Filter-Accelerated Candidate Management

• Filter-accelerated candidate management is a class of methods that decouples candidate selection into filtering and generation stages to boost efficiency.
• It is widely applied in vector retrieval, recommender systems, join algorithms, and screening pipelines to reduce computational bottlenecks.
• Learning-based dynamic filtering integrates selectivity estimation and query planning to adaptively optimize pre- and post-filtering, achieving up to 4× acceleration with high recall.

Filter-accelerated candidate management refers to a class of methods, frameworks, and algorithmic architectures that apply explicit filtering mechanisms—either as standalone filtering, in combination with candidate generation, or through fusions of both stages—to efficiently manage the set of entities (candidates) considered for downstream processing, matching, or recommendation. These techniques are prominent in vector retrieval, multi-criteria recommender systems, streaming assignment, entity joining, and large-scale screening, and they address the computational and practical bottlenecks created by naïve candidate enumeration, unrestrained similarity search, or slow post hoc filtering.

1. Foundational Paradigms: Filtering and Candidate Generation

At the core of filter-accelerated candidate management is the partition of the workload into filtering and candidate selection, whose execution order and integration profoundly affect computational efficiency and output quality.

• In filtered approximate nearest-neighbor (ANN) search, the two canonical execution orders are:
  • Pre-filtering: Apply the predicate first to prune the search space to a subset, then run ANN on this filtered set.
  • Post-filtering: Run ANN on the full dataset to retrieve a (possibly enlarged) candidate set, then apply the filter to return the desired candidates (Gan et al., 20 Feb 2026).

The relative merits of these strategies depend on parameters such as dataset scale, predicate selectivity, underlying ANN index efficiency, and the interplay between filtering cost and candidate processing. Pre-filtering can restrict computation to the relevant subset but may incur prohibitively high costs for indexing or linear scanning under low selectivity. Post-filtering can result in wasted ANN computation and recall loss, especially under stringent filtering (Gan et al., 20 Feb 2026).

This paradigm extends to other domains:

• In multi-stage screening (fair candidate promotion pipelines), explicit statistical filtering (by group and pass/fail outcome) is interposed at each stage to efficiently promote suitable candidates and enforce fairness constraints (Blum et al., 2022).
• In set similarity joins, filtering is tightly integrated with verification to avoid the explicit enumeration of candidate pairs, compressing both steps into single-stage algorithms (Feng et al., 4 Jun 2025).

2. Learning-Based Dynamic Filtering: Query Planning and Selectivity-Aware Strategies

Recent advancements leverage lightweight prediction and query-planning to choose or blend pre- and post-filtering dynamically. For filtered ANN search, a learning-based framework operates as follows:

• Selectivity Estimation: Categorical predicates use frequency tables and co-occurrence statistics; range predicates estimate selectivity using fine-grained attribute histograms. Mixed predicates combine these features, refined with a gradient-boosted machine (Gan et al., 20 Feb 2026).
• Core Planner: A two-layer MLP classifier, trained on synthetic query grids, maps query and corpus features (embedding dimensionality, corpus size, mean inter-point distance, predicate type, estimated selectivity) to the optimal execution plan ("pre" or "post").
• Execution Workflow: At query time, the system selects and executes the optimal plan in microseconds, never modifying the ANN index internals, and applies adaptive enlargement in post-filtering for recall guarantees.

This approach yields up to $4\times$ acceleration at $\geq 90\%$ recall, construction times $1.24\times-20.2\times$ faster than index retraining baselines, and traces the Pareto-optimal latency–recall front for selectivity regimes $s\in [1\%, 20\%]$ (Gan et al., 20 Feb 2026).

A plausible implication is that learning-based, selectivity-informed filter planners generalize well to any context where the operational cost of filtering and candidate generation is query-distribution dependent.

3. Integrated Filtering and Verification in Join and Matching Algorithms

Classic two-stage set similarity joins use filter-and-verify frameworks: a filtering pass produces candidate pairs, and a verification pass computes exact similarities. This can result in candidate explosion and excessive intermediate I/O.

The CF-RS-Join family avoids explicit candidate generation. The key innovation is the Filter-and-Verification Tree (FVT) and its Linear variant (LFVT) (Feng et al., 4 Jun 2025):

• FVT: Compresses all set–element associations into a trie-like structure, where tree traversal for a query set allows simultaneous filtering (via set size bounds from the Jaccard threshold) and in-place intersection accumulation. Each element points directly to its deepest tree node, supporting $O(1)$ access per element.
• LFVT: Linearizes maximal non-branching FVT paths to optimize memory usage and cache efficiency. Candidate verification and filtering are interleaved along linear-chain traversals.
• Single-Stage Processing: For both FVT/LFVT, candidate verification occurs only during traversal—no explicit candidate set is materialized. Pruning is achieved via length filtering, and survivors are directly output as true matches.
• Scalability: These approaches integrate with MapReduce (MR-CF-RS-Join/LFVT), using load-aware partitioning to optimize parallelization and minimize data movement. Each reducer independently builds its FVT/LFVT and processes local queries, with partitions computed to balance workload.

Empirical results show that these algorithms are $2$–$20\times$ faster than two-stage MR baselines, with lower memory and disk footprint and near-linear scalability to large cluster sizes (Feng et al., 4 Jun 2025).

