Target-Decoupled SASRec: Methodologies and Applications

• The paper introduces target decoupling in SASRec by separating output projections or training objectives to reduce semantic and temporal interference.
• It employs target-specific representations, using dedicated output heads and soft-target calibration to isolate category semantics in long behavior sequences.
• Empirical results show improved GMV, AUC, and LogLoss metrics, validating the benefits of decoupling mechanisms in large-scale, multi-objective recommendation systems.

Searching arXiv for the specified papers and closely related SASRec work. Target-Decoupled SASRec denotes recent adaptations of SASRec that separate target-dependent signals rather than forcing all objectives, categories, or feedback semantics into a single coupled representation. In the literature represented here, the term is used in two distinct but related senses. In Tencent Advertising, it refers to a matching model that preserves a shared SASRec self-attentive backbone while introducing target-category-specific output heads to eliminate target-wise semantic interference in unified long behavior trajectories (Hu et al., 10 Sep 2025). In a separate line of work on soft-target optimization, it refers to training SASRec with a decoupled objective that separates target confidence from the latent interest distribution over non-target items, thereby modifying the learning signal rather than the encoder topology (Zhang et al., 2024).

1. Conceptual scope and forms of decoupling

The common premise behind Target-Decoupled SASRec is that a single coupled target representation can be inadequate when recommendation data are heterogeneous, sparse, or multi-objective. In Tencent Advertising, the motivating setting is long-sequence modeling over unified commercial behavior trajectories that span advertising scenarios and content domains; the central difficulty is target-wise interference in temporal and semantic patterns when optimizing for different advertising targets (Hu et al., 10 Sep 2025). In the soft-target framework, the motivating setting is multi-class optimization under sparse positive feedback, where standard softmax treats all unobserved items as equally negative and therefore fails to distinguish plausible-but-unobserved positives from truly negative items (Zhang et al., 2024).

A concise comparison of the two uses of the term is as follows.

Variant	Decoupling locus	Primary purpose
Tencent Target-Decoupled SASRec	Output heads on top of a shared SASRec backbone	Eliminate target-wise semantic interference in matching
Soft-target Target-Decoupled SASRec	Training objective	Separate target calibration from non-target ranking

These two formulations are methodologically different. The Tencent design changes how SASRec maps a user sequence to a target-conditioned user vector, while the soft-target design changes how logits are optimized. This suggests that “target decoupling” is not a single mechanism but a broader design principle that can be instantiated either in representation space or in supervision space.

2. Problem setting in unified commercial behavior trajectories

In Tencent Advertising, targets are the business objectives a model optimizes and predicts, including click (CTR), conversion (pCVR with multiple conversion types such as download, order, payment), and downstream GMV. When unified commercial behavior trajectories are constructed, user sequences span ad scenarios such as Moments, Channels, Official Accounts, News, and Video, and also extend into content domains. They include multiple action types, including impressions, clicks, plays, likes, follows, and comments (Hu et al., 10 Sep 2025).

This consolidation substantially lengthens sequences, reaching tens of thousands of behaviors, but it also introduces target-wise interference. The paper distinguishes two forms. Temporal interference arises because different target categories exhibit different decay patterns in time; hobbies such as Hiking decay slowly, whereas time-sensitive topics such as Breaking News decay fast. A single global temporal encoding therefore entangles category-specific temporal dynamics. Semantic interference arises because heterogeneous fields and domains mix semantics that matter differently to distinct target categories, and a single user representation forces incompatible category semantics into the same embedding subspace (Hu et al., 10 Sep 2025).

The interference is exacerbated by the cross-domain, cross-scenario setting. Taxonomy gaps and diverse action types inject many irrelevant interactions into self-attention, and richer content-domain behaviors can dominate sparse ad-domain signals. The resulting motivation is explicit target decoupling: produce category-specific user representations and temporal encodings so that each target category activates only the parameters tailored for its semantics and temporal dynamics (Hu et al., 10 Sep 2025).

3. Architectural formulation in matching

The Tencent formulation starts from standard SASRec as a Transformer-style self-attention model over a behavior sequence. Let $b_i$ denote the $i$-th behavior and $E(\cdot)$ the behavior embedding. With embedding dimension $d$, the input embedding is $x_i = E(b_i) \in \mathbb{R}^d$, with positional encoding $p_i \in \mathbb{R}^d$, and the first-layer input is $h_i^{(0)} = x_i + p_i$. For each layer $\ell$, with $$H^{(\ell-1)} = [h_1^{(\ell-1)}, \ldots, h_n^{(\ell-1)}] \in \mathbb{R}^{n \times d}$$, SASRec uses

$$Q^{(\ell)} = H^{(\ell-1)} W_Q^{(\ell)}, \quad K^{(\ell)} = H^{(\ell-1)} W_K^{(\ell)}, \quad V^{(\ell)} = H^{(\ell-1)} W_V^{(\ell)},$$

and

$i$0

With residuals, layer normalization, and a position-wise feed-forward module $i$1, the layer output is

$i$2

In vanilla SASRec retrieval, a user embedding $i$3 and an ad embedding $i$4 are combined via the dot product $i$5 (Hu et al., 10 Sep 2025).

