Asymmetric Diffusion Rec Model

• The paper introduces a novel asymmetric diffusion mechanism that models realistic missing features via discrete dropout and latent denoising.
• It employs a lightweight two-layer MLP denoiser to reconstruct clean representations, preserving user-specific signals for improved CTR and ranking.
• Empirical results on the Douyin Music App show significant improvements in AUC and engagement metrics with only a marginal increase in inference time.

Searching arXiv for the primary paper and closely related diffusion-based recommendation work. Asymmetric Diffusion Recommendation Model (AsymDiffRec) is a diffusion-based recommendation framework that redefines the forward and reverse diffusion processes for discrete recommendation data by making them explicitly asymmetric. Rather than applying symmetric Gaussian corruption and denoising in a continuous space, AsymDiffRec simulates realistic missing features in raw discrete feature space and performs denoising in latent representation space, with an auxiliary task-oriented objective designed to preserve personalized information for ranking and CTR prediction. The model was proposed for industrial recommendation settings with heterogeneous features and serving-time feature missingness, and was reported to have been implemented in the Douyin Music App (Zhu et al., 18 Aug 2025).

1. Motivation and problem formulation

AsymDiffRec is motivated by two limitations attributed to prior diffusion-based recommenders. First, recommendation samples inhabit a discrete data space composed of heterogeneous fields such as meta features and ID-based personalized features, so adding continuous Gaussian noise to latent vectors does not correspond to a realistic perturbation of the original sample. Second, symmetric Gaussian noise can dominate denoiser training and thereby corrupt fine-grained user-specific signals that are central to CTR and ranking quality (Zhu et al., 18 Aug 2025).

The formulation used in AsymDiffRec is a binary classification setting within a ranking task. A sample has label $y \in \{0,1\}$ and raw features $\mathbf{x} = \{x_1,\dots,x_N\}$, where $N$ is the number of raw features. A feature extractor $h(\cdot)$ maps raw features to a latent representation $\mathbf{z} = h(\mathbf{x})$, and a predictor $f(\cdot)$ maps $\mathbf{z}$ to a scalar probability $\hat{y} = f(\mathbf{z})$. The main task objective is the standard cross-entropy loss

$$\mathcal{L}_{main} = -y \log \hat{y} - (1-y) \log (1 - \hat{y}).$$

A central design premise is that industrial logs frequently contain missing features at serving time. AsymDiffRec therefore treats missingness not as an incidental nuisance but as the forward corruption process itself. This suggests that the model is less a generic denoising module than a recommendation-specific robustness mechanism aligned with real serving conditions.

2. Asymmetric diffusion processes

The asymmetry in AsymDiffRec is defined at the level of data space, noise type, schedule, and conditioning. The forward process operates on raw discrete features by dropping features, while the reverse process operates in continuous latent space and reconstructs a clean representation directly rather than reconstructing raw features and then re-encoding them (Zhu et al., 18 Aug 2025).

Let $\mathbf{x}_0$ be the original feature set, and let $\mathbf{x} = \{x_1,\dots,x_N\}$0 be a binary mask in which $\mathbf{x} = \{x_1,\dots,x_N\}$1 indicates that feature $\mathbf{x} = \{x_1,\dots,x_N\}$2 is missing. The forward process iteratively flips zeros to ones in the mask and drops the corresponding features. For $\mathbf{x} = \{x_1,\dots,x_N\}$3,

$\mathbf{x} = \{x_1,\dots,x_N\}$4

where $\mathbf{x} = \{x_1,\dots,x_N\}$5 is the unit vector for feature index $\mathbf{x} = \{x_1,\dots,x_N\}$6. The induced transition over features is

$\mathbf{x} = \{x_1,\dots,x_N\}$7

One feature is dropped per step, and $\mathbf{x} = \{x_1,\dots,x_N\}$8.

The reverse process is not a time-reversal of that raw-space Markov chain. Instead, the incomplete sample $\mathbf{x} = \{x_1,\dots,x_N\}$9 is encoded as $N$0, and a denoiser $N$1 receives the concatenation of the missing-feature indicator and the latent representation:

$N$2

This one-shot latent reconstruction is trained against the clean latent representation $N$3 through

$N$4

The paper characterizes the asymmetry in three explicit ways: different noise type and space, different schedules, and different conditionings. The forward process is a discrete feature-drop mechanism in raw space; the reverse process is a single denoising map in latent space; and the reverse map conditions jointly on $N$5 and the missing-position mask $N$6.

