Masked Diffusion Generative Recommendation

Traditional recommenders assign independent IDs to items and therefore make limited use of multimodal content. Generative recommendation encodes content into a continuous space, quantizes it into semantic IDs, and generates those IDs with a compact vocabulary. TIGER introduced RQ-VAE plus autoregressive Transformer decoding; Cobra added continuous vectors; RPG used optimized product quantization. Despite these advances, most systems still inherit fixed-order language-model decoding.

Given user history su, the task is to generate a target SID c=(c1,…,cL), learning pθ(c|su). Autoregressive modeling factorizes this as ∏ℓpθ(cℓ|c<ℓ,su), restricting each decision to left-prefix information and weakening global consistency. A shared order also conflicts with heterogeneous user attention. One-shot parallel prediction improves efficiency but can overlook local constraints and fine-grained token correlations.

• ��Parallel codebook: OPQ splits the item representation into independently quantized subspaces, removing the fixed hierarchy induced by residual codebooks.
• ��Dynamic training noise: curriculum scheduling gradually increases mask difficulty, while history-aware allocation targets rare and difficult tokens.
• ��Two-stage inference: single-position warm-up captures coarse semantic anchors; parallel prediction then fills token groups, with beam search producing multiple candidates.

• ��Encoding: a pretrained encoder produces ei; projection fℓ(ei) yields subspace vectors, and cℓi=argminj||eℓi−Cℓ[j]||² selects codewords.
• ��Corruption: target SIDs are partially replaced with [MASK], producing xτ. The decoder receives the corrupted SID, user history, timestep, and a difficulty embedding.
• ��Curriculum: with τ=min(1,n/N), the paper defines δ=(√(1−cos²(π/2(1−τ))))γ and samples mask count k from a δ-dependent distribution.
• ��Allocation: frequency fℓ counts how often the target token appears in the user history; lower frequency receives higher masking priority.
• ��Denoising: bidirectional attention predicts masked tokens. Inference begins with warm-up, switches to parallel updates, and uses beam search to retrieve items.

The paper evaluates MDGR on two public datasets and one industrial dataset against ten baselines, including TIGER and Cobra, covering generative and discriminative recommendation. Training minimizes masked-position cross-entropy, LDDM=E[−∑ℓ∈Mτlog pθ(xℓ0|xτ,τ,su)]. The authors also conduct an online A/B test on a large advertising platform and examine the roles of parallel codebooks, curriculum noise, history-aware allocation, and two-stage decoding.

MDGR achieves the best performance across the reported comparison settings. The abstract states a maximum gain of 10.78%, while the contribution summary reports improvements of 7.17%–10.78% over discriminative and generative baselines. Online deployment produced a 1.20% revenue increase and a 3.69% GMV increase. Because the supplied text lacks per-dataset metric tables and complete ablations, it cannot establish exact gains for each component or a precise acceleration factor.

MDGR is suited to candidate generation in e-commerce, advertising, short-video, and content platforms where catalogs are large, item semantics are rich, and latency matters. Deployment requires a stable multimodal encoder, an OPQ codebook, and a fast SID-to-item index. Operators must tune warm-up steps, beam width, and parallel group size to balance retrieval quality, duplication, memory, and tail latency.

The framework depends heavily on SID quality: quantization errors or weak content representations can produce internally consistent but incorrect semantic combinations. History-aware masking may be unreliable for cold-start users, short histories, or rapidly shifting interests. Although diffusion reduces strict autoregressive dependence, it still performs multiple decoding rounds. Production studies should report tail latency, GPU memory, codebook refresh stability, and robustness under domain or distribution shifts.