Preference Evolution Tracking Methods

• Preference Evolution Tracking (PET) is a framework that models evolving latent states, capturing how user tastes, strategic preferences, or imaging signals change over time.
• It integrates methodologies from session-based recommendation, evolutionary models, LLM inference, and molecular imaging to reconcile dynamic behavior with structured state evolution.
• PET approaches impose temporal consistency using domain-specific mechanisms like attention-guided updates, equilibrium-based matching, or continuity equations, enabling more interpretable and stable inference.

Preference Evolution Tracking (PET) denotes a set of research formulations in which the central inferential target is an evolving latent state rather than a static endpoint. In the cited literature, the term is used most directly for preference modeling in recommendation, where the objective is to recover how user taste changes across a session or over longer horizons; it is also used to summarize evolutionary models in which preferences, partner choice, equilibrium play, and material payoffs co-determine long-run selection. Separately, the same acronym appears in positron emission tomography research, where dynamic source distributions or longitudinal lesion identities are tracked across time. This suggests a common organizing idea: PET methods replace framewise, one-shot, or end-to-end list generation with temporally coupled inference over a structured state space (Wang et al., 2023, Hu et al., 2021, Zhang et al., 29 Sep 2025, Schmitzer et al., 2019, Yousefirizi et al., 24 Mar 2026).

1. Terminological scope and research strands

The cited literature uses PET in several distinct but related senses. In recommendation, PET names models for session-level preference evolution and LLM-based distributional preference inference. In evolutionary economics and game theory, PET is a natural shorthand for tracking the feedback loop from preferences to matching, play, payoffs, and back to preference evolution. In molecular imaging, PET retains its standard imaging meaning, but the associated tracking problems are formally similar because they also infer time-varying hidden structure from sparse observations.

Strand	Object tracked	Core mechanism
Evolutionary preference models	Preference types over time	Stable matching with Nash / Bayes-Nash equilibrium
Session-based recommendation	Evolving session preference	Two-stage retrieval and attention-guided recurrent fusion
LLM-based recommendation	Probability distribution over preference clusters	Logit-probing, generative classification, hierarchical probing
Molecular imaging and theranostics	Radionuclide density or lesion correspondence	Optimal transport regularization; segmentation plus matching

A recurrent source of confusion is that PET is not a single standardized algorithmic family. The 2025 recommendation paper introduces an explicit framework called “PET: Preference Evolution Tracking with LLM-Generated Explainable Distribution” (Zhang et al., 29 Sep 2025), whereas the 2021 session-based recommendation paper and the 2023 evolutionary model study do not present the acronym as a formal framework name in their titles, even though both are organized around preference evolution as the central process (Hu et al., 2021, Wang et al., 2023). The 2026 theranostics paper uses the label more loosely, identifying longitudinal lesion tracking as the missing computational layer for PSMA theranostics (Yousefirizi et al., 24 Mar 2026).

2. Preference evolution as endogenous matching and evolutionary selection

In the evolutionary model of preference evolution with endogenous matching, PET is the feedback system

$$\text{preferences} \to \text{matching} \to \text{equilibrium play} \to \text{material payoffs} \to \text{preference evolution}.$$

Agents are matched in pairs to play a symmetric two-person simultaneous game $\Gamma$ with finite action set $X$, material payoffs $\pi(x,y)$, and preference types $\theta \in \Theta$ represented by utility functions $U_\theta(x,y,t)$, where utility may depend on own action, opponent’s action, and opponent’s preference type. The framework combines stable matching with Nash equilibrium under complete information and Bayes-Nash equilibrium under incomplete information, so the “unit of selection” is a preference type that shapes both sorting and play (Wang et al., 2023).

Under complete information, a matching profile $(\mu(\varepsilon),\Sigma)$ must satisfy internal stability and external stability. Internal stability requires Nash equilibrium within every matched pair: $$x^* \in \arg\max_{x\in\Delta(X)} U_a(x,y^*,b), \qquad y^* \in \arg\max_{y\in\Delta(X)} U_b(y,x^*,a).$$ External stability excludes blocking pairs: two agents cannot rematch and choose a Nash equilibrium that makes both strictly better off. A key refinement is the loser-best Nash equilibrium,

$\NE^\lb_\theta = \arg\max_{(x,y)\in\NE_\theta}\min\{U_\theta(x,y,\theta),U_\theta(y,x,\theta)\},$

which implies that stable matching selects same-type equilibria that maximize the welfare of the worse-off participant. The associated fitness quantities are the average material payoffs $G_\theta(\mu(\varepsilon),\Sigma)$ and $\Gamma$0, and evolutionary stability is defined by payoff dominance of the incumbent type over mutants across Nash stable profiles (Wang et al., 2023).

