Cold-Start Relevance Matching

• Cold-Start Relevance Matching (CSRM) is a framework that infers ranking relevance in the absence of historical interactions by using side information, metadata, and transferred behavioral signals.
• CSRM methods typically decouple the relevance model from the evidence-completion module, employing strategies such as representation completion, content-to-behavior transfer, and uncertainty-aware updating.
• Empirical studies demonstrate that CSRM techniques improve key metrics like click-through rates and engagement by reconstructing latent behavior features for cold-start scenarios.

Cold-Start Relevance Matching (CSRM) denotes ranking or matching settings in which the strongest relevance signals—historical interactions, mature labels, or behavior-heavy features—are missing for one side of the match, so relevance must be inferred from side information, content, metadata, or transferred behavioral structure (Wang et al., 1 Sep 2025, Feng et al., 2021). In the literature summarized here, CSRM spans user cold-start recommendation, item cold-start recommendation, system cold-start recommendation, expert routing for new questions, and emerging-market search relevance. The shared technical problem is not merely sparsity in the abstract, but missing evidence at the point where a scoring function must still assign a usable relevance estimate.

1. Scope and formalization

A recurring formulation treats CSRM as scoring a pair under missing evidence. In new-user recommendation, old users are represented with full inputs

$$\mathbf{x}_s=\{\mathbf{a}_s,\mathbf{v}_s,\mathbf{c}_s,\mathbf{t}_s\},$$

whereas new users are represented only by

$$\mathbf{x}_o=\{\mathbf{a}_o,\mathbf{c}_o,\mathbf{t}_o\},$$

because the behavior term is absent; the task is then to estimate a ranking score $\hat y$ for CTR/CVR/CTCVR-style prediction (Feng et al., 2021). In emerging-market search, the target is a query-product relevance classifier

$\mathcal{P}(y \mid i, q; \Theta),$

with $y$ taking the business-defined values exact match, substitute, or irrelevant (Wang et al., 1 Sep 2025). In system cold-start recommendation, the defining assumption is even stronger: the target domain has an empty interaction matrix,

$\mathbf{R}_{\mathrm{tgt}}=\emptyset,$

so neither training nor inference can rely on target-domain interactions (Wu et al., 2023).

This range of settings makes clear that CSRM is broader than item cold start alone. Some work is explicitly query-side or user-side, as in new-user recommendation and cold-question routing; some is candidate-side or item-side, as in cold song, movie, or product matching; and some treats the full deployment context as cold, as in system cold-start recommendation and new-market search (Sun et al., 2018, Zhao et al., 2023, Wang et al., 1 Sep 2025). A common implication is that CSRM methods usually separate a relevance model from an evidence-completion, transfer, or alignment mechanism.

A second shared feature is that CSRM rarely reduces to generic semantic similarity. Several papers state, in different ways, that content semantics and behavioral relevance are not identical: MAIL reconstructs latent user-interest regions rather than literal click sequences (Feng et al., 2021); SaviorRec centers the mismatch between semantic space and user behavior space (Yao et al., 2 Aug 2025); ColdRoute shows that tags can be more informative than semantic text matching for cold questions (Sun et al., 2018); and CSRM-LLM argues that emerging-market search is not solved by generic zero-shot multilingual transfer alone because label semantics are domain-specific (Wang et al., 1 Sep 2025).

2. Canonical architectures and recurring design patterns

Across the literature, four architectural patterns recur. First, some methods perform representation completion: they infer a missing user-side or query-side latent state from attributes or a few interactions. Second, some perform content-to-behavior transfer: they learn item representations from content but shape them to approximate collaborative relevance. Third, some introduce uncertainty-aware online updating, so cold-start relevance is treated as posterior inference rather than static estimation. Fourth, some use prompted or language-model-based scoring in place of conventional recommender parameterizations.

Representative system	Cold-start side	Core mechanism
MAIL (Feng et al., 2021)	New user	Cross-modal reconstruction from attributes to virtual behavior
Music Cold-Start Matching (Zhao et al., 2023)	New item	Content embedding, warm-item retrieval, audience transfer
BayesCNS (Ardywibowo et al., 2024)	New or drifting item	Bayesian priors over interaction features with Thompson sampling
ColdRoute (Sun et al., 2018)	New question	Factorization Machines over sparse tags and IDs
CSRM-LLM (Wang et al., 1 Sep 2025)	New market search	Multilingual LLM with MT, RQA, and self-distillation

This suggests that CSRM is best understood as a family of auxiliary relevance-induction strategies rather than a single model class. Some systems generate a missing latent behavior vector; some propagate preference through similarity graphs or prototype banks; some discretize semantic content into trainable IDs; and some verbalize the pairwise task and let a LLM score a label token. What remains stable is the requirement that the final score still behave as a relevance score usable for ranking.