4. Filtering for Candidate Recommendation: Graph Filtering and Multi-criterion Aggregation

Filter-accelerated management is fundamental in multi-criteria recommender systems. The CA-GF methodology reframes candidate curation as the application of polynomial graph filters over user-item similarity graphs, with the following architecture (Park et al., 13 Feb 2025):

• User-Expansion Graph: Each user is expanded into $C+1$ criterion-views, with the ratings matrix stacked across criteria and normalized.
• Item–Item Similarity Matrix: Derived by projecting the expanded matrix to obtain a unified similarity graph.
• Criterion-specific Polynomial Filters: For each criterion, a second-order polynomial filter (linear, inward, or outward) is selected via validation. The filter is applied to smooth user signals separately in each criterion-view, with aggregation guided by learned or user-extracted preference weights.
• End-to-End Candidate Management: Candidate scoring reduces to $1$–$2$ sparse matrix multiplications per criterion, requiring no iterative model training or eigendecomposition. Candidate ranking for top-$\geq 90\%$0 selection is immediate from the aggregated scores.

CA-GF achieves sub-second recommendation on million-scale graphs, strong accuracy (up to $\geq 90\%$1 over baselines), and explicit per-criterion score attributions, facilitating interpretability (Park et al., 13 Feb 2025).

Extensions to multi-view and multi-behavior logs employ the same scheme—unified adjacency construction, per-view filter tuning, per-view scoring, and explicit aggregation.

5. Streaming, Training-Aided, and Multi-Stage Filtered Screening

In online candidate assignment and screening, filter-accelerated management combines statistical filtering with irrevocable decisions to minimize false positives and workload, while providing strong guarantees of optimal assignment.

• No-training regime: Greedy algorithms skip an initial batch, then retain candidates only when they belong to the current optimum matching, leading to $\geq 90\%$2 expected retentions (Cohen et al., 2019).
• Training-aided regime: By learning attribute thresholds on prior samples, thresholds policies can achieve exponential reductions, with $\geq 90\%$3 expected retentions, due to rapidly vanishing empirical error in calibrating fixed thresholds (Cohen et al., 2019).

The result is substantially more efficient initial filtering—in expectation, exponentially fewer candidates are advanced to final matching, given suitable training data.

In multi-stage screening pipelines, especially under fairness constraints (Equality of Opportunity), per-stage filter-promotion rates are optimized:

• Policies are parameterized by promotion probabilities $\geq 90\%$4, $\geq 90\%$5 decided per group $\geq 90\%$6 at each stage.
• Structural lemmas yield efficient enumeration and FPTAS algorithms to maximize linear combinations of precision and recall, subject to fairness constraints. Nonconvexity of the feasible set requires policy expertization (Blum et al., 2022).

6. LLM-Based and Embedding-Driven Filtering in HR Screening

Modern resume and document screening frameworks employ filter-accelerated candidate management through a hybrid of embedding-based pre-filtering, representative sample selection for in-context learning, and prompt-based evaluation by local LLMs (Xu et al., 19 Mar 2026):

• Pipeline Stages: Raw documents are ingested, embedded with an open-source encoder, and stored in a vector database. A sample selector (similarity, diversity, or clustering-based) extracts $\geq 90\%$7 examples for few-shot prompting.
• Evaluation: For each resume, a prompt including persona, evaluation rubric, and the $\geq 90\%$8 examples is constructed and evaluated by an LLM. Dimension scores and a binary decision are output, and a threshold is applied for advancement.
• Efficiency and Privacy: Embedding and sample selection are performed offline; only per-resume LLM inference (typically $\geq 90\%$9–$1.24\times-20.2\times$0 s) is required online. All computation can be performed on-premise, avoiding exposure of personally identifiable information.
• Performance: With ground-truths defined by commercial GPT models, open-source LLM evaluators (Qwen3-8B, Llama-3.1-8B-Instruct) equipped with this filter-accelerated pipeline achieve up to $1.24\times-20.2\times$1 higher accuracy than GPT-5-nano and substantial reductions in per-resume latency (Xu et al., 19 Mar 2026).

Filtering is further refined by controlling the selection strategy of few-shot examples and adjusting the shot count for accuracy-latency trade-offs.

7. Comparative Table: Representative Filter-Accelerated Frameworks

Domain	Technique/Framework	Distinctive Feature	arXiv ID
ANN search	Learning-based query planning	Per-query dynamic plan based on selectivity estimation	(Gan et al., 20 Feb 2026)
Set similarity join	FVT/LFVT single-stage join	Eliminates candidate generation, implicit filtering	(Feng et al., 4 Jun 2025)
Recommendations	CA-GF graph filtering	Parallelizable, criterion-specific polynomial filtering	(Park et al., 13 Feb 2025)
Screening pipelines	Multi-stage fairness optimization	Non-convex fairness constraints, combinatorial FPTAS	(Blum et al., 2022)
Resume screening	AutoScreen-FW	LLM + embedding+few-shot in-context, privacy-safe filter	(Xu et al., 19 Mar 2026)
Online assignment	Thresholds-based policies	Exponential candidate reduction with training	(Cohen et al., 2019)

These frameworks exemplify the integration of filter-accelerated processing into candidate management pipelines across a wide range of computational settings.

---

Filter-accelerated candidate management encompasses a broad spectrum of techniques grounded in explicit or learned filtering strategies, unified by the principle of minimizing downstream workload while preserving or optimizing output quality. The surveyed methods demonstrate substantial, often order-of-magnitude, improvements in runtime, scalability, and/or fairness, with technical guarantees and robust empirical validation (Feng et al., 4 Jun 2025, Gan et al., 20 Feb 2026, Park et al., 13 Feb 2025, Blum et al., 2022, Cohen et al., 2019, Xu et al., 19 Mar 2026).