The main complete-information result is that a $\Gamma$1-homophilic efficient type,

$\Gamma$2

for any $\Gamma$3, and a parochial efficient type,

$\Gamma$4

are neutrally stable and evolutionarily stable against any type $\Gamma$5 that exhibits same-type inefficiency. The mechanism is positive assortative matching together with efficient play within same-type matches. By contrast, selfish preferences are stable only under restrictive conditions, and if a preference type does not exhibit plasticity and efficient outcomes are asymmetric, the type is evolutionarily unstable (Wang et al., 2023).

Incomplete information changes the stability logic because deviations themselves reveal information. The state becomes $\Gamma$6, where $\Gamma$7 specifies the fractions of observable and unobservable agents, and Bayes-Nash blocking includes complete-information blocking among observable types, one-sided incomplete-information blocking, two-sided conditional blocking, and two-sided strong-incentive blocking. The central conclusion is sharper than in the complete-information case: weak homophily is no longer enough, and if there exists a strictly efficient strategy pair $\Gamma$8 with $\Gamma$9, the $X$0-homophilic efficient type is evolutionarily unstable for any $X$1. The parochial efficient type remains neutrally stable and evolutionarily stable against any type with same-type inefficiency (Wang et al., 2023).

3. Session-based recommendation and sequence-level preference trajectories

In session-based recommendation, PET refers to explicit modeling of how user intent changes over the course of an anonymous interaction sequence. The task is: given a session

$X$2

predict the next clicked item $X$3 from the item set $X$4. The PEN4Rec framework argues that previous methods often either assume adjacent behaviors are equally informative and sequentially dependent or compress the whole session into one attention-pooled preference. Its counterclaim is that users may express multiple preferences, each preference has its own evolving track, and recent clicks can both reveal and distort current intent. The architecture therefore combines a session graph embedding stage with a two-stage preference retrieval process (Hu et al., 2021).

The first stage uses the recent $X$5 items as a local preference query,

$X$6

and computes attention weights

$X$7

followed by the initial preference vector

$X$8

The query is then updated as

$X$9

The second stage models the preference evolving trajectory dynamically. A bidirectional GRU yields context-aware representations

$\pi(x,y)$0

which are fused via

$\pi(x,y)$1

Preference evolution is then tracked by attention-guided recurrent updates,

$\pi(x,y)$2

The final state $\pi(x,y)$3 is fused into the session representation $\pi(x,y)$4, and prediction uses

$\pi(x,y)$5

This formulation makes preference evolution an explicit sequential control problem rather than a single attention pass (Hu et al., 2021).

The reported training setup uses Adam, batch size 100, dropout 0.5, $\pi(x,y)$6 penalty $\pi(x,y)$7, latent dimension 100, one GGNN layer, Gaussian initialization with mean 0 and std 0.1, and a search over $\pi(x,y)$8 with best values 3 on Yoochoose, 2 on Diginetica, and 4 on LastFM. Evaluation uses P@20 and MRR@20. PEN4Rec achieves P@20 = 71.53 and MRR@20 = 31.71 on Yoochoose, P@20 = 52.50 and MRR@20 = 18.56 on Diginetica, and P@20 = 28.82 and MRR@20 = 11.33 on LastFM, improving over the strongest baseline by 1.2% / 2.5%, 1.5% / 5.4%, and 4.6% / 6.7%, respectively. Ablations further isolate the PET mechanism: PEN4Rec-Non exceeds GNN-Last and AGNN-Last, the full model exceeds PEN4Rec-Non, PEN4Rec exceeds PEN4Rec-GRU and PEN4Rec-ATT, and it also exceeds PEN4Rec-ATT-GRU (Hu et al., 2021).

Conceptually, this strand of PET treats preference drift as structured noise that must be suppressed without discarding long-range relevance. It differs from plain GRU-like models, which can be distracted by unimportant clicks, and from pure attention models, which can identify relevance but do not model the trajectory of preference evolution over time (Hu et al., 2021).

