Multi-task Learning to Rank (MTL2R)

• Multi-task Learning to Rank (MTL2R) is a framework that jointly optimizes various relevance criteria by integrating primary signals like clicks and purchases with auxiliary signals such as dwell time and ratings.
• It employs techniques like explicit multi-loss optimization, collapsed supervision, and ranking-aware auxiliary losses to balance trade-offs and mitigate negative transfer among tasks.
• MTL2R has been successfully applied in domains such as e-commerce search, ads ranking, and spatial networks, demonstrating improvements in key metrics like NDCG and conversion rates.

Multi-task Learning to Rank (MTL2R) denotes ranking formulations in which a model must optimize multiple relevance criteria or behavioral objectives jointly rather than fit a single relevance label. In the literature represented here, the tasks include clicks, dwell time, streams, sales, ratings, add-to-carts, purchases, query types, and auxiliary geometric or structural signals, with the ranked object ranging from documents and products to paths and even candidate training subsets (Dong et al., 16 Feb 2026, Morishetti et al., 13 Aug 2025, Ding et al., 2023). The area therefore encompasses both explicit multi-loss rankers and systems in which multiple objectives are encoded through auxiliary tasks, residual transfer, or collapsed supervision into a single ranking function (Hadash et al., 2018, Gaydhani et al., 26 Feb 2026).

1. Scope and formal definitions

A concise formalization appears in DeepMTL2R, where each item $x_i$ is associated with multiple labels

$y_i = \left(y_i^1, \ldots, y_i^K\right),$

and training minimizes a vector-valued objective

$$\mathcal{L}(\bm{\theta}) = \big[\mathcal{L}_1(\bm{\theta}), \mathcal{L}_2(\bm{\theta}), \ldots, \mathcal{L}_K(\bm{\theta})\big].$$

This formulation makes Pareto optimality central, because realistic ranking systems often cannot minimize all task losses to their individual optima simultaneously (Dong et al., 16 Feb 2026).

Within that broad definition, MTL2R is not tied to a single ranking unit. In recommendation, Rank and Rate treats ranking as the primary task over user–item interactions and rating prediction as a downstream auxiliary task (Hadash et al., 2018). In industrial product search, the training unit may be an impression-level triplet $(Q,I,C)$ with click, add-to-cart, and transaction labels, or a query–item–week instance with hierarchical engagement counts and multi-channel provenance (Morishetti et al., 13 Aug 2025, Gaydhani et al., 26 Feb 2026). In spatial networks, PathRank ranks candidate paths for a source–destination pair by regressing a similarity-based ranking score and augmenting it with auxiliary path-property prediction (Yang et al., 2019). In active learning, RAMBO ranks candidate labeled subsets by downstream utility rather than ranking documents for queries, but it still uses shared representations, pairwise ranking loss, and auxiliary supervision in a recognizably MTL2R-style formulation (Ding et al., 2023).

This diversity suggests that MTL2R is better understood as a family of ranking problems unified by joint optimization over multiple ranking-relevant signals, rather than as a single architectural template.

2. Supervision design and objective construction

A first major axis of variation is how multiple tasks enter the objective. In explicit multi-task formulations, the losses remain separate. Rank and Rate optimizes a ranking loss and a rating-prediction loss jointly,

$\begin{split} O = \min_{U,I,I^d,\theta} \; &\alpha\cdot L_R(\mathcal{R},D;U,I) \ &+ (1-\alpha)\cdot L_P(\mathcal{P},D;FC_\theta(U),FC_\theta(I+I^d)) \ &+ \lambda(\|U\|^2+\|I\|^2+\|I^d\|^2+\|\theta\|^2), \end{split}$

while modeling rating as a post-consumption transformation of the ranking representation through $q_i^{post}=q_i+q_i^d$ and tied fully connected mappings $p_u^P=FC_\theta(p_u)$, $q_i^P=FC_\theta(q_i^{post})$ (Hadash et al., 2018). Personalized product search likewise uses a weighted sum of binary cross-entropy losses for click, add-to-cart, and transaction prediction, and extends the setup with an auxiliary relevance-label task trained by categorical cross-entropy (Morishetti et al., 13 Aug 2025).

