DimeRec: A Unified Framework for Enhanced Sequential Recommendation via Generative Diffusion Models

1 Introduction

• Fixed user embeddings and discriminative multi-interest vectors can limit representation of interest diversity and uncertainty, while deterministic methods can disadvantage less-interacted items.The exposure-bias concern arises when interacted items are assumed to be the most relevant.Source: §1 · p. 2
• Diffusion models applied directly to recommendation face discrete, time-varying item spaces, sparse interactions, and conflicting recommendation and reconstruction objectives.End-to-end training can also leave item representations disorganized, making direct diffusion difficult.Source: §1 · p. 2
• DimeRec shifts generation from individual items to users’ next interests, using stationary interest guidance to support diffusion reconstruction and representation learning.Its guidance comes from the user’s non-stationary interaction history.Source: §1 · pp. 2–3
• DimeRec introduces a new noise space and loss function with a guidance loss intended to stabilize representation learning alongside diffusion.The guidance loss is independent of the diffusion module.Source: §1 · p. 3
• DimeRec improves on multiple baselines across three public datasets and online testing reports gains in user time spent and recommendation diversity.The reported online evaluation is on an industrial short-video platform.Source: §1 · p. 3

2.1 Sequential Recommendation

• GRU4Rec uses GRUs, Caser uses convolutional filters, SASRec uses a transformer, and BERT4Rec uses bidirectional transformers with Cloze training to model behavioral sequences.Earlier methods FPMC and Fossil use Markov chains for dynamic transitions.Source: §2.1 · p. 3

2.2 Diffusion Models for Recommendation

• DiffRec maps discrete items to continuous space, while DiffuRec corrupts and iteratively reconstructs the target item embedding with scheduled Gaussian noise.Wang et al. extend diffusion to latent space and model interaction-sequence probability with a time-sensitive approach.Source: §2.2 · p. 3
• DreamRec uses a learning-to-generate paradigm but is described as challenging to scale industrially without negative-sampling generalization; DCDR is limited to reranking, not retrieval.DreamRec reports comparable results on some specific datasets.Source: §2.2 · p. 3

3 Preliminary

• The forward process gradually adds Gaussian noise to a sample x_0 over T steps, with each transition controlled by a pre-designed variance schedule β_t.The multi-step transition is described as tractable.Source: §3 · pp. 3–4
• In the reverse process, a parameterized Markov chain learns the joint distribution over the noisy states and the original sample.The reverse process starts from a standard Gaussian, and its mean and variance are parameterized at each step.Source: §3 · p. 4
• Because the ground-truth x_0 is unavailable during reverse generation, the model substitutes using x_t and trains a denoiser to predict noise ϵ_t.The text notes that x_t is available at timestep t.Source: §3 · p. 4
• Diffusion models can be trained with a simplified MSE loss instead of the original variational lower bound.Source: §3 · p. 4

4.1 Framework Overview

• The framework combines GEM, which extracts abstract guidance from behavior sequences, with DAM, which aggregates that guidance into a final user embedding.DAM conditions its process on the guidance sequence, the noised target-item embedding, and the diffusion step embedding.Source: §4.1 · pp. 4–5

4.2 Guidance Extraction Module

• Rule-based GEM can apply business-designed selection rules or select the latest K interactions and encode them as guidance.The paper gives popularity, empirical XTR statistics, and balancing exploration and exploitation as examples of rules.Source: §4.2 · pp. 5–6
• Model-based GEM applies self-attention to encoded behavior history to produce K guidance vectors, trained to match guidance with the target-item embedding using sampled softmax.The guidance length K is much smaller than the behavior-sequence length N.Source: §4.2 · p. 6

4.3 Diffusion Aggregation Module

• DAM predicts the target embedding with a conditional MLP, combining an MSE reconstruction objective and sampled-softmax retrieval loss in joint GEM–DAM training.The coefficients λ and μ balance the diffusion reconstruction and diffusion sampled-softmax losses.Source: §4.3 · pp. 6–7
• Spherical item representations and Geodesic Random Walk address the conflict between recommendation and reconstruction losses, enabling their joint optimization.On a sphere, item-vector norms are constant; the paper describes Euclidean updates as creating divergent optimization directions for the two losses.Source: §4.1 · p. 5, §4.3 · p. 7

4.4 Serving Strategy

• Serving samples an initial user embedding from a standard Gaussian, iteratively applies the reverse process, then retrieves Top-N candidates with an independent ANN search over the global item pool.The passage describes the final embedding as the input to candidate retrieval.Source: §4.4 · p. 7

4.5 Complexity Analysis

• Parameters scale as O(Nd), while forward-network and self-attention inference costs are O(nd^2) and O(n^2d); full-link online inference stays below 1 ms.The paper says increasing reverse steps raises offline inference time linearly but has little online efficiency impact.Source: §4.5 · p. 7

5.1.1 Dataset; 5.1.2 Training and Evaluation Protocal

• Evaluation uses YooChoose, KuaiRec, and ML-10M, timestamp-ordered next-item prediction, an 8:1:1 user split, and HR and NDCG metrics.Training samples are truncated at length 10 for YooChoose and KuaiRec and 70 for ML-10M; HR measures retrieval and NDCG ranking.Source: §5.1.1 · p. 8, §5.1.2 · p. 8

