MoodArchive: Mood-Based Data Archives

• MoodArchive is a framework of mood-indexed systems that transform raw media into structured emotional repositories.
• The image domain version includes over 8 million images with 27 GoEmotions-based labels, employing both automated annotation and human validation.
• Diverse methodological strategies—from discrete emotion labels to continuous valence–arousal modeling—enable robust affective computing and sentiment tracking.

Searching arXiv for the cited MoodArchive-related papers to ground the article in current arXiv records. I’ll look up the relevant arXiv entries for "MoodArchive", "Moodifier", and closely related mood-archive systems. MoodArchive denotes a family of archive-oriented systems, datasets, and pipelines for organizing, inferring, and operationalizing mood-related information across heterogeneous media. In the most specific recent usage, MoodArchive is an image dataset introduced in connection with "Moodifier: MLLM-Enhanced Emotion-Driven Image Editing," where it comprises 8 million+ images with hierarchical emotional annotations generated by LLaVA and partially validated by human evaluators (Ye et al., 18 Jul 2025). In a broader technical sense, the term also functions as a design pattern for mood-centric repositories in music information retrieval, affective computing, public-sentiment monitoring, and longitudinal mental-health analysis, where the archive stores mood annotations, embeddings, trajectories, or aggregate indices derived from audio, images, text, interaction logs, or sensor streams (Korzeniowski et al., 2020, Sasaki et al., 2021, Tsakalidis et al., 2022, Pellert et al., 2020).

1. Conceptual scope and domain variants

MoodArchive, as instantiated across the cited literature, is not a single canonical architecture. Rather, it refers to archive structures in which mood is the primary indexing, retrieval, or analytical variable. In the image domain, MoodArchive is a large-scale corpus for emotion-conditioned vision-language modeling and image editing; the dataset contains 8,240,312 images, four image contexts per emotion, and 27 discrete emotion labels derived from GoEmotions (Ye et al., 18 Jul 2025). In music-oriented work, analogous mood archives are built either from expert annotations linked to audio collections, as in the AllMusic Mood Subset with 66,993 tracks and 188 mood tags (Korzeniowski et al., 2020), or from social tags projected into interpretable affective spaces, as in Affective Circumplex Transformation over Last.fm mood tags (Saari et al., 2013).

The same label is also used in design summaries for systems that archive collective or personal mood over time. Nation-wide Mood aggregates individual mood estimates from web-search queries and mobile sensor data into a Nation-wide Mood Score over more than 11,000,000 users (Sasaki et al., 2021). Dashboard-based public monitoring of sentiment in Austrian social media provides a self-updating, file-based archive of daily emotional indicators during COVID-19 (Pellert et al., 2020). Longitudinal user-text systems store mood-change annotations such as Switch and Escalation events over 500 manually annotated timelines and 18.7K posts (Tsakalidis et al., 2022). This suggests that MoodArchive is best understood as an umbrella category for mood-indexed data infrastructures rather than as a single modality-specific resource.

A recurrent architectural theme is the transformation of raw observations into compact affect representations that support retrieval, prediction, clustering, or trend analysis. Depending on the domain, those representations may be discrete emotion labels, multi-label mood vectors, valence–arousal coordinates, latent embeddings derived from listening behavior, or longitudinal event annotations (Korzeniowski et al., 2020, Ye et al., 18 Jul 2025, Saari et al., 2013).

2. Image-domain MoodArchive dataset

The image-domain MoodArchive introduced with Moodifier contains 8 million+ images with emotion annotations, and the detailed dataset summary reports an overall size of 8,240,312 images (Ye et al., 18 Jul 2025). The collection spans four image contexts per emotion: facial expressions, natural scenery, urban scenery, and object classes such as fashion items, jewelry, and home décor. Images were collected from permissively licensed or royalty-free web sources including unsplash.com, pexels.com, pixabay.com, elements.envato.com, openverse.org, burst.shopify.com, stocksnap.io, foodiesfeed.com, and freenaturestock.com (Ye et al., 18 Jul 2025).

Its annotation taxonomy is based on GoEmotions and contains 27 discrete emotion labels: admiration, amusement, approval, caring, desire, excitement, gratitude, joy, love, optimism, pride, relief, anger, annoyance, disappointment, disapproval, disgust, embarrassment, fear, nervousness, grief, remorse, sadness, confusion, curiosity, realization, and surprise (Ye et al., 18 Jul 2025). Annotation is generated automatically with LLaVA-NeXT using a prompt that requests a concise global summary, three distinct local emotional stimuli, and an overall emotion assessment. The resulting per-image output is a structured 5-sentence caption comprising one summary sentence, three stimulus phrases, and one overall label (Ye et al., 18 Jul 2025).