3. Query-side and user-side CSRM

MAIL is a canonical user-side CSRM architecture. It frames new-user cold start as a zero-shot problem and introduces a two-tower design: a zero-shot tower infers missing behavior information for cold-start users from attributes, and a ranking tower performs the actual recommendation scoring (Feng et al., 2021). Its zero-shot tower compresses user attributes and behavior into attention-weighted summaries, maps them through dual autoencoders, reconstructs both modalities from either latent code, and aligns attribute-derived and behavior-derived latent distributions with an MMD penalty. For a new user, the reconstructed behavior representation is

$\hat{\mathbf{v}_o = D2(E1(\mathbf{h}_o^a)),$

which the paper explicitly calls virtual behavior data (Feng et al., 2021). The important clarification is that these are not explicit synthetic click logs or generated discrete item sequences; they are reconstructed dense behavior feature vectors in the same representation space used by the downstream recommender.

The same paper also makes a more general CSRM point: the ranking tower is model-agnostic and can be attached to any embedding-based deep ranking model (Feng et al., 2021). That design reappears in later practical systems. In CSRM terms, the explicit matching function remains in the ranker, while cold-start handling is delegated to a representation-completion module.

LLAE provides a different user-side formulation by recasting cold-start recommendation as zero-shot learning. It learns a low-rank linear autoencoder between a user behavior space $\mathbf{X}$ and an auxiliary attribute space $\mathbf{S}$, with the prediction rule

$\mathbf{X}_{new}=\mathbf{W}^\top \mathbf{S}_{new}$

for new users (Li et al., 2019). The paper emphasizes bilateral mapping and low-rank structure to address domain shift, spurious correlations, and efficiency. This suggests a CSRM interpretation in which side information is not merely appended to a ranker but decoded directly into a relevance vector over candidates.

metaCSR moves from side information to few-behavior adaptation. It targets user cold-start sequential recommendation with three components: a Diffusion Representer over the interaction graph, a masked self-attention Sequential Recommender, and a Meta Learner that optimizes for fast adaptation after one or a few gradient updates (Huang et al., 2021). The score is written as

$$\mathbf{x}_o=\{\mathbf{a}_o,\mathbf{c}_o,\mathbf{t}_o\},$$0

where $$\mathbf{x}_o=\{\mathbf{a}_o,\mathbf{c}_o,\mathbf{t}_o\},$$1 is the sequence-derived user preference embedding (Huang et al., 2021). Rather than imputing missing behavior from metadata, metaCSR treats CSRM as few-shot relevance estimation from sparse interaction prefixes.

VM-Rec is closely related in spirit but uses a frozen pretrained recommender and a variational mapping from the first $$\mathbf{x}_o=\{\mathbf{a}_o,\mathbf{c}_o,\mathbf{t}_o\},$$2 interactions to an expressive user embedding (Zheng et al., 2023). Its decoder is a sparse linear combination of warm-user embeddings,

$$\mathbf{x}_o=\{\mathbf{a}_o,\mathbf{c}_o,\mathbf{t}_o\},$$3

with spike-and-slab latent variables over the coefficients $$\mathbf{x}_o=\{\mathbf{a}_o,\mathbf{c}_o,\mathbf{t}_o\},$$4 (Zheng et al., 2023). The paper’s main observation is that users with similar initial interactions tend to end up with similar warm embeddings, so a few early actions can place a new user near the correct region of the mature embedding space.

ColdRoute shows that query-side CSRM also appears outside recommender systems. It treats a newly posted Stack Exchange question as a cold query and ranks candidate answerers with a second-order Factorization Machine,

$$\mathbf{x}_o=\{\mathbf{a}_o,\mathbf{c}_o,\mathbf{t}_o\},$$5

over sparse one-hot and multi-hot features such as question ID, asker ID, answerer ID, and tags (Sun et al., 2018). The central empirical finding is that tags are the most critical feature, while asker information is less important. This is a useful corrective to a common misconception: in some CSRM regimes, structured side information can dominate richer semantic text features.

