{"id":"3b4b4816-3b54-4b49-9875-f925ca62b15f","arxiv_id":"2509.08337","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Using 15 years of Last.fm listening data, the study shows that intra-user musical diversity increases with age while inter-user variation decreases, and that after age 40 listening becomes increasingly nostalgic.","lead":"This paper tracks how music listening habits change as people age, using 15 years of Last.fm data from over 42,000 users. It finds that younger listeners favor current popular songs, while older listeners develop more personal and nostalgic tastes, which could help music apps tailor recommendations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Diversity metric U_O is confounded by playcount volume; the age-related diversity increase may be an artifact of lower listening volume.","rationale":"The reader's weakest assumption focuses on age accuracy. While that is a legitimate limitation and is acknowledged in the paper (Section 4.1), I find a more load-bearing threat to the central claim in the construction and interpretation of the diversity measures. The primary metric U_O = unique_tracks/playcounts rises when playcounts fall, even if unique tracks are flat or declining. Figure 1 shows playcounts decline monotonically from age 19 onward, so the increasing U_O trend in Fig. 2 may simply reflect lower listening volume, not broader taste. This directly undermines the statement that 'intra-user diversity increases' with age. The inter-user diversity claim is similarly unsupported: the Gini index in Fig. 3 measures inequality of U_O values within an age group, not overlap in musical tastes, so it cannot substantiate 'inter-user diversity decreases.' These issues are independent of age-measurement noise. The proposed rarefaction test would separate volume effects from true breadth. The paper remains potentially insightful, but the empirical foundation for the headline claims is weaker than the age-accuracy concern. The conditional verdict stands, but the condition should be a robustness check on the diversity metric rather than only improved age data.","tokens_in":7443,"tokens_out":8733,"duration_ms":95107,"concrete_test":"Recompute the age-diversity curves after rarefying each user-year's plays to a fixed count (e.g., 500 plays) and estimating expected unique tracks/artists; if the monotonic increase from ages 15 to 40 disappears or reverses, the headline diversity claim is a playcount-volume artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that intra-user diversity increases with age rests on U_O = unique_tracks / playcounts (Section 3.2). Because playcounts decline sharply with age (Fig. 1: R drops from 2318 at age 19 to ~560 at age 58), U_O rises even when the number of unique tracks is flat or declining (S peaks at 805 at age 22 and then falls). The paper interprets the U_O increase as 'broader listening preferences' (Section 3.3), but the metric largely captures decreasing play volume, not taste breadth. A user who halves playcount while keeping the same unique set doubles U_O. The same volume confound taints the Gini analysis in Fig. 3, which measures dispersion of U_O values across users—not overlap of musical taste—so the statement 'inter-user diversity decreases' (Section 4) is not supported by the presented evidence. These concerns are independent of age-measurement noise: even with perfect ages, the diversity conclusions may be artifacts.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the LFM-2b listening-event dataset to study how music preferences and consumption behavior change with user age. The authors adjust a static age variable to event time, restrict the sample to users aged 10–64, and compute annual playcounts, unique tracks, unique artists, a track-based diversity ratio, an artist-based ratio, the Gini index of diversity, and a song-specific age (SSA) metric. They report that intra-user diversity increases with age, inter-user diversity decreases, younger users favor contemporary releases, and after age 40 a secondary 'nostalgic' peak appears. The paper draws recommender-system implications from these trends.","tokens_in":7716,"tokens_out":3777,"duration_ms":47903,"significance":"If the reported trends are valid, the paper addresses an under-studied longitudinal dimension in recommender-systems research using a rare large-scale dataset. Its descriptive contribution is potentially useful: the cleaning procedure is clearly described, the age-adjustment idea is sensible, and the SSA analysis connects directly to prior music-psychology findings. However, the central metric used to support the main claims is confounded by listening volume, and no statistical uncertainty is reported. The paper's value therefore depends on whether the main conclusions survive additional analysis.","major_comments":[{"comment":"The central diversity measure U_O = |S|/R conflates taste breadth with listening volume. Fig. 1 shows median annual playcounts falling from 2,318 at age 19 to 560 at age 58, while unique tracks peak at 805 (age 22) and then decline. A user who halves playcounts while keeping the same unique-track set doubles U_O. The rise in U_O between ages 15 and 40 is therefore not by itself evidence of 'broader listening preferences' (Section 3.3). The same confound propagates to the Gini index in Fig. 3, which measures dispersion of this ratio across users, not overlap or distance between users' musical tastes. The statement in Section 4 that 'inter-user diversity decreases (Fig. 3)' is not supported by the presented evidence. Please report at least one volume-independent diversity measure (e.g., entropy, Simpson diversity, or unique tracks conditioned on playcount) and, for inter-user diversity, a","section":"§3.2, §3.3, Fig. 1, Fig. 2, Fig. 3"},{"comment":"The paper uses 'significantly' in describing the diversity increase between ages 15 and 40, but no confidence intervals, significance tests, or effect sizes are reported for any of the median trends. The regression lines in Fig. 1 are fitted without stating the procedure or uncertainty. Because the main claims are about monotone age trends in a noisy, long-tailed dataset, bootstrap confidence intervals on the medians by age would be a minimal and necessary addition.","section":"§3.3, Figs. 1–5"},{"comment":"The entire analysis depends on the adjusted age variable M = M + year difference from a fixed reference date. The paper itself lists 'improbable user ages' as a limitation. The static age is retrieved in 2013–2014, and age at each event is computed by adding the year difference; if the static age is misreported or systematically biased, all age-trend conclusions are affected. A sensitivity analysis is needed: report the distribution of static ages before cleaning, show the effect of excluding users with suspicious or inconsistent ages, and check whether the results hold when restricting to users whose activity is consistent with a single reported age.","section":"§3.1, §4.1"}],"minor_comments":[{"comment":"The equation for the adjusted age is garbled: 'ω𝑁 (𝑁𝑀 ,𝐿 𝐿 )' and related symbols appear corrupted. The quartile-based outlier formula also has a typo ('𝑂1 ↑ 1.5 · 𝑃𝑂𝑄' should presumably be Q1 − 1.5·IQR).","section":"§3.1"},{"comment":"The text says 'Table 1 shows the number of users in each age group post-cleaning,' but Table 1 only gives global counts. Please include per-age-group user counts or remove the cross-reference.","section":"Table 1 and §3.1"},{"comment":"The caption lists '(b) Number of Unique Artists' and '(c) Number of Unique Tracks,' while the main text describes the subfigures in the opposite order. Please make the caption consistent with the panels.","section":"Fig. 1 caption"},{"comment":"The 'algorithmic confounding' paragraph is speculative and not tied to any data or prior result in the paper. It may be better framed as future work.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"This is a short descriptive paper with a potentially useful longitudinal dataset, but the main diversity conclusions rest on a metric confounded with volume, and the 'inter-user diversity decreases' claim is not supported by a Gini index of a per-user ratio. These are fixable with additional analyses, so I recommend major revision rather than rejection. The authors should also be encouraged to provide code or per-age summary tables to make the descriptive findings reusable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look for the longitudinal LFM-2b analysis and the SSA (song-specific age) curves, but the central diversity claims have a load-bearing measurement problem.\n\nWhat's new: this is the first analysis of LFM-2b that assigns an event-level age by shifting the static age against a reference date, which lets them track users across 15 years. That's a reasonable and clever use of the dataset's timestamp structure. The SSA finding – a second listening peak around age 15-20 for users over 40 – is a genuine, interpretable pattern that aligns with the music psychology literature and is the most convincing part of the paper. The descriptive plots are clear, and the cleaning procedure is reproducible in principle.\n\nThe soft spots. The stress-test note is right and it's a serious one. U_O = unique_tracks / playcounts. The paper shows playcounts drop steeply with age (Fig. 1a), while unique tracks peak around 22 and then decline gently. So U_O increases with age for most of the range even if actual variety is flat or falling. The text literally says diversity grows 'significantly' between 15 and 40, but with no confidence intervals or significance tests anywhere, 'significantly' is doing no statistical work. Worse, the Gini index in Fig. 3 is computed on U_O values across users within an age group, not on the overlap of their actual music choices. That is a measure of dispersion of a volume-confounded ratio, not inter-user preference overlap. The conclusion that 'inter-user diversity decreases' (Section 4) is therefore not supported by the evidence as presented. The paper's own limitation section admits the metric is basic, but doesn't flag the playcount-volume confound, which is more serious than a nuance.\n\nThe age-accuracy limitation is real but secondary, and the paper already admits it. The survivor bias (users active in 2013-2014) is also acknowledged.\n\nNet: as a descriptive, exploratory study, it's a decent contribution and a useful pointer for recommender-system people who want age-aware diversity or nostalgia handling. The SSA results alone justify a serious referee. The intra-user diversity increase claim needs either a better metric (e.g., entropy, or unique tracks controlled for playcount) or at least a clear caveat that it's partly a volume artifact. The inter-user diversity claim should be retracted or re-measured with actual taste overlap.\n\nI'd send it to peer review, but with a strong steer that the diversity analysis needs rethinking before it can be published as-is.","headline":"Useful descriptive longitudinal study, but the diversity metric is volume-confounded and the Gini-index based inter-user diversity claim is not supported by the presented evidence.","tokens_in":8177,"tokens_out":658,"would_cite":false,"duration_ms":9328,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Using 15 years of Last.fm logs, this paper argues that individual musical taste broadens with age even as listeners become more alike, and that after about age 40 nostalgia for teenage-era tracks resurges.","keywords":["music recommender systems","longitudinal user modeling","age and music taste","listening diversity","song-specific age","nostalgia","Last.fm listening logs","LFM-2b"],"falsifier":"A decisive check would be to rerun Figures 2, 3, and 5 on the subset of users whose reported age does not end in 0 or 5 and who have listening events in both 2005–2008 and 2016–2020; if the post-40 nostalgia peak and the falling Gini index shrink to noise, the findings would be artifacts of age adjustment and sample attrition rather than true aging effects.","tokens_in":7398,"feed_emoji":"🎵","tokens_out":9339,"duration_ms":101708,"temperature":0.7,"pith_summary":"This paper tries to establish that musical preferences change with age in systematic, measurable ways that recommender systems should model. It works with the LFM-2b dataset, covering about 542 million listening events from 42,883 Last.fm users across 15 years, and adjusts each user's one-time reported age to the moment of each listen. The analysis finds that as a person ages, their own listening diversity increases, while differences between listeners of the same age shrink. A song-specific age analysis shows that younger listeners strongly favor contemporary releases, while after about age 40 a second, nostalgia-driven peak appears for music from the listener's adolescence. If the findings hold, they give recommender systems a concrete age-based dial for tuning how much diversity or personalization to serve.","feed_headline":"Diversity grows with age; at 40, nostalgia kicks in","feed_subtitle":"15 years of Last.fm logs show taste widening inside users while listeners converge—clues for age-aware recommenders.","key_machinery":"The argument is carried by four instruments. (1) An age-adjusted longitudinal sample: each user's static age from LFM-2b is shifted per listening event by the year difference from Oct 31, 2013, so every event gets an adjusted age M. (2) Two [0,1] diversity ratios—track diversity U_O = unique tracks ÷ playcounts and artist diversity U_A = unique artists ÷ unique tracks—following the diversity definition of Schedl and Hauger [15] and inverted per Spear et al. [18]. (3) The Gini index applied to within-age-group diversity, measuring the 'diversity of diversity' and quantifying inter-user convergence. (4) Song-specific age (SSA), the listener's age at a track's release, with playcounts log-norma","core_discovery":"The paper's central claim: once users are tracked for years rather than weeks, age visibly reshapes musical taste. Track diversity (unique tracks per playcount) grows from about 0.32 at age 15 to 0.60 at age 40 and then plateaus, so each person's listening widens. Over the same span, the Gini index of diversity within age groups falls from about 0.33 to 0.17, so same-age listeners become more alike. Song-specific age—the listener's age when a track was released—peaks