Hierarchical Diffusion-Based Ad Recommendation with Variational Graph Attention and Adversarial Refinement

1. Introduction

Sequential advertisement recommendation means learning how users interact with ads over time. These interactions come from different types of data, such as text, time, and embeddings. Traditional models like RNNs or attention-based methods use fixed patterns. They often do not work well when data is sparse or new users appear. They also have trouble suggesting a wide range of ads. Because these models process steps one by one, they are slower in real-time use.

Diffusion models provide a new way by treating data as a process of removing noise. But most current diffusion models use too much computation and are not designed for structured recommendation. To solve this, we propose GDAR. GDAR is a model that combines a multi-level diffusion process with noise steps that can be trained. It also uses parallel denoising to speed up sequence generation.

GDAR also uses a variational graph attention network to learn how ads are related and how these relations change over time. It uses uncertain embeddings to do this. The model also adds an adversarial part with contrastive learning to improve the meaning and variety of results. The training includes several types of loss: diffusion, variational, adversarial, and contrastive. It also uses data augmentation and feature normalization. These parts make GDAR strong and efficient for large-scale recommendation tasks.

2. Related Work

Building on the challenges highlighted in the introduction, we now survey recent advances in diffusion-based and graph-based recommendation methods to position our contribution. Diffusion models have been used in sequential recommendation tasks. Chen et al.[1] propose a coarse-to-fine SLAM-integrated multi-view reconstruction with Transformer-based matching and parallel bundle adjustment that improves correspondence robustness and reprojection error, which can directly strengthen GDAR’s cross-modal alignment. These methods gave good results but did not handle graph structures well.Zhang and Bhattacharya [2] propose an iteratively trained Recurrent Neural Operator as a transferable surrogate for history-dependent multiscale models, improving accuracy and efficiency, which can inform GDAR’s history-aware sequence modeling and accelerate inference. Luo et al.[3] Gemini-GraphQA pairs a graph encoder with an executable-graph solver and an execution-correctness loss; adopting its graph-encoder and execution-loss could tighten GDAR’s structural reasoning over user–item graphs.

Some researchers tried to improve control in generation. Jiang et al.[4] propose RobustKV, a KV-cache eviction that prunes low-attention tokens to block jailbreaks and can be applied to GDAR’s cross-attention to improve robustness. Li et al.[5] added social graph signals to the model. Wang et al.[6] propose a latent-variable multitask LVMTL with a QOIEM estimator to model dependent, heterogeneous degradation and handle missing data, which could inform GDAR’s modeling of user heterogeneity and robustness to incomplete interaction logs.

Other works focused on adding randomness and faster generation. Luo et al. [7]propose TriMedTune, a triple-branch fine-tuning (HVPI, DATA, MKD-UR) that improves multimodal image–text alignment and robustness; adopting HVPI and MKD-UR could tighten GDAR’s cross-modal fusion. Ren et al.[8] and Fan et al.[9] used non-step-by-step generation and random attention to speed things up. But these models do not combine different types of data well and do not include adversarial learning.

Earlier work made progress in diffusion and contrastive methods. But many models do not use dynamic graphs or combine generation with refinement. Our method brings these parts together into one generative model.

3. Methodology

We propose Generative Diffusion-based Advertisement Recommendation (GDAR), which reframes ad recommendation as a conditional sequence generation problem by unifying diffusion models, variational graph attention, and cross-modal fusion. Its key components are:

• Hierarchical Diffusion Process: Adaptive noise schedules that reflect user behavior.
• Variational Graph Attention Network: Latent-space modeling of dynamic item–item relationships.
• Cross-Modal Fusion Network: Integration of textual metadata, pretrained embeddings, and historical interactions.
• Adversarial Refinement Module: Enforcement of realistic temporal dynamics in generated sequences.
• Contrastive Alignment Mechanism: Alignment of generated outputs with user preferences.

As shown in Figure 1, the GDAR framework unifies diffusion models, variational graph attention, and cross-modal fusion to reframe ad recommendation as a conditional sequence generation problem, enabling efficient parallel denoising and interpretable preference evolution.