Target-Decoupled SASRec modifies only the output stage. A shared SASRec encoder computes base user features from the sampled behavior sequence, and then $i$6 target-specific output heads are introduced, one for each target category $i$7. Only the head matching the target’s category $i$8 is activated at inference. For backbone output features $i$9 or a final pooled feature, the $E(\cdot)$0-th head computes a category-specific user vector $E(\cdot)$1, and the matching score is

$E(\cdot)$2

The decoupling locus is therefore narrow and explicit. Parameters are shared in the backbone self-attention and feed-forward layers, while the output-head projection is target-specific. The paper explicitly states that it does not introduce category-specific $E(\cdot)$3 projections, nor MoE or gating in the backbone. This confines target decoupling to the final projection space, which isolates category semantics without materially increasing inference compute, because only one head is executed for a request (Hu et al., 10 Sep 2025).

The model operates within a two-tower matching architecture consisting of a user tower, implemented as the SASRec backbone plus category head, and an ad tower, implemented as the ad embedding. The score remains a dot product. The paper does not specify an explicit loss for this component; it notes that retrieval models commonly use sampled softmax or pairwise logistic losses, but presents such a formulation only as generic and not as a claimed training objective (Hu et al., 10 Sep 2025).

4. Temporal target decoupling and its relation to ranking

Tencent’s target-decoupling program is not restricted to matching. The same paper introduces Target-Decoupled Positional Encoding (TdPE) in the ranking backbone, implemented with TIN, to address target-wise temporal interference (Hu et al., 10 Sep 2025).

Let $E(\cdot)$4 denote the semantic embedding of behavior $E(\cdot)$5. Let $E(\cdot)$6 be a target-aware relative-position encoding and $E(\cdot)$7 encode the time interval $E(\cdot)$8 in days. Then the temporally encoded behavior is

$E(\cdot)$9

The target representation is shifted by category-specific temporal offsets,

$d$0

where $d$1 are learned per category. The additive target-side offsets allow downstream attention and representation modules to perceive category-tailored temporal shifts (Hu et al., 10 Sep 2025).

The paper discusses two variants. TIN+TdPE uses target-side decoupling with shared global behavior-side positional encodings and therefore uses fewer parameters. TIN+dPE learns multiple sets of timing information jointly for target and behavior, achieving slightly better performance but with more parameters (Hu et al., 10 Sep 2025).

TdPE is positioned alongside Decoupled Side Information TIN (DSI-TIN), which mitigates inter-field interference by using multiple TINs with different side-information subsets in attention while fusing all side information in representation, and alongside Stacked TIN, which captures high-order behavioral correlations. In the two-stage framework, the search stage first retrieves or samples behaviors, and the sequence modeling stage then applies Target-Decoupled SASRec in matching and TdPE in ranking so that category-specific signals are present and can be isolated (Hu et al., 10 Sep 2025).

5. Objective-level target decoupling with soft targets

A different formulation of Target-Decoupled SASRec appears in work on learning recommender systems with soft targets. Here the starting point is the standard SASRec objective for next-item prediction. Given an item universe $d$2 with $d$3, a user sequence $d$4, and SASRec hidden state $d$5 at timestep $d$6, with item embedding matrix $d$7, the logit for item $d$8 is $d$9, the softmax probability is

$x_i = E(b_i) \in \mathbb{R}^d$0

and the standard training loss is $x_i = E(b_i) \in \mathbb{R}^d$1, where $x_i = E(b_i) \in \mathbb{R}^d$2 is the true next item (Zhang et al., 2024).

The paper identifies the limitation that this multi-class softmax treats all unobserved items as equally negative. To address this, it introduces a soft target distribution $x_i = E(b_i) \in \mathbb{R}^d$3 and decouples optimization into two parts. For a target item $x_i = E(b_i) \in \mathbb{R}^d$4, the target confidence is $x_i = E(b_i) \in \mathbb{R}^d$5, and the non-target latent interest distribution is

$x_i = E(b_i) \in \mathbb{R}^d$6

With model outputs $x_i = E(b_i) \in \mathbb{R}^d$7, Bernoulli distributions $x_i = E(b_i) \in \mathbb{R}^d$8 and $x_i = E(b_i) \in \mathbb{R}^d$9, and conditional non-target model distribution

$p_i \in \mathbb{R}^d$0

the decoupled loss is

$p_i \in \mathbb{R}^d$1

where $p_i \in \mathbb{R}^d$2 is a free trade-off parameter (Zhang et al., 2024).

