder $$X=\{x_i\}_{i=1}^N,\qquad Y=\{y_i\}_{i=1}^N,\qquad y_i\in\{0,1\}^M,$$3 such as ResNet-50 or ViT produces spatial feature maps

$$X=\{x_i\}_{i=1}^N,\qquad Y=\{y_i\}_{i=1}^N,\qquad y_i\in\{0,1\}^M,$$4

to which 2D position embeddings are added. A Transformer-decoder $$X=\{x_i\}_{i=1}^N,\qquad Y=\{y_i\}_{i=1}^N,\qquad y_i\in\{0,1\}^M,$$5 with $$X=\{x_i\}_{i=1}^N,\qquad Y=\{y_i\}_{i=1}^N,\qquad y_i\in\{0,1\}^M,$$6 layers of self-attention, cross-attention to $$X=\{x_i\}_{i=1}^N,\qquad Y=\{y_i\}_{i=1}^N,\qquad y_i\in\{0,1\}^M,$$7, and feed-forward networks then processes the masked token embeddings. During training, a fraction $$X=\{x_i\}_{i=1}^N,\qquad Y=\{y_i\}_{i=1}^N,\qquad y_i\in\{0,1\}^M,$$8 of the $$X=\{x_i\}_{i=1}^N,\qquad Y=\{y_i\}_{i=1}^N,\qquad y_i\in\{0,1\}^M,$$9 tokens in $y_{ij}=1$0 are replaced by [MASK]; the framework also uses the Attribute Query Network’s own occasional mispredictions as “free” masks. The decoder is trained to reconstruct all $y_{ij}=1$1 tokens, not only the masked ones. The masked-token formulation is

$y_{ij}=1$2

At inference time, the model first runs an Attribute Query Network (AQN) to predict a pseudo-label vector $y_{ij}=1$3 from $y_{ij}=1$4 binary logits thresholded at $y_{ij}=1$5. The pseudo-label vector is then fed into the IC-MLM decoder without masking, and the final per-attribute probabilities are thresholded at $y_{ij}=1$6. No beam search is needed since reconstruction is per-token.

The empirical results reported for three benchmark tasks are summarized below.

Task	Prior method	Label2Label
Facial attributes (error ↓)	PS-MCNN*: 12.64%	12.49%
Pedestrian attributes (mA, F1 ↑)	SSC: 81.87, 86.87	82.24, 87.08
Clothing attributes (accuracy ↑)	MG-CNN: 92.82%	92.87%

The ablations are equally informative. On LFWA, AQN layers $y_{ij}=1$7 and IC-MLM layers $y_{ij}=1$8 are reported as optimal; a mask ratio $y_{ij}=1$9 gives the best error, $x_i$0 versus $x_i$1 at $x_i$2 or $x_i$3; attribute-specific learnable [MASK] embeddings outperform a single shared token or a zero vector; and removing masking, yielding pure label reconstruction, raises the error to approximately $x_i$4. Within this formulation, masking is not merely a regularizer but the mechanism that forces the network to internalize instance-wise attribute relations.

4. Common-to-special decomposition and disentangled generative variants

A distinct MAL line addresses personalized opinions over visual attributes. In the common-to-special model, each personalized predictor is decomposed into a shared component, an attribute-specific component, and a user-specific component:

$x_i$5

The complete objective combines least squares with sparsity-inducing penalties,

$x_i$6

The cited analysis states that the problem is jointly convex, that its smooth part has Lipschitz-continuous gradient, and that an accelerated proximal-gradient method with Nesterov acceleration yields an $x_i$7 global convergence rate. Closed-form soft-thresholding updates are obtained for $x_i$8, $x_i$9, and $a_j$0, and high-probability bounds are proved under a Restricted Eigenvalue-type assumption (Yang et al., 2017).

The empirical study evaluates this formulation on simulated regression and on the Shoes dataset with both binary and relative attributes. For binary attributes, MAL achieves $a_j$1 versus the best baseline $a_j$2 with $a_j$3 training, and $a_j$4 versus $a_j$5 with $a_j$6 training. For relative attributes, MAL achieves user-averaged ranking accuracy $a_j$7 versus $a_j$8 at $a_j$9 training, and $M$0 versus $M$1 at $M$2 training. The paper also reports recovery of the true sparsity patterns of $M$3, $M$4, and $M$5, and that Nesterov acceleration halves the number of iterations to convergence.

