G2Rec: Graph interest tokens for generative recommendation¶

简体中文

Fidelity: complete core pipeline. Optimizes sparse soft modularity and trains the interleaved autoregressive Transformer with both losses.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	Meta
First public release	2026-06-18 (arXiv v1)
Author code	No official author code identified (checked 2026-08-09).
Local adapter or method key	`g2rec`
Local reproduction code	`src/auto_research/reproductions/g2rec/`

Paper method and evidence¶

Item-only sequences cannot express overlapping interest clusters or co-engagement from other users. G2Rec builds a sparse item graph, learns soft membership in multiple interest prototypes through differentiable modularity, and alternates item and continuous interest-profile tokens in the decoder. An auxiliary profile loss complements next-item prediction.

flowchart LR
  A[All user histories] --> B[Sparse co-engagement graph]
  B[Sparse co-engagement graph] --> C[Soft cluster profiles]
  C[Soft cluster profiles] --> D[Interleave item and interest tokens]
  D[Interleave item and interest tokens] --> E[Autoregressive decoder and profile loss]

Original paper figure¶

Figure 2 shows the method and its main components. From the original paper; copyright remains with the authors.

Core equation¶

\[ Q_{soft}(P)=\frac1{|E|}\sum_{(i,j)\in E}p_i^Tp_j-\gamma\frac{\|P^Tk\|_2^2}{|E|^2}. \]

\[ v_a=\frac{\sum_i p_{i,a}x_i}{\sum_i p_{i,a}},\qquad y_i=\sum_a p_{i,a}v_a, \]

\[ \mathcal L^t=\underbrace{-\log F(i_{t+1}\mid R_{u,\le2t+1})}_{\mathcal L_{item}^t} +\lambda\underbrace{\left[-\sum_a p_{i_t,a}\log F(a\mid R_{u,\le2t})\right]}_{\mathcal L_{profile}^t}. \]

\(P\in\mathbb R^{|I|\times C}\) stores soft memberships, \(k\) is the degree vector and \(|E|\) counts edges with repeated directions. The decoder receives \(R_u=[BOS,x_{i_1},y_{i_1},\ldots,x_{i_N},y_{i_N}]\).

Reported results¶

On Beauty, Sports, Toys and Yelp, the authors use 5-core leave-one-out data, 99 negatives, Llama 2 13B LoRA and three epochs. G2Rec leads on all six metrics in all four datasets.

Dataset	Best baseline NDCG@5	G2Rec NDCG@5	G2Rec NDCG@10	MRR
Beauty	0.2848	0.3035	0.3334	0.3034
Sports	0.2497	0.2869	0.3254	0.2867
Toys	0.2820	0.2931	0.3225	0.2942
Yelp	0.4252	0.4398	0.4927	0.4180

Soft modularity vs Leiden is 0.419→0.499, 0.365→0.452, 0.437→0.550 and 0.691→0.757, respectively. Added train/inference time is 0.043s/0.0027s per batch. Meta's seven-day and longer online A/B reports in-session lift >+0.03% and engagement (time spent, likes and shares) +0.06% to +0.19%.

Local reproduction¶

The runner downloads Amazon Beauty 5-core (22,363 users, 12,101 items) and builds 265,133 sparse edges using a three-step window. It optimizes modularity across 12 clusters. Matched item-only and interleaved decoders use 96 dimensions, two causal Transformer layers and 240 steps, adding a 0.1 profile-loss weight for G2Rec. Test fixes 1,000 users with one positive and 99 random negatives each.

Variant	Hit@10	NDCG@10	Head share@10
Item-only decoder	0.1100	0.0474	0.1107
Item + interest decoder	0.1130	0.0530	0.1177

Soft modularity reaches 0.3286; final losses are 9.8684 vs 11.2813. NDCG improves 11.92%, with +0.3 percentage points on Hit@10 but +0.7 on head share. This verifies the joint tokenization/loss pipeline, not a DIN comparison, and possible popularity bias and the small decoder preclude inferring production lift. Stable metrics. Downloaded data and raw runs remain ignored; only code, tests, documentation and sanitized metrics are committed.

pip install -e '.[neural-recs]'
auto-research reproduce --paper g2rec --dataset-dir data --seed 42