3. Architecture and optimization

AsymDiffRec augments a production DCN-V2-based multi-task ranking model. The feature extractor $N$7 consists of embedding layers and deep networks mapping raw features to a latent representation $N$8. The denoiser $N$9 is a lightweight two-layer MLP operating in latent space, and the predictor $h(\cdot)$0 is the base model’s standard CTR or ranking head (Zhu et al., 18 Aug 2025).

The model is optimized end-to-end with three loss terms. The first is the main ranking loss on the clean representation:

$h(\cdot)$1

The second is the latent reconstruction loss:

$h(\cdot)$2

The third is a task-oriented auxiliary loss evaluated on the denoised representation:

$h(\cdot)$3

The total objective is

$h(\cdot)$4

No additional weights or schedules are introduced beyond the random choice of $h(\cdot)$5. The paper does not report a variance-preserving or variance-exploding design; the forward schedule is purely discrete feature dropout. It argues qualitatively that discrete dropout is better matched to the discrete nature of recommendation features, while the auxiliary task loss counteracts erosion of personalized information.

This optimization design is central to the model’s identity. A plausible implication is that AsymDiffRec treats diffusion not as a standalone generative objective but as a task-coupled representation refinement procedure. In that sense, the auxiliary loss is not merely regularization; it is the mechanism by which denoising is made recommendation-aware.

4. Serving-time denoising and system integration

A distinguishing property of AsymDiffRec is that the same missing-feature semantics used during training are preserved at serving time. For an incoming request, the system constructs a mask $h(\cdot)$6 in which $h(\cdot)$7 indicates that feature $h(\cdot)$8 is missing. The incomplete feature vector is encoded as $h(\cdot)$9, the denoiser computes $\mathbf{z} = h(\mathbf{x})$0, and the predictor outputs $\mathbf{z} = h(\mathbf{x})$1 for ranking (Zhu et al., 18 Aug 2025).

Because $\mathbf{z} = h(\mathbf{x})$2 operates on latent representations, the model is described as model-agnostic. The reported compatibility envelope includes feature-MLP models such as DCN-V2 and DNNs, matrix factorization or embedding-based rankers in which $\mathbf{z} = h(\mathbf{x})$3 maps IDs to embeddings, and sequence models in which sequences are encoded into $\mathbf{z} = h(\mathbf{x})$4 via an RNN or Transformer and then denoised conditionally on sequence-specific masks.

The computational overhead is reported as small. Average inference time increases from 12.1 ms for the baseline to 12.6 ms with AsymDiffRec. The denoiser is therefore positioned as suitable for real-time serving. This serving-stage symmetry between training corruption and deployment corruption is one of the model’s most distinctive industrial characteristics: the raw sample with missing features is explicitly treated as a noisy input to be denoised, rather than as a malformed sample to be handled only by the base recommender.

5. Empirical results and ablations

The main empirical evidence reported for AsymDiffRec consists of online A/B tests in the Douyin Music App and offline experiments on the DouyinMusic-40B dataset (Zhu et al., 18 Aug 2025).

In the online setting, the model was evaluated in a one-month A/B test at 100% traffic on more than 20 million users using a DCN-V2-based multi-task ranking model. The reported main metrics are Active Day at +0.131% and Duration at +0.166%. Constraint metrics are Like at +0.285%, Finish at +0.353%, Play at +0.256%, and DisLike at $\mathbf{z} = h(\mathbf{x})$5. The paper states that the improvements in Active Days and Duration are statistically significant and interprets the engagement improvements in Like, Finish, and Play as indications of better user satisfaction.

In the offline setting, the industrial dataset contains more than 40B samples collected over 10 weeks, with more than 200 features per sample. The reported evaluation metrics are AUC and UAUC, together with Relative Improvement over the base model. The base DCN-V2 model achieves AUC 0.92267 and UAUC 0.61578. The reported baselines are CDAE with AUC 0.92220 and UAUC 0.61600, MultiVAE with AUC 0.92285 and UAUC 0.61834, SSL with AUC 0.92297 and UAUC 0.62009, and DiffRec with AUC 0.92307 and UAUC 0.62152. AsymDiffRec reports AUC 0.92359 and UAUC 0.62614.