A related generative direction studies disentangled and generalizable representations for visual attributes. The architecture uses $M$6 separate convolutional encoders, one encoder per specified attribute and one residual encoder, together with a decoder and a PatchGAN discriminator. A pre-trained classifier $M$7 is associated with each attribute encoder, and the learning objective combines reconstruction loss, adversarial loss in WGAN-GP form, latent-classification loss, and a disentanglement regularizer that forces the posterior over every irrelevant attribute to be uniformly distributed. The resulting latent spaces are trained by randomly shuffling attribute codes across a mini-batch and reconstructing synthesized images, so that the model learns continuous, transferable, and generalizable codes (Oldfield et al., 2019).

This generative formulation is not a predictor of binary attribute vectors in the same sense as Label2Label, but it advances a closely related multi-attribute objective: semantic decomposition of the latent space into disentangled “slots,” information-invariance through the uniform-posterior regularizer, and label-free, intensity-preserving multi-attribute transfer at test time. Quantitatively, the reported Fréchet Inception Distance values are $M$8 on BU-3DFE and $M$9 on RaFD, compared with StarGAN’s $l_j(x)=\sigma(w_j^\top f(x)+b_j)$0 and $l_j(x)=\sigma(w_j^\top f(x)+b_j)$1 respectively; on MultiPIE, the reported value is $l_j(x)=\sigma(w_j^\top f(x)+b_j)$2, versus StarGAN’s $l_j(x)=\sigma(w_j^\top f(x)+b_j)$3 (Oldfield et al., 2019). A plausible implication is that MAL in visual domains spans both discriminative label prediction and controllable attribute-structured generation.

5. MAL for conversion rate prediction under multiple attribution mechanisms

In industrial CVR prediction, MAL is formulated directly in terms of labels produced by multiple attribution mechanisms. One concrete setup defines a dataset

$l_j(x)=\sigma(w_j^\top f(x)+b_j)$4

where for each user–ad interaction $l_j(x)=\sigma(w_j^\top f(x)+b_j)$5, $l_j(x)=\sigma(w_j^\top f(x)+b_j)$6 is the fractional conversion label under mechanism $l_j(x)=\sigma(w_j^\top f(x)+b_j)$7, $l_j(x)=\sigma(w_j^\top f(x)+b_j)$8 is the sample weight, and one mechanism $l_j(x)=\sigma(w_j^\top f(x)+b_j)$9 is designated as the Primary Attribution Mechanism. The goal is simultaneously to exploit all $$L_{BCE} = -\sum_{i=1}^N \sum_{j=1}^M \left[y_{ij}\log l_j(x_i)+(1-y_{ij})\log(1-l_j(x_i))\right].$$0 attribution signals and to produce a calibrated CVR estimate $$L_{BCE} = -\sum_{i=1}^N \sum_{j=1}^M \left[y_{ij}\log l_j(x_i)+(1-y_{ij})\log(1-l_j(x_i))\right].$$1 under the primary mechanism for deployment (Chen et al., 21 Aug 2025).

The proposed architecture consists of two core components: the Attribution Knowledge Aggregator (AKA) and the Primary Target Predictor (PTP). AKA is a multi-task learner with a shared feature-interaction module and one MLP tower per mechanism. From the penultimate layer of each tower it extracts a conversion knowledge vector $$L_{BCE} = -\sum_{i=1}^N \sum_{j=1}^M \left[y_{ij}\log l_j(x_i)+(1-y_{ij})\log(1-l_j(x_i))\right].$$2, and the vectors are concatenated into

$$L_{BCE} = -\sum_{i=1}^N \sum_{j=1}^M \left[y_{ij}\log l_j(x_i)+(1-y_{ij})\log(1-l_j(x_i))\right].$$3