A second pattern is the collapsed multi-task objective, where several behavioral signals are compressed into one scalar supervision target. Unified multi-channel retrieval fusion in large-scale e-commerce search is exemplary. It does not train separate click, add-to-cart, and purchase heads; instead it constructs a single engagement label

$$L(q,i,w) = a P_{qiw} + b A_{qiw} + c C_{qiw} + d V_{qiw},$$

with $a=1$, $y_i = \left(y_i^1, \ldots, y_i^K\right),$0, $y_i = \left(y_i^1, \ldots, y_i^K\right),$1, $y_i = \left(y_i^1, \ldots, y_i^K\right),$2, and then normalizes it within query to $y_i = \left(y_i^1, \ldots, y_i^K\right),$3. The paper explicitly characterizes this as joint optimization achieved through label engineering and feature integration rather than through separate task heads or a multi-loss architecture (Gaydhani et al., 26 Feb 2026).

A third pattern inserts ranking-aware auxiliary losses into otherwise conventional multi-task predictors. PWiseR adds a task-specific pairwise ranking term to BCE in CTR/CVR learning,

$y_i = \left(y_i^1, \ldots, y_i^K\right),$4

and uses margin-based pairwise constraints so that conversion-associated clicked samples should score above click-only and non-click samples (Durmus et al., 2024). RAMBO takes the ranking view more directly: it learns pairwise preferences between candidate subsets using a RankNet-style cross-entropy

$y_i = \left(y_i^1, \ldots, y_i^K\right),$5

while jointly predicting optimal transport distance as an auxiliary target (Ding et al., 2023).

Across these formulations, the common theme is not a single loss family but the use of multiple supervision channels to shape a ranking function.

3. Architectural families

The most classical MTL2R design is shared-bottom with task-specific heads. In email search, QC-MTLRM shares lower neural layers between pairwise ranking and auxiliary query-cluster prediction, then branches into a sigmoid ranking head and a softmax cluster head, optimizing

$y_i = \left(y_i^1, \ldots, y_i^K\right),$6

Its contribution is not merely multi-task training, but the use of automatically induced query-cluster labels as auxiliary supervision for query-dependent ranking (Shen et al., 2018). Early-stage ads ranking uses the same broad template with a shared DLRM-based two-tower backbone and task-specific heads for CTR, Consolidated Quality Score, and teacher distillation, again emphasizing ranking alignment through auxiliary objectives rather than architectural novelty (Wang et al., 2023).

A more elaborate family combines shared bottoms with expert routing. Personalized product search uses a hybrid shared-bottom + MMoE design in which DCN-V2 or FT-Transformer provides the shared representation, TinyBERT supplies query–product matching features, and MMOE experts plus task-specific gates and towers predict click, add-to-cart, transaction, and optionally auxiliary relevance (Morishetti et al., 13 Aug 2025). DeepMTL2R systematizes this design space as a research framework rather than a single model: it couples a shared Transformer encoder over ranked lists with 21 multi-task learning algorithms and 9 LTR losses, allowing controlled comparison of scalarization-based, gradient-manipulation, and Pareto-front-oriented optimization strategies (Dong et al., 16 Feb 2026).

Recent work pushes beyond encoder-plus-head decomposition. ResFlow replaces generic sharing with directed residual transfer between task towers: $y_i = \left(y_i^1, \ldots, y_i^K\right),$7 so that later, sparser tasks learn residual corrections on top of earlier, denser tasks. The design is explicitly motivated by progressive behavior chains such as click $y_i = \left(y_i^1, \ldots, y_i^K\right),$8 add-to-cart $y_i = \left(y_i^1, \ldots, y_i^K\right),$9 order and is intentionally lightweight enough for pre-rank deployment (Fu et al., 2024). OneRank goes further by making the Transformer itself task-native: it introduces task-private tokens, structured masking, candidate-aware contextualization, controlled cross-task attention, and dynamic matching-based scoring

$$\mathcal{L}(\bm{\theta}) = \big[\mathcal{L}_1(\bm{\theta}), \mathcal{L}_2(\bm{\theta}), \ldots, \mathcal{L}_K(\bm{\theta})\big].$$0

thereby internalizing task specialization, cross-candidate reasoning, and score computation inside the Transformer stack rather than attaching static MLP towers on top of a task-agnostic encoder (Tang et al., 15 Jun 2026).

