Greatest Hits Dataset Overview

• Greatest Hits Dataset is a curated collection that integrates hit items from music, network, and astronomical domains using authoritative external standards.
• It combines extensive feature engineering—ranging from audio descriptors to tensor-based network features—to facilitate predictive modeling and data-driven insights.
• The dataset underpins ranking methodologies like Random Forests and MD-HITS, achieving high prediction accuracies and robust centrality measures across applications.

The term "Greatest Hits Dataset" encompasses a range of curated, annotated, or engineered datasets used to identify, model, and predict the most significant entities—most commonly musical hits but also notable nodes or events—in a given domain. In research spanning music information retrieval, web network analysis, and astronomical surveys, the concept has been operationalized both as a collection of items of exceptional notability (e.g., hit songs, influential papers, principal asteroids), and as a structured labeling or ranking problem enabling advanced machine learning and network science applications.

1. Dataset Construction and Characteristics

The construction of a Greatest Hits Dataset typically involves aggregating items recognized as "hits" by established external standards (such as Billboard Hot 100 charts, citation lists, or detection events in scientific surveys) and integrating them with comprehensive feature representations and ground-truth annotations.

In music hit prediction research (Herremans et al., 2019, Middlebrook et al., 2019, Dimolitsas et al., 2023, Tseng et al., 2024), datasets are compiled by:

• Collecting track listings from authoritative sources (Billboard, Official Charts Company).
• Obtaining rich audio descriptors from platforms and APIs (The Echo Nest, Spotify Web API).
• Merging audio features (e.g., timbre, rhythm, loudness, danceability, energy, valence) with metadata such as artist history and release date.
• Creating balanced subsets for statistical modeling by randomly sampling non-hit tracks to match the number of hits (Middlebrook et al., 2019, Dimolitsas et al., 2023).
• For memorability research, extracting structurally meaningful 5-second clips and annotating these via human experiments to obtain objective memorability scores (Tseng et al., 2024).

Outside music, the term applies to multi-dimensional ranking or event datasets. For network analysis, the MD-HITS model (Arrigo et al., 2018) uses complex, multi-layered adjacency tensors representing nodes, layers (contexts), and time stamps—enabling simultaneous ranking of the most significant entities across dimensions, often interpreted as the "greatest hits" in network activity.

In astronomical surveys (Peña et al., 2020), serendipitous observations of large numbers of events (e.g., asteroid detections) are linked and characterized to construct a dataset in which the principal bodies (e.g., those most frequently or distinctly observed) become the greatest hits for analysis.

2. Feature Engineering and Annotation Strategies

Greatest Hits Datasets contain not only the item identities but also a broad spectrum of features designed for predictive modeling:

Dataset Domain	Features	Ground Truth
Song Prediction	Audio descriptors (timbre, energy, tempo), meta-info (artist, year), temporal statistics	Chart position, memorability labels
Network Ranking	Adjacency tensor, node-layer-time statistics	Centrality measures, rank scores
Asteroid Survey	Observation time series, color (g', r'), orbital elements	Detection frequency, physical properties

Music-based datasets explicitly leverage both "static" features (mean values over tracks) and "temporal" features (beat difference, time-evolving timbre vectors) (Herremans et al., 2019). Advanced annotation protocols employ experimental setups (e.g., "music memory games" for memorability) and consistent ground-truth metrics (e.g., proportion of recognitions, chart appearances).

In network analytics, the model engineering is driven by tensor operations and the definition of nonlinear, multi-dimensional centrality vectors (Arrigo et al., 2018). For scientific surveys, feature extraction involves transformations (e.g., barycenter coordinates for celestial bodies), power-law fitting, and discovery of intrinsic correlations or their absence (e.g., color-size relations) (Peña et al., 2020).

3. Ranking Methodologies and Algorithms

Identifying "greatest hits" within complex datasets is mathematically formalized through ranking models and classification algorithms:

• Music Hit Prediction: Decision Trees, Rulesets, Naive Bayes, Logistic Regression, Random Forests, Support Vector Machines (with kernels and cross-validation), and Multilayer Perceptrons are employed. Random Forests consistently yield top accuracies (up to 88% (Middlebrook et al., 2019), 86% (Dimolitsas et al., 2023)), with logistic regression sometimes prevailing for robustness (Herremans et al., 2019).
• Memorability Prediction: Support Vector Regression and end-to-end deep spectrogram models (SSAST) are compared, with ablation studies revealing optimal feature sets (e.g., top 25 handcrafted descriptors, pitch-shift augmentation) (Tseng et al., 2024).
• Multi-dimensional Network Ranking: The MD-HITS model defines five centrality vectors (node-hubs, authorities; layer-broadcasting, receiving; time stamps) via the Perron eigenvector of a multi-homogeneous order-preserving map. A globally convergent, parallelizable power-iteration algorithm computes these vectors regardless of network connectivity (Arrigo et al., 2018).
• Web Graph Efficiency: For large web datasets rife with dangling nodes, lumping and similarity transformations yield smaller hub matrices, accelerating and simplifying centrality computation (Dong et al., 2021).

