Overcoming Data Sparsity in Sequential Recommendation

264 concepts237 papers2 proposals

Sequential recommendation systems operate on a simple premise: a user's future actions can be predicted by modeling the chronological trajectory of their past behaviors. Modern architectures like Recurrent Neural Networks (RNNs) and Transformers excel at this by learning transition probabilities between items.

However, this paradigm breaks down in the face of data sparsity. In real-world applications, the vast majority of users are "tail users" with very short interaction histories, and many sessions consist of only a few clicks. When a sequence contains only two or three items, the self-attention matrix in a Transformer is nearly empty, and an RNN's hidden state cannot stabilize. The model lacks the statistical evidence required to distinguish between a user's long-term preferences, immediate goals, or random noise.

To prevent performance degradation on sparse data, recent research focuses on three core strategies: densifying the training signal via self-supervision, retrieving external context, and grounding abstract item tokens in rich semantic priors.

Densifying the Signal via Self-Supervision

Standard sequential recommendation often uses a "many-to-one" training paradigm, where an entire sequence is mapped to a single next-item target. As Siciliano et al. argue, this creates an inefficient learning process due to the sparsity of feedback. By shifting to a sequence-to-sequence paradigm, models can extract supervision signals at every step of the interaction history.

When the raw data itself cannot be expanded, contrastive learning frameworks can artificially densify the supervision signal. Sun et al. introduced this concept to sequential recommendation with CL4SRec. By applying random perturbations—such as cropping, masking, or reordering items within a user's history—the model learns to maximize the similarity between different augmented views of the same sequence. This forces the network to learn robust, structural properties of user behavior rather than memorizing exact item sequences.

Raw Sequence:    [Item A] -> [Item B] -> [Item C]
                   /                         \
Augmented View 1: [Item A] -> [Mask]   -> [Item C]  \__ Maximize Mutual
Augmented View 2: [Item A] -> [Item B]              /   Information

However, purely random data augmentations can introduce noise or create unrealistic user profiles. To address this "Goldilocks problem" of synthetic data, Dang et al. developed BASRec, which mixes item representations both within and across different user sequences to balance relevance and diversity. To bypass data manipulation entirely, Liu et al. proposed CT4Rec, which enforces consistency between different dropout masks applied to the same sparse sequence. This provides a clean, regularized supervision signal without the risk of generating low-quality synthetic sequences.

Retrieving External Context

If a local sequence is too short to provide sufficient context, models can borrow information from elsewhere. Wang framed short-session recommendation as a few-shot learning problem. The proposed INSERT network uses a retrieval mechanism to pull "prior knowledge" from highly similar sessions across the entire global dataset, effectively filling in the blanks of a brief user interaction.

This retrieval strategy can extend beyond the boundaries of the target platform. Wu et al. developed RUEL, a framework that enriches sparse user profiles by retrieving similar behavior patterns from massive, anonymous browser logs. By using contrastive learning to align these external web-scale logs with internal item sequences, the model bypasses the need for explicit, shared user IDs across platforms while successfully mitigating local data scarcity.

Similarly, cross-domain sequential recommendation transfers behavioral patterns from a data-rich source domain (e.g., movies) to a data-sparse target domain (e.g., books). Li et al. introduced RecGURU, which employs adversarial learning to map user distributions into a generalized representation space, allowing the model to transfer sequential dynamics even when there is zero overlap in user identities between the two domains.

Grounding Abstract Tokens in Semantic Priors

Traditional collaborative filtering and sequential models represent items as abstract, orthogonal IDs. For long-tail or cold-start items with few interactions, these ID embeddings cannot be trained effectively.

To overcome this, models can ground item representations in multimodal or textual priors. Song et al. developed IFCDSR, which fuses traditional item IDs with visual features extracted via a frozen CLIP model. When a user interacts with a new or rare item, the model can rely on the shared "visual vibe" of the product category to maintain accurate recommendations, bypassing the lack of historical interaction data.

Taking this concept to its logical conclusion, Guo et al. discarded item IDs entirely in their TCPLP framework. By representing items solely through universal natural language descriptions and using a co-attention prompt learning strategy, the authors enabled zero-shot knowledge transfer across completely non-overlapping domains.

Finally, structural relationships can substitute for missing behavioral data. Di et al. introduced the TRR network, which uses explicit semantic relationships—such as whether items are "complements" or "substitutes"—to regulate how a target item matches against a sparse purchase history. By treating the target item and its historical relatives as a unified query, the model filters out noise and infers user intent even when direct interaction evidence is missing.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_data_sparsity_data_sparsity_in_sequentia_792ab9fd,
  title = {data sparsity (data sparsity in sequential recommendation)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-sparsity-data-sparsity-in-sequential-recommendation/rdr_792ab9fddc484161bf15b7ce7bec97b6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_data_sparsity_data_sparsity_in_sequentia_792ab9fd_v1,
  title = {data sparsity (data sparsity in sequential recommendation)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-sparsity-data-sparsity-in-sequential-recommendation/rdr_792ab9fddc484161bf15b7ce7bec97b6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake