Prior-Based Data Filtering: Methods and Applications

• Prior-based data filtering is a design pattern that uses learned or specified priors to control filtering actions across diverse applications.
• It constructs priors from data evidence to generate weights, masks, or kernels that enhance tasks like reconstruction fidelity, fairness, or predictive accuracy.
• This approach is applied in various domains such as image deblurring, depth super-resolution, sequential forecasting, and language model curation, yielding notable improvements.

Prior-based data filtering denotes a family of methods in which a prior—learned from data, estimated from corpus statistics, induced from metadata, or specified through a probabilistic model—controls what a filter suppresses, preserves, propagates, or emphasizes. In recent work, the term spans spatially varying restoration priors for image deblurring, scene priors for depth super-resolution, predictive priors in particle filtering, token priors for LLM corpus curation, and edited graph-signal priors for fairness-constrained recommendation. Across these settings, the common structure is not a single estimator but a recurring pipeline: construct a prior from available evidence, convert it into weights, kernels, masks, or execution decisions, and use that object to bias filtering toward a task-specific objective such as reconstruction fidelity, forecast accuracy, recall, fairness, or capability suppression (Liu et al., 18 Jul 2025, Wang et al., 2024, Luo et al., 10 Aug 2025, Seo et al., 23 Sep 2025, Rathi et al., 29 Jan 2026).

1. Conceptual scope and recurring structure

Across the recent literature, “prior” has several technically distinct meanings. In unpaired image deblurring it is a spatially varying texture prior $z_p \in \mathbb{R}^{H\times W\times C}$ that modulates adaptive filters inside attention blocks. In depth super-resolution it is a set of scene priors—RGB, normal, semantic, and depth—used to build similarity-conditioned kernels. In probabilistic forecasting it is a predictive prior $p(x_t \mid x_{t-1}, S_t)$ that injects forward-looking signals into state evolution. In graph signal denoising it is partial prior information such as graph topology and temporal window length. In corpus curation it is the token prior $p_{\mathrm{prior}}(t)$ estimated from corpus-level term frequency. In graph fairness it is the editable graph signal $q$ whose filtered posterior is $r = H(W)q$ (Liu et al., 18 Jul 2025, Wang et al., 2024, Luo et al., 10 Aug 2025, Yan et al., 11 Sep 2025, Seo et al., 23 Sep 2025, Krasanakis et al., 2021).

This diversity suggests that prior-based data filtering is best understood as a design pattern rather than a narrow probabilistic doctrine. Some papers use the prior to parameterize a linear or nonlinear filter directly; others use it to choose among filtering plans, to regularize posterior inference, or to delete or mask training signals before a model is learned. A common misconception is that the phrase refers only to Bayesian conjugate priors. The cited work shows a broader usage: corpus-frequency priors, metadata priors, graph-signal priors, and learned neural priors all function as filtering controls when they determine which data are retained, downweighted, or transformed (Mao et al., 2023, Granstrom et al., 2016, Gan et al., 20 Feb 2026).

Domain	Prior object	Filtering action
Unpaired deblurