PaperlayerView exact revision on arXiv

Source-linked AI summary

Nearest but Not Dearest: Shared Curator-Feedback Infrastructure for Content-Only Search and Recommendation

Matt Sandler

arXiv:2609.30568v1cs.IRcs.SD

TL;DR

In content-only music discovery, curator judgment is the principal feedback signal, but acoustically close neighbors can still be unsuitable. The paper routes sound and context failures to separate mechanisms in a shared retrieval stack. The combined intervention reduces rejection by 24.5% relative, with the authors treating the deployment as an industrial case study rather than a validated general method.

  • Problem

    In a content-only stack without listener interactions, curator judgment is the principal available feedback, while acoustically close retrieved tracks can still be rejected for suitability.

    Source: §1 · p. 1

  • Method

    The paper partitions curator rejections into sound and context failures and routes them to a reweighting head and a candidate-generation constraint filter, respectively.

    Source: §1 · p. 2

  • Results

    The combined intervention reduces nearest-neighbor rejection from 38.17% to 28.83%, a −24.5% relative drop (McNemar χ² = 22.37, p = 2.2 × 10−6).

    Source: §1 · p. 2

  • Takeaways & Limitations

    Because the failure partition is orthogonal to the search–recommendation split, corrections in shared infrastructure can structurally propagate across entry points.

    Source: §5 · p. 6

    The unblinded, compound deployment on one system and catalog does not provide a clean per-mechanism causal estimate or establish generalization to other catalogs and encoders.

    Source: §5 · p. 6

Abstract

from arXiv · show

A deployed B2B music-discovery platform serves both query-driven search (text prompts, vibe tags) and seed-driven recommendation (seed-track and artist stations) over one licensed catalog, one LAION-CLAP joint audio-text embedding space, one candidate-generation filter, and one ranking head -- and neither path consumes end-listener behavioral signal. In this content-only regime, curator judgment is the principal feedback signal available, and offline cosine similarity predicts it poorly: 38% of cosine-nearest neighbors are rejected by curators. The rejections reveal a clean partition: a majority (55%) are sound failures the encoder could address (style, tempo, mood mismatch), and a substantial minority (37%) are context failures orthogonal to the waveform (wrong language, holiday content, devotional content, rights and lyric flags). We deploy this sound-vs-context decomposition as feedback infrastructure, routing each failure mode to the layer that can absorb it: context failures to a constraint filter at candidate generation, sound failures to an embedding reweighting head at the representation layer -- both below the search/recommendation split, so a single curator loop maintains both experiences. On 1,200 curator judgments collected over two production rounds one month apart, the combined intervention reduces rejection rate from 38.17% to 28.83% (-24.5% relative, McNemar chi-squared = 22.4, p = 2.2e-6). An accounting decomposition attributes 4.08 pp of the drop to filter-eligible categories and 5.25 pp to the rest; the deployment was unblinded and compound, so this is a production accounting bound, not a causal estimate. We present this as an industrial case study rather than a validated general method, and close with lessons for content-only discovery: the failure partition is orthogonal to the paradigm partition, and corrections land in layers shared by both paradigms.

1 Introduction

In a content-only music-discovery stack, curator judgment is the principal available feedback, yet acoustically close neighbors can still fail curator suitability. The paper partitions these failures and deploys separate corrections in layers shared by search and recommendation.

1 Introduction

  • The failure pattern arises when acoustically close neighbors violate non-audio suitability judgments, such as language or seasonal context.The examples include cross-language pop matches and a Christmas standard retrieved for an electro-funk seed; curators reject them despite their acoustic similarity.Source: §1 · p. 1
  • The deployed stack serves query-driven and seed-driven discovery from one catalog, one LAION-CLAP space, one candidate filter, and one ranking head, without using listener interactions.Curator judgment is the principal judgment signal available in this content-only regime.Source: §1 · p. 1
  • Among 458 thumbs-down judgments, 55.2% were sound failures and 36.9% were context failures, requiring different engineering levers.Sound failures include style, tempo, and mood mismatch; context failures include language, holiday, devotional, children’s catalog, and rights restrictions.Source: §1 · p. 2
  • The combined intervention reduces nearest-neighbor rejection from 38.17% to 28.83%, a −24.5% relative drop (McNemar χ² = 22.37, p = 2.2 × 10−6).The accounting decomposition attributes 4.08 pp to filter-eligible categories and 5.25 pp to non-filter categories; the authors describe this as an accounting bound, not a causal estimate.Source: §1 · p. 2
  • The authors frame the work as an industrial case study, not a validated general method, based on one compound production deployment measured on the seed-track path.The claim is limited to a routing criterion for expert feedback in a shared content-only stack.Source: §1 · p. 2

2 Related Work

Prior work studies embedding quality, recommender accuracy, and unified search–recommendation systems, often with behavioral signals. This paper’s related-work framing centers on operational retrieval corrections in a content-only setting.

Embedding-based music retrieval

  • Audio encoders are commonly evaluated on tagging benchmarks, while direct nearest-neighbor usefulness against curator or listener judgment receives less attention.The cited tagging benchmarks include; production nearest-neighbor retrieval is increasingly common [19].Source: §2 · p. 2

Beyond accuracy in recommender evaluation

  • The paper’s operational contribution is a partition of music-retrieval failures according to which engineering lever resolves them.This framing responds to the imperfect relationship between recommender accuracy metrics and what users want, alongside music-specific genre, language, and context failures [19].Source: §2 · p. 2

Unified search and recommendation

  • Prior unified search–recommendation work unifies at the model or user-behavior layer and generally presumes logged interactions, unlike this deployment’s shared lower-level stack without behavioral logs.The related work cites joint models [28], an industrial contextual model [1], and generative retrieval [18].Source: §2 · p. 2

Joint and multimodal search–recommendation models

  • The system is closest to multimodal recommenders using pretrained encoders, but lacks the user-behavior sequences modeled by USER, SESRec, and UniSAR.It uses a bi-encoder with a learned projection; cross-encoders and late-interaction models are identified as future work [15] [9].Source: §2 · p. 2

Industry deployments and metadata-driven lift

  • The paper complements industry work on behavioral audio recommenders and metadata-driven lift by addressing upstream retrieval correction before metadata enrichment, without listener-level retrieval signals.The related-work discussion cites deployed audio recommenders [6] [16] and the DDEX/MEAD report.Source: §2 · p. 2

3 The Decomposition

The section describes a shared retrieval stack, an operational sound-versus-context partition, and feedback routing to distinct correction mechanisms. Both mechanisms sit in infrastructure shared by the system’s entry points.

3.1 Deployed system: one stack, four entry points

  • Four entry points differ only in query-vector construction and share nearest-neighbor retrieval over one catalog and candidate filter in a joint audio–text space.The curator evaluation uses seed-track queries exclusively; contextual metadata also feeds an existing pre-filter.Source: §3.1 · p. 3

3.2 The sound-vs-context partition

  • The partition treats context failures as resolvable by non-acoustic metadata and assigns uncleanly attributed codes to an explicit residual group.For accounting, Cother is conservatively grouped with Csound in the non-filter category.Source: §3.2 · p. 3
  • The two failure groups call for different work: representation changes for sound failures and catalog filtering or candidate restriction for context failures.Without distinguishing them, a system risks spending its budget on the wrong engineering levers.Source: §3.2 · p. 3

