Binary Five Factor Model (BIG5)

• Binary Five Factor Model (BIG5) is a framework that converts continuous personality traits into binary representations for simplified, interpretable analysis.
• It employs mathematical methods such as binary loading matrices and tangle-based clustering to unveil trait hierarchies and enhance prediction accuracy.
• The model integrates techniques like zero-shot LLM prompting and regression binarization on text and behavior data, achieving up to 91.2% accuracy in trait prediction.

The Binary Five Factor Model (BIG5) refers to a set of methods, data encodings, and evaluation protocols for personality trait analysis where each of the five canonical Big Five traits—Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism—is represented in binary form: present (1) or absent (0). In contrast to classical continuous factor scores, the binary formulation is operationalized for interpretability, model comparability, and practical utility in computational personality prediction, particularly for text-driven, social signal, and language modeling applications.

1. Canonical Binary Formalism and Trait Definitions

In the binary BIG5 paradigm, each input (e.g., text instance, behavior trace) receives a label vector $y\in\{0,1\}^5$, with each entry indicating the predicted presence or absence of the corresponding trait. The formal definition used in automatic personality prediction from text (APPT) tasks is:

• $y_{i,t}=1$ if trait $t$ is present in sample $x_i$; $y_{i,t}=0$ otherwise.
• Traits and operational cues:
  • Openness ($O$): 1 if the input shows "rich ideas, novelty, abstract thinking".
  • Conscientiousness ($C$): 1 if "planning, order, detail orientation" is manifested.
  • Extraversion ($E$): 1 if "outgoing tone, social engagement" is detected.
  • Agreeableness ($A$): 1 if "compassion, warmth, collaborative language" is present.
  • Neuroticism ($N$): 1 if "worry, mood swings, anxiety" appears.

Each trait is thus a separate, interpretable classification target, aligning with prior APPT research conventions for cross-study comparability (Cursi et al., 28 Nov 2025).

2. Mathematical and Algorithmic Foundations

2.1 Binary Factor Model Structures

Traditionally, continuous models posit $f_1,\dots,f_5$ as real-valued latent factors for each individual, with item responses modeled as:

$$P(r_i(q)=1)=\sigma(\beta_q + \sum_{k=1}^5 \lambda_{qk}f_k(i))$$

In the strictly binary formalization, items ("questions") index a binary loading matrix $L\in\{0,1\}^{Q\times 5}$, where $L_{qk}=1$ if item $q$ is intended to load on factor $k$, and $L_{qk}=0$ otherwise. Each participant's response vector $r_i\in\{0,1\}^Q$ induces a vector of binary trait endorsements $f(i)\in\{0,1\}^5$ (Bergen et al., 2024).

2.2 Tangle-Based Validation and Refinement

A recent mathematically rigorous approach replaces latent-continuous modeling with tangle-based clustering (Bergen et al., 2024). Here, personality traits are conceptualized as robust clusters within the item set, discovered through the computation of tangles—orientations of item-set partitions (separations) that pass a set of combinatorial intersection tests. The method reveals a trait hierarchy:

• At low resolution, traits like Neuroticism and Conscientiousness emerge first.
• Combined or supertraits (e.g., Agreeableness–Extraversion) and subtraits are discovered as the resolution parameter $k$ increases.

This binary tangle structure allows one to define and validate binary factors at varying levels of granularity and test for the cohesion and distinctness of classical trait groupings.

3. Modeling and Operationalization From Behavioral and Linguistic Data

3.1 Text-Based Prediction Protocols

Under the binary BIG5 scheme, prediction is typically performed as five parallel binary classification tasks per input, using the following approaches:

• Zero-shot LLM prompting: LLMs receive minimal or enriched prompts defining the trait and return 0/1 per trait per sample. Enriched prompts (containing trait definitions, facet cues, and adjective lists) reduce invalid outputs and improve class balance, but introduce a positive bias toward predicting trait presence (Cursi et al., 28 Nov 2025).
• Regression with post hoc binarization: Trait scores predicted as continuous values (e.g., $Y^T \in [1,5]$ via neural regression, as in BIG5-TPoT (Le et al., 12 Nov 2025)) can be thresholded, e.g., $1[Y^T \geq 3.0]$, to yield binary labels.

3.2 Next-Token Probability and Latent Factor Discovery

Log-probabilities of trait-descriptive adjectives in LLMs can be decomposed (SVD) to uncover latent directions matching the Big Five, with projections of new data yielding trait scores whose signs provide binary high/low trait labels. This method achieves trait prediction accuracies up to 91.2%, outperforming direct prompt-based LLM scoring by 21 percentage points (Suh et al., 2024).

3.3 Behavioral Coding from Video and Signals

Low-level features extracted from crowd videos (speed, orientation, collectivity, socialization) are mapped via deterministic functions to OCEAN scores, with binarization enabled by quantization steps. The majority of items in these models relate to Extraversion, affecting trait-specific sensitivity (Favaretto et al., 2019).

4. Evaluation Protocols and Metrics

Binary BIG5 systems are evaluated with metrics suited for dichotomous outcomes:

• Accuracy: Proportion of correctly predicted 0/1 labels.
• Per-class precision/recall/F1: Calculated for both presence (1) and absence (0), macro-averaged to account for class imbalance:

$$\text{Macro-F1} = \frac{1}{2}(\mathrm{F1}_0+\mathrm{F1}_1)$$

with $\mathrm{Recall}_0$, $\mathrm{Recall}_1$ diagnosing class-wise biases.

• “Within-0.5” accuracy: Fraction of samples where regression-predicted trait scores fall within 0.5 of ground truth.

Reporting per-class recall and invalid output rates is critical, as accuracy and macro-F1 can mask asymmetries and trivial solutions (e.g., always predicting the majority class) (Cursi et al., 28 Nov 2025).

5. Empirical Findings and Trait-Detection Asymmetries

Experimental evaluation across essays, social media text, and Reddit data exhibits the following patterns (Cursi et al., 28 Nov 2025):

• Openness and Agreeableness are consistently easier to predict in binary form (F1 scores for both classes often $\geq0.5$).
• Extraversion and Neuroticism present persistent difficulty: few models/datasets achieve F1$_1\geq0.5$.
• Open-source LLMs can approach GPT-4 performance with enriched prompts, but no zero-shot configuration is universally reliable.
• Enriched prompts shift predictions positively (increasing Recall$_1$ by 10–30 percentage points, decreasing Recall$_0$), sometimes correcting under-prediction but risking false positives. Reporting only aggregate metrics (accuracy, macro-F1) is insufficient.

Empirical studies with text filtering (BIG5-TPoT) show that semantically focusing input on trait-relevant material increases both accuracy and mean absolute error gains across all five traits. Binarized predictions from these models are directly usable for binary BIG5 tasks (Le et al., 12 Nov 2025).

6. Hierarchies, Refinements, and Theoretical Extensions

Tangle analysis reveals that the Big Five trait structure is neither strictly flat nor uniquely defined at one resolution:

• Each trait, as a binary cluster, becomes clearly defined at specific levels of data granularity.
• Traits may combine (supertraits) or split (subtrait refinement) as resolution is varied (e.g., an Agreeableness–Extraversion supercluster, two Openness subtraits centering on vocabulary/abstract ideation versus imagination).
• The tangle-tree structure provides a binary factor model where each item or feature is optimally assigned to an internal or leaf trait, sidestepping requirements for orthogonality or linearity (Bergen et al., 2024).
• Test construction, evaluation, and refinement can proceed by identifying which items are core or peripheral to a binary trait cluster.

7. Best Practices, Limitations, and Open Issues

Current evidence indicates that binary BIG5 prediction from high-dimensional behavioral or language data remains sensitive to configuration:

• Prompt engineering: Careful enrichment and calibration are required to avoid invalid outputs and trait over-prediction.
• Evaluation: Disaggregated metrics (per-class recall, invalid rates) are essential for diagnostic interpretability.
• Bias and imbalance: Trait class skew and dataset composition heavily impact efficacy and generalization.
• Model validity: Zero-shot LLMs, even when anchor-enriched, are not yet fully robust for binary BIG5 APPT; mild improvements over regression binarization or classical methods are context-dependent (Cursi et al., 28 Nov 2025).

Further research should emphasize calibrated decision thresholds, lightweight adaptation methods (e.g., parameter-efficient fine-tuning), and systematic reporting of dataset and evaluation properties for reproducible comparison and model advancement. Tangle-based binary representation and semantic text preselection (e.g., TPoT) provide theoretically motivated, empirically validated advances in both the definition and application scope of the binary Five Factor Model (Bergen et al., 2024, Le et al., 12 Nov 2025, Cursi et al., 28 Nov 2025).