Variable Length Plackett-Luce (VLPL) Model

• Variable Length Plackett-Luce (VLPL) is a ranking model that extends the classic Plackett-Luce framework to support arbitrary, partial rankings with query-specific lengths.
• It enables joint optimization over both the ordering of items and their allocated presentation lengths, improving metrics like expected attractiveness in web search and recommendations.
• The model employs robust estimation techniques such as MLE, composite likelihood, and policy gradients, and is applied in fields like learning-to-rank, social choice, and feature-driven preference modeling.

The Variable Length Plackett-Luce (VLPL) model is a generalization of the classic Plackett-Luce (PL) framework in which each observed ranking can be of arbitrary, query- or agent-specific length. This class subsumes a family of discrete choice, learning-to-rank, and preference learning models where the output space is expanded to partial or truncated permutations. Recent research on VLPL further extends the model to accommodate joint optimization over both the ordering of items and associated presentation lengths, as in the variable-length display setting for web search or recommendation, where the allocation of presentation space per item is itself part of the ranking decision (Knyazev et al., 29 Jun 2025).

1. Model Definition and Extensions

The standard PL model, parameterized by a positive-worth vector $\theta = (\theta_1, \ldots, \theta_n)$, generates full rankings via a multi-stage choice rule where, at each stage, the next item is selected proportionally to its worth among survivors. VLPL extends this to partial rankings of arbitrary length $k \leq n$, sampled without replacement, by assigning:

$$P(i_1 \succ \cdots \succ i_k \mid \theta) = \prod_{j=1}^k \frac{\theta_{i_j}}{\sum_{\ell=j}^k \theta_{i_\ell}},$$

where only the top-$k$ prefix is observed (Han et al., 2023, Turner et al., 2018). VLPL models can be further enriched along several axes:

• Ties and Partial Rankings: The Davidson–Luce extension models arbitrary-order ties by writing each ranking as a sequence of tie-sets, normalizing over all possible ties at each stage (Turner et al., 2018).
• Feature-Driven Parameters: In modern variants, $\theta_i$ can be instantiated as functions of features, e.g., $\theta_i = \exp(\beta^\top x_i)$ where $x_i$ denotes item features, and agent- or context-dependence is introduced (Zhao et al., 2020).
• Presentation Length Joint Modeling: For display optimization, the output space is expanded to $$y = \big[(d_1, l_1), \ldots, (d_{|y|}, l_{|y|})\big]$$ under slot budget $\sum_i l_i = K$, requiring simultaneous selection of both permutation and length allocation (Knyazev et al., 29 Jun 2025).

2. Probabilistic Structure and Inference

In the general case with presentation lengths, the variable-length PL defines a stochastic policy $\pi(y)$ over valid $k \leq n$0 (rankings and length assignments) via sequential sampling:

$k \leq n$1

$k \leq n$2

where $k \leq n$3 is a learned score, and $k \leq n$4 is the starting slot for document $k \leq n$5 (Knyazev et al., 29 Jun 2025).

For preference learning and social choice, VLPL enables observation models corresponding to arbitrary-length top-$k \leq n$6 lists (i.e., each agent or judge outputs a list of variable length), or, if needed, explicitly models the distribution over list lengths $k \leq n$7 (Zhao et al., 2020).

3. Training Objectives and Statistical Properties

The canonical objective is to maximize expected utility over the ranking-length pairs. In web ranking, expected attractiveness (EA) is formalized as:

$k \leq n$8

where $k \leq n$9 is the attractiveness (e.g., click probability) of document $$P(i_1 \succ \cdots \succ i_k \mid \theta) = \prod_{j=1}^k \frac{\theta_{i_j}}{\sum_{\ell=j}^k \theta_{i_\ell}},$$0 at length $$P(i_1 \succ \cdots \succ i_k \mid \theta) = \prod_{j=1}^k \frac{\theta_{i_j}}{\sum_{\ell=j}^k \theta_{i_\ell}},$$1, and $$P(i_1 \succ \cdots \succ i_k \mid \theta) = \prod_{j=1}^k \frac{\theta_{i_j}}{\sum_{\ell=j}^k \theta_{i_\ell}},$$2 is an exposure/observation probability model (e.g., based on DCG or $$P(i_1 \succ \cdots \succ i_k \mid \theta) = \prod_{j=1}^k \frac{\theta_{i_j}}{\sum_{\ell=j}^k \theta_{i_\ell}},$$3) (Knyazev et al., 29 Jun 2025).

For pure statistical settings, the likelihood is given by

$$P(i_1 \succ \cdots \succ i_k \mid \theta) = \prod_{j=1}^k \frac{\theta_{i_j}}{\sum_{\ell=j}^k \theta_{i_\ell}},$$4

and estimation proceeds via MLE (full or marginal), with mathematical analysis establishing uniform consistency and asymptotic normality under minimal graph expansion conditions (Han et al., 2023).

The VLPL with features involves a strictly concave log-likelihood with closed-form expressions for gradient and Hessian, with identifiability contingent on the full-rankness of the normalized feature matrix $$P(i_1 \succ \cdots \succ i_k \mid \theta) = \prod_{j=1}^k \frac{\theta_{i_j}}{\sum_{\ell=j}^k \theta_{i_\ell}},$$5 (Zhao et al., 2020).