This objective separates a target-confidence term from a non-target latent-interest term. The gradient structure is central: the non-target term has zero gradient with respect to the target logit $p_i \in \mathbb{R}^d$3, while the target term contributes the Bernoulli-regression gradient $p_i \in \mathbb{R}^d$4. The paper presents this as the key consequence of decoupling, because target calibration can then be adjusted via $p_i \in \mathbb{R}^d$5 without direct perturbation from non-target ranking signals (Zhang et al., 2024).

Soft labels are generated via neighbor label propagation. Users are clustered with K-means over pretrained user representations obtained by pretraining the backbone recommender with standard softmax. Within a cluster, neighbors are defined by cluster membership, similarity is $p_i \in \mathbb{R}^d$6, and row-normalized propagation weights are

$p_i \in \mathbb{R}^d$7

With one-hot initialization $p_i \in \mathbb{R}^d$8, the update rule is

$p_i \in \mathbb{R}^d$9

The paper notes that $h_i^{(0)} = x_i + p_i$0 is small, gives practical ranges for $h_i^{(0)} = x_i + p_i$1, the number of clusters $h_i^{(0)} = x_i + p_i$2, and top-$h_i^{(0)} = x_i + p_i$3 truncation of sparse label distributions, and also describes a sampled-softmax approximation when full softmax over the item universe is expensive (Zhang et al., 2024).

6. Empirical evidence, limitations, and synthesis

The Tencent architectural variant reports online A/B gains in production. In WeChat Channels, Target-Decoupled SASRec in matching yields a $h_i^{(0)} = x_i + p_i$4 GMV lift relative to a vanilla SASRec-based user tower when using 39 target categories plus a default category. For ranking, TdPE achieves a $h_i^{(0)} = x_i + p_i$5 GMV lift in WeChat Moments relative to TIN without decoupling. The paper also reports t-SNE visualization in which outputs from the vanilla SASRec user tower show weak category clustering, whereas Target-Decoupled SASRec produces clear clusters aligned with ad categories. Overall deployments of multiple components deliver a $h_i^{(0)} = x_i + p_i$6 GMV lift in WeChat Channels and a $h_i^{(0)} = x_i + p_i$7 GMV increase in WeChat Moments, with the text explicitly noting that these are cumulative gains from multiple components rather than from Target-Decoupled SASRec alone (Hu et al., 10 Sep 2025).

For temporal decoupling in ranking, the reported offline results compare TIN, TIN+TdPE, and TIN+dPE. On Alipay, the TIN baseline reports AUC $h_i^{(0)} = x_i + p_i$8 and LogLoss $h_i^{(0)} = x_i + p_i$9; TIN+TdPE reaches AUC $\ell$0 and LogLoss $\ell$1; TIN+dPE reaches AUC $\ell$2 and LogLoss $\ell$3. On Taobao, the TIN baseline reports AUC $\ell$4 and LogLoss $\ell$5; TIN+TdPE reaches AUC $\ell$6 and LogLoss $\ell$7; TIN+dPE reaches AUC $\ell$8 and LogLoss $\ell$9 (Hu et al., 10 Sep 2025).