Human validation was conducted on 10,000 randomly sampled images in a Mechanical Turk study comparing original alt-text against the LLaVA-generated caption. The reported results are that 85% of LLaVA captions were preferred over alt-text and 15% were rejected, with cited failure modes including mis-detections, inappropriate wording, and cultural misalignment; inter-annotator agreement is not reported (Ye et al., 18 Jul 2025). Automated filtering removes NSFW content and corrupted files, computes CLIP cosine similarity between each image and its target-emotion text prompts, and drops the bottom 20% of images by CLIP similarity to reduce label noise. Formally, for image $I_i$ and target-emotion text $T_i$, the CLIP similarity is

$s_i = \cos(E_v(I_i), E_t(T_i)),$

and retained items satisfy $s_i \ge \tau$, where $\tau$ is the 20th percentile of all $s_i$ values in the preliminary crawl (Ye et al., 18 Jul 2025).

The released format is described as original-resolution image files in JPEG or PNG and a JSON annotation record with fields file_path, global_summary, stimulus_phrases, emotion_label, clip_similarity_score, source_url, and license (Ye et al., 18 Jul 2025). No canonical train/validation/test split is published; instead, a human-verified subset, MoodArchive-5k, containing 5,000 image–caption pairs is held out for retrieval evaluation (Ye et al., 18 Jul 2025). Intended use cases include pretraining and fine-tuning of vision-LLMs, zero-shot emotion classification and image-text retrieval, and downstream applications such as emotion-driven image editing, affective image captioning, and emotion-conditioned generative modeling (Ye et al., 18 Jul 2025).

3. Annotation schemas and affect representations

A central distinction across MoodArchive variants is the choice of affect representation. The image MoodArchive uses 27 discrete emotion labels with structured language descriptions and local stimulus phrases (Ye et al., 18 Jul 2025). By contrast, MooD for affective image editing directly models emotion in continuous Valence–Arousal space, with valence and arousal each in $[-1,1]$, and learns an affect encoder $f_{\mathrm{aff}}: I \to (v,a)$ trained by an $\ell_2$ regression loss,

$$L_{\mathrm{aff}} = \|(\hat v,\hat a) - (v_{\mathrm{gt}},a_{\mathrm{gt}})\|_2^2.$$

Its AffectSet dataset contains 63,387 VA-annotated images, with quadrant counts of 38,604 for $T_i$0, 4,011 for $T_i$1, 13,088 for $T_i$2, and 7,684 for $T_i$3 (Yin et al., 4 May 2026). This contrast between discrete labels and continuous VA coordinates highlights two complementary strategies: explicit categorical labeling and low-dimensional affective geometry.

Music-focused archives similarly vary in representational choice. The AllMusic Mood Subset unfolds album-level AllMusic annotations to tracks, producing a multi-label inventory of 188 distinct moods over 66,993 tracks, with a mean of 9.1 moods per track, standard deviation $T_i$4, and median $T_i$5 (Korzeniowski et al., 2020). An earlier music architecture maps songs into Thayer’s two-dimensional emotion space using audio-derived Intensity, Timbre, and Rhythm, supplemented by a one-dimensional lyric valence score; the quadrant assignment yields Exuberance, Frantic/Anxious, Depression, and Contentment (Singh et al., 2012). ACT, in turn, builds an explicit valence–arousal–tension representation from Last.fm social mood tags by aligning a data-driven MDS term space to reference emotion coordinates through Procrustes transformation (Saari et al., 2013).

Text and behavioral archives often represent mood temporally rather than taxonomically. In longitudinal user text, the key labels are abrupt Switches and gradual Escalations. A Switch at time $T_i$6 is defined by a sign change between consecutive mood scores, $T_i$7, while an Escalation is a contiguous monotonic run culminating in a local extremum (Tsakalidis et al., 2022). In nation-scale search and sensor analysis, self-reported mood on a 7-point Likert scale is collapsed into three classes $T_i$8 for negative, neutral, and positive mood (Sasaki et al., 2021). These choices illustrate that mood archives can target categorical, dimensional, multi-label, or event-centric formulations depending on the downstream task.