3.3 Routing failures to mechanisms

  • The constraint filter applies metadata-based Cctx rules at candidate generation, before cosine ranking, and updates can ship the same day they are reviewed.Because it acts upstream of query-vector construction, seed-track feedback can constrain text-prompt and vibe-tag retrieval too.Source: §3.3 · p. 3
  • The reweighting head learns seed-relative sonic matches from accepted and Csound-rejected neighbors using a linear projection over frozen LAION-CLAP embeddings.It is available to any entry point resolving to a CLAP vector, but whether it improves search-shaped paths remains a measurement question.Source: §3.3 · pp. 3–4
  • The Round-2 production comparison uses the same evaluation workload that supplied Round-1 feedback to the head, so it is not a held-out generalization benchmark.A seed-level validation split was used for model selection, but generalization to never-judged seeds remains future work.Source: §3.3 · p. 4
  • The decomposition makes routing well-defined by assigning each thumbs-down to the intervention suited to its failure mode.Figure 1 depicts Cctx routing to the constraint filter and Csound ∪ Cother to the reweighting head.Source: §3.3 · p. 4, Figure 1 · p. 4

4 Evaluation

Across two rounds of curator evaluation, the study examines rejection patterns and a combined intervention in a production music-recommendation system. The evaluation reports a significant paired-seed change, while its category-level accounting is not a controlled estimate of each mechanism’s effect.

4.1 Curator panel and protocol

  • Four curators judged 1,200 recommendations per round across 400 seeds, with Round 2 targeting the same seed set about one month later.Each seed returned three neighbors; the paired-seed analysis covered only 269 seeds with complete judgments in both rounds.Source: §4.1 · p. 4

4.2 Baseline failure-mode breakdown

  • Round 1 recorded 458 coded rejections among 1,200 recommendations (38.17%); 55.2% were sound failures, 36.9% context failures, and 7.9% other.The interface did not disaggregate sound-related from contextual selections in the “other” category.Source: §4.2 · p. 4, Table 1 · p. 5
  • Cosine similarity weakly separated accepted from rejected recommendations (AUC = 0.644), and the category-specific similarity means overlapped.Accepted and rejected means were 0.897 and 0.881, respectively; the sound- and context-category means occupied similar ranges.Source: §4.2 · pp. 4–5, Figure 2 · p. 5

4.3 Round-2 intervention

  • 76 seeds improved and 27 worsened; the paired-seed test found χ2 = 22.37, df=1, p= 2.25 × 10−6.The paired subset included 269 seeds judged in both rounds; the comparison treats the seed, rather than each of its three recommendations, as the unit.Source: §4.3 · p. 5, Table 2 · p. 5

4.4 Inter-rater agreement: a brief note

  • All four curators’ rejection rates moved within ±5 pp of the aggregate change, while 75% of 36 disagreements were sound-related.The study reports Cohen’s κ = 0.451 for 134 shared seed-neighbor pairs and suggests accounting for greater label noise in reweighting-head training data.Source: §4.3 · p. 5, §4.4 · p. 6

4.3 Round-2 intervention

  • The category accounting assigns −4.08 pp to filter-eligible categories and −5.25 pp to non-filter categories, but is not a controlled per-mechanism estimate.The mechanisms shipped together; their category contributions sum to the observed 9.33 pp drop. The filter requires days of catalog work, versus weeks for the reweighting head.Source: §4.3 · pp. 5–6, Table 3 · p. 5, Figure 3 · p. 6

5 Lessons and Limitations

The paper argues that curator feedback should be routed by failure type, not by discovery entry point: sound and context failures call for different mechanisms in the shared stack. The reported cost advantage favors the filter, while the deployment and evaluation leave causal attribution, generalization, and listener impact unresolved.

