Recommender systems frequently struggle with two fundamental bottlenecks: data sparsity (the long-tail problem where most items and users have few interactions) and interaction noise (accidental clicks or shifting contexts that pollute user histories). Traditional collaborative filtering models rely on flat, single-vector ID embeddings and raw interaction logs, making them highly sensitive to these issues.
To achieve superior recommendation performance, a new paradigm has emerged. This direction focuses on enriching the geometry of representation spaces, explicitly denoising interaction signals, and generating robust self-supervision to bridge data gaps.
Expanding the Geometry of Representation Spaces
Standard embedding techniques compress a user's multi-faceted interests or an item's complex attributes into a single static vector. When data is sparse, these vectors easily overfit or fail to capture the nuances of user behavior. Modern approaches solve this by expanding the mathematical structure of the embedding space.
Instead of a uniform vector space across an entire user-item graph, Purificato et al. introduced Sheaf4Rec, which leverages Sheaf Neural Networks. Sheaves attach localized vector spaces to nodes and restriction maps to edges. This allows the model to capture asymmetric, contextual relationships along graph edges that a flat dot-product would smooth away, preserving distinct facets of user-item interactions.
Similarly, treating embeddings as fixed points ignores the inherent uncertainty of user preferences. In the RVRec framework, Mi et al. model embeddings as probability distributions rather than deterministic vectors. By representing preferences as distributions, the framework naturally encodes interaction uncertainty and employs game-theoretic value interactions to enhance both explainability and accuracy.
Standard Embedding: [ User Vector ] ───(Dot Product)───> [ Item Vector ] (No Uncertainty)
Probabilistic Embedding: [ User Distribution ] ───(Overlap)───> [ Item Distribution ] (Encodes Uncertainty)
When rich multimodal data is available, large language models (LLMs) can be leveraged to unify these spaces. Lei et al. developed UniRec, which uses a schema-aware triplet representation and a hierarchical Q-Former to map text, images, and categories into a unified space, allowing LLMs to capture deeply nested interaction structures. For dynamic environments like online forums, Srivastava shows in SITRec that modeling evolving user interests as continuous trajectories using coupled Recurrent Neural Networks prevents the representation from degrading when interaction data is sparse.
Denoising and Aligning Behavioral Signals
Raw interaction sequences are inherently noisy. A user might click an item by mistake, out of brief curiosity, or due to aggressive algorithmic promotion. Standard sequential models mistake these actions for core preferences, leading to poor recommendations.
To isolate true user intent, Sheng et al. proposed EICD-Rec, which performs denoising in the frequency domain. By transforming behavior sequences into frequency components, the model filters out high-frequency noise (like sporadic, accidental clicks) and retains low-frequency signals that represent stable, category-based user intent.
[Raw Sequence] ───(Fourier Transform)───> [Frequency Domain] ───(Filter High-Freq Noise)───> [True Intent Signal]
Noise also complicates multi-behavior recommendation, where systems try to predict rare target actions (like purchases) using abundant auxiliary actions (like clicks). Xin et al. addressed this challenge with the Multi-Behavior Alignment (MBA) framework. MBA aligns user preferences across different behavior types while explicitly denoising the implicit feedback. This ensures that dense but noisy auxiliary behaviors reinforce, rather than mislead, the target prediction model.
Synthesizing Self-Supervision for Sparse Data
When interaction data is extremely sparse—such as for new users or long-tail items—models lack sufficient signal to learn robust representations. Self-supervised learning (SSL) solves this by creating alternative "views" of the data to train the model via contrastive objectives. However, standard data augmentation (like randomly masking items) can create unrealistic sequences that confuse the model.
To combat this, Lee et al. introduced SCONE, which uses score-based generative models to synthesize personalized contrastive views and hard negative samples. By generating high-quality, realistic negatives rather than relying on random sampling, the model learns sharper decision boundaries for low-activity users.
Graph structures can also be leveraged for global self-supervision. Xia et al. developed the Self-Supervised Hypergraph Transformer (SHT), which uses hypergraphs to model non-pairwise, global collaborative relationships. The self-supervised component aligns these global hypergraph views with local graph views, providing a powerful regularizer that protects the model from data sparsity and localized noise.
Generalizing to Open-World and Cross-Domain Scenarios
The ultimate test of robustness is the cold-start or "open-world" setting, where the model encounters entirely unseen users or domains. Traditional collaborative filtering is transductive, meaning it requires a complete model retrain to accommodate new users.
To bypass this limitation, Wu et al. proposed Inductive Collaborative Filtering (IDCF). IDCF learns "meta latents" from a core set of existing users and applies an attention-based relation model to map new users into the latent space on-the-fly based on their initial interactions.
Transductive: [New User Appears] ───> Requires Full Model Retraining
Inductive: [New User Appears] ───(Attention + Meta Latents)───> Immediate Latent Embedding
When transferring knowledge across different domains, models often suffer from "negative transfer," where noisy or irrelevant data from a source domain degrades performance in the target domain. In the SyNCRec framework, Park et al. introduced a cooperative "pacer and runner" mechanism. Single-domain models act as pacers to guide cross-domain runner models, adaptively weighting gradients to suppress cross-domain noise while maximizing the transfer of clean, relevant preference signals.