A plausible implication is that interoperability across MoodArchive systems depends less on a universal label set than on explicit mappings among representations such as discrete emotions, VA coordinates, and longitudinal state transitions.

4. Modeling and retrieval mechanisms

The modeling layer of a MoodArchive system typically converts raw content or behavior into embeddings or scores usable for retrieval and prediction. In the AllMusic Mood Subset, listening-based track embeddings are obtained by weighted matrix factorization of the listener–track play-count matrix with embedding dimension $T_i$9:

$s_i = \cos(E_v(I_i), E_t(T_i)),$0

The rows of $s_i = \cos(E_v(I_i), E_t(T_i)),$1 serve as 200-dimensional TP-L track embeddings (Korzeniowski et al., 2020). The same work compares these listening embeddings to audio-derived baselines from Musicnn and a Short-Chunk CNN, and then trains a transfer-learning MLP $s_i = \cos(E_v(I_i), E_t(T_i)),$2 with four hidden layers, ReLU activations, dropout $s_i = \cos(E_v(I_i), E_t(T_i)),$3, a 188-dimensional sigmoid output, and binary cross-entropy loss (Korzeniowski et al., 2020).

In Moodifier, MoodArchive supports a training-free image-editing pipeline through MoodifyCLIP and multimodal LLMs (Ye et al., 18 Jul 2025). The dataset’s structured annotations are intended to translate abstract emotions into specific visual attributes. Closely related work in MooD shows how retrieval and generative control can be coupled in a continuous affect space. Its VA-Aware retrieval first filters candidate images within Euclidean radius $s_i = \cos(E_v(I_i), E_t(T_i)),$4 of the target $s_i = \cos(E_v(I_i), E_t(T_i)),$5 in VA space, then re-ranks them by cosine similarity of OpenCLIP ViT-bigG/14 image embeddings to the source image, returning the semantically compatible reference $s_i = \cos(E_v(I_i), E_t(T_i)),$6 (Yin et al., 4 May 2026). MooD’s editor conditions SDXL through two branches: visual transfer via an Image Adapter and semantic guidance via an Affective Semantic Projection Network that maps affect encoder features into learned token sequences aligned with CLIP text-embedding spaces (Yin et al., 4 May 2026).

Music archives built on semantic tags employ a different retrieval logic. ACT converts sparse mood-tag vectors into coordinates in an interpretable valence–arousal–tension space by computing TF–IDF weights, projecting through SVD and non-metric MDS, and applying a Procrustes transform. A track’s mood coordinates are then obtained by weighted centroids of term coordinates, and specific mood strength is estimated by projection onto the unit vector of a mood term (Saari et al., 2013). Earlier architecture-based systems instead combine audio and lyrics, define a fused valence score $s_i = \cos(E_v(I_i), E_t(T_i)),$7, keep arousal from audio, and assign one of four quadrant clusters (Singh et al., 2012).

Behavioral and textual MoodArchive systems rely on sessional or temporal fusion. DeepMood processes three views of keyboard metadata—alphanumeric keystrokes, special-character events, and accelerometer readings—using per-view bidirectional GRUs followed by fully connected, factorization-machine, or multi-view-machine fusion layers; the best reported held-out HDRS performance is 90.31% accuracy, and YMRS regression yields RMSE $s_i = \cos(E_v(I_i), E_t(T_i)),$8 (Cao et al., 2018). Nation-wide Mood trains a Random Forest sensor mood model on approximately 113 sensor features and then a regularized linear or logistic query mood model over an approximately 81k query vocabulary, using the sensor model to augment labels for sessions lacking questionnaires (Sasaki et al., 2021). In longitudinal social-media analysis, context-aware sequential modeling stacks a BiLSTM over BERT representations to capture rare mood-change events more effectively than post-level baselines (Tsakalidis et al., 2022).

5. Evaluation regimes and empirical findings

MoodArchive systems are evaluated according to their operational role: classification accuracy for mood inference, correlation with human ratings for semantic mood coordinates, retrieval quality for archive search, fidelity and controllability for generative editing, or temporal sensitivity for event detection.