Lessons and Limitations

  • The failure partition, rather than the SQL filter or metric-learning head, is the contribution: it assigns curator feedback to distinct mechanisms.A single retraining loop can wash the two signals together and dilute both.Source: §5 · p. 6
  • 4.08 pp of the rejection-rate reduction came from filter-eligible categories at days-of-engineering cost, versus 5.25 pp from non-filter categories at weeks-of-engineering cost.The authors recommend routing Cctx first when modeling capacity is limited.Source: §5 · p. 6
  • Because both mechanisms sit below the search/recommendation split, feedback collected on one entry point structurally propagates to the other.That transfer follows from shared infrastructure, but has not yet been measured on text-prompt or vibe-tag paths.Source: §5 · pp. 6–7
  • The unblinded, compound deployment supports an accounting bound, not a clean per-mechanism causal estimate.A counterfactual filter replay and staged rollout are proposed to estimate mechanisms separately.Source: §5 · pp. 6–7
  • The evaluation used the seed-track path and workload that fed the feedback; it does not establish gains on unseen seeds or listener satisfaction.Curator-approved retrieval remains an unvalidated proxy for listener preference.Source: §5 · pp. 6–7
  • An LLM auditor could scale discovery of context rules from metadata, while sound labels require an audio-capable judge calibrated against curator disagreements.The authors frame LLM judgment as a possible way to scale the cheaper context side first.Source: §5 · p. 7
  • A cross-encoder or late-interaction model could rescore a short bi-encoder shortlist, but whether a few hundred curator pairs suffice to train one remains open.The paper notes an 85 ms mean latency budget and rules out cross-encoding the full catalog.Source: §5 · p. 7
  • Listener skip-rate comparisons and replication across encoders are proposed to test whether curator improvements reach listeners and whether the partition generalizes.Replication across encoders, including MERT, is an open empirical question.Source: §5 · p. 7

6 Conclusion

The paper reports a deployed sound-vs-context feedback decomposition for content-only music retrieval, routing each failure type to a different mechanism in shared infrastructure. The compound intervention reduced curator rejection, but the paper remains a single production case study with per-mechanism and cross-paradigm claims still unmeasured.

6 Conclusion

  • The combined intervention reduced production rejection rate by 24.5% relative on 1,200 judgments per round (p = 2.2 × 10^-6).The paper reports accounting reductions on both filter-eligible and non-filter category groups, with a cost-per-percentage-point advantage for the filter.Source: §6 · p. 7
  • The decomposition routes context failures to a constraint filter and sound failures to an embedding reweighting head below the search/recommendation split.The mechanisms operate at shared stack layers and on distinct engineering cadences.Source: §6 · p. 7
  • The evidence is a single production case study, and per-mechanism and cross-paradigm claims remain to be measured.The conclusion characterizes the pattern as low-overhead for content-only systems without behavioral retrieval signal.Source: §6 · p. 7

Ethics Statement

The evaluation relied on in-house professional curators rather than external or crowdsourced raters, and used no end-listener data. The authors caution that context categories could risk bias if applied asymmetrically and recommend auditing outcomes by seed context.

Ethics Statement

  • Four in-house professional music curators participated; no external or crowdsourced raters took part, and no end-listener data was used.The curators were compensated as part of regular employment.Source: Ethics Statement · p. 7
  • Asymmetric use of categories such as wrong_language, devotional, or childrens could risk bias against non-dominant languages, cultural traditions, or listener communities.The taxonomy is context-aware: a category fires only when seed context and a retrieved neighbor's properties mismatch.Source: Ethics Statement · p. 7
  • The authors recommend auditing protocol outputs by seed context to detect and correct systematic under-service of non-dominant contexts.Source: Ethics Statement · p. 7

AI Usage Statement

AI assistants were used for grammar, style, copy-editing, and LaTeX formatting. The authors state that research ideas, experimental design, analyses, statistical computations, and figures were their own.

AI Usage Statement

  • AI assistants supported grammar, style, copy-editing, and LaTeX formatting tasks.The formatting tasks included table layout, figure placement, and bibliography hygiene.Source: AI Usage Statement · p. 8
  • The authors state that all research ideas, experimental design choices, data analyses, statistical computations, and figures were their own.Source: AI Usage Statement · p. 8
  • The stated AI use covered prose editing and formatting rather than the listed research and analysis work.Source: AI Usage Statement · p. 8
Loading 2609.30568v1…