Cold-Start Problem in Recommender Systems

• The cold-start problem is a challenge where new users or items lack historical interaction data, leading to sparse collaborative embeddings and reduced recommendation quality.
• Hybrid and probabilistic approaches, including content-based models and variational frameworks, mitigate cold-start issues by leveraging side information and regularizing latent representations.
• Graph neural networks and submodular optimization methods offer efficient strategies to model cold entities, enhancing diversity and user engagement across large-scale systems.

The cold-start problem encompasses a family of foundational challenges in recommender systems and related machine learning domains, characterized by sharply reduced predictive accuracy when new users, items, or contexts enter the system without sufficient historical interaction data to reliably estimate their latent characteristics or affinity to existing system entities. Cold-start phenomena impede collaborative filtering and are known to trigger degraded recommendation quality, reduced content diversity, delayed exposure for new content, and suboptimal user engagement across industrial-scale platforms, e-commerce, advertising, time series, serverless computing, and group or context-driven systems.

1. Formal Definitions and Taxonomy

Cold-start arises in any system that leverages historical interaction data to infer user or item latent features. For user $u$ and item $i$, the classical matrix completion problem operates on a sparse observation matrix $R\in \mathbb{R}^{|\mathcal U|\times |\mathcal I|}$, where $R_{u,i}$ is known only if $u$ has interacted with $i$. A cold-start user or item has $\forall j, R_{u,j}=0$ or $\forall j, R_{j,i}=0$ respectively, yielding undefined collaborative embeddings in traditional latent factor, GNN, or nearest-neighbor frameworks (Qian et al., 2019). The cold-start problem further generalizes to contextual, time-varying, or group settings, including:

• User cold-start: Predicting recommendations for new users with zero or few item interactions (Biswas et al., 2017).
• Item cold-start: Recommending newly arrived items with no feedback (Yuan et al., 2016, Wang et al., 12 Jul 2025).
• Continuous cold-start (CoCoS): Users or items that repeatedly re-enter cold-state due to infrequency, volatility, multiple personas, or identity fragmentation (Bernardi et al., 2015).
• Extreme cold-start: Group or bundle recommendation scenarios where group–item interactions are so sparse as to preclude direct learning (linxin et al., 2022).
• Pure cold-start: Cold users/items with neither interaction data nor side information (Meng et al., 2020).

In time series and serverless/FaaS systems, "cold-start" refers to lack of historical data for forecasting or initialization delays when no prior container is available, respectively (Mouen et al., 15 Apr 2025, Fatemi et al., 2023).

2. Classical and Contemporary Solution Strategies

2.1 Hybrid and Content-Based Methods

Classical methods address cold-start via attribute-based models (demographics, text, images), mapping side information into the latent space. For items, deep document embedding (doc2vec) models can pair each new item with a similar, well-connected existing item, allowing recommendations to "ride" on warm items' collaborative footprints without retraining the collaborative filtering engine (Yuan et al., 2016). Attribute Graph Neural Networks construct attribute graphs (rather than interaction graphs) for users/items, allowing embedding inference via extended variational autoencoding on side features, and aggregating multi-modal attributes via gated message passing (Qian et al., 2019).

2.2 Causal, Probabilistic, and Variational Approaches

Modern systems increasingly adopt probabilistic representations, where each user/item embedding is parameterized as a distribution, e.g. a Gaussian in a variational framework. The Variational Embedding Learning Framework (VELF) regularizes cold ID representations by (a) learning posterior distributions with data-dependent variance, and (b) incorporating trainable priors informed by side information (e.g., campaign, demographics), further regularized by hyper-prior penalties (Xu et al., 2022). Bayesian exploration controls, leveraging Beta–Bernoulli conjugacy, support filtering and ranking cold-start items by posterior lower confidence bounds during the exploration phase (Wang et al., 12 Jul 2025). In time-series, cold-start forecasting is mitigated by (i) extracting sparse causal graphs (e.g., via VARLiNGAM), directing GNN–LSTM pipelines to model dependencies robustly, and (ii) transferring knowledge from similar entities via clustering (Fatemi et al., 2023).

2.3 Graph Neural Networks and Patching Architectures

Canonical GNNs cannot represent cold-start nodes (degree zero in the interaction graph). GPatch and Graph Neural Patching (GNP) frameworks separate modeling into two branches: (a) a highly efficient GWarmer (random-walk, self-adaptive pooling on warm nodes; pre-computable), and (b) lightweight patching networks (MLPs) mapping auxiliary features to the collaborative space for cold entities (Chen et al., 2024, Chen et al., 2022). Training involves simulating cold entities via random dropout among warm nodes, ensuring patching networks mimic the collaborative embeddings distributionally, without degrading warm-node performance. This separation of concerns allows both faithful collaborative modeling of the warm core and high-quality recommendations for cold-start cases.

3. Algorithmic and Theoretical Foundations

3.1 Submodular Optimization for Cold-Start Item Selection

In pure cold-start settings—especially new user cold-start without any feedback or side information—the challenge can be reformulated as a set selection problem to maximize coverage/regret reduction. Meng et al. cast the item selection task as a monotone submodular maximization—finding a subset of $M$ items whose maximum affinity over all plausible user vectors is as close as possible to their true favorite—yielding a greedy algorithm with $1-1/e$ approximation guarantees (Meng et al., 2020). Relatedly, Biswas et al. formalize the optimal cold-start interview in model-based CF as minimization of the trace of the Fisher information matrix inverse, showing this to be NP-hard and not submodular, leading to fast (accelerated) greedy heuristics via rank-one updates (Biswas et al., 2017).