4. LLM-based PET as explainable distributional preference inference

The 2025 framework “PET: Preference Evolution Tracking with LLM-Generated Explainable Distribution” reframes recommendation as estimating a dynamic probability distribution over a stable, interpretable set of preference clusters, rather than directly generating the next item or a top-$\pi(x,y)$9 list. If a user’s history up to time $\theta \in \Theta$0 is $\theta \in \Theta$1, PET defines the preference state as

$\theta \in \Theta$2

The paper assumes a latent utility vector $\theta \in \Theta$3 and a softmax ground-truth distribution

$\theta \in \Theta$4

Because the future true distribution is unobserved, PET constructs a proxy from the future interaction window: $\theta \in \Theta$5 Learning minimizes a distributional distance

$\theta \in \Theta$6

where $\theta \in \Theta$7 can be Jensen–Shannon divergence, and ranking sorts clusters by $\theta \in \Theta$8 (Zhang et al., 29 Sep 2025).

The stable and interpretable lattice of preference clusters is central. Clusters are stable because categories change far less than individual items, interpretable because they correspond to human-understandable concepts such as genres, business categories, or content topics, and scalable because the model reasons over structured semantic labels rather than volatile item IDs. For large label spaces, the framework uses a hierarchical cluster lattice. In Yelp, 1,311 categories are organized into 26 coarse L1 clusters and fine-grained L2 subcategories under a selected L1 branch. The training pipeline is PT + SFT with LoRA fine-tuning, using expanding history windows $\theta \in \Theta$9 paired with future distributions $U_\theta(x,y,t)$0 (Zhang et al., 29 Sep 2025).

Preference inference uses logits rather than free-form decoding. In likelihood-based probing, the model is asked a yes/no question for each cluster, affirmative and negative verbalizer scores are averaged,

$U_\theta(x,y,t)$1

converted to

$U_\theta(x,y,t)$2

and normalized: $U_\theta(x,y,t)$3 Generative classification instead maps each cluster to a unique token, reads one-step logits $U_\theta(x,y,t)$4, sets $U_\theta(x,y,t)$5, and normalizes with the same softmax. For extreme multi-label settings, hierarchical probing with strategic exploration first computes

$U_\theta(x,y,t)$6

selects L1 branches, conditionally probes L2 children, composes by

$U_\theta(x,y,t)$7

and finally normalizes to $U_\theta(x,y,t)$8 (Zhang et al., 29 Sep 2025).

The framework is motivated by three limitations of direct generation baselines: limited personalization and opaque user modeling, popularity bias, and poor transparency and auditability. PET claims an isotonicity condition,

$U_\theta(x,y,t)$9

under which sorting inferred probabilities yields the same order as sorting true latent attractiveness scores, and that this ranking is optimal for common order-aware metrics such as NDCG@k, Recall@k, and Precision@k. Experiments use MovieLens, Yelp, and a short-video platform dataset, with Qwen3-8B and DeepSeek-Distill-Llama-8B. On MovieLens, with Qwen3-8B, context window = 8, and long-term prediction, Logit-Probing reaches NDCG@10 = 0.863, reported as +55% over Direct Generation and +20% over Qwen3-Reranker-8B, with JS divergence reduced to 0.316 from 0.406. On Yelp L1, Qwen3-8B achieves NDCG@1 = 0.980, NDCG@5 = 0.879, and NDCG@10 = 0.915. On the short-video platform, the production model’s long-tail NDCG@20 is 0.0243, whereas PET with Logit-Probing reaches 0.1971, a +711% improvement (Zhang et al., 29 Sep 2025).

In this strand, PET is explicitly not direct next-item generation. It is a transparent distributional preference mapping designed for recommendation, analysis, and auditing, with claimed benefits for long-tail ranking, fairness, diversity, and interpretability (Zhang et al., 29 Sep 2025).

5. PET in molecular imaging and theranostics

In imaging, PET retains its standard association with positron emission tomography, yet the tracking problems are structurally close to preference-evolution models because they also infer latent trajectories from sparse sequential evidence. The 2019 paper on dynamic cell imaging proposes a dynamic image reconstruction method from PET listmode data for tracking single or small numbers of radiolabelled cells. The inferential targets are the dynamic radionuclide density $(\mu(\varepsilon),\Sigma)$0 and associated mass flux $(\mu(\varepsilon),\Sigma)$1, constrained by the continuity equation

$(\mu(\varepsilon),\Sigma)$2

Temporal consistency is enforced by the Benamou–Brenier action

$(\mu(\varepsilon),\Sigma)$3

which chooses, among all temporally evolving radionuclide distributions consistent with the PET measurement, the one with least kinetic motion energy. The expected counts satisfy $(\mu(\varepsilon),\Sigma)$4, measurements are Poisson, $(\mu(\varepsilon),\Sigma)$5, and the paper minimizes a penalized Poisson log-likelihood with regularization weight $(\mu(\varepsilon),\Sigma)$6, including a scatter-reweighted form with $(\mu(\varepsilon),\Sigma)$7. The problem is jointly convex in $(\mu(\varepsilon),\Sigma)$8 and is solved with a primal-dual implicit gradient descent/ascent algorithm of Chambolle–Pock type (Schmitzer et al., 2019).