4. Empirical Results and Performance Analysis

Empirical studies consistently report robust performance using Greatest Hits Datasets, with the following findings:

• Hit Song Prediction: High accuracy is achievable using only audio features (Herremans et al., 2019, Dimolitsas et al., 2023), with ensemble models capturing nonlinear dynamics and mitigating overfitting (Middlebrook et al., 2019).
• Memorability: Predictive models reach Spearman’s rank correlations up to 0.2988 using SVR on the top-25 handcrafted features and competitive results via SSAST with mel-spectrogram input and pitch augmentation (Tseng et al., 2024).
• Network Analysis: MD-HITS ensures the uniqueness of centrality scores even for disconnected or weakly connected graphs, with convergence in 10–30 iterations and robust scaling (Arrigo et al., 2018).
• Asteroid Surveys: A steeper size distribution slope (~0.9) is found, distinct from previous surveys; cadence issues in color measurement are identified as critical for future work (Peña et al., 2020).

Notably, temporal features (e.g., evolving timbre vectors, beat statistics) are emphasized for their importance in musical prediction, while network paradigms rely on multi-dimensional nonlinear mapping to extract robust global rankings.

5. Application Domains and Utility

Greatest Hits Datasets have enabled significant advances and cross-domain applications:

• Music Industry: Hit prediction models inform A&R decisions, algorithmic marketing, adaptive recommendations, and research in "hit song science" (Herremans et al., 2019, Middlebrook et al., 2019, Dimolitsas et al., 2023).
• Music Memorability: New measures of cognitive "stickiness" refine recommender systems, music style transfer, and marketing targeting, with future prospects for personalization and broader analytics (Tseng et al., 2024).
• Network Science: MD-HITS applicability ranges from multiplex citation analysis and trade network evaluation to urban infrastructure modeling, identifying key agents ("greatest hits") in temporal and layered systems (Arrigo et al., 2018).
• Astronomy: Large-scale surveys using event-based Greatest Hits Datasets support population characterization and data-driven guidelines for survey design (e.g., LSST cadence planning) (Peña et al., 2020).
• Web Analytics: Efficient computation on massive digital networks allows ranking of principal items, products, or pages, reducing computational costs (Dong et al., 2021).

6. Methodological Limitations and Future Research

Current limitations include scale restrictions (particularly in memorability datasets), the impact of evolving trends and features not captured by static audio descriptors, class imbalance challenges, and the dependency on available ground-truth standards.

Future research directions articulated in the literature encompass:

• Dataset expansion to more fully represent musical structure and broader temporal patterns (Tseng et al., 2024).
• Advances in transfer learning and feature engineering for enhanced generalizability and adaptation (Tseng et al., 2024, Middlebrook et al., 2019).
• Algorithmic refinements for multi-dimensional, multi-modal network ranking—including the further application of lumping methodology and its extension to other models beyond HITS (Dong et al., 2021, Arrigo et al., 2018).
• Personalization and XAI interpretability for memorability prediction and music curation (Tseng et al., 2024).
• Optimization of survey cadence and observing strategies for astronomical event datasets (Peña et al., 2020).

7. Technical and Mathematical Foundations

Underlying the construction and exploitation of Greatest Hits Datasets are precise mathematical frameworks:

• Nonlinear Eigenvector Equations:

$$F^{(\alpha)}(\mathbf{c}) = \lambda \otimes \mathbf{c}$$

where $F^{(\alpha)}$ is a multi-homogeneous nonlinear map acting on tuples of centrality vectors, and $\lambda$ is a positive vector of scaling factors (Arrigo et al., 2018).

• Network Tensor Slicing and Multi-modal Centrality: Each centrality component is computed by slicing the adjacency tensor along a corresponding mode and applying power-law nonlinearity for global fixed-point existence and uniqueness (Arrigo et al., 2018).
• Classification Metrics: Statistical metrics such as accuracy, precision, recall, area under the ROC curve, and Spearman’s correlation rank the effectiveness of models (Herremans et al., 2019, Dimolitsas et al., 2023, Tseng et al., 2024).
• Data Transformation and Feature Selection: PCA, feature scaling, and selection protocols (e.g., CfsSubsetEval, GeneticSearch, SHAP) optimize model input and interpretability (Dimolitsas et al., 2023, Tseng et al., 2024).

This ensemble of technical, methodological, and empirical innovations renders the Greatest Hits Dataset a foundational resource for predictive, ranking, and analytical tasks in contemporary computational research.