3.1. Variational Graph Neural Network for Item Modeling

We represent advertisements as nodes in a dynamic graph $\mathcal{G}=(\mathcal{V},\mathcal{E})$, where edge weights combine co-occurrence and temporal decay:

$$w_{ij}={\displaystyle \frac{\mathrm{count}(i,j)}{\sqrt{\mathrm{count}\left(i\right)\mathrm{count}\left(j\right)}}}exp\left(-|\Delta t_{ij}|/\tau \right).$$

(1)

A Variational Graph Attention Network (VGAT) then encodes each node $v_i$ with a Gaussian posterior

$$q(z_i\mid \mathcal{G})=\mathcal{N}\left({\mu }_i,diag\left({\sigma }_i^2\right)\right),$$

(2)

where ${\mu }_i$ and ${\sigma }_i$ are produced by attention layers with coefficients

$${\alpha }_{ij}={\displaystyle \frac{exp\left(\mathrm{LeakyReLU}\left(a^{\top }[W{h}_i\parallel W{h}_j]\right)\right)}{{\sum }_{k\in \mathcal{N}\left(i\right)}exp\left(\mathrm{LeakyReLU}\left(a^{\top }[W{h}_i\parallel W{h}_k]\right)\right)}}.$$

(3)

We apply the reparameterization $z_i={\mu }_i+{\sigma }_i\odot \epsilon$, $\epsilon \sim \mathcal{N}(0,I)$, to allow end-to-end training. This variational formulation regularizes representations, quantifies uncertainty and improves generalization (Figure 2).

3.2. Hierarchical Diffusion Model for Sequence Generation

Adaptive Noise Schedule Learning. Traditional diffusion models use fixed schedules, which may not suit varied user behaviors. We introduce a learnable schedule:

$${\beta }_t=\sigma \left({\mathbf{w}}_{\beta }^T\mathrm{MLP}\left([t/T;{\mathbf{u}}_{\mathrm{context}}]\right)\right)({\beta }_{max}-{\beta }_{min})+{\beta }_{min},$$

(4)

allowing faster or slower diffusion depending on user consistency.

Conditional Denoising Architecture. At each reverse step, the model predicts the noise component conditioned on the noisy sequence and user context:

$${ϵ}_{\theta }({\mathbf{x}}_t,t,\mathbf{c})=\mathrm{TransformerBlock}\left({\mathbf{x}}_t+{\mathbf{PE}}_t,\mathrm{CrossAttn}\left(\mathbf{c}\right)\right),$$

(5)

incorporating timestep embeddings, cross-attention, and combined absolute/relative positional encodings (Figure 3).

4. Training Objectives

4.1. Diffusion Noise Prediction

The primary loss component trains the denoising network to predict the noise added during the forward process. This loss is computed over randomly sampled timesteps and noise levels:

$${\mathcal{L}}_{\mathrm{diffusion}}={\mathbb{E}}_{t\sim U(1,T),{\mathbf{x}}_0\sim p_{\mathrm{data}},ϵ\sim \mathcal{N}(0,I)}{∥ϵ-{ϵ}_{\theta }({\mathbf{x}}_t,t,\mathbf{c})∥}^2.$$

(6)

To improve training efficiency, we employ importance sampling for timestep selection, giving higher weight to poorly predicted steps. Additionally, we use an SNR-weighted variant:

$${\mathcal{L}}_{\mathrm{diffusion}}^{\mathrm{weighted}}={\mathbb{E}}_{t,{\mathbf{x}}_0,ϵ}\left[{\displaystyle \frac{\mathrm{SNR}(t-1)-\mathrm{SNR}\left(t\right)}{\mathrm{SNR}\left(0\right)}}{\parallel ϵ-{ϵ}_{\theta }({\mathbf{x}}_t,t,\mathbf{c})\parallel }^2\right],$$

(7)

where $\mathrm{SNR}\left(t\right)={\overline{\alpha }}_t/(1-{\overline{\alpha }}_t)$.

4.2. Variational ELBO Regularization

The VGNN component uses a modified ELBO to prevent posterior collapse:

$${\mathcal{L}}_{\mathrm{VAE}}={\mathbb{E}}_{q(z\mid \mathcal{G})}[logp(\mathcal{G}\mid z)]-\beta \left(t\right)\mathrm{KL}\left(q(z\mid \mathcal{G})\parallel p\left(z\right)\right).$$

(8)

The annealing factor $\beta \left(t\right)$ follows a cyclical schedule:

$$\beta \left(t\right)=min\left(1,(tmod{T}_{\mathrm{cycle}})/T_{\mathrm{warmup}}\right){\beta }_{max}.$$

(9)

4.3. Adversarial & Contrastive Objectives

We combine a least-squares adversarial loss with a contrastive alignment term.

