Modern recommendation systems often suffer from "tunnel vision." When a user has only clicked on two or three items, or when a specific product has rarely been viewed, the system lacks the context necessary to make a meaningful prediction. This is the sparsity problem, and it is particularly acute in session-based recommendation (SBR), where models must predict the next item based on a short, anonymous sequence of interactions.
To overcome this, researchers have moved beyond treating individual sessions or user histories as isolated islands. Instead, they construct a global graph—a massive map of item-to-item transitions and user-item relationships aggregated from the entire community. By transferring knowledge from this global structure back to local contexts, systems can "borrow" the wisdom of the crowd to fill in the blanks for sparse data.
The Tunnel Vision of Local Sequences
Traditional sequential models, such as those using Recurrent Neural Networks (RNNs) or Transformers, focus heavily on the temporal order of a single user's actions. While effective for long histories, these models struggle with "short sessions" (often five items or fewer) where internal preference patterns haven't yet emerged. Shoujin Wang et al. (Next-item Recommendations in Short Sessions) frame this as a few-shot learning problem, noting that local data alone is often insufficient.
Furthermore, Shu Wu et al. (Session-Based Recommendation with Graph Neural Networks.) point out that modeling sessions as isolated graphs fails to capture universal item-transition trends that persist across the entire dataset. Without a global perspective, a system cannot distinguish between a genuine user intent and a noisy, accidental click.
Building the Global Map
The core idea is to aggregate all observed interactions into a unified structure. This can take several forms:
- Transition Graphs: Andre Busche et al. (Mining Music Playlogs for Next Song Recommendations) demonstrated that simple first-order Markov chains trained on massive collaborative "playlogs" could outperform complex baselines by capturing how the wider community moves between items.
- Bipartite and Heterogeneous Graphs: Daniel Dajun Zeng et al. (A Link Analysis Approach to Recommendation under Sparse Data) applied link analysis (like PageRank) to global bipartite graphs to discover transitive relationships—hidden similarities between users who have no direct overlap in history but share common neighbors.
- Multi-Scenario Fusion: In large-scale production, Fan Zhang et al. (Multi-Graph based Multi-Scenario Recommendation in Large-scale Online Video Services) showed that integrating data from different apps (e.g., a browser and a video feed) into one graph allows a system to borrow interest transitions from one context to alleviate exposure bias in another.
Mechanisms of Knowledge Transfer
Once a global graph is built, the challenge is transferring that macro-level knowledge to a micro-level prediction.
1. Message Passing and Aggregation Graph Convolutional Networks (GCNs) are frequently used to facilitate this transfer. Xinlin Xia et al. (SearchGCN: Powering Embedding Retrieval by Graph Convolution Networks for E-Commerce Search) use GCNs to aggregate information from high-frequency "head" nodes to enrich the representations of "long-tail" (rare) items. Similarly, Ruihong Qiu et al. (Exploiting Cross-Session Information for Session-based Recommendation with Graph Neural Networks) sample multi-hop neighbors from a global graph to provide collaborative context for anonymous sessions.
2. Meta-Path Expansion In extremely sparse environments, such as virtual gift-sending in live streams, Jiaxin Deng et al. (MMBee: Live Streaming Gift-Sending Recommendations via Multi-Modal Fusion and Behaviour Expansion) use metapaths to traverse heterogeneous graphs. This expands a user's limited direct actions by incorporating the behaviors of structurally similar neighbors. Zhiqiang Wang 0005 (Counterfactual Task-augmented Meta-learning for Cold-start Sequential Recommendation) further leverages these global structural relationships to uncover latent preferences that are invisible in local sequential data.
3. Global Regularization and Denoising Global patterns also serve as a "sanity check" for local noise. Yongqiang Han et al. (GUESR: A Global Unsupervised Data-Enhancement with Bucket-Cluster Sampling for Sequential Recommendation) use global transition patterns as a regularizer to help models distinguish between genuine intent and accidental interactions. Chi Zhang et al. (SSDRec: Self-Augmented Sequence Denoising for Sequential Recommendation) take this further by "self-augmenting" sparse sequences with global knowledge before denoising, ensuring that the model doesn't accidentally remove useful but rare signals.
Intent Discovery via Global Prototypes
Beyond simple item transitions, global graphs can reveal high-level user motivations. Ding Zou et al. (Knowledge Enhanced Multi-intent Transformer Network for Recommendation) use a Graph Transformer to refine "intent prototypes" across the entire interaction graph. This allows the system to align a specific user's embedding with a global behavioral pattern, such as "shopping for a home renovation" or "looking for upbeat music," even if the user's current session is too brief to explicitly signal that intent.
By anchoring individual predictions in these collective patterns, global-to-local transfer transforms recommendation from a narrow search for sequence patterns into a robust analysis of a shared behavioral ecosystem.
Key Foundational Work
A Link Analysis Approach to Recommendation under Sparse Data (2004) — contributed 1 concept elements to this direction.
Mining Music Playlogs for Next Song Recommendations (2010) — contributed 1 concept elements to this direction.