The ablation study isolates the two added training terms. Removing the reconstruction loss yields AUC 0.92276 and UAUC 0.62025. Removing the auxiliary loss yields AUC 0.92248 and UAUC 0.62069. The paper highlights that removing either reconstruction or auxiliary loss hurts performance, and that the “w/o aux” configuration reduces AUC below the base model. That result is used to support the claim that task-oriented optimization is necessary for preserving personalization during denoising.

The paper further attributes the observed robustness to training on realistic missing-feature corruption and to denoising conditioned on the explicit missing-position mask. Cold-start performance, however, is not specifically reported.

6. Position within diffusion-based recommendation and broader uses of “asymmetry”

Within diffusion-based recommendation, AsymDiffRec occupies a specific design point: discrete forward corruption in raw feature space, latent reverse denoising, and task-oriented optimization for ranking. This differs from symmetric Gaussian diffusion models that operate in continuous latent or input spaces in both forward and reverse processes, such as DiffRec and DDRM-style recommenders as characterized in the AsymDiffRec paper itself (Zhu et al., 18 Aug 2025).

The broader recommendation literature uses “asymmetry” in several related but non-identical senses. In slate recommendation, the diffusion framework in “Diffusion Model for Slate Recommendation” does not introduce a model named AsymDiffRec, but it explicitly sketches how an asymmetric variant could use position-dependent noise, non-symmetric transition operators, position-conditioned denoisers, and position-weighted utility terms for slate generation (Tomasi et al., 2024). In sequential recommendation, “Sequential Recommendation with Diffusion Models” describes an asymmetric process at the sequence level because only the target item is noised, while preceding positions remain fixed, and because the model bridges discrete item IDs and continuous latent embeddings through additional forward and reverse transitions (Du et al., 2023). In another sequential formulation, “Bridging User Dynamics: Transforming Sequential Recommendations with Schrödinger Bridge and Diffusion Models” makes the process asymmetric through user-conditioned boundary distributions, replacing the Gaussian prior with the user’s current preference state and modeling a directed path from current state to target item (Xie et al., 2024). A further extension, “Towards A Tri-View Diffusion Framework for Recommendation,” associates asymmetry with anisotropy-preserving denoising, free-energy optimization, and timestep-adaptive hard-negative sampling (Chen et al., 25 Nov 2025).

A common misconception is to treat all of these models as instances of the same algorithmic object. The evidence does not support that equivalence. The survey “A Survey on Diffusion Models for Recommender Systems” explicitly notes that it does not mention a model named “Asymmetric Diffusion Recommendation (AsymDiffRec)” under its recommender taxonomy, even though it discusses anisotropic noise, piece-wise variance schedules, and conditional mechanisms that can imply asymmetry in practice (Lin et al., 2024). Accordingly, the primary technical meaning of AsymDiffRec refers to the raw-to-latent asymmetric denoising architecture introduced for industrial ranking, whereas adjacent literature uses the term more broadly for directional, anisotropic, or non-uniform diffusion mechanisms.

7. Limitations and prospective extensions

The reported limitations of AsymDiffRec are tied to reproducibility, reverse-process granularity, and incomplete coverage of recommendation regimes (Zhu et al., 18 Aug 2025). The approach is evaluated only on the industrial DouyinMusic-40B dataset, and the authors note a lack of suitable public datasets, which may limit external reproducibility. The reverse process is one-shot in latent space rather than a multi-step time reversal, and the paper identifies multi-step latent denoising or explicit probabilistic modeling of $\mathbf{z} = h(\mathbf{x})$6 as possible future directions. The combined loss uses equal weighting in the reported experiments, and adaptive weighting or a curriculum over $\mathbf{z} = h(\mathbf{x})$7 is named as another prospective direction. Dedicated cold-start analysis is also not reported.

These limitations delimit the current scope of the method. AsymDiffRec is not presented as a general theory of asymmetric diffusion in recommendation, but as a practical industrial design in which discrete feature dropout models realistic corruption and a lightweight latent denoiser repairs the resulting representation. The broader literature suggests multiple extension paths—position-aware asymmetry in slates, user-conditioned boundary asymmetry in sequential recommendation, and anisotropy-aware denoisers in graph or tri-view settings—but those should be understood as related design families rather than as properties already established for the original AsymDiffRec model (Tomasi et al., 2024).