Adversarial Loss:

$$\begin{array}{cc}\hfill {\mathcal{L}}_D& =\frac{1}{2}{\mathbb{E}}_{\mathbf{x}\sim p_{\mathrm{data}}}\left[{(D_{\varphi }(\mathbf{x},\mathbf{c})-1)}^2\right]+\frac{1}{2}{\mathbb{E}}_{\mathbf{x}\sim p_{\theta }}\left[D_{\varphi }{(\mathbf{x},\mathbf{c})}^2\right],\hfill \end{array}$$

(10)

$$\begin{array}{cc}\hfill {\mathcal{L}}_G& =\frac{1}{2}{\mathbb{E}}_{\mathbf{x}\sim p_{\theta }}\left[{(D_{\varphi }(\mathbf{x},\mathbf{c})-1)}^2\right].\hfill \end{array}$$

(11)

Contrastive Alignment:

$$Z=exp\left(\frac{f({\mathbf{x}}_{\mathrm{gen}},\mathbf{c})}{\tau }\right)+\sum _{i\in \mathrm{HardNeg}\left(K\right)}exp\left(\frac{f({\mathbf{x}}_i^{-},\mathbf{c})}{\tau }\right),$$

(12)

$${\mathcal{L}}_{\mathrm{align}}=-log{\displaystyle \frac{exp\left(f({\mathbf{x}}_{\mathrm{gen}},\mathbf{c})/\tau \right)}{Z}}.$$

(13)

Figure 4 shows training dynamics of these objectives.

4.4. Data Preprocessing

Figure 5 illustrates the preprocessing pipeline, which unifies sequence augmentation, multimodal feature handling, and temporal encoding.

We extract overlapping context–target pairs by sliding a window of size W over each sequence $\mathbf{S}$:

$${\mathcal{D}}_{\mathrm{train}}=\bigcup _{i=k}^{L-1}\bigcup _{j=1}^{min(i-k+1,W)}\left\{({\mathbf{S}}_{i-k-j+1:i-j+1},{\mathbf{S}}_{i-j+2:i+1})\right\},$$

(14)

then apply 15% stochastic masking and, for long sequences, combine the most recent k items with k items sampled by importance:

$${\mathbf{S}}_{\mathrm{sampled}}={\mathbf{S}}_{\mathrm{recent}}\cup \mathrm{ImportanceSample}({\mathbf{S}}_{\mathrm{history}},k).$$

(15)

Text is tokenized into subwords (vocab. size 30000), padded or truncated to $l_{max}=128$:

$${\mathbf{C}}_{\mathrm{proc}}=\mathrm{Truncate}(\mathrm{Pad}(\mathrm{Tokenize}\left(text\right),l_{max}),l_{max}).$$

(16)

Pretrained vectors ${\mathbf{v}}_{\mathrm{orig}}$ are projected and layer-normalized:

$${\mathbf{v}}_{\mathrm{norm}}=\mathrm{LayerNorm}\left(W_{\mathrm{proj}}{\displaystyle \frac{{\mathbf{v}}_{\mathrm{orig}}}{\parallel {\mathbf{v}}_{\mathrm{orig}}{\parallel }_2}}+b_{\mathrm{proj}}\right).$$

(17)

Temporal features—hour-of-day ${\mathbf{e}}_{\mathrm{hour}}$, day-of-week ${\mathbf{e}}_{\mathrm{dow}}$, and inter-action gap $\Delta t$—are fused with item embeddings via a gating mechanism:

$${\mathbf{h}}_{\mathrm{temp}}=g_t\odot {\mathbf{h}}_{\mathrm{item}}+(1-g_t)\odot \mathrm{MLP}\left([{\mathbf{e}}_{\mathrm{hour}};{\mathbf{e}}_{\mathrm{dow}};log(1+\Delta t)]\right).$$

(18)

5. Experiment Results

We evaluate GDAR against state-of-the-art baselines on the Baidu advertisement dataset. All experiments were conducted on NVIDIA V100 GPUs using PaddlePaddle framework with mixed precision training. And the changes in model training indicators are shown in Figure 6.

.

As shown in Table 1, GDAR significantly outperforms all baseline methods across multiple metrics. The improvement is particularly notable in coverage, demonstrating the benefit of our generative approach in promoting diversity.

The ablation study in Table 2 reveals that the VGNN component contributes most significantly to performance, highlighting the importance of modeling item relationships. The adversarial and cross-modal components also provide substantial improvements.

Table 3 demonstrates GDAR’s robustness across different user segments. Notably, the model maintains competitive performance even for cold-start users, attributed to the generative approach and graph-based item modeling.

6. Conclusion

This paper introduced GDAR, a novel generative framework for sequential advertisement recommendation that successfully combines diffusion models with graph neural networks and multimodal fusion. Our comprehensive experiments demonstrate that GDAR achieves state-of-the-art performance. The generative paradigm not only improves recommendation accuracy but also enhances diversity and provides better support for cold-start scenarios. Future work will explore extending GDAR to multi-task learning scenarios and investigating more efficient inference strategies for real-time deployment.