Contextual Diffusion Purifier for CTR

• Contextual Diffusion Purifier (CDP) is a diffusion-based technique that purifies noisy, category-filtered user behavior into a context-aware latent interest for CTR prediction.
• It combines category filtering, adaptive conditioning via a Mixture-of-Experts, and multi-step denoising to refine user representations based on query, user, item, and context signals.
• Empirical evaluations on industrial data show that CDP significantly outperforms baseline models in AUC, UAUC, and GAUC, while remaining practical in latency-sensitive deployments.

Contextual Diffusion Purifier (CDP) is a diffusion-based method for click-through rate (CTR) prediction in industrial search and recommendation systems. It treats category-filtered user behavior sequences as “contaminated observations” of latent interest, applies a forward noising and conditional reverse denoising process guided by Query × User × Item × Context signals, and produces a context-aware purified interest representation for downstream CTR estimation. In the formulation reported for industrial search, CDP reframes user-interest modeling from static history aggregation to context-conditioned latent purification (Zhao et al., 24 Sep 2025).

1. Problem formulation and conceptual basis

CDP is defined for the CTR objective

$p(\text{click}\mid q,u,i,s),$

where $q$ is the query, $u$ the user, $i$ the candidate item, and $s$ contextual metadata such as hour, page, and client. A central premise is that historical behavior sequences—clicks, views, favorites, and purchases—are not direct measurements of stable preference. The reported sources of contamination include exposure bias, position bias, category drift, accidental or low-quality clicks, and dynamic context. The method therefore treats observed history as an imperfect proxy for intent rather than as a faithful representation of preference (Zhao et al., 24 Sep 2025).

This position is presented as a departure from the “identify-aggregate” paradigm. In that paradigm, a model identifies salient historical events and aggregates them into a fixed representation. CDP instead first narrows the sequence by category and then models the resulting latent as noisy. The raw behavior sequence is

$$\mathcal{H}=\{e_1,\dots,e_T\},\qquad e_t\in\mathbb{R}^d,$$

and category filtering retains

$$\mathcal{H}_{\text{cat}}=\{e_t\in\mathcal{H}\mid \text{category}(e_t)=\text{category}(i)\}.$$

The filtered sequence is mean pooled to form the initial latent

$$\mathbf{z}_0=\frac{1}{|\mathcal{H}_{\text{cat}}|}\sum_{e_t\in\mathcal{H}_{\text{cat}}} e_t.$$

The paper explicitly characterizes this category-filtered, mean-pooled latent as a contaminated latent observation rather than as a final interest embedding (Zhao et al., 24 Sep 2025).

The category filter is described as the first purification step. Its stated purpose is to suppress cross-category noise before diffusion. This suggests that CDP assumes same-category history is the most relevant substrate for subsequent denoising, while the remaining contamination is delegated to the diffusion process itself.

2. Architecture and conditional denoising mechanism

CDP is organized into three stages: category-gated sequence pooling and noise injection, adaptive condition construction via Mixture-of-Experts (MoE), and conditional denoising followed by downstream CTR prediction. The final purified interest embedding is denoted $\mathbf{z}^*$, and the CTR head computes

$$p(\text{click}\mid q,u,i,s)=\sigma\Big(f_{\text{DNN}}\big(\mathbf{z}^* \,\|\, \phi_{\text{sparse}}(q,u,i,s)\big)\Big).$$

The purified latent is therefore not a standalone recommender state; it is concatenated with the original sparse feature embedding and consumed by a downstream DNN (Zhao et al., 24 Sep 2025).

The conditioning signal is a fused vector $q$0 built from three components. The collaborative condition combines query, user, and item embeddings with explicit pairwise interactions:

$q$1

The context and scene conditions are encoded separately as

$q$2

A gating network then produces mixture weights

$q$3

with $q$4, and the final condition vector is

$q$5

This MoE fusion is presented as the mechanism that adaptively emphasizes different conditioning sources under different search situations (Zhao et al., 24 Sep 2025).

The denoiser is a multi-layer MLP,

$q$6

which takes the noisy latent $q$7, diffusion step $q$8, and condition vector $q$9. The reverse process is written as

$u$0

and after $u$1 denoising steps,

$u$2

Because the condition vector depends on the current $u$3, the same user history can produce different purified embeddings in different scenarios. The paper identifies this as the key novelty of the method: denoising is guided by a current-task-specific condition rather than producing one static user vector (Zhao et al., 24 Sep 2025).

3. Diffusion formulation, objectives, and implementation

CDP applies diffusion to the category-pooled latent $u$4, not to the raw sequence. The reported schedule uses 100 diffusion steps for both noising and denoising, a linear schedule, and noise intensity ranging from 0.005 to 0.01. The forward process is

$u$5

with

$u$6

In the paper’s interpretation, mean pooling functions as a low-pass filter, but still leaves a noisy representation that must be purified through conditional reverse denoising (Zhao et al., 24 Sep 2025).

Training combines a diffusion reconstruction term with CTR supervision:

$u$7

The reconstruction loss is

$u$8

and the BCE term is written as

$u$9

The training objective therefore couples noise prediction and click prediction in a single optimization criterion (Zhao et al., 24 Sep 2025).

The reported implementation uses Adam, 4 NVIDIA A10 GPUs, batch size 4096 per GPU, learning rate $i$0 for dense parameters, and $i$1 for sparse parameters. Although training uses 100 forward and reverse steps, online inference uses 20 denoising steps for an efficiency/effectiveness trade-off. The paper also states that CDP can be integrated into existing ranking architectures without structural modifications, which is presented as an operational advantage for industrial deployment (Zhao et al., 24 Sep 2025).

