From Sentences to Structure: Co-Purchase Graphs for Next-Purchase Recommendation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Customers often buy several products in one order. Our earlier study described such co-purchases to a language model, one sentence per order, and improved next-purchase prediction. A sentence, however, cannot express that many customers buy two products together. We therefore represent co-purchase as structure: an item–item co-purchase graph, summarised by LightGCN, serves as an item representation alongside a frozen text embedding, and a Transformer that reads each order as a set scores all items at once. With only the item representation varied, the combination outperforms either alone on the Korean platform Katcher’s and on Instacart and matches or exceeds learned ID embeddings. On Katcher’s, among items new to the user, the graph helps those co-purchased with an item in the user’s history (+0.0216 NDCG@10 under sampled-candidate ranking), most when that item is in the latest order (+0.0288), and not those without such a link. Under the same ranking, degree-preserving rewiring retains less than half of this gain on both platforms: the gain depends on which items are bought together, not only on how many co-purchase partners an item has. Co-purchase structure thus complements item content where co-purchases link a user’s purchases to items the user has not bought.