Three measurement levels joined to the same stimuli
The dataset's contribution is alignment rather than a new predictive result. Every video in the stimulus set carries laboratory brain recordings, subjective evaluations, and platform-level engagement counts, so a researcher can ask how an individual's neural and psychological response to a clip relates to how that same clip performed online. The paper frames this as a cross-level resource: intra-individual EEG and ratings, individual-difference measures such as demographics and psychological traits, and population-level engagement metrics.
The gap the authors identify is that widely used EEG datasets built on complex audiovisual material, such as DEAP, MAHNOB-HCI, SEED, SEED-IV, DREAMER, and DECAF, were designed mainly for emotion recognition and label clips with low-dimensional affective variables like valence and arousal. Those datasets do not include higher-level social and cognitive evaluations such as empathy, trust, or sharing intention, and they do not carry the real-world dissemination performance of their stimuli. The new dataset is positioned against exactly that absence.
Fifteen videos arranged by theme and narrative frame
Stimuli were drawn from Bilibili and screened down to 15 videos: 14 experimental clips plus one control and practice video. The 14 experimental clips sit in a 7 × 2 framework, with seven semantic themes (urban development, transportation, technology, history and culture, multi-ethnic integration, humanistic life, and ecological environment) and two narrative perspectives, grand versus individual, giving one video per perspective within each theme.
Inclusion rules kept the set comparable: duration between 110 and 160 seconds, release between 2019 and 2023, closed-loop narrative logic with high-quality audio, Mandarin Chinese with Chinese subtitles, and live-action rather than animated footage. Three experts classified candidate videos as grand, individual, or ambiguous with substantial agreement (Fleiss' κ = 0.755), and a video entered a category only when all three raters agreed. The authors state plainly that this design introduces comparable thematic and narrative contrasts rather than a fully stimulus-general manipulation.
One video per cell makes narrative effects stimulus-bound
Because each theme contains exactly one grand-narrative and one individual-narrative video, narrative perspective is partly confounded with whatever else distinguishes those two clips, such as pacing, visual composition, narration style, or emotional intensity. The authors therefore describe grand-versus-individual comparisons as exploratory and stimulus-bound rather than general tests of narrative framing.
A second design constraint compounds this: the manipulation is partially between-subject. Within each theme, a participant viewed only one version of the stimulus, so narrative comparisons are not fully within-subject contrasts. The paper recommends mixed-effects models or between-subject frameworks when analyzing narrative perspective, and it treats the resulting differences as hypotheses about the selected stimuli rather than settled effects of narrative style.
Ratings split emotional resonance from cognitive trust
Subjective evaluations covered 13 dimensions spanning affective, cognitive, and social domains, and the analysis tested theme, narrative perspective, and their interaction on each. Theme produced a significant main effect on every dimension, with effect sizes from ηp² = .065 for familiarity to .189 for quality. Narrative perspective was significant on most dimensions but with much smaller effects on some, including agreement (ηp² = .009) and trust (ηp² = .013).
The pattern the authors draw out is dimension-specific. Individual narratives held an advantage on emotional dimensions such as empathy, arousal, and willingness to share, while grand narratives had a slight edge on cognitive dimensions including agreement, relevance to life, trustworthiness, and familiarity. The theme-by-perspective interaction was significant in 11 of 13 dimensions, with the largest interactions on trustworthiness, empathy, and agreement (ηp² > .14), and familiarity was the one dimension without a significant interaction. Simple-effect follow-ups located the pattern in specific themes: for ecological environment, individual narratives scored higher on 10 dimensions; for transportation, grand narratives were higher on agreement, relevance, trustworthiness, and familiarity.
EEG distinguishes themes more reliably than narrative frames
Power spectrum density analysis compared band power across delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–25 Hz), and gamma (25–49 Hz) bands at four representative electrodes: Fz, Cz, Pz, and Oz. The theme main effect was significant in most regions for delta, theta, alpha, and beta, which the authors read as neural-level discriminability of the stimulus themes. Gamma showed no significant theme effect in any region.
Narrative perspective behaved differently, appearing only in specific band-and-region combinations: significant in the delta band at Cz and Oz, and in the theta band only at Pz. The theme-by-perspective interaction was significant across multiple bands and regions, strongest in theta at Pz and Oz. Outlier handling was conservative, with 6 participants excluded for collection failures or excessive noise and 1.44% to 3.52% of values excluded per dependent variable (mean 1.95%), and BH-FDR correction was applied separately across the 20 tests for each effect family.
Platform counts measure reach, not content quality
The engagement metrics were collected on 18 November 2023 and span a wide gradient, with play counts from 856 to 10,451,354. Selection deliberately covered low, medium, and high popularity within each theme and balanced official institutional accounts against individual creators, so the set is not a sample of typical videos but a spread across reach levels.
The authors are explicit that these counts are ecological indicators of collective behavior rather than intrinsic measures of content quality. Recommendation algorithms, uploader popularity and credibility, posting time, exposure duration, topic familiarity, and platform traffic allocation all shape them. The recommended use is modeling large-scale behavioral tendencies and relative engagement patterns, not absolute content evaluation, and any link drawn between neural or subjective measures and these counts should be read as exploratory and correlational.
Start with theme contrasts, treat forecasting as preliminary
The strongest evidence in the paper supports using the dataset for stimulus-level contrasts. Theme differences appeared across nearly all subjective dimensions and across four of five EEG bands in most regions, and the theme-by-perspective interaction was widespread in both ratings and EEG. Narrative perspective, by contrast, produced smaller and more localized effects, and its one-video-per-cell design means those effects cannot be cleanly separated from clip-specific properties.
That ordering should guide use. A researcher asking whether naturalistic videos elicit distinguishable neural and evaluative responses has direct support here; a researcher asking whether individual brain responses forecast population engagement has aligned data but no validated predictive framework, and the authors describe neuroforecasting applications as preliminary investigations of potential associations. The dataset also ships raw and preprocessed EEG, stimulus annotations, and MNE-Python preprocessing scripts, while the original videos are withheld for copyright reasons, so stimulus content must be reconstructed from the provided annotations.
Sources used in this answer
A multimodal dataset bridging brain activity, cognition, and real-world collective behavior
Provides 64-channel EEG from 65 participants watching 15 naturalistic videos, 13-dimension subjective ratings, trait measures, and Bilibili engagement metrics, enabling exploratory cross-level analysis rather than validated prediction.