4. Empirical evaluation and reported performance

The evaluation is conducted on industrial user logs collected because existing public datasets were said not to simultaneously satisfy the needs for dense behavior sequences, search queries, and rich contextual information. The split uses the first 60 days for training and the last 1 day for testing. The scale reported is 300 million samples per day, with average user sequence length around 2000. Offline evaluation uses AUC, UAUC, and GAUC, with baselines SIM, QIN, TWIN, HSTU, DDRM, and DiffRec (Zhao et al., 24 Sep 2025).

Model	AUC / UAUC / GAUC
SIM	0.7298 / 0.6425 / 0.6283
QIN	0.7314 / 0.6437 / 0.6301
TWIN	0.7329 / 0.6451 / 0.6312
HSTU	0.7332 / 0.6455 / 0.6310
DDRM	0.7321 / 0.6429 / 0.6295
DiffRec	0.7327 / 0.6447 / 0.6315
CDP	0.7402 / 0.6514 / 0.6362

The offline results place CDP above both classical sequence models and the listed diffusion-based recommenders. The paper interprets this as evidence that conditioning plus purification is stronger than plain generative denoising in this setting (Zhao et al., 24 Sep 2025).

An ablation study compares no condition, no collaborative condition, no context-and-scene conditions, no MoE fusion, and full CDP. The reported examples include full CDP at AUC 0.7402, UAUC 0.6514, GAUC 0.6362; no condition at AUC 0.7341, UAUC 0.6467, GAUC 0.6336; and no MoE fusion at AUC 0.7355, UAUC 0.6470, GAUC 0.6341. The paper states that full CDP is best on all offline metrics, that removing MoE fusion causes the largest performance drop, and that the collaborative condition is the most influential single signal (Zhao et al., 24 Sep 2025).

The online study is a 14-day A/B test over billions of requests on an e-commerce search platform. Relative to a base SIM-like model, the reported gains are +1.240% GMV, +1.032% total orders, and +0.745% CTR, with about +2 ms latency. These figures are used to support the claim that the method remains practical in a latency-sensitive production system (Zhao et al., 24 Sep 2025).

5. Relation to diffusion purification and adjacent research areas

The word “purifier” in CDP does not denote adversarial image purification. In adversarial-defense work, diffusion purification typically means adding noise to an attacked input and reversing the diffusion process so that the result returns toward the clean data manifold. “Guided Diffusion Model for Adversarial Purification” uses partial forward diffusion and a guided reverse process conditioned on the adversarial image itself, with the guidance implemented through distance-based conditioning using MSE or negative SSIM (Wang et al., 2022). In image restoration, diffusion purification has also been used as a refinement phase decoupled from data-consistency optimization, alternating reconstruction and purification rather than embedding likelihood gradients inside every reverse step (Li et al., 2024). By contrast, Contextual Diffusion Purifier purifies latent interest representations derived from user history, and its conditioning signal is the Query × User × Item × Context quadrangle rather than an attacked image or a measurement operator (Zhao et al., 24 Sep 2025).

The acronym “CDP” is also overloaded in other research areas. In robotics, CDP refers to “Causal Diffusion Policy,” a transformer-based diffusion model for visuomotor policy learning that conditions on historical action sequences and uses causal temporal attention, chunk-wise autoregressive inference, and cache sharing (Ma et al., 17 Jun 2025). In print-security research, CDP authentication refers to copy detection pattern authentication recast as multi-class out-of-distribution detection, where a class-conditional ControlNet is trained only on authentic printer families and counterfeits are detected by class-conditioned reconstruction error (Atoki et al., 1 Jul 2026). These usages are conceptually distinct from Contextual Diffusion Purifier, though all of them use diffusion to enforce or exploit contextual structure.

This broader landscape suggests that “purification” has become a flexible diffusion motif rather than a single technical recipe. In CDP for CTR prediction, purification means reconstructing a context-aware latent interest vector from noisy behavioral evidence; in adversarial defense it means attempting to remove perturbations; in restoration it means enforcing a generative prior; and in CDP authentication it means testing whether a sample is well explained by an authentic manifold under known context.

6. Interpretation, scope, and points of caution

The reported qualitative interpretation of CDP is that user histories are “interest fossils”: they preserve genuine intent but are eroded by contextual and exposure-related noise. Category filtering removes cross-category clutter, forward noising formalizes residual contamination, and conditional denoising reconstructs a purified latent that can shift with the current query and scene (Zhao et al., 24 Sep 2025). Within this framing, the method’s central claim is not merely that diffusion can model user behavior, but that diffusion can serve as a purification operator under contextual control.

At the same time, the evaluation relies on industrial logs rather than a public benchmark, because the paper states that existing public datasets do not jointly provide dense behavior sequences, search queries, and rich contextual information. A plausible implication is that independent reproduction depends on access to comparable industrial-scale data and serving infrastructure. Another plausible implication is that the effectiveness of CDP depends on the relevance of same-category history and on the informativeness of the conditioning signals supplied by query, user, item, context, and scene.

Broader diffusion-purification literature also provides a cautionary contrast. In adversarial defense, “DiffBreak: Is Diffusion-Based Purification Robust?” argues that diffusion-based purification should not be assumed robust under adaptive white-box attacks, especially when evaluation uses only a single random purification path; it proposes majority-vote evaluation and shows that low-frequency attacks can still break the defense (Kassis et al., 2024). That critique is directed at stochastic purification defenses for adversarial examples rather than at CTR interest modeling, but it underscores that “purification” is not a guarantee of robustness across domains.

In its own domain, CDP is best understood as a context-conditioned latent denoiser for ranking. Its stated novelty lies in treating filtered history as a contaminated latent observation, denoising it under Query × User × Item × Context guidance, and using the resulting $i$2 as a scenario-dependent interest representation rather than as a static summary of user behavior (Zhao et al., 24 Sep 2025).