PTP then reuses the exact production CVR model stack, computes its own raw representation $$L_{BCE} = -\sum_{i=1}^N \sum_{j=1}^M \left[y_{ij}\log l_j(x_i)+(1-y_{ij})\log(1-l_j(x_i))\right].$$4, aligns $$L_{BCE} = -\sum_{i=1}^N \sum_{j=1}^M \left[y_{ij}\log l_j(x_i)+(1-y_{ij})\log(1-l_j(x_i))\right].$$5 into the production semantic space, fuses the aligned vector additively with $$L_{BCE} = -\sum_{i=1}^N \sum_{j=1}^M \left[y_{ij}\log l_j(x_i)+(1-y_{ij})\log(1-l_j(x_i))\right].$$6, and predicts the primary label through a binary-classification head. The joint objective is

$$L_{BCE} = -\sum_{i=1}^N \sum_{j=1}^M \left[y_{ij}\log l_j(x_i)+(1-y_{ij})\log(1-l_j(x_i))\right].$$7

A distinctive auxiliary component is Cartesian-based Auxiliary Training (CAT). The $$L_{BCE} = -\sum_{i=1}^N \sum_{j=1}^M \left[y_{ij}\log l_j(x_i)+(1-y_{ij})\log(1-l_j(x_i))\right].$$8 binary labels are combined into a single integer label

$$L_{BCE} = -\sum_{i=1}^N \sum_{j=1}^M \left[y_{ij}\log l_j(x_i)+(1-y_{ij})\log(1-l_j(x_i))\right].$$9

and an additional tower predicts this $y \subset A$00-class target. The CAT penultimate embedding is appended into $y \subset A$01, yielding richer cross-attribution patterns.

The deployment rationale is explicit. PTP reuses the exact production model architecture, is trained with the primary-target cross-entropy objective, and therefore produces outputs that are directly usable in existing deployment pipelines and retain their calibration properties. The system further applies the existing online calibration table, using binning plus isotonic regression or Platt scaling as in production. The reported implementation adds only one additional shared-bottom flow, with inference latency and memory overhead remaining within production budgets.

On Taobao display-ad logs, with Last-Click as primary, the reported offline results are: Base, GAUC $y \subset A$02 and AUC $y \subset A$03; MAL, GAUC $y \subset A$04 and AUC $y \subset A$05; and the best multi-task baseline PLE, GAUC $y \subset A$06 and AUC $y \subset A$07. The stated gain is $y \subset A$08 GAUC improvement over the single-attribution baseline. In the ablation study, MAL without CAT yields GAUC $y \subset A$09, while MAL without multi-attributions, meaning all towers are trained on the primary labels, gives approximately GAUC $y \subset A$10 and AUC $y \subset A$11. The online A/B test over $y \subset A$12 days reports $y \subset A$13 GMV, $y \subset A$14 BuyCnt, and $y \subset A$15 ROI versus the baseline (Chen et al., 21 Aug 2025).

6. Benchmarking, architectural principles, and the MoAE line

The 2026 MAC benchmark systematizes CVR MAL as a public benchmark featuring labels from multiple attribution mechanisms. MAC is constructed from $y \subset A$16 consecutive days of anonymized ad-click logs from Taobao’s production system, with stratified sampling that yields $y \subset A$17M users, $y \subset A$18M total clicks, and $y \subset A$19M distinct items. Each click contains $y \subset A$20 user features, $y \subset A$21 item/shopping context features, and a behavior sequence of up to $y \subset A$22 past items with IDs, shop, category, and visual-similarity embedding. For each click, MAC provides continuous or binary labels under four mechanisms: Last-Click, First-Click, Linear, and Data-Driven Attribution from the CausalMTA causal model. The final day is used as test, and the main evaluation metrics are AUC and

$y \subset A$23

The reported positive-label ratios are Last $y \subset A$24, First $y \subset A$25, DDA $y \subset A$26, and Linear $y \subset A$27 (Wu et al., 2 Mar 2026).

To standardize comparison, the accompanying PyMAL library implements BASE, Shared-Bottom, MMoE, PLE, HoME, and NATAL. Common architectural patterns include embedding layers for categorical features, target attention for sequences, SimTier for multimodal fusion, MLP towers per mechanism, and joint training with $y \subset A$28. This benchmarking infrastructure is used to extract three empirical regularities. First, MAL consistently beats BASE across all four primary targets. Second, the performance lift grows for users with longer or more complex multi-touch conversion paths. Third, auxiliary objectives and architecture design both matter: under most primary settings, adding more auxiliary attribution losses steadily improves performance, but when First-Click is primary, indiscriminately adding all auxiliaries can hurt, and the best auxiliary is Last-Click alone.