3.2 Provable Preference Transfer

The Pretender algorithm empowers end users to actively solve cold-start by optimal selection of $i$0 items to label in the target system, minimizing a divergence (e.g., MMD or Wasserstein) between their old and new preference distributions (Sato, 18 Feb 2025). Pretender adopts a quadrature optimization—solving a convex program to choose sample weights, followed by randomized rounding—yielding finite-sample bounds on the discrepancy, with theoretical error decaying as $i$1 with $i$2 (MMD) and matching lower bounds. This approach gives rigorous guarantees that the "user-side" preference transfer can approximate a fully warmed recommender experience within explicit statistical limits.

4. Extensions: Diversity, Popularity Bias, and Large-Scale Validation

Cold-start solutions increasingly address systemic biases—particularly suppression of long-tail content and popularity bias in exploration. Item-centric exploration policies filter user–item pairs so that new items are only exposed to users for whom the model is sufficiently confident in the item's intrinsic quality, improving new-item success rates, recommendable corpus size, and content diversity, while reducing exploration waste (Wang et al., 12 Jul 2025). Prompt-tuning methods for item cold-start leverage "pinnacle feedback" (high-value interaction traces) as soft prompts, yielding substantial gains over baseline content-based or text-based prompts in collaborative scenarios, and mitigating positive-feedback dominance effects that bias recommenders toward popular/warm items (Jiang et al., 2024). Industrial deployments on billion-user platforms substantiate these findings in actual click, dwell, and corpus metrics (Wang et al., 12 Jul 2025, Jiang et al., 2024).

5. Cold-Start Beyond Classical Recommendation: Groups, Sessions, Time, and Serverless

Cold-start phenomena appear in multiple modalities:

• Group and bundle recommendation: EXTRE introduces a fixed-point GCN formalism to propagate user–item–group affinities via infinite-layer meta-path reasoning, deriving interpretable "consistency" and "discrepancy" scores for contrastive training even when group–item edges are fully missing. This enables near-baseline Recall/NDCG@10 in extreme cold-start regimes (linxin et al., 2022).
• Session and context volatility: Continuous Cold-Start (CoCoS) is characterized by users/items becoming "cold" repeatedly due to infrequency, volatility, or multi-persona drift. Models incorporating temporal factorization, contextual bandits, and online intent prediction are necessary to track and adapt to dynamic latent profiles (Bernardi et al., 2015).
• Time-series and serverless FaaS: Cold-start in demand forecasting for a new time series leverages causal GNNs guided by VARLiNGAM and inter-entity model borrowing (Fatemi et al., 2023). In cloud computing, cold-start latency is mitigated by Transformer-based forecasting of invocation patterns to dynamically pre-warm containers and control idle windows, reducing both frequency and duration of cold-starts by over 50–70% in production-scale traces (Mouen et al., 15 Apr 2025).

6. Minimal Activity, User Cooperation, and No-Side-Info Scenarios

Višňovský et al. empirically derive the minimal number of ratings a cold-start user must supply for reliable cluster assignment: in MovieLens, 20–50 ratings generally suffice; for Jester, 68–80 are needed. Cluster-quality metrics (Davies–Bouldin index) and assignment success rates provide operational criteria for resolving cold-start (Visnovsky et al., 2021). In completely data-free scenarios, Zipfian matrix factorization (ZeroMat, DotMat) infers plausible user/item embedding distributions by leveraging power-law assumptions, producing surprisingly competitive results—especially when hybridized as pre-filling for standard matrix factorization (Wang, 2022). Visual preference modeling solves new-user cold-start in image-rich domains by inferring individual feature weights for clustering and retrieval based only on a handful of user-supplied images (Verma et al., 2020).

7. Future Directions and Open Challenges

Emerging research targets more robust and versatile cold-start solutions:

• Meta-learning and mixture-of-expert models (e.g., Cold & Warm Net) can separate cold and warm profiles, combine them via gate networks, and activate dynamic knowledge distillation for transfer (Zhang et al., 2023).
• Continuous, session-level, and context-aware adaptation is needed to address CoCoS and extreme volatility (Bernardi et al., 2015).
• Preference transfer and user-initiated cold-start bootstrapping (Pretender) allow individuals to improve their recommendations without provider-side cooperation, under explicit error bounds (Sato, 18 Feb 2025).
• Cold-start in group and cross-domain settings (EXTRE), as well as in time-series and cloud resource orchestration (linxin et al., 2022, Mouen et al., 15 Apr 2025), motivates further development of universal, scalable, and interpretable cold-start paradigms.
• Contextual bandits, session-based recommendation, and dynamic factorization are advancing to address multi-persona and volatile-user/item states (Bernardi et al., 2015).

Persistent challenges include achieving high-quality cold-start without heavy side information, maintaining warm-user/item performance, rigorous characterization of minimal interview complexity, and extending provable solutions to highly distributed, adversarial, or privacy-constrained environments. The cold-start problem thus persists as a central, evolving barrier at the intersection of large-scale machine learning, user modeling, and adaptive decision systems.