The soft-target optimization variant reports consistent improvements for SASRec on three public datasets. On Diginetica, SASRec improves from Base $$H^{(\ell-1)} = [h_1^{(\ell-1)}, \ldots, h_n^{(\ell-1)}] \in \mathbb{R}^{n \times d}$$0, $$H^{(\ell-1)} = [h_1^{(\ell-1)}, \ldots, h_n^{(\ell-1)}] \in \mathbb{R}^{n \times d}$$1, $$H^{(\ell-1)} = [h_1^{(\ell-1)}, \ldots, h_n^{(\ell-1)}] \in \mathbb{R}^{n \times d}$$2, and $$H^{(\ell-1)} = [h_1^{(\ell-1)}, \ldots, h_n^{(\ell-1)}] \in \mathbb{R}^{n \times d}$$3 to Ours $$H^{(\ell-1)} = [h_1^{(\ell-1)}, \ldots, h_n^{(\ell-1)}] \in \mathbb{R}^{n \times d}$$4, $$H^{(\ell-1)} = [h_1^{(\ell-1)}, \ldots, h_n^{(\ell-1)}] \in \mathbb{R}^{n \times d}$$5, $$H^{(\ell-1)} = [h_1^{(\ell-1)}, \ldots, h_n^{(\ell-1)}] \in \mathbb{R}^{n \times d}$$6, and $$H^{(\ell-1)} = [h_1^{(\ell-1)}, \ldots, h_n^{(\ell-1)}] \in \mathbb{R}^{n \times d}$$7. On MovieLens, Base $$H^{(\ell-1)} = [h_1^{(\ell-1)}, \ldots, h_n^{(\ell-1)}] \in \mathbb{R}^{n \times d}$$8, $$H^{(\ell-1)} = [h_1^{(\ell-1)}, \ldots, h_n^{(\ell-1)}] \in \mathbb{R}^{n \times d}$$9, $$Q^{(\ell)} = H^{(\ell-1)} W_Q^{(\ell)}, \quad K^{(\ell)} = H^{(\ell-1)} W_K^{(\ell)}, \quad V^{(\ell)} = H^{(\ell-1)} W_V^{(\ell)},$$0, and $$Q^{(\ell)} = H^{(\ell-1)} W_Q^{(\ell)}, \quad K^{(\ell)} = H^{(\ell-1)} W_K^{(\ell)}, \quad V^{(\ell)} = H^{(\ell-1)} W_V^{(\ell)},$$1 improve to Ours $$Q^{(\ell)} = H^{(\ell-1)} W_Q^{(\ell)}, \quad K^{(\ell)} = H^{(\ell-1)} W_K^{(\ell)}, \quad V^{(\ell)} = H^{(\ell-1)} W_V^{(\ell)},$$2, $$Q^{(\ell)} = H^{(\ell-1)} W_Q^{(\ell)}, \quad K^{(\ell)} = H^{(\ell-1)} W_K^{(\ell)}, \quad V^{(\ell)} = H^{(\ell-1)} W_V^{(\ell)},$$3, $$Q^{(\ell)} = H^{(\ell-1)} W_Q^{(\ell)}, \quad K^{(\ell)} = H^{(\ell-1)} W_K^{(\ell)}, \quad V^{(\ell)} = H^{(\ell-1)} W_V^{(\ell)},$$4, and $$Q^{(\ell)} = H^{(\ell-1)} W_Q^{(\ell)}, \quad K^{(\ell)} = H^{(\ell-1)} W_K^{(\ell)}, \quad V^{(\ell)} = H^{(\ell-1)} W_V^{(\ell)},$$5. On Last.FM, Base $$Q^{(\ell)} = H^{(\ell-1)} W_Q^{(\ell)}, \quad K^{(\ell)} = H^{(\ell-1)} W_K^{(\ell)}, \quad V^{(\ell)} = H^{(\ell-1)} W_V^{(\ell)},$$6, $$Q^{(\ell)} = H^{(\ell-1)} W_Q^{(\ell)}, \quad K^{(\ell)} = H^{(\ell-1)} W_K^{(\ell)}, \quad V^{(\ell)} = H^{(\ell-1)} W_V^{(\ell)},$$7, $$Q^{(\ell)} = H^{(\ell-1)} W_Q^{(\ell)}, \quad K^{(\ell)} = H^{(\ell-1)} W_K^{(\ell)}, \quad V^{(\ell)} = H^{(\ell-1)} W_V^{(\ell)},$$8, and $$Q^{(\ell)} = H^{(\ell-1)} W_Q^{(\ell)}, \quad K^{(\ell)} = H^{(\ell-1)} W_K^{(\ell)}, \quad V^{(\ell)} = H^{(\ell-1)} W_V^{(\ell)},$$9 improve to Ours $i$00, $i$01, $i$02, and $i$03 (Zhang et al., 2024).

Both strands of work also make their limitations explicit. In the Tencent deployment, data sparsity per category can limit head quality, motivating retention of a default head for tail categories and cold-start cases. Category proliferation increases parameter footprint and can raise concerns about memory and overfitting, especially under category drift. The paper further notes that decoupling is most beneficial when categories have heterogeneous temporal patterns or distinct semantics, and less impactful when targets share similar dynamics or behavior distributions (Hu et al., 10 Sep 2025). In the soft-target framework, limitations include static, sequence-agnostic propagation, the computational and memory cost of maintaining an $i$04-dimensional soft target distribution, sensitivity to graph quality induced by clustering, and sampling bias under sampled softmax (Zhang et al., 2024).

Taken together, the two formulations show that Target-Decoupled SASRec is best understood as a family of SASRec adaptations that isolate target-dependent information at different points in the learning pipeline. One version decouples the final projection space of the user tower to avoid semantic averaging across categories; the other decouples target and non-target supervision to avoid conflicting gradients under sparse feedback. This suggests a broader research direction in which sequence encoders, temporal encodings, side-information pathways, and learning objectives are all treated as possible loci for target-aware disentanglement.