Semantic-ID-enhanced paths for knowledge graph-based generative recommendation

Recommender systems suggest to a user the items in a catalog, such as movies, songs, or products, that the user is most likely to be interested in. They commonly learn this from the items the user has engaged with in the past, although recommendations can also come from sources that are not personalized at all, such as the most popular items of a catalog or the items most recently added to it. A more recent line of research casts personalized recommendation as a generative task: instead of scoring every item in the catalog, the model generates the identifier of the next item directly, token by token, much as a large language model generates text. Going beyond the raw interaction data, recent generative recommender systems enrich this task with additional information about the items, and there two complementary directions have emerged. The first improves how the items are expressed: Semantic IDs (SIDs) replace each item's arbitrary identifier with a short sequence of hierarchical, discrete codes derived from metadata and content, for example from the text describing an item or from a audio embeddings of a music track, so that similar items share parts of their representation. The second improves the signal which the model reasons about: Knowledge Graph (KG) path reasoning generates explicit paths through a graph of items, attributes, and users, for example linking two movies that share a director, or two songs sharing a common genre, so that each recommendation comes with a chain of relationships that produced it. These two directions have complementary limitations: SID-based models describe what an item is but not how it relates to others, while KG-based generative recommenders capture these relationships but still give every item its own arbitrary identifier, which enlarges the model and prevents it from sharing information across similar items. In this thesis we investigate whether the two can be combined, so that a single generative recommender benefits from the relational grounding of KGs and the compact, content-driven item representation of SIDs. As a concrete realization of this idea, we propose SPRIG, which expresses the items inside KG paths as SIDs. We evaluate this model on movie, music and product review datasets against sequential, KG-augmented, and SID-based baselines, and show that it matches or improves prior generative recommenders while using fewer parameters and less computing power.