5.1.3 Baseline Methods; 5.1.4 Implementation Details

• The comparisons cover traditional sequence, multi-interest, and generative recommenders, with ComiRec used as DimeRec’s GEM module.The reported setup uses 64-dimensional embeddings, batch size 256, 10 sampled negatives per positive, four interests, and Adam at learning rate 0.005.Source: §5.1.3 · p. 8, §5.1.4 · pp. 8–9

5.2.1 Recommendation metrics evaluation

• 18.29%, 22.14%, and 39.13% are DimeRec’s HR@50 improvements over the second-best method on YooChoose, KuaiRec, and ML-10M; NDCG@50 gains are 20.75%, 9.81%, and 32.90%.DimeRec significantly outperforms all baselines on HR@20, HR@50, NDCG@20, and NDCG@50 across all datasets; NDCG@10 is only comparable on YooChoose.Source: Table 1 · p. 9, §5.2.1 · p. 9

5.2.2 Validation for DimeRec’s effectiveness

• DimeRec significantly outperforms other models in linear-probe classification accuracy and, excluding random-like DreamRec, recommends the highest number of categories.The probe trains a classifier on frozen item embeddings using ML-10M item categories; DreamRec’s diversity is discounted because its recommendations approach random performance.Source: §5.2.2 · p. 10, Table 2 · p. 10

5.3.1 Ablation for GEM

• SASRec and MLP GEM alternatives underperform the multi-interest ComiRec variants, and ComiRec-SA performs better than ComiRec-DR.The comparison reports Recall@20 and supports multi-interest guidance for diffusion-based recommendation.Source: §5.3.1 · p. 10, Table 3 · p. 11

5.3.2 Ablation for DAM and Loss Function

• Removing Lgem significantly harms DimeRec, while removing Lssm also substantially reduces performance.Lgem supplies guidance independent of diffusion to stabilize item-representation learning; Lssm scores the generated interest against the target item using sample softmax.Source: §5.3.2 · p. 11
• Without GRW, accuracy falls from 49.85% to 43.13% and reconstruction loss rises late in training; with GRW, all three losses decrease together.The authors attribute the loss incompatibility in Euclidean space to the conflict between reconstruction and recommendation objectives, which spherical space and GRW noise resolve.Source: §5.3.2 · p. 11, Figure 4 · p. 12
• Removing each individual module lowers Recall@10 on ML-10M, while HR@50 gradually increases as inference uses more denoising steps.The one-component-at-a-time results are summarized in Table 4; intermediate-embedding retrieval is examined across ML-10M and YooChoose, with the same trend on KuaiRec.Source: §5.3.2 · p. 12, Table 4 · p. 12, Figure 5 · p. 12

5.4 Hyper-parameter Sensitivity Analysis

• 20 diffusion steps yielded the best HR@20 results among the tested settings and were one to two orders of magnitude fewer than DreamRec’s settings.The authors attribute this to generating user-interest rather than item representations, which they say is easier for diffusion to learn.Source: §5.4 · p. 12, Table 5 · p. 13

5.5 Online Results

• DimeRec achieved significant business gains in average time spent and category-count diversity over effectively watched videos.Table 6 summarizes the online A/B metric gains.Source: §5.5 · p. 13, Table 6 · p. 13
• The online test used 10% traffic, about 13 million users, and a 7-day observation period; DimeRec was deployed to main-scene traffic serving hundreds of millions of active users.Source: §5.5 · p. 13

• DimeRec captures historical interests with a multi-interest model, generates a next interest through diffusion on a spherical representation space, and retrieves the closest item.The paper reports extensive offline experiments on three public datasets and online experiments on a short-video platform.Source: §6 · p. 13

• Training takes a historical sequence, target item embedding, diffusion step limit and noise schedule, and initializes GEM and DAM.Source: §A.1 · p. 16
• Within the training loop, the target embedding is mapped onto a sphere and Gaussian noise is sampled in its tangent space.Source: §A.1 · p. 16
• The pseudocode optimizes θ and ϕ with loss L via Eq. (23) using Adam.Source: §A.1 · p. 16

• The reversing procedure outputs the user’s next-interest embedding.Source: §A.2 · p. 16
• Serving takes the historical sequence, diffusion step limit and noise schedule, together with trained GEM and DAM modules.Source: §A.2 · p. 16
• At a reverse step, the procedure updates x_t to x_{t−1} using the stated mean and noise term, then applies L2 normalization.Source: §A.2 · p. 16

B Hyper-parameter Sensitivity Analysis

• λ = 0.1 achieves the best HR@50 results on all three datasets, with limited sensitivity to changes in λ.Table 7 varies λ for Lrecon while holding µ at 10.Source: B Hyper-parameter Sensitivity Analysis · p. 16, Table 7 · p. 17
• µ = 1.0 gives the best HR@50 on YooChoose and KuaiRec, while µ = 10 performs best on ML-10M.Table 8 varies µ for Lssm while holding λ at 0.1.Source: B Hyper-parameter Sensitivity Analysis · p. 17, Table 8 · p. 17
• The authors recommend choosing µ between 1 and 10 when using DimeRec.Source: B Hyper-parameter Sensitivity Analysis · p. 17