Preprocessing Recommender Agent

• Preprocessing recommender agents are modules that transform raw interactions into structured inputs, effectively constructing state representations for downstream models.
• They leverage techniques such as histogram equalization, feature construction via LLMs, and token compression to improve recommendation accuracy and efficiency.
• By mediating instructions and reordering candidate lists, these agents enhance personalization and fairness while protecting user interests.

A preprocessing recommender agent is a pre-recommendation module that converts raw interactions, user histories, item content, instructions, and environmental signals into structured inputs, transformed targets, or reranked candidate sets before a downstream recommendation policy acts. Recent work uses this role in several distinct but convergent senses: as a model-agnostic transformation of the user–item matrix, as an automated feature-construction loop over unstructured text, as an offline user-modeling layer that caches personas for later retrieval, as a token-selection encoder that compresses multi-behavior context for LLMs, and as a user-side shielding agent that mediates between platform recommendations and user interests (Wang, 2023, Škrlj et al., 25 Jul 2025, Shi et al., 4 Mar 2025, Mukande et al., 6 Dec 2025, Liu et al., 12 Sep 2025, Xu et al., 20 Feb 2025).

1. Conceptual scope and system role

In the narrowest sense, preprocessing denotes an input-side transformation applied before a recommender is trained or queried. The histogram-equalization formulation treats the raw user–item rating matrix $R$ as the input to a preprocessing stage that reshapes the rating distribution while leaving the downstream algorithm unchanged; the paper emphasizes that such a module is “model-agnostic,” “drop-in,” and “cheap to deploy” (Wang, 2023). In a broader agentic formulation, RecoWorld describes a preprocessing recommender agent as the component that turns “messy, heterogeneous environment signals” into a compact state representation $s_t$, clean item and user features, and structured feedback signals usable by RL or supervised modules (Liu et al., 12 Sep 2025).

This broader interpretation shifts preprocessing from mere data cleaning to state construction. In Rec4Agentverse, the “Agent Recommender” is not only a selector of “Agent Items” but also an orchestrator that aggregates user preference summaries, injects objectives into downstream agents, and routes cross-agent collaboration (Zhang et al., 2024). In iAgent, the same idea becomes explicitly user-protective: the agent stands between the user and the platform, receives the platform’s candidate list, and re-ranks it according to user instructions $\Omega_u$, history $S_u$, and external environment $E$, yielding an indirect-exposure paradigm rather than direct platform control (Xu et al., 20 Feb 2025).

A plausible implication is that “preprocessing recommender agent” now names a layer rather than a single algorithm. Depending on system design, that layer may operate on rating values, sparse matrices, long behavior sequences, multimodal item representations, natural-language instructions, or candidate lists returned by another recommender.

2. Matrix-level transformation and densification

A classical line of work treats preprocessing as an operation on the interaction matrix itself. In “Enhancing Recommender System Performance by Histogram Equalization,” the observed ratings $\mathcal{D}=\{R_{u,i}\}$ are globally equalized through the mapping

$$\tilde{R}(r_k)=R_{\max}F_R(r_k)=R_{\max}\sum_{j=1}^{k}P(R_x=r_j),$$

so that each observed entry becomes $\tilde{R}_{u,i}=\tilde{R}(R_{u,i})$ before matrix factorization or KL-Mat is trained (Wang, 2023). The paper applies a single global histogram rather than per-user or per-item equalization, keeps MAE and RMSE on the transformed scale, and reports that histogram equalization “enhances vanilla matrix factorization accuracy performance by a large margin, and is on par with the vanilla algorithms on fairness metrics” on MovieLens 1M and LDOS-CoMoDa (Wang, 2023).

A second matrix-centric variant is the two-step ProbS+ILLS pipeline. The method first constructs a sparse user–item matrix from behavior, then applies Probabilistic Spreading with

$$W_{\alpha\beta}=\frac{1}{k_\beta}\sum_{j=1}^{m}\frac{a_{\alpha j}a_{\beta j}}{k_j}, \qquad f'=Wf,$$

to convert sparse binary interactions into a denser resource matrix, and then applies Iterative Local Least Squares to estimate missing values (Liu et al., 2012). The local regression stage solves

$\min_x \|B^T x-w\|,$

iteratively reselecting neighbors as imputations improve. On MovieLens, Netflix, and RYM, the paper reports that ProbS+ILLS improves AUC over sparse-matrix ILLS, with the clearest gain on the sparsest dataset, where ProbS reaches $s_t$0 AUC and ProbS+ILLS reaches $s_t$1 (Liu et al., 2012).

These two methods instantiate a matrix-first preprocessing agent. One family reshapes the value distribution before optimization; the other densifies and imputes missing structure. In both cases, the downstream recommender is largely unchanged, and the preprocessing layer carries the burden of combating skew or sparsity.

3. Feature construction, memory formation, and state encoding

A more recent line of work relocates preprocessing from matrices to representations. Agent0 frames preprocessing as automated feature discovery from raw, noisy text. Its Architect–Sentinel–Oracle loop searches over prompts that instruct a lightweight LLM to emit “multi-value categorical features,” while a fixed DCNv2-based oracle evaluates the resulting feature by $s_t$2 and stores prompt-score pairs in shared memory (Škrlj et al., 25 Jul 2025). The best discovered prompt achieved more than $s_t$3 RIG over baseline, and the system’s core contribution is not a new ranker but a closed-loop mechanism for extracting compact, high-signal fields from unstructured documents (Škrlj et al., 25 Jul 2025).

PersonaX similarly treats preprocessing as a separate user-modeling layer. It first embeds a long behavior sequence $s_t$4, clusters it with a distance threshold $s_t$5, allocates a sampling budget across clusters, and then selects a sub-behavior sequence by balancing prototypicality and diversity. For each selected subset $s_t$6, it generates a persona $s_t$7, yielding a cached multi-persona set $s_t$8 (Shi et al., 4 Mar 2025). At inference time, the system retrieves only the most relevant persona(s) rather than regenerating a profile online. Using only $s_t$9 to $\Omega_u$0 of behavioral data with sequence length $\Omega_u$1, PersonaX reports gains of $\Omega_u$2 to $\Omega_u$3 for AgentCF and $\Omega_u$4 to $\Omega_u$5 for Agent4Rec, while halving online time relative to summarization-plus-relevance baselines (Shi et al., 4 Mar 2025).

HGLMRec moves preprocessing into token compression. Multi-behavior user–item data are encoded as a hypergraph $\Omega_u$6, propagated by a lightweight hypergraph convolution, and compressed through attention-based adaptive readout into graph tokens $\Omega_u$7, which are fused with prompt tokens $\Omega_u$8 by

$\Omega_u$9

The paper states that the model “retrieves only the relevant tokens during inference, reducing computational overhead while enriching the retrieval context” (Mukande et al., 6 Dec 2025). This is preprocessing in a strict operational sense: the LLM never sees raw logs, only a compact, structured context selected by the hypergraph encoder.

Collectively, these systems suggest an “encoder-preprocessor” pattern: the preprocessing recommender agent discovers feature fields, compresses histories into personas, or turns large structured contexts into a small number of high-value tokens before the main recommender reasons over them.

4. Instruction mediation, shielding, and task decomposition

When recommendations are interactive, preprocessing also becomes intent mediation. iAgent formalizes this through the user–agent–platform paradigm. The platform still produces a candidate list, but the user-specific agent re-ranks it according to

$S_u$0

where $S_u$1 is the user’s instruction and $S_u$2 is user-specific rather than globally shared (Xu et al., 20 Feb 2025). InstructAgent parses instructions into internal knowledge $S_u$3, tool decisions $S_u$4, keywords $S_u$5, and optional external knowledge $S_u$6; Instruct$S_u$7Agent adds a profile generator $S_u$8 and a dynamic extractor that derives $S_u$9 and $E$0 for instruction-conditioned reranking (Xu et al., 20 Feb 2025). The self-reflection prompt constrains the LLM to return only permutations of the original candidate IDs, and the paper reports that self-reflection reduces hallucination rates by at least $E$1 (Xu et al., 20 Feb 2025).

TAIRA addresses a related but distinct preprocessing problem: complex natural-language intent must often be decomposed before item retrieval. Its manager agent uses Thought Pattern Distillation to retrieve a task template, generates a JSON plan, and orchestrates SearcherAgent, ItemRetriever, PlannerAgent, and InteractorAgent across phases (Yu et al., 30 Jun 2025). In difficult cases such as ambiguous or multi-scenario requests, the preprocessing step is not ranking but scenario discovery and attribute derivation. The paper’s results emphasize that planning quality, measured by Success Rate, improves most on harder tasks, and ablations show the largest drop when thought-pattern matching is removed (Yu et al., 30 Jun 2025).

These systems make explicit a recurrent misconception: preprocessing in recommender systems is not only a numeric operation. It can also be the conversion of vague user language into structured constraints, scenario-specific subqueries, or candidate-list permutations that better reflect user intent than the platform’s original ordering.

5. Simulation, evaluation, and synthetic environments

A preprocessing recommender agent also appears as infrastructure for simulated training and evaluation. RecoWorld provides a Gym-style environment with a dual-view architecture: an LLM-based user simulator and an agentic recommender interact over multi-turn trajectories, with retention-oriented rewards and multiple content representation modes—text-based, multimodal, and semantic ID (Liu et al., 12 Sep 2025). In this setting, the preprocessing layer implements

$E$2

aggregates per-item behavior into reward components, and packages natural-language instructions and feedback into state variables usable by the policy (Liu et al., 12 Sep 2025).

GGBond extends this logic into a socially aware simulation platform. Its Sim-User Agents combine episodic memory, affective state, adaptive preference learning, and the ICR$E$3 motivational engine. The motivational score

$E$4

and the watch probability

$E$5

allow the system to model how intimacy, novelty, reciprocity, and risk shape responses to recommendations under evolving social ties (Zhong et al., 27 May 2025). This makes preprocessing not only a filter over inputs but also a predictive user-response layer that can reweight or rerank platform outputs according to social-cognitive state.

AgentCF uses LLM-based user and item agents to simulate behavior directly. User agents choose between positive and sampled negative items, then both user and item memories are updated through collaborative reflection if the simulated choice disagrees with the real log (Zhang et al., 2023). The framework produces user–item, user–user, item–item, and collective interactions, thereby generating synthetic logs, review-like rationales, and cold-start item descriptions that can precede or augment a downstream recommender (Zhang et al., 2023).

Evaluation itself requires preprocessing. In e-commerce settings, reliable CTR, ATC-TR, BTR, and conversion revenue depend on filtering out B2B customers, outlier customer-days, bounced traffic, ambiguous visitor identities, duplicate events, combo-product attribution errors, and invalid UI exposures (Chaudhary et al., 2019). The paper prescribes exact attribution windows—$E$6 minutes for clicks, $E$7 minutes for add-to-cart, and $E$8 hours for buys—and a robust Cust ID construction from cookie_id and user_id (Chaudhary et al., 2019). This evaluation-centric perspective shows that a preprocessing recommender agent may govern not only training inputs but also the validity of the metrics used to judge recommender quality.

6. Limitations, trade-offs, and emerging directions

The literature also delineates several tensions. Histogram equalization can improve MAE and remain competitive on fairness metrics, but it may cause “loss of calibration,” is presented only for explicit ratings, and may have limited effect when the original distribution is already balanced or too coarse (Wang, 2023). Agent0’s oracle is tied to DCNv2 and $E$9, so feature discovery is biased toward one downstream model and one metric; the paper also notes convergence and exploration issues when all agents share a single memory store (Škrlj et al., 25 Jul 2025). PersonaX explicitly does not address incremental or streaming updates, so persona validity over time remains open (Shi et al., 4 Mar 2025).

Current work also points toward more autonomous preprocessing. STARec models each user as an agent with fast ranking and slow reflective profile updates, trained by an anchored SFT-plus-RL pipeline; notably, it reports substantial gains while using only $\mathcal{D}=\{R_{u,i}\}$0 of the full training data (Wu et al., 26 Aug 2025). CogRec couples an LLM perception layer with Soar, converting textual inputs into symbolic working memory and using chunking to turn impasse resolutions into reusable rules, thereby framing preprocessing as neuro-symbolic state construction rather than pure embedding or prompting (Hu et al., 30 Dec 2025). EvoRec goes further by proposing a multi-agent framework that can evolve “all parts of a recommender pipeline, including the early, data-centric stages,” and reports offline gains up to $\mathcal{D}=\{R_{u,i}\}$1 plus an online A/B lift of $\mathcal{D}=\{R_{u,i}\}$2 revenue and $\mathcal{D}=\{R_{u,i}\}$3 CTR (Mu et al., 15 Jun 2026). This suggests that future preprocessing recommender agents may not be static modules at all, but self-improving components with persistent memory, skill libraries, and explicit diagnosis of failed data-centric interventions.

A plausible synthesis is that the preprocessing recommender agent is becoming a first-class recommender subsystem. In older formulations it densified matrices or removed evaluation artifacts; in current agentic systems it also constructs state, retrieves profiles, compresses tokens, interprets instructions, filters platform outputs, and simulates user response. The unifying principle is invariant across these implementations: recommendation quality is shaped not only by the ranking model, but by the agent that decides what the model is allowed to see, optimize, and expose.