4. Ranking settings and application domains

E-commerce search and recommendation dominate the current MTL2R landscape, but the underlying ranking settings differ substantially. Unified multi-channel retrieval fusion reformulates heterogeneous candidate-source merging as one ranking problem over the union of channel outputs, with a single scalar score $$\mathcal{L}(\bm{\theta}) = \big[\mathcal{L}_1(\bm{\theta}), \mathcal{L}_2(\bm{\theta}), \ldots, \mathcal{L}_K(\bm{\theta})\big].$$1 learned by LambdaMART from channel-aware features, engagement history, and temporal signals (Gaydhani et al., 26 Feb 2026). Personalized product search treats each impression as a triplet $$\mathcal{L}(\bm{\theta}) = \big[\mathcal{L}_1(\bm{\theta}), \mathcal{L}_2(\bm{\theta}), \ldots, \mathcal{L}_K(\bm{\theta})\big].$$2 and induces ranking from joint click, add-to-cart, and purchase prediction over mixed tabular and text-derived features (Morishetti et al., 13 Aug 2025). ResFlow and OneRank sit later in the industrial ranking pipeline, targeting Shopee pre-rank or recommendation stages where multiple behavioral objectives must be estimated under strict latency budgets and then fused into a single serving score (Fu et al., 2024, Tang et al., 15 Jun 2026).

Other domains show that the framework is not limited to commerce. Email search ranking uses six candidate emails per query and one clicked document, with query-cluster prediction as auxiliary supervision to address query heterogeneity in a private corpus where aggregated cross-user click signals are unavailable (Shen et al., 2018). Early-stage ads ranking uses CTR, a consolidated quality objective, and final-stage teacher signals to improve recall of ads that would be valuable in downstream ranking and auction, treating multi-task learning as a solution to cross-stage ranking inconsistency (Wang et al., 2023). PathRank applies the same general logic to route selection in spatial networks, where the main ranking task predicts path preference scores derived from weighted Jaccard similarity to historical trajectories and the auxiliary tasks reconstruct path distance, travel time, and fuel consumption (Yang et al., 2019).

RAMBO extends the idea even further. There the ranked objects are candidate labeled subsets in active learning, the relevance notion is downstream validation utility, and the model uses set encoders plus pairwise ranking and auxiliary OT prediction to rank batches for acquisition (Ding et al., 2023). A plausible implication is that MTL2R principles are transferable to any setting where selection quality is better expressed as a relative ranking over structured alternatives than as direct regression of noisy utility.

5. Evaluation, deployment, and production constraints

MTL2R evaluation is correspondingly heterogeneous. Some systems report classical ranking metrics such as NDCG@8, MRR, Kendall’s $$\mathcal{L}(\bm{\theta}) = \big[\mathcal{L}_1(\bm{\theta}), \mathcal{L}_2(\bm{\theta}), \ldots, \mathcal{L}_K(\bm{\theta})\big].$$3, or Spearman’s $$\mathcal{L}(\bm{\theta}) = \big[\mathcal{L}_1(\bm{\theta}), \mathcal{L}_2(\bm{\theta}), \ldots, \mathcal{L}_K(\bm{\theta})\big].$$4; others emphasize AUC/GAUC for pointwise task prediction; industrial papers often add online business metrics such as conversion, order-per-user, GMV/UU, Total Value, or Bad Query Rate (Gaydhani et al., 26 Feb 2026, Yang et al., 2019, Tang et al., 15 Jun 2026). Ads ranking adds replay-based recall and Total Value Divergence to measure cross-stage consistency rather than relying only on AUC or normalized entropy (Wang et al., 2023). ResFlow proposes Weighted Recall@K as an offline metric aligned with online OPU, arguing that order-volume-weighted retrieval mass is more predictive of business impact than binary Recall@K, NDCG, or List AUC in its pre-rank setting (Fu et al., 2024).