For the AllMusic Mood Subset, the primary metric is macro-averaged Average Precision over 188 tags,

$s_i = \cos(E_v(I_i), E_t(T_i)),$9

Overall test results are 0.51 for TP-L listening embeddings, 0.57 for proprietary P-L embeddings, 0.40 for MCN-MSD-A, 0.35 for SCC-A, and 0.30 for MCN-MTT-A (Korzeniowski et al., 2020). Listening embeddings significantly outperform the strongest audio baseline, MCN-MSD-A, with $s_i \ge \tau$0 in a paired $s_i \ge \tau$1-test over track-wise AP. Only 20 tags favor the audio model over TP-L, and those tags cluster into a single “audio-friendly” mood subspace (Korzeniowski et al., 2020). Pairwise Pearson correlations of $s_i \ge \tau$2 show that audio models correlate above $s_i \ge \tau$3 with one another, whereas audio-versus-listening correlations are at most about $s_i \ge \tau$4, indicating complementary information (Korzeniowski et al., 2020).

ACT evaluates mood predictions against listener ratings on a held-out set of 600 tracks. Median Spearman correlations for ACT are $s_i \ge \tau$5 for Valence, $s_i \ge \tau$6 for Arousal, and $s_i \ge \tau$7 for Tension, exceeding VSM, SVD, NMF, and PLSA baselines; all reported comparisons are significant at $s_i \ge \tau$8 by Wilcoxon tests (Saari et al., 2013). Robustness experiments show that even with a single tag per track, ACT reaches median correlations of $s_i \ge \tau$9 for Valence, $\tau$0 for Arousal, and $\tau$1 for Tension (Saari et al., 2013).

For MoodArchive in image editing, the direct empirical claims are attached to Moodifier at a high level: extensive experimental evaluations show that Moodifier outperforms existing methods in both emotional accuracy and content preservation (Ye et al., 18 Jul 2025). More granular measurements are available for the related continuous-VA framework MooD. Against EmoEdit, Make Me Happier, SDEdit, ControlNet, and InstructPix2Pix, MooD reports Time/MP $\tau$2, PSNR $\tau$3, SSIM $\tau$4, LPIPS $\tau$5, CLIP-I $\tau$6, V-Err $\tau$7, and A-Err $\tau$8 (Yin et al., 4 May 2026). Ablations indicate that full VA-Aware retrieval balances affective control and semantic fidelity better than VA-only or CLIP-only retrieval, and that adding ASP-Net reduces V-Err from $\tau$9 to $s_i$0, A-Err from $s_i$1 to $s_i$2, while increasing CLIP-I from $s_i$3 to $s_i$4 (Yin et al., 4 May 2026).

Temporal and large-scale mood archives use task-specific metrics. DeepMood reports 90.31% accuracy on depression score prediction from session-level typing dynamics (Cao et al., 2018). Nation-wide Mood reports 72.0% accuracy and macro-$s_i$5 for the sensor mood model, with query mood model accuracy improving from 87.0% without sensor-based augmentation to 94.2% with it (Sasaki et al., 2021). In longitudinal user-text analysis, BiLSTM-bert achieves the best reported post-level macro-$s_i$6 of $s_i$7, along with window precision $s_i$8, window recall $s_i$9, coverage precision $[-1,1]$0, and coverage recall $[-1,1]$1 at window size $[-1,1]$2 (Tsakalidis et al., 2022). These results suggest that mood archives are most effective when their evaluation protocol matches the archive’s intended operational semantics: tag ranking for multi-label corpora, coordinate correlation for semantic projections, and temporally tolerant metrics for change detection.

6. Infrastructure, curation, and reproducibility

MoodArchive implementations differ substantially in storage, update cadence, and release policy. The image MoodArchive stores original-resolution image files and JSON annotations, with planned release via the project website and annotations under CC-BY-NC; end users must also comply with original source terms (Ye et al., 18 Jul 2025). The AMS music subset provides code, data splits, and pre-computed embeddings publicly at a GitHub repository for reproducible research (Korzeniowski et al., 2020). Earlier song-clustering systems use XML as an intermediate “live” database and combine Matlab plus MIRToolbox for audio, Perl for lyric processing, and C#/.NET for orchestration and presentation (Singh et al., 2012).

Public-monitoring archives often adopt incremental, file-based pipelines. The Austrian COVID-19 dashboard stores raw JSON transformed into TSV or CSV files under version control, re-renders an HTML dashboard daily via a cronjob, deploys through GitHub Pages, and is automatically archived by the Austrian National Library’s web crawler (Pellert et al., 2020). The daily workflow includes data retrieval from derstandard.at, Crimson Hexagon Twitter counts, and a student chat API; LIWC-based scoring; computation of relative changes against weekday-conditioned baselines; generation of plotly time series and word clouds; and automated git commit and push (Pellert et al., 2020). The system is reported to run in under 15 minutes per day on a 2 GHz CPU with 4 GB RAM (Pellert et al., 2020).