4. Learning Algorithms and Estimation Techniques

VLPL models admit several estimation procedures:

• Full/Marginal Likelihood MLE: Direct maximization (often via Newton-Raphson or quasi-Newton methods, e.g., L-BFGS) of the log-likelihood; guarantees of strict concavity and closed-form updates for certain parameter blocks (e.g., length distribution $$P(i_1 \succ \cdots \succ i_k \mid \theta) = \prod_{j=1}^k \frac{\theta_{i_j}}{\sum_{\ell=j}^k \theta_{i_\ell}},$$6) (Zhao et al., 2020, Turner et al., 2018).
• Composite Marginal Likelihood: For scalability, especially with feature-rich or large candidate sets, reduce top-$$P(i_1 \succ \cdots \succ i_k \mid \theta) = \prod_{j=1}^k \frac{\theta_{i_j}}{\sum_{\ell=j}^k \theta_{i_\ell}},$$7 fitting to composite likelihood over pairwise comparisons via rank-breaking (Zhao et al., 2020).
• Iterative Scaling / MM: Hunter-style minorization-maximization (MM) algorithms; directly applicable to worth parameters and tie-sets (Turner et al., 2018).
• Stochastic Gradient and Policy Gradients: For complex objectives with structured exposure (e.g., in search), apply REINFORCE-style list-wise gradient estimation, with variance reduction by policy importance reweighting and reward sharing techniques (e.g., in VLPL-2) (Knyazev et al., 29 Jun 2025).

A concise pseudocode structure is provided in (Knyazev et al., 29 Jun 2025), leveraging sampling of extended rankings, deterministic filtering for budgeted slot assignment, helper arrays for empirical gradient computation, and optimizing via Adam.

5. Theoretical Results and Distinguishing Properties

VLPL introduces fundamental differences from standard PL/LTR models:

• Probability Ranking Principle (PRP) Breakdown: The optimal order may not follow the monotonic order of document relevance if presentation length/exposure interacts with position, even when $$P(i_1 \succ \cdots \succ i_k \mid \theta) = \prod_{j=1}^k \frac{\theta_{i_j}}{\sum_{\ell=j}^k \theta_{i_\ell}},$$8 is monotone in $$P(i_1 \succ \cdots \succ i_k \mid \theta) = \prod_{j=1}^k \frac{\theta_{i_j}}{\sum_{\ell=j}^k \theta_{i_\ell}},$$9 (Knyazev et al., 29 Jun 2025).
• No Separability: There does not exist a decomposition such that $k$0; joint modeling over order and length is required (counterexamples provided) (Knyazev et al., 29 Jun 2025).
• Identifiability: For feature-parameterized VLPL, identifiability is equivalent to $k$1; empirical proportions determine $k$2, while feature degeneracy impedes uniqueness of $k$3 (Zhao et al., 2020).
• MLE Consistency & Efficiency: Under random hypergraph designs for observations, both MLE and composite likelihood estimators are uniformly consistent and asymptotically normal, with computational-statistical trade-offs controlled by list length, likelihood structure, and underlying graph properties (Han et al., 2023).

6. Practical Implementations and Applications

VLPL is realized in several practical contexts:

• Web Search and Recommendation: Simultaneous optimization over ranking and snippet length for document presentation in search engines or recommender systems, where variable presentation length reallocates exposure and user attention (Knyazev et al., 29 Jun 2025).
• Social Choice and Preference Aggregation: Analysis of partially observed rankings from surveys, experiments, or competitions; VLPL models variable-length lists and accommodates ties, disconnected datasets (pseudo-comparisons with ghost items), and covariate-driven subgroup detection (via Plackett-Luce trees) (Turner et al., 2018).
• Feature-Driven Preference Models: Integration of agent/item features enables personalized ranking and inference in domains such as product recommendation or opinion aggregation (Zhao et al., 2020).

The R package PlackettLuce provides practical utilities for fitting VLPL, estimating comparison intervals via quasi-variances, and performing model-based recursive partitioning (Turner et al., 2018).

Empirical studies report substantial gains: in LTR settings, variable-length baselines outperform fixed-length by 20–25% in expected attractiveness, with joint VLPL models further exceeding these by over 10% absolute EA (significantly, $k$4), and highly sample-efficient policy gradient variants that scale to large datasets (Knyazev et al., 29 Jun 2025).

7. Limitations and Future Directions

Current VLPL methodologies exhibit several limitations:

• Continuous Presentation Lengths: Models thus far address discrete snippet lengths; extension to continuous (pixel-level) allocations is an open direction (Knyazev et al., 29 Jun 2025).
• Multi-Objective Extensions: Incorporation of additional criteria such as fairness or explainability has not been fully explored (Knyazev et al., 29 Jun 2025).
• User Study Calibration: Exposure functions $k$5 typically use heuristic choices; empirical calibration via real user behavior is needed (Knyazev et al., 29 Jun 2025).
• Generalization Across Devices: Systematic study of length-allocation models and their fit to mobile vs. desktop presentation formats remains to be addressed (Knyazev et al., 29 Jun 2025).

In summary, the Variable Length Plackett-Luce family provides a rigorous, extensible, and empirically validated foundation for ranking problems with partial, feature-driven, and joint combinatorial outputs. The model admits provably consistent statistical estimators, efficient learning algorithms, and practical adaptations for learning-to-rank, social choice, and feature personalized ranking tasks (Knyazev et al., 29 Jun 2025, Turner et al., 2018, Han et al., 2023, Zhao et al., 2020).