The simulation setup uses Siemens Biograph mCT PET/CT geometry in GATE, an XCAT torso phantom, a 1 mm point source, $(\mu(\varepsilon),\Sigma)$9 with half-life $$x^* \in \arg\max_{x\in\Delta(X)} U_a(x,y^*,b), \qquad y^* \in \arg\max_{y\in\Delta(X)} U_b(y,x^*,a).$$0, a circular trajectory of radius $$x^* \in \arg\max_{x\in\Delta(X)} U_a(x,y^*,b), \qquad y^* \in \arg\max_{y\in\Delta(X)} U_b(y,x^*,a).$$1 in the $$x^* \in \arg\max_{x\in\Delta(X)} U_a(x,y^*,b), \qquad y^* \in \arg\max_{y\in\Delta(X)} U_b(y,x^*,a).$$2-$$x^* \in \arg\max_{x\in\Delta(X)} U_a(x,y^*,b), \qquad y^* \in \arg\max_{y\in\Delta(X)} U_b(y,x^*,a).$$3 plane, $$x^* \in \arg\max_{x\in\Delta(X)} U_a(x,y^*,b), \qquad y^* \in \arg\max_{y\in\Delta(X)} U_b(y,x^*,a).$$4, $$x^* \in \arg\max_{x\in\Delta(X)} U_a(x,y^*,b), \qquad y^* \in \arg\max_{y\in\Delta(X)} U_b(y,x^*,a).$$5 frames by default, a $$x^* \in \arg\max_{x\in\Delta(X)} U_a(x,y^*,b), \qquad y^* \in \arg\max_{y\in\Delta(X)} U_b(y,x^*,a).$$6 spatial grid with mesh size $$x^* \in \arg\max_{x\in\Delta(X)} U_a(x,y^*,b), \qquad y^* \in \arg\max_{y\in\Delta(X)} U_b(y,x^*,a).$$7, and Gaussian positron-range kernel width $$x^* \in \arg\max_{x\in\Delta(X)} U_a(x,y^*,b), \qquad y^* \in \arg\max_{y\in\Delta(X)} U_b(y,x^*,a).$$8. Four particles moving at velocity $$x^* \in \arg\max_{x\in\Delta(X)} U_a(x,y^*,b), \qquad y^* \in \arg\max_{y\in\Delta(X)} U_b(y,x^*,a).$$9 achieve reconstruction errors of approximately $\NE^\lb_\theta = \arg\max_{(x,y)\in\NE_\theta}\min\{U_\theta(x,y,\theta),U_\theta(y,x,\theta)\},$0 mm at $\NE^\lb_\theta = \arg\max_{(x,y)\in\NE_\theta}\min\{U_\theta(x,y,\theta),U_\theta(y,x,\theta)\},$1 cps, $\NE^\lb_\theta = \arg\max_{(x,y)\in\NE_\theta}\min\{U_\theta(x,y,\theta),U_\theta(y,x,\theta)\},$2 mm at $\NE^\lb_\theta = \arg\max_{(x,y)\in\NE_\theta}\min\{U_\theta(x,y,\theta),U_\theta(y,x,\theta)\},$3 cps, and $\NE^\lb_\theta = \arg\max_{(x,y)\in\NE_\theta}\min\{U_\theta(x,y,\theta),U_\theta(y,x,\theta)\},$4 mm at $\NE^\lb_\theta = \arg\max_{(x,y)\in\NE_\theta}\min\{U_\theta(x,y,\theta),U_\theta(y,x,\theta)\},$5 cps, with error near the discretization floor for $\NE^\lb_\theta = \arg\max_{(x,y)\in\NE_\theta}\min\{U_\theta(x,y,\theta),U_\theta(y,x,\theta)\},$6 cps. For four cells with total detection rate $\NE^\lb_\theta = \arg\max_{(x,y)\in\NE_\theta}\min\{U_\theta(x,y,\theta),U_\theta(y,x,\theta)\},$7 cps, the per-cell rate is about $\NE^\lb_\theta = \arg\max_{(x,y)\in\NE_\theta}\min\{U_\theta(x,y,\theta),U_\theta(y,x,\theta)\},$8 cps, tracking efficiency about $\NE^\lb_\theta = \arg\max_{(x,y)\in\NE_\theta}\min\{U_\theta(x,y,\theta),U_\theta(y,x,\theta)\},$9, and tracking accuracy about $G_\theta(\mu(\varepsilon),\Sigma)$0. The method is described as oblivious to the number of tracked cells, able to reconstruct one or multiple cells without explicit particle-path parameterization (Schmitzer et al., 2019).

