GENERATIVE · 2026-11-03

Generative — Paper Daily, 2026-11-03

The generative track has only 2 papers today. Neither studies a recommender itself, but both use a familiar move from generative recommendation: compress objects into discrete codes, much like SIDs, and then generate or predict over those codes. LatentDiff compresses a whole knowledge graph into a set of vectors and uses a diffusion model to go step by step from the incomplete graph to a completed one, recovering relations that subgraph-based methods cannot reach. Ant International's ID2Pattern ignores the card and device ID values and records only how they show up in a transaction sequence, then groups that record into a few kinds of discrete behavior patterns, so it works just as well on IDs it has never seen. Teams that care about cold start for new IDs can start with ID2Pattern.

2.85

1. From Local Recovery to Unified Generation: Discovering Triple Set in Knowledge Graphs via Observation-Anchored Latent Diffusion

Missing relations in a knowledge graph can be filled in without first cutting the graph into small subgraphs and completing each one separately. If the two ends of a missing relation land in different subgraphs, no subgraph can recover it, and on the Wiki79k dataset over a fifth of the relations to be recovered are like this. LatentDiff uses an autoencoder to compress the whole graph into a set of vectors: one vector per entity per relation, from which the entities it links to through that relation must be recoverable. During training, the vectors of the complete graph are pushed step by step toward those of the incomplete graph, with a little noise added at each step. The diffusion model learns to walk the other way, from the incomplete graph back to the complete one. As a result, some cross-subgraph relations are recovered: Wiki79k has 3,391 of them, the subgraph methods find none, and LatentDiff finds 178.

2.60

2. ID2Pattern: Pretraining Generalizable Identity Behavior Patterns for Transaction Risk Foundation Models

This Ant International paper lets payment risk models use card, device and account IDs. It ignores the ID value and records only how each ID shows up in the transaction sequence: in what order it first appeared, whether this is its first appearance, how many times it has appeared, and how many transactions ago it last appeared. The values themselves mean nothing, and new cards and devices appear every day, so learning one embedding per ID works poorly. The behavior, though, carries over: when someone who always pays with the same card and device suddenly switches to a new device, that switch is a risk signal whatever the device ID is. On payment risk data with more than 100 million sequences, AUC is 4.13% higher in relative terms than the production model that leaves these fields out. Replacing every test ID with a value never seen in training leaves the results unchanged, because the model never stored the IDs themselves.