4. Item-side CSRM and content-to-behavior transfer

On the item side, a frequent pattern is to learn content-based item representations that approximate collaborative or behavioral relevance. Music Cold-Start Matching formalizes this as cold item to warm item relevance matching: a new song has only content features—metadata, audio, and lyrics—and must be matched to existing songs whose audiences are likely to overlap with its future audience (Zhao et al., 2023). The encoder produces a song representation $$\mathbf{x}_o=\{\mathbf{a}_o,\mathbf{c}_o,\mathbf{t}_o\},$$6, trained with BPR over song-song interactions and a contrastive regularizer called Bootstrapping Contrastive Learning (BCL). At serving time, the system retrieves top-$$\mathbf{x}_o=\{\mathbf{a}_o,\mathbf{c}_o,\mathbf{t}_o\},$$7 nearest warm songs with Faiss and then applies Clustering-based Audience Targeting (CAT) to rank users more precisely (Zhao et al., 2023). The key transferable pattern is two-stage: item-to-item relevance matching, then audience transfer.

NeuHash-CF addresses item cold start through content-aware binary hashing. Every item, seen or unseen, is encoded from its content vector $$\mathbf{x}_o=\{\mathbf{a}_o,\mathbf{c}_o,\mathbf{t}_o\},$$8 into a binary code $$\mathbf{x}_o=\{\mathbf{a}_o,\mathbf{c}_o,\mathbf{t}_o\},$$9; users are also assigned binary codes, and relevance is computed in Hamming space (Hansen et al., 2020). The central identity,

$\hat y$0

makes Hamming distance equivalent to binary inner-product ranking (Hansen et al., 2020). The important design choice is that cold and warm items are represented in exactly the same way—directly from content—rather than learning warm-item codes first and only later approximating cold-item codes.

A graph-based version of the same idea appears in the hybrid path model for cold-start items. It builds a tripartite user-item-feature graph and learns feature weights so that the content-based path item $\hat y$1 feature $\hat y$2 item approximates the collaborative path item $\hat y$3 user $\hat y$4 item (Bernardis et al., 2018). The learning objective is

$\hat y$5

which makes feature importance a distilled form of collaborative relevance (Bernardis et al., 2018). MARec pushes this alignment idea further by constructing metadata-derived item-item similarities $\hat y$6, mixing them across feature groups, and injecting them into collaborative backbones such as EASE and VAE (Monteil et al., 2024).

PROMO and SaviorRec represent two later, production-oriented extensions. PROMO argues that content descriptions may be suboptimal prompt information for item cold-start recommendation because of semantic gaps and warm-item bias, and instead uses pinnacle feedback, selected from exceptionally positive users for the item,

$\hat y$7

together with item-wise personalized prompt networks (Jiang et al., 2024). SaviorRec, by contrast, begins with image and text features, learns a behavior-aware multimodal representation $\hat y$8 from co-click pairs, discretizes it into a residual quantized semantic ID, and then uses the MBA module to produce a residual alignment vector $\hat y$9 so that

$\mathcal{P}(y \mid i, q; \Theta),$0

It then performs four target-attention operations across behavior and semantic spaces before CTR prediction (Yao et al., 2 Aug 2025). In both cases, the item representation is explicitly engineered to become behavior-aligned rather than merely descriptive.

5. Search-system and market-level CSRM

BayesCNS addresses item cold start and non-stationarity in search systems that already rely on interaction-heavy learning-to-rank models. It treats user-interaction features $\mathcal{P}(y \mid i, q; \Theta),$1 as uncertain random variables, predicts prior parameters from contextual features $\mathcal{P}(y \mid i, q; \Theta),$2 through a neural network, and updates those priors online with a discounted Gamma-Poisson mechanism (Ardywibowo et al., 2024). Scoring is performed by the incumbent ranker after posterior sampling of interaction features, so exploration is ranker-guided rather than heuristic. This is a distinct CSRM interpretation: cold-start relevance matching becomes Bayesian imputation of missing behavioral features under contextual guidance.

PromptRec targets the stronger system cold-start setting in which the target interaction matrix is empty. It converts recommendation into a sentiment-analysis style language-model problem by verbalizing the user-item pair as natural language and scoring positive versus negative label tokens: $\mathcal{P}(y \mid i, q; \Theta),$3 To make small LLMs competitive, it adds a data-centric pipeline consisting of refined-corpus pretraining (RCMP) and decomposed prompt pretraining (TPPT) (Wu et al., 2023). In a CSRM reading, this is a pointwise relevance scorer or reranker for settings with no target-domain interactions at all.

