Explicit Factor Models (EFM)

• Explicit Factor Models are approaches that replace latent dimensions with measurable, domain-specific attributes, ensuring model transparency.
• They integrate data-driven feature extraction and optimization methods to improve prediction accuracy, evidenced by RMSE reductions and enhanced CTR in recommendation.
• EFMs directly map predictions to explicit features, facilitating actionable insights and attribution in applications such as recommender systems and asset return modeling.

Explicit Factor Models (EFM) are a class of interpretable factorization approaches wherein the traditionally latent dimensions of factor models are anchored to observable, domain-specific, and semantically meaningful attributes. Unlike classical latent factor models—widely applied in recommender systems and financial risk modeling, where the factors lack inherent interpretability—EFMs embed human-interpretable features, enabling both accurate prediction and direct attribution of outcomes to explicit explanatory variables. This attribute-level transparency facilitates model explainability, result justification, and actionable insights in application domains ranging from user-item recommendation to cross-sectional asset return modeling (Zhang, 2017, Grotmol et al., 2022).

1. Motivation and Conceptual Framework

The central motivation for EFMs arises from the limitations of latent-factor models (e.g., SVD, PMF, PCA), where the learned factors represent abstract dimensions without direct semantic correspondence to observable variables. In many settings, such as recommender systems and portfolio risk analysis, the lack of intuitive explanations undermines user trust, hinders effective system auditing, and restricts actionable deployment. EFMs address this deficit by grounding factor dimensions in explicitly measured or extracted features—such as product attributes mined from user reviews or firm-specific characteristics sourced from market and accounting data.

This approach is characterized by constructing the factor-loading matrices and exposure vectors using data-driven extraction (e.g., sentiment analysis in recommendation, firm fundamentals in finance), thus ensuring coherence between the mathematical representation and the domain concepts of interest (Zhang, 2017, Grotmol et al., 2022).

2. Mathematical Formalizations

2.1 Recommender Systems

Let $R\in\mathbb{R}^{m\times n}$ denote the observed user-item rating matrix. Classical latent factorization seeks to approximate $R\approx\widehat R=UV^\top$, with $U\in\mathbb{R}^{m\times K}$ (user factors) and $V\in\mathbb{R}^{n\times K}$ (item factors):

$$\min_{U,V} L(U,V) = \sum_{(u,i)\in\mathcal D}\Bigl(r_{u,i}-\sum_{k=1}^K U_{u,k}V_{i,k}\Bigr)^2 + \lambda(\|U\|_F^2+\|V\|_F^2).$$

EFMs enhance interpretability by enforcing that the $k$ dimensions correspond to explicit features $F = \{f_1, ..., f_p\}$, extracted, for example, via phrase-level sentiment analysis from reviews. They introduce an additional penalty to ensure that the product $U_{u,f}V_{i,f}$ aligns with the sentiment $s_{u,i,f}$ the user has expressed for feature $f$ in item $R\approx\widehat R=UV^\top$0:

$R\approx\widehat R=UV^\top$1

yielding the full optimization objective:

$R\approx\widehat R=UV^\top$2

2.2 Financial Factor Models

In asset return modeling, Exabel's EFM expresses the return vector $R\approx\widehat R=UV^\top$3 at time $R\approx\widehat R=UV^\top$4 for $R\approx\widehat R=UV^\top$5 companies as:

$R\approx\widehat R=UV^\top$6

where $R\approx\widehat R=UV^\top$7 encodes explicit factor exposures (style, country, industry), $R\approx\widehat R=UV^\top$8 denotes factor returns, and $R\approx\widehat R=UV^\top$9 represents idiosyncratic residuals. Each factor is constructed from economic or readily observable quantities (e.g., volatility, dividend yield, country of domicile).

Explicitness in this setting requires all matrix entries in $U\in\mathbb{R}^{m\times K}$0 to derive from transparent and economically meaningful characteristics, contrasting with statistical factor models (e.g., PCA) where factor loadings are opaque linear combinations lacking interpretability (Grotmol et al., 2022).

3. Feature Extraction, Data Processing, and Integration

EFMs require robust procedures for extracting and aggregating explicit features:

• In recommender systems, phrase-level sentiment analysis is performed on user-generated review corpora. NLP pipelines tokenize text, perform POS-tagging and dependency parsing, extract $U\in\mathbb{R}^{m\times K}$1feature, opinion$U\in\mathbb{R}^{m\times K}$2 pairs, and assign sentiment scores (polarity). The resulting user-feature "concern" matrices ($U\in\mathbb{R}^{m\times K}$3) and item-feature "quality" matrices ($U\in\mathbb{R}^{m\times K}$4) are constructed as frequency-normalized and sentiment-aggregated values, respectively (Zhang, 2017).
• In finance, explicit style, country, and industry exposures—such as volatility (price standard deviation over 91 days), book-to-price ratio, or revenue growth—are calculated directly from historical time-series and fundamental databases (Grotmol et al., 2022).

This direct mapping from raw data to factor structure is a defining property of EFMs, ensuring that all analytic outputs can be attributed to human-interpretable domain concepts.

4. Model Training, Optimization, and Computational Aspects

Optimization in EFMs typically employs alternating least squares or stochastic gradient descent (SGD) for matrix factorization, with additional penalties and constraints to enforce alignment with explicit features. For recommendation:

• Training iterates over observed ratings and feature-level sentiments.
• Each SGD pass scales with the sum of observed rating events and relevant sentiment entries.
• Bordered Block Diagonal Form (BBDF) decomposes the user-item matrix into parallelizable blocks, facilitating multi-core acceleration (4–6× speedup on 8 cores reported) (Zhang, 2017).

In financial models, weighted least squares (WLS)—with weights proportional to market capitalization—ensures that larger entities have greater influence in parameter estimation. Identifiability constraints (zero-centering of style exposures; zero-sum country and industry returns) guarantee that factor returns capture systematic effects and not arbitrary drift (Grotmol et al., 2022).

5. Explanatory Mechanisms and Empirical Performance

5.1 Natural Language and Attribution

EFMs facilitate explicit, instance-level explanation. Recommendations articulate the contribution of specific features, e.g., "We recommend item $U\in\mathbb{R}^{m\times K}$5 because it has strong $U\in\mathbb{R}^{m\times K}$6 (battery-life) and you've shown positive opinion on that feature." Negative recommendation logic explicitly states feature-level deficiencies (Zhang, 2017).

In finance, explicit factor decomposition allows portfolio attribution to style, country, and industry effects, with risk and return breakdowns directly mapped to observable company characteristics (Grotmol et al., 2022).

5.2 Empirical Evaluation

Recommender Systems

• EFM reduces RMSE by 5–8% versus best latent-only models; dynamic extensions (Fourier-assisted ARIMA on feature popularity) achieve further 3% reduction.
• NDCG improves 10% over NMF, with an additional 10% from dynamic modeling.
• A/B tests yield click-through rate (CTR) increase from 3.2% to 4.3% when using feature-level explanations.
• Approximately 15% reduction in undesirable purchases when flagging negative recommendations.
• RMSE benefits plateau as the number of explicit factors increases to 30–60, with interpretability and accuracy degrading if too many (noisy) features are included (Zhang, 2017).

Financial Models

• Full explicit factor models (market, 11 styles, 50 countries, 14 industries) explain ~41.5% (in-sample) and ~37.6% (cross-validated) of 1-day return variance, and ~44.7%/41.2% for 90-day returns.
• Country factors contribute the largest marginal increment to explained variance, with style and industry factors further enhancing explanatory power.
• Constructed factor-based portfolios (e.g., Value, Momentum, Growth) exhibit characteristic performance profiles, with some outperforming the market cap benchmark over extended periods but also displaying cyclicality across years (Grotmol et al., 2022).

6. Strengths, Limitations, and Usage Considerations

Domain	Advantages	Limitations
Recommendation	- High interpretability<br>- Actionable explanations<br>- Dynamic adaptivity (time series extension)	- Reliant on quality of review analysis<br>- Possible overfitting with many noisy features
Finance	- Each factor interpretable<br>- Stable exposures<br>- Attribution clarity<br>- Portfolio construction transparency	- May omit latent sources of risk<br>- Factor collinearity<br>- Nonstationarity<br>- Residual risk (60% of daily variance unexplained)

In recommender systems, EFMs present a pragmatic trade-off between predictive accuracy and explanation fidelity, as interpretability grows with the number of explicit features up to a threshold, beyond which overfitting to noisy or unstable features degrades utility.

In finance, the explicit enumeration of style and fundamental factors delivers stable and interpretable exposures, but may overlook latent structures not captured in the explicit taxonomy. Correlation among style factors reduces marginal benefit from adding further explicit dimensions, and a significant portion of risk remains unexplained without recourse to statistical methods or stochastic volatility models (Zhang, 2017, Grotmol et al., 2022).

7. Extensions and Research Directions

EFMs admit extension to temporal adaptation, as in dynamic user preference modeling using methods such as Fourier-assisted ARIMA (FARIMA), which model daily evolution of feature popularity for time-aware recommendation. In portfolio applications, explicit factor models provide a framework for systematic risk evaluation, stress testing, and targeted portfolio tilting.

A plausible implication is that further hybridization of explicit and latent approaches may mitigate coverage limitations while preserving explanatory power, although this raises new challenges in balancing interpretability, model complexity, and generalization (Zhang, 2017, Grotmol et al., 2022).