Nation-wide Mood operates at production scale every 3 hours by extracting query logs, forming binary query vectors, scoring users with a linear query mood model, and computing

$[-1,1]$3

over active users (Sasaki et al., 2021). Its deployment notes emphasize anonymization, aggregate-only reporting, and coverage under a commercial privacy policy (Sasaki et al., 2021). Longitudinal clinical-style archives place stronger emphasis on consent, deletion, auditability, and storage minimization; the depression-oriented Facebook mood-profile design proposes relational tables for users, posts, daily sentiment, and windowed features, plus opt-in and data-portability safeguards (Chen et al., 2020).

A persistent curation issue is the relationship between annotation richness and noise. MoodArchive filters images by CLIP similarity percentiles (Ye et al., 18 Jul 2025); AMS inherits album-level AllMusic moods unfolded to tracks (Korzeniowski et al., 2020); ACT filters mood tags through a curated lexicon and removes terms with fewer than 100 track associations (Saari et al., 2013); GlobalMood avoids predefined categories during elicitation and then consolidates language-specific descriptors with native-speaker validation (Lee et al., 14 May 2025). This suggests that reproducible mood archives depend not only on scale, but also on transparent filtering, alignment, and release protocols.

7. Limitations, controversies, and future directions

Across domains, MoodArchive systems inherit nontrivial annotation and representational limitations. In AMS, AllMusic tags are proprietary, assigned at the album level, and potentially noisy when unfolded to tracks; overlapping or near-synonym moods are not merged; only 30-second mp3 previews are used; and moods are modeled as flat binary labels rather than continuous arousal/valence (Korzeniowski et al., 2020). In MoodArchive for images, human validation reports an 85% acceptance rate on sampled captions, but inter-annotator agreement is not reported, and rejected outputs include cases of cultural misalignment (Ye et al., 18 Jul 2025). MooD’s AffectSet relies on Qwen2.5-VL-72B for VA annotations, with quality control on a stratified subset showing Pearson $[-1,1]$4 for Valence and $[-1,1]$5 for Arousal against few-shot human raters, but that still places annotation quality within a model-mediated pipeline rather than a purely human-labeled one (Yin et al., 4 May 2026).

A second issue is the tension between universal and culture-specific mood semantics. GlobalMood explicitly argues that existing music-emotion datasets focus predominantly on Western songs and English-derived terms, and its cross-cultural benchmark reveals both a shared valence–arousal structure and substantial divergences in the perception of dictionary-equivalent terms across cultures (Lee et al., 14 May 2025). For example, the mean inter-language correlation for “happy” is reported as $[-1,1]$6, compared with within-culture reliability $[-1,1]$7; equivalent terms can occupy different quadrants in MDS analyses (Lee et al., 14 May 2025). A plausible implication is that large archives using a single English-centered mood ontology may encode systematic semantic drift when transferred across languages or domains.

Future directions recur across several papers. AMS suggests multimodal fusion of audio, listening, lyrics, and metadata; mood clustering or embedding to reduce sparsity; continuous emotion modeling; personalization through listener embeddings; interpretability of latent listening factors; and zero-shot or meta-learning for unseen moods (Korzeniowski et al., 2020). MoodArchive’s intended use cases already include pretraining and fine-tuning of affective vision-LLMs, zero-shot retrieval, and emotion-conditioned generation (Ye et al., 18 Jul 2025). MooD advances fine-grained semantic control through continuous VA conditioning and efficient retrieval (Yin et al., 4 May 2026). GlobalMood recommends bottom-up lexicon elicitation, storage of original-language tags with verified English translations, continuous ratings, and open release of raw and processed data (Lee et al., 14 May 2025). Longitudinal mood archives emphasize temporal windows, majority-vote gold labels, focal loss for rare classes, and span-aware evaluation (Tsakalidis et al., 2022).

Taken together, the literature indicates that MoodArchive is evolving from static mood-labeled collections toward multimodal, temporally aware, semantically grounded infrastructures. Some variants prioritize scale and weak supervision, others prioritize interpretability or temporal sensitivity, and still others aim at controllable generation. The common objective is the same: to turn heterogeneous observational data into structured, queryable, and analytically meaningful mood representations (Ye et al., 18 Jul 2025, Korzeniowski et al., 2020, Sasaki et al., 2021).