The 2026 PSMA theranostics paper uses PET more loosely, describing longitudinal lesion segmentation and tracking across serial PSMA PET/CT and post-therapy SPECT/CT as essential infrastructure for lesion-centric response assessment and dosimetry. RECIP 1.0 is defined through whole-body PSMA-derived tumor volume and the appearance of new lesions, with categories RECIP-CR, RECIP-PR, RECIP-PD, and RECIP-SD. The paper emphasizes that new-lesion detection and consistent longitudinal correspondence are fundamental to applying RECIP 1.0 correctly. Its workflow is: segment lesions on each scan; extract lesion attributes such as centroid, volume, uptake, morphology, and texture; build a correspondence cost function using spatial and feature similarity; solve matching across time points, often as an assignment problem or graph-matching problem; propagate lesion identities across serial scans and therapy cycles; and compute lesion-level volume change, uptake change, appearance of new lesions, and lesion-level dosimetry. Registration-based matching and pairwise feature-based matching are the two main method families, with the latter using cost matrices built from centroid distance, volume change, uptake change, radiomics or texture features, and possibly topology-aware spatial context, sometimes solved with spectral graph matching (Yousefirizi et al., 24 Mar 2026).

Although this imaging usage is not a preference model in the recommender or evolutionary-theoretic sense, it preserves the same formal intuition: temporal coupling is introduced to recover an evolving hidden state that cannot be reliably inferred frame by frame. This suggests a methodological rather than semantic continuity with other PET formulations (Schmitzer et al., 2019, Yousefirizi et al., 24 Mar 2026).

6. Common design principles, distinctions, and recurrent misconceptions

Across the cited literature, PET methods reject the assumption that the relevant latent state can be inferred from one time step in isolation. In the evolutionary model, behavior alone is insufficient because preferences also determine partner choice and thereby reshape the payoff environment (Wang et al., 2023). In session-based recommendation, one attention-pooled preference is insufficient because different items in the same session may correspond to different latent preferences and some clicks from further back can remain more relevant than recent clicks (Hu et al., 2021). In the LLM framework, direct ranked-list generation is insufficient because it obscures user profiling, exacerbates popularity bias, and makes fairness or diversity analysis difficult (Zhang et al., 29 Sep 2025). In theranostic imaging, patient-level or limited target-lesion criteria are insufficient because they ignore inter-lesion heterogeneity and cannot support lesion-level dosimetry (Yousefirizi et al., 24 Mar 2026).

A second common principle is that temporal consistency is imposed through explicit structure rather than post hoc smoothing. The relevant structures are stable matching plus Nash or Bayes-Nash equilibrium, attention-modulated hidden-state updates, probability distributions over a stable lattice of preference clusters, the continuity equation with optimal transport action, and longitudinal lesion correspondence constrained by anatomy and feature similarity (Wang et al., 2023, Hu et al., 2021, Zhang et al., 29 Sep 2025, Schmitzer et al., 2019, Yousefirizi et al., 24 Mar 2026). This suggests that PET is best understood as a modeling stance: the latent state is made inspectable, and evolution is constrained by a domain-appropriate dynamic law.

Several misconceptions are therefore directly addressed by the literature. PET is not synonymous with nearest-neighbor matching in longitudinal imaging; topology-aware cost functions and graph matching are explicitly emphasized (Yousefirizi et al., 24 Mar 2026). It is not ordinary sequential modeling in recommendation; the PEN4Rec ablations are designed to show that attention alone, GRU alone, or simple combinations do not recover the same evolving preference trajectory (Hu et al., 2021). It is not direct next-item generation in LLM recommenders; the 2025 PET framework is defined by exposing a belief state as a distribution over clusters (Zhang et al., 29 Sep 2025). Nor is preference evolution reducible to static game play; endogenous matching and informational frictions change which preference types are stable (Wang et al., 2023).

Taken together, these works portray PET as a family of temporally structured inference problems in which latent evolution is itself the object of study. Where they differ is in ontology: preference type, session intent, cluster distribution, radionuclide mass distribution, and lesion identity are not interchangeable objects. Where they converge is in method: temporally linked data are used to infer a structured evolving state that is more informative, more stable, and often more interpretable than a sequence of isolated predictions.