Heterogeneous graph collaborative contrastive learning for cross-platform information 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).

Cross-platform recommender systems are gaining popularity for addressing cold start and data sparsity issues. However, existing research often assumes that data from the auxiliary platform can be fully leveraged, overlooking the inherent differences in services and the sensitivity of user-item interactions across platforms. Given the large-scale, noisy cross-platform data in this scenario, accurately identifying consistency and variability in user preferences is crucial. We propose the heterogeneous graph collaborative contrastive learning algorithm, which distinguishes users’ platform-core interests from platform-specific ones while capturing both global consistencies and local variations across platforms. Heterogeneous graph collaborative contrastive learning leverages the local and global views of a heterogeneous information network to model various general interests, while using two enhanced gated recurrent unit (GRU) networks to represent and transfer dynamic interests across platforms. In addition, heterogeneous graph collaborative contrastive learning introduces a collaborative contrastive learning mechanism that integrates and contrasts cross-platform and cross-view features to reduce redundancy and noise in cross-platform data while enriching the semantic depth of interest representations. We evaluate heterogeneous graph collaborative contrastive learning on the Weibo-Zhihu datasets, which contain 18,383 matched users, 17,802 items, and 768,539 user-item interaction records. Comprehensive experiments demonstrate that heterogeneous graph collaborative contrastive learning outperforms the best baseline model by 7.35% to 11.65% in hit rate @K and 10.18% to 26.12% in mean reciprocal rank @K.