Characteristic Characteristics

While five-factor models of personality are widespread, there is still not universal agreement on this as a structural framework. Part of the reason for the lingering debate is its dependence on factor analysis. In particular, derivation or refutation of the model via other statistical means is a worthwhile project. In this paper we use the methodology of spectral clustering to articulate the structure in the dataset of responses of 20,993 subjects on a 300-item item version of the IPIP NEO personality questionnaire, and we compare our results to those obtained from a factor analytic solution. We found support for five- and six-cluster solutions. The five-cluster solution was similar to a conventional five-factor solution, but the six-cluster and six-factor solutions differed significantly, and only the six-cluster solution was readily interpretable: it gave a model similar to the HEXACO model. We suggest that spectral clustering provides a robust alternative view of personality data.

💡 Research Summary

The paper investigates whether spectral clustering (SC), a network‑based unsupervised learning technique, can provide an alternative view of personality structure to the traditional factor‑analytic (FA) approach. Using a large online sample of 20,993 respondents who completed a 300‑item International Personality Item Pool (IPIP) version of the NEO‑PI‑R, the authors first construct a weighted adjacency matrix A from the item‑item correlation matrix. Correlations are transformed into Euclidean distances on a unit sphere, then into similarities via a Gaussian kernel with scale parameter σ. By varying σ between 0.1 and 1, they explore how edge strength changes and how the resulting graph’s connectivity evolves.

From A they compute the symmetric normalized Laplacian L = I – D⁻¹ᐟ² A D⁻¹ᐟ², where D is the degree matrix. The eigenvalues of L are examined; small non‑zero eigenvalues correspond to well‑separated clusters. An ad‑hoc “spectral scree” method is used to select the number of eigenvectors l to retain. Across the examined σ range, the first four eigenvalues consistently separate from the bulk, leading the authors to fix l = 4 and restrict σ to values where the graph remains connected (σ ≥ 0.35).

To determine the optimal number of clusters k, the authors introduce a “cluster consistency” procedure. For each σ and a fixed l, they repeatedly (100–200 times) draw a random subset of 150 items, run SC with a candidate k, and compute the proportion of items whose cluster assignments differ from the full‑data solution. The k that yields the lowest misclassification rate is deemed most stable. This analysis shows that for σ between roughly 0.4 and 1, k = 5 or k = 6 consistently produce the smallest error (≈30 % misclassification), while larger k values do not improve stability.

Having identified plausible 5‑ and 6‑cluster solutions, the authors compare them to varimax‑rotated factor solutions derived from the same data. For the five‑factor model, 286 of 300 items (95 %) receive identical assignments under both methods, indicating strong concordance between SC and FA at the domain level. In contrast, the six‑factor FA yields a tiny sixth factor (only seven items, largely from a single facet) that lacks substantive meaning. The six‑cluster SC solution, however, isolates a robust sixth cluster of 34 items, primarily drawn from Agreeableness (especially the Morality and Modesty facets) and Conscientiousness (Dutifulness). This pattern mirrors the Honesty‑Humility factor of the HEXACO model, suggesting that SC uncovers a latent dimension that FA fails to capture with the same data.

The authors note that the sixth SC cluster aligns with HEXACO’s predictions: it consists mainly of items reflecting moral integrity, modesty, and truthfulness, and its correlations with the other clusters resemble those reported for HEXACO’s H factor. They also observe that the Emotionality cluster absorbs items from Openness‑O3, and that the traditional Neuroticism items (particularly Angry Hostility) remain grouped with Emotionality rather than Agreeableness, a deviation from the HEXACO mapping.

Importantly, the SC analysis does not reveal a clear facet‑level structure (≈30 clusters) within the data; the misclassification curve shows no pronounced drop around 30, implying that the 300‑item IPIP‑NEO may not contain sufficient granularity for reliable facet detection using this method.

In discussion, the authors argue that SC offers a complementary perspective to FA: while FA maximizes explained variance, SC minimizes inter‑cluster edge weight, thereby emphasizing community cohesion in the correlation network. The convergence of SC on both five‑ and six‑cluster solutions suggests that the underlying personality data may simultaneously support the Five‑Factor Model (FFM) and a six‑dimensional structure akin to HEXACO. This duality underscores the importance of employing multiple analytic lenses when evaluating personality taxonomy.

The paper concludes that spectral clustering is a robust, scalable alternative for personality data analysis, capable of revealing dimensions that traditional factor analysis may overlook. Its application to a large, heterogeneous sample validates the presence of both the classic five‑factor structure and an additional Honesty‑Humility dimension, providing empirical support for the HEXACO model and encouraging further network‑based investigations of personality architecture.