Deployment evidence is one of the strongest features of the area. Unified ranking for multi-channel retrieval improves NDCG@8 from 0.6620 under Weighted Interleaving to 0.7994 with engagement features and conversion-weighted labeling, and reports a $$\mathcal{L}(\bm{\theta}) = \big[\mathcal{L}_1(\bm{\theta}), \mathcal{L}_2(\bm{\theta}), \ldots, \mathcal{L}_K(\bm{\theta})\big].$$5 online conversion lift with p95 latency under 50 ms on Target.com (Gaydhani et al., 26 Feb 2026). Early-stage ads ranking reports $$\mathcal{L}(\bm{\theta}) = \big[\mathcal{L}_1(\bm{\theta}), \mathcal{L}_2(\bm{\theta}), \ldots, \mathcal{L}_K(\bm{\theta})\big].$$6 recall, $$\mathcal{L}(\bm{\theta}) = \big[\mathcal{L}_1(\bm{\theta}), \mathcal{L}_2(\bm{\theta}), \ldots, \mathcal{L}_K(\bm{\theta})\big].$$7 Total Value, and $$\mathcal{L}(\bm{\theta}) = \big[\mathcal{L}_1(\bm{\theta}), \mathcal{L}_2(\bm{\theta}), \ldots, \mathcal{L}_K(\bm{\theta})\big].$$8 total CPU relative to a separate CTR + CQS baseline, showing that multi-task consolidation can improve both ranking consistency and serving efficiency (Wang et al., 2023). ResFlow reports a $$\mathcal{L}(\bm{\theta}) = \big[\mathcal{L}_1(\bm{\theta}), \mathcal{L}_2(\bm{\theta}), \ldots, \mathcal{L}_K(\bm{\theta})\big].$$9 increase in OPU without additional system latency in Shopee Search, and its WR@100 reaches a Pearson correlation of $(Q,I,C)$0 with online OPU uplift in the 3-target setting (Fu et al., 2024). OneRank reports online gains of GMV/UU $(Q,I,C)$1, Paid GMV/UU $(Q,I,C)$2, AR/UU $(Q,I,C)$3, and Bad Query Rate $(Q,I,C)$4, supporting the claim that Transformer-native MTL2R can be deployed at scale (Tang et al., 15 Jun 2026).

Benchmarking frameworks increasingly expose trade-offs rather than a single scalar winner. DeepMTL2R evaluates all $(Q,I,C)$5 bi-objective task pairs on MSLR-WEB30K with 10 different reference vectors per pair, uses NDCG@30 and hypervolume-oriented Pareto analysis, and reports that Pareto-front-finding methods such as Weighted Chebyshev recover clearer trade-off structure while WC and LOG_MGDA perform best among six highlighted methods in aggregate comparison (Dong et al., 16 Feb 2026).

6. Boundaries, limitations, and open directions

A recurring misconception is that any multi-signal ranker is automatically a canonical MTL2R architecture. The literature itself draws a sharper distinction. Unified multi-channel retrieval fusion is explicitly described as having only a weak but meaningful connection to MTL2R because it collapses clicks, add-to-carts, and purchases into a single scalar label and uses one LambdaMART ranker, with no separate task heads, no mixture-of-experts, and no explicit multi-loss decomposition (Gaydhani et al., 26 Feb 2026). PWiseR is similarly adjacent rather than canonical: it is best understood as ranking-based auxiliary supervision inside CTR/CVR learning rather than classical listwise learning to rank over ranked lists (Durmus et al., 2024).

Another limitation is that many systems remain pointwise in optimization even when used for ranking. Personalized product search, early-stage ads ranking, ResFlow, and PathRank all derive rankings from pointwise classification or regression outputs, even though their deployment objective is ranking quality (Morishetti et al., 13 Aug 2025, Wang et al., 2023, Fu et al., 2024, Yang et al., 2019). By contrast, other works make ranking more explicit through pairwise or listwise objectives, such as RankNet-style subset ranking in RAMBO, listwise-targeted LambdaMART in multi-channel retrieval fusion, or hybrid listwise-plus-pointwise optimization in OneRank (Ding et al., 2023, Gaydhani et al., 26 Feb 2026, Tang et al., 15 Jun 2026). This suggests that one major fault line in MTL2R is not only architectural but also whether ranking is optimized directly or only induced from calibrated task prediction.

Task interference and task priority remain unresolved core issues. Several papers report negative transfer, sensitivity to loss weights, or the seesaw phenomenon (Wang et al., 2023, Tang et al., 15 Jun 2026). Prioritized MTL via Lagrangian Differential Multiplier Methods addresses this by reframing lower-priority optimization as constrained optimization under higher-priority protection, a design explicitly motivated by search and recommendation ranking where metrics such as relevance, pay, or order volume must not regress while auxiliary objectives improve (Cheng et al., 2024). At a more structural level, low-rank multi-task regularization that penalizes the exactly $(Q,I,C)$6 smallest singular values is not itself a ranking method, but it is explicitly proposed as transferable to MTL2R by replacing regression loss with pairwise or listwise ranking losses, making shared latent-subspace learning a plausible direction for future ranking models (Chang et al., 2021).

Taken together, these works indicate that MTL2R is no longer a narrow recipe of shared-bottom plus multiple heads. It now includes auxiliary-task rankers, collapsed multi-objective rankers, Pareto-oriented benchmarking frameworks, residual transfer models, prioritized optimization schemes, and Transformer-native architectures. The common technical question is unchanged: how to share information across related ranking objectives without destroying the task-specific structure that makes each objective valuable.