The paper distills two design principles from its ablations: fully learn multi-attribution knowledge via flexible sharing, and fully exploit that knowledge in a main-task-prioritized manner. MoAE, or Mixture of Asymmetric Experts, is proposed to satisfy both. Its bottom is a PLE-style Mixture-of-Experts block with one shared expert $y \subset A$29 and one expert $y \subset A$30 per attribution mechanism; its top is a main-task-centric asymmetric feature-transfer module that routes auxiliary representations into the primary tower with learned gating; and it uses one MLP head per mechanism. Formally, with shared input embedding $y \subset A$31 and experts $y \subset A$32, each mechanism computes

$y \subset A$33

while the primary task enriches its feature by

$y \subset A$34

The authors argue that the MoE bottom captures shared patterns and idiosyncrasies without destructive gradient interference, while the asymmetric top focuses auxiliary knowledge solely on boosting the primary signal (Wu et al., 2 Mar 2026).

The benchmark gains reported for the best MAL method are as follows.

Primary target	Best MAL lift vs. BASE (GAUC)	Best method
Last-Click	+2.12 pt	MoAE
DDA	+1.74 pt	MoAE
Linear	+0.80 pt	MoAE
First-Click	+0.34 pt	MoAE

The corresponding main-result GAUC values for MoAE are $y \subset A$35 for Last-Click, $y \subset A$36 for First-Click, $y \subset A$37 for DDA, and $y \subset A$38 for Linear, each exceeding NATAL. The ablations further report that parameter scaling alone, with auxiliary-weight set to $y \subset A$39, changes GAUC by less than $y \subset A$40 pt, implying that the gains arise from multi-attribution signals rather than merely from additional parameters. Auxiliary-task learning methods such as GCS or PCGrad provide modest gains on weak MTL baselines, but no benefit, and sometimes slight harm, on strong models such as NATAL or MoAE.

7. Relation to multi-touch attribution and unresolved issues

The multi-touch attribution literature provides an important precursor to later CVR MAL. DNAMTA models a customer journey as a sequence of touchpoints

$y \subset A$41

embeds touchpoints, encodes the sequence with an LSTM, computes attention weights

$y \subset A$42

or, with time decay,

$y \subset A$43

and fuses the resulting path representation with static control variables before final conversion prediction. It then derives attribution scores either by incremental removal of a channel or directly from normalized attention weights. On a dataset of $y \subset A$44 customer paths, the reported test metrics are Accuracy $y \subset A$45 and AUC $y \subset A$46 for DNAMTA with fusion, exceeding Last-Touch Attribution, HMM, logistic regression, plain LSTM, and attention-only DNAMTA variants (Li et al., 2018).

This earlier sequence-based work differs from later MAL frameworks in that it predicts a single conversion label and then infers per-touchpoint contributions, whereas MAC and the industrial MAL framework directly learn from labels generated by multiple attribution mechanisms. Even so, the connection is substantive: both lines treat attribution as a structured learning problem involving channel interaction, time dependency, or multiple supervisory views.

Several unresolved issues recur across the literature. In Label2Label, the vocabulary scales linearly in $y \subset A$47, and multi-valued attributes require expanding $y \subset A$48 further; the authors suggest future work on joint pre-training on large image-attribute corpora, dynamic masking schedules, richer side information such as object category, and beam-search-style decoding for hierarchical attributes (Li et al., 2022). In CVR MAL, the MAC study shows that auxiliary choice must respect the noisiness and complementarity of signals, since adding all auxiliaries can be counterproductive when First-Click is the primary target (Wu et al., 2 Mar 2026). The same paper suggests richer auxiliary signals such as loyalty and lifetime value, the use of large-scale foundation models for causal reasoning, advanced auxiliary-task learning algorithms for MAL, and benchmark extensions with real-time and cross-device attributions. Taken together, these issues indicate that the main open problem is no longer whether multiple attribution signals can help, but how to select, represent, and transfer them without inducing negative transfer or deployment mismatch.