P-Arpan/Mood_Based_Music_Search

Dataset Card for Acoustic_Features_From_Spotify

This dataset is created by merging multiple Kaggle datasets of Spotify audio features and track metadata into a unified, clean, and deduplicated collection of 2,577,667 unique tracks. Each record is indexed by Spotify track_id (with planned support for ISRC identifiers in future iterations).

Dataset Details

Dataset Description

Acoustic_Features_From_Spotify consolidates acoustic properties and descriptive track metadata extracted via the Spotify Web API across five distinct public Kaggle datasets.

The dataset underwent an extensive offline multi-stage ETL and cleaning pipeline:

1. Column schema homogenization across disparate data formats.
2. Two-tier deduplication resolving identical entries and conflicting duplicates using tolerance-based heuristics.
3. Stringified list cleanup for collaborative artist names.
4. Statistically backed imputation for missing time signatures.
5. Removal of entries missing all audio features.

This dataset represents the clean, unscaled tabular feature set stored in high-performance Apache Parquet format, optimized for Music Information Retrieval (MIR), recommendation algorithms, and clustering tasks.

• Curated by: Arpan Paul (P-Arpan)
• Language(s): Multilingual (Track metadata covers global music releases)
• License: Open Database License (ODbL) v1.0 / Spotify Developer Terms of Service

Dataset Sources

• Repository: Moody GitHub Repository
• Hugging Face Hub: P-Arpan/Mood_Based_Music_Search
• Source Kaggle Repositories:
  • thedevastator/spotify-tracks-genre-dataset
  • joebeachcapital/30000-spotify-songs
  • amitanshjoshi/spotify-1million-tracks
  • maharshipandya/-spotify-tracks-dataset
  • rodolfofigueroa/spotify-12m-songs

Uses

Direct Use

• Music Recommendation Systems: Collaborative and content-based filtering using vector embeddings or nearest-neighbor searches (e.g., k-NN, ANN).
• Audio Classification & Clustering: Unsupervised grouping of tracks based on acoustic similarities (e.g., energy, valence, danceability, tempo).
• Mood & Emotion Recognition: Modeling relationships between valence/energy quadrants and perceived emotional valence.
• Data Analytics & MIR Research: Long-term historical trend analysis in music production attributes (loudness wars, track duration compression, tempo shifts).

Out-of-Scope Use

• Audio Synthesis / Generative Audio: This dataset does not contain raw audio files (.wav, .mp3) or spectrogram representations.
• Commercial Audio Reselling: Violating Spotify API Terms of Service or rights owner copyrights.
• Absolute Ground Truth for Emotion: Acoustic features (e.g., valence, danceability) are Spotify algorithmic approximations and should not be treated as objective psychological ground truths.

Dataset Structure

The dataset consists of 2,577,667 rows and 19 columns stored in a single .parquet file.

Field Name	Type	Missing Count	Description
track_id	string	0	Unique 22-character alphanumeric Spotify track identifier (Base-62).
artists	string	0	Name(s) of contributing artist(s). Multiple artists are joined with semicolons (;).
album_name	string	1,099,519	Name of the containing album (unrecorded in select source subsets).
track_name	string	0	Title of the track.
duration_ms	float64	0	Track length measured in milliseconds.
explicit	float64	1,099,531	Binary flag: 1.0 if track contains explicit lyrics, 0.0 otherwise.
danceability	float64	0	Describes how suitable a track is for dancing (0.0 to 1.0).
energy	float64	0	Perceptual measure of intensity and activity (0.0 to 1.0).
key	float64	0	Estimated overall musical key using standard Pitch Class notation (0 = C, 1 = C♯/D♭, ..., 11 = B).
loudness	float64	0	Overall loudness of the track in decibels (dB), averaged across duration (typically -60 to 0 dB).
mode	float64	0	Modality of the track: 1.0 = Major, 0.0 = Minor.
speechiness	float64	0	Detects the presence of spoken words (0.0 to 1.0; >0.66 indicates spoken word/podcasts).
acousticness	float64	0	Confidence score that the track is acoustic (0.0 to 1.0).
instrumentalness	float64	0	Predicts whether a track contains no vocals (0.0 to 1.0; >0.5 represents instrumental).
liveness	float64	0	Detects audience presence in the recording (0.0 to 1.0; >0.8 indicates strong likelihood of live performance).
valence	float64	0	Musical positiveness conveyed by a track (0.0 = sad/depressed to 1.0 = cheerful/euphoric).
tempo	float64	0	Overall estimated tempo in beats per minute (BPM).
time_signature	float64	0	Estimated overall time signature (meter, typically 3 to 5). Missing values imputed to 4.0.
dataset_name	string	0	Provenance tag indicating the source CSV file origin during multi-dataset merging.

Dataset Creation

Curation Rationale

Public Spotify datasets on Kaggle are fragmented, frequently repeat tracks with subtle floating-point variations, contain disparate column headers (id vs track_id, artist_name vs artists), and often represent collaborative artists as escaped python strings ("['Artist A', 'Artist B']"). This dataset consolidates, reconciles, and cleans those disparate sources into a single, standardized, deduplicated parquet dataset of over 2.57 million tracks.