CSRM-LLM applies the term CSRM explicitly to emerging e-commerce markets. It studies query-product relevance classification in a new market with very limited human labels, sparse or absent user behavior, and cross-lingual domain mismatch (Wang et al., 1 Sep 2025). Its framework uses a multilingual LLM, auxiliary machine translation tasks, retrieval-based query augmentation, and multi-round self-distillation; because the LLM is too expensive for online serving, the final model is distilled into a twin-tower XLM-R–based online ranker (Wang et al., 1 Sep 2025). A key lesson is that generic multilingual capability is insufficient when label semantics are marketplace-specific.

Although not a cold-start paper, CSRS is relevant methodologically because it combines a relevance matching module $\mathcal{P}(y \mid i, q; \Theta),$4 with a co-attention semantic matching module in code search (Cheng et al., 2022). This suggests that hybrid lexical and semantic matching remains useful even when CSRM systems are dominated by metadata, side information, or behavioral transfer: exact or soft lexical evidence can remain the stable branch when higher-level semantic transfer is weak.

6. Evaluation, misconceptions, and open problems

The empirical evidence across these papers is unusually deployment-oriented. MAIL was deployed on NetEase Cloud Music and reported a click-through rate improvement of 13% to 15% and a 3% to 4% relative CTCVR uplift in A/B testing (Feng et al., 2021). Music Cold-Start Matching reported +38.47% effective play count, +1.42% full play rate, and +1.95% effective red rate in online experiments (Zhao et al., 2023). BayesCNS reported 1.05% overall success improvement and 10.60% increased new-item interactions or impressions in a month-long live test (Ardywibowo et al., 2024). CSRM-LLM reported a 45.8% reduction in defect ratio and a 0.866% uplift in session purchase rate after distillation to an online twin-tower model (Wang et al., 1 Sep 2025). PROMO reported +3.2% Click Rate, +4.8% Video Play Time, +3.9% Video Like, and +4.0% Video Collecting in a billion-user-scale short-video application (Jiang et al., 2024). SaviorRec reported 0.83% offline AUC improvement and online 13.21% clicks increase with 13.44% orders increase in the abstract (Yao et al., 2 Aug 2025). MARec reported gains from +8.4% to +53.8% on cold-start ranking metrics, and its semantic-feature ablations reported additional gains from +46.8% to +105.5% (Monteil et al., 2024).

Several misconceptions are repeatedly corrected in the literature. First, CSRM is not synonymous with item cold start: MAIL, LLAE, metaCSR, VM-Rec, and ColdRoute are primarily user-side or query-side methods (Feng et al., 2021, Li et al., 2019, Huang et al., 2021, Zheng et al., 2023, Sun et al., 2018). Second, naive content similarity is often insufficient: MARec emphasizes metadata alignment to behavioral structure, SaviorRec centers continuous semantic-behavior alignment, and ColdRoute shows semantic text baselines can trail tag-based sparse interaction models (Monteil et al., 2024, Yao et al., 2 Aug 2025, Sun et al., 2018). Third, cold-start completion modules do not necessarily generate literal missing interactions: MAIL explicitly reconstructs dense behavior feature vectors rather than synthetic click sequences (Feng et al., 2021). Fourth, a multilingual LLM alone does not solve market cold start when the target business labels are domain-specific (Wang et al., 1 Sep 2025).

The main unresolved issues are also consistent across papers. Many methods depend heavily on informative side information, multimodal content, or reliable metadata at onboarding. Several are one-sided: user cold-start methods do not solve item cold start, and item cold-start methods do not solve user cold start. Many rank by pointwise CTR or reconstruction losses rather than direct pairwise or listwise ranking objectives. Some industrial systems omit detailed latency or systems measurements even when deployment claims are strong. This suggests that the central open problem in CSRM is not only learning a cold-start representation, but doing so with robust calibration, low serving cost, and stable transfer under distribution shift.

Viewed synthetically, the field converges on a clear principle: CSRM works best when side information is transformed into a relevance-ready latent object rather than appended as raw metadata. That latent object may be a virtual behavior vector, a sparse prototype mixture, a behavior-aware multimodal embedding, a semantic ID, a Bayesian posterior over interaction features, or a prompt-conditioned label-token score. The specific mechanism varies by domain, but the operational goal remains the same: recover enough of the missing relevance structure to let an existing or newly deployed ranker behave as if cold-start entities already had usable interaction evidence.