DyGen predicts dynamic text-graph links by generating semantic IDs, giving history-free source nodes a ranking signal
Synopsis
Link prediction on dynamic text-attributed graphs has long relied on scoring candidate targets one by one, and structural evidence nearly vanishes when a source node has no interaction history. DyGen compresses node text into short semantic identifiers (SIDs) via residual quantization, then an encoder-decoder Transformer autoregressively generates the target SID conditioned on the source SID, recent interaction targets and relative time gaps, reusing one encoded query context to rank candidates at inference. Across nine datasets performance is broadly competitive; on Amazon Movies DyGen beats TPNet on Hits@1, Hits@5 and MRR, while Hits@10 is close to TPNet and Hits@20 is lower.
Interpretation
The prediction object shifts from an independent score per candidate target to token-by-token generation of a target's text-derived semantic identifier, and the same generator produces a query representation even when the source node has no interaction history. Earlier dynamic graph methods mostly follow a discriminative paradigm that scores sampled or enumerated candidate targets directly, and such models struggle when structural evidence is weak or absent for cold sources. On Wikirevision, IMDB and Weibo-Daily, DyGFormer's Hits@10 for sources with no training or validation interactions is only 54%, 17% and 2% of that for sources with more than ten prior interactions.
Beyond the source SID, the generator reads the identities of recent interaction targets together with their relative time gaps, and when history is empty the input reduces to the source SID with no synthetic neighbor or imputed interaction. History is no longer compressed into only its length or a pooled activity vector; the identity, order and relative time of interacted targets are retained. On Amazon Movies, randomizing historical target identities while preserving their count, order, positions and timestamps lowers Hits@10 from 0.1085 to 0.0695 and MRR from 0.0635 to 0.0387, significant under paired bootstrap.
At inference the query context is encoded once, candidates are ranked by SID likelihood, and three structural signals for repetition, co-occurrence and popularity are added as calibration. Structural calibration changes neither the generator nor its training objective; the three weights are selected on validation data, and setting all to zero recovers generation-only prediction. Relative to DyGFormer, DyGen cuts per-epoch training time by about 59% on Wikirevision and 51% on Sephora, peak GPU memory by about 76% and 75%, and per-query inference time by about 80% at the evaluated candidate count.
Perspective
Evaluation covers nine datasets, and the main experiments are sampled ranking of 2000 test queries against 4999 uniformly sampled negatives rather than full-catalog retrieval; Amazon Movies uses a fixed manifest of 2000 queries and 5000 target-only candidates. New-item results are limited to ranking 186 new items within the cold-item universe on Amazon Movies, and Hits@10 is 0 under a manifest mixing warm and cold candidates.
The authors treat mixed-catalog calibration as an unresolved deployment problem, and the Hits@10 of 0 with MRR of 0.00024 indicates a score-calibration gap between targets with abundant interaction evidence and first-arrival targets. A finite suffix vocabulary can leave residual full-SID collisions in extreme collision buckets, and full-SID distinctness stays slightly below 100% on Dianping, Sephora and Wikilife. The authors also note the absence of a continuous-target retrieval control, so the results support the usefulness of the SID interface rather than proving discretization is universally superior to continuous representations.