for current releases at every age, but after 40 a second, nostalgia peak appears at ages 15–20. The authors interpret this as a shift from broad, contemporary, socially shared listening to narrower, personalized, nostalgia-infus","pith_inferences":["Inference: The paper stops at descriptive evidence; a direct algorithmic extension would be to feed adjusted age (or SSA band) into a recommender baseline and measure whether the diversity/personalization trade-off actually shifts, turning the observation into a deployable design rule.","Inference: Because the age variable is self-reported and snapshotted in 2013–2014, the longitudinal curves implicitly assume one user per account and stable truthful reporting; a robustness check on users with non-heaped ages and continuous fifteen-year histories would separate genuine aging effects from cohort and attrition artifacts.","Inference: The same song-specific-age machinery could be applied to films or books, where release-year-relative-to-age may produce an analogous nostalgia bump; finding one would suggest a general reminiscence effect rather than a music-specific quirk.","Inference: The diversity metric is a playcount-concentration ratio, so the exact magnitudes of age differences likely depend on metric choice; entropy- or coverage-based measures could shrink or grow the reported gaps even if the direction holds."],"forward_implications":["Recommender systems can use age as a diversity/personalization dial: younger users benefit from broad, contemporary-heavy suggestions, while older users benefit from tighter, nostalgia-weighted ones.","Around age 40 is a preference tipping point; beyond it, consumption of current releases declines and adolescent-era music dominates, so age-aware models should weight historical taste more heavily.","Because inter-user diversity shrinks with age, older users' tastes are easier to model from their peers, while younger users need more exploration and less conformity.","Algorithmic feedback loops matter: if recommenders feed young users varied content and older users narrow content, the systems may amplify the very age divergence the paper describes."],"supporting_citations":[{"why":"Provides the LFM-2b dataset—15 years of Last.fm listening events with user ages—on which every analysis runs.","marker":"[14]"},{"why":"Defines song-specific age (SSA), the measure that exposes the post-40 nostalgia peak.","marker":"[8]"},{"why":"Supplies the track-diversity definition (unique tracks over playcounts) used as the main proxy.","marker":"[15]"},{"why":"Contributes the inverted [0,1] diversity scale and prior evidence that children's taste diversity diverges with age.","marker":"[18]"},{"why":"Earlier age-group comparison on genre preferences that this study extends to longitudinal, within-user behavior.","marker":"[5]"},{"why":"Prior children/adolescent listening study whose use of means the paper explicitly replaces with medians.","marker":"[13]"},{"why":"Replication of the song-specific-age hypothesis that motivates RQ2 and the authors' integration of a recommendation perspective.","marker":"[10]"}],"fun_headline_variants":["As we age, playlists widen—then nostalgia kicks in at 40","Music taste broadens with age, then nostalgia peaks after 40","Age 40 marks a shift: from broad listening to nostalgia","How your music tastes evolve: broadening, then nostalgic after 40"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The clearest vulnerable premise is the age data: Last.fm users report their age once (2013–2014), and the study shifts that one number forward and backward by calendar year for every listening event, so if users misreport, round off, or share accounts, all the age curves would be distorted.","fun_headline_variants_meta":{"raw":{"variants":["As we age, playlists widen—then nostalgia kicks in at 40","Music taste broadens with age, then nostalgia peaks after 40","Age 40 marks a shift: from broad listening to nostalgia","How your music tastes evolve: broadening, then nostalgic after 40"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1200,"prompt_tokens":747,"completion_tokens":453,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":377}},"tokens_in":491,"tokens_out":453,"duration_ms":5407,"temperature":1.0,"reasoning_tokens":377,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T20:44:23.871165+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check would be to rerun Figures 2, 3, and 5 on the subset of users whose reported age does not end in 0 or 5 and who have listening events in both 2005–2008 and 2016–2020; if the post-40 nostalgia peak and the falling Gini index shrink to noise, the findings would be artifacts of age adjustment and sample attrition rather than true aging effects.","supporting_citations":[],"review_version":1}