Source Data

Data Collection and Processing

The data processing pipeline is implemented across modular Python scripts and Jupyter notebooks:

1. Acquisition (data_download.py): Automated download of five primary Spotify feature datasets using kagglehub.
2. Schema Normalization & Initial Merge (merging.py):
   • Standardized column headers into snake_case.
   • Unified aliases: id → track_id, artist_name/track_artist → artists, album → album_name, name → track_name.
   • Filtered out non-acoustic columns while tracking source provenance via dataset_name.
3. a. Conflict Resolution & Deduplication (duplicate_cleaning.ipynb):
   • Identified duplicate track_id entries with varying values.
   • Categorized differences into rounding errors vs. true conflicts based on defined thresholds:
     • Δ duration_ms > 5000 ms
     • Δ tempo > 3.0 BPM
     • Δ key ≥ 1 or Δ mode ≥ 1
     • Δ (danceability, energy, valence, acousticness) > 0.05
   • Minor discrepancies were resolved by rounding audio features to 4 decimal places.
   • For real conflicts, entries with the highest metadata completeness (least NaN count) were prioritized and kept.
4. b. Recording-level deduplication:
   • The same recording often has many track_ids (one per album or playlist).
   • Rows with the same lowercased track_name + artists were treated as one recording if:
     • duration_ms differed by ≤ 1000 ms,
     • danceability/energy/valence/acousticness by ≤ 0.01,
     • loudness by ≤ 0.5 dB,
     • tempo by ≤ 1 BPM,
     • and key and mode matched exactly.
   • One row per recording was kept (the one with the fewest missing values). Remasters, live versions, and re-recordings outside these tolerances are kept as separate tracks.
5. Metadata & Feature Cleaning (cleaning.ipynb):
   • Artist String Cleanup: Extracted stringified Python lists ("['Artist 1', 'Artist 2']") via ast.literal_eval and serialized them into semicolon-delimited strings ("Artist 1;Artist 2"), preserving literal bracketed artist names (e.g., "[Alexandros]").
   • Explicit Flag Normalization: Converted boolean flags to binary float representation (1.0 / 0.0).
   • Missing Audio Features: Dropped entries where all acoustic features were NaN.
   • Time Signature Imputation: Evaluated 15,704 missing time_signature entries against the broader population. Because the median/mean acoustic vector of the missing subset closely matched tracks with time_signature == 4.0 (standard 4/4 common time), missing values were imputed to 4.0.

Who are the source data producers?

The acoustic features were algorithmically generated by the Spotify Audio Analysis API (formerly Echo Nest algorithms). The tabular datasets were initially collected and posted on Kaggle by community members: The Devastator, Joe Beach Capital, Amitansh Joshi, Maharshi Pandya, and Rodolfo Figueroa.

Annotations

Annotation process

No manual subjective human annotation was performed. All acoustic descriptors are machine-derived measurements from Spotify's audio analysis pipeline.

Who are the annotators?

Not applicable (algorithmic feature extraction).

Personal and Sensitive Information

The dataset contains only publicly available commercial metadata (track names, artist names, album titles, and acoustic features). It contains no personally identifiable information (PII) or user listening histories.

Bias, Risks, and Limitations

• Algorithmic Black-Box: Features like danceability, energy, and valence are generated by proprietary Spotify models. Their internal weights and exact acoustic definitions are closed-source.
• Missing Album & Explicit Data: Approximately 1.10M rows lack album_name or explicit ratings because certain contributing Kaggle datasets omitted these columns during their original scraping.
• Imputed Time Signatures: A total of 15,704 tracks had missing time_signature values imputed as 4.0 based on statistical feature similarity.
• Geographic & Genre Representation: The underlying scraping leans toward catalog availability on Spotify at the time of each Kaggle release, which historically skews toward Western commercial music.

Recommendations

Users should be aware that Spotify's acoustic features are algorithmic estimations. When training models sensitive to scale (such as SVMs, neural networks, or distance-based k-NN), consider standardizing unbounded features (duration_ms, loudness, tempo) or consult the standardized companion parquet file (standardasised_acoustic_features.parquet).

Licensing and Terms of Use

• Dataset license: This compiled database is released under the Open Database License (ODbL) v1.0, because it is derived from source databases that carry ODbL, DbCL, and CC0 terms (see Dataset Sources). If you share a database built from this one, ODbL requires you to keep the same license and credit the sources.
• Origin of the audio features: The audio features (danceability, energy, valence, etc.) were generated by Spotify's audio analysis and collected by third-party Kaggle contributors. This dataset does not claim ownership of that underlying data.
• Spotify terms: Use of Spotify-derived data is subject to the Spotify Developer Terms of Service. Users are responsible for checking that their use complies with those terms.
• Source licensing gap: One of the five source datasets does not declare a license on Kaggle.
• Purpose: This dataset is shared for research and educational purposes. It is not legal advice, and no warranty is given.

Citation

BibTeX:

@misc{paul2026spotifyfeatures,
  author       = {Arpan Paul},
  title        = {Acoustic_Features_From_Spotify (2.57M Deduplicated Tracks)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/P-Arpan/Mood_Based_Music_Search}}
}

APA:

Paul, A. (2026). Acoustic_Features_From_Spotify (2.57M Deduplicated Tracks) [Data set]. Hugging Face. https://huggingface.co/datasets/P-Arpan/Mood_Based_Music_Search

Glossary

• BPM (Beats Per Minute): Tempo measure reflecting the speed of the track's rhythmic pulse.
• ISRC: International Standard Recording Code, the international standard code for uniquely identifying sound recordings.
• Pitch Class: An integer mapping (0–11) representing musical semitones (0 = C, 1 = C♯/D♭, ..., 11 = B).
• Valence: A metric from 0.0 to 1.0 describing the musical positiveness conveyed by a track. High valence sounds positive (e.g., happy, cheerful, euphoric), while low valence sounds negative (e.g., sad, depressed, angry).

More Information

This dataset forms the core offline tabular features database for the Moody music search and mood recommendation engine.

Dataset Card Authors

Arpan Paul (P-Arpan)

Dataset Card Contact

For questions or updates regarding the preprocessing pipeline, please submit an issue on the Moody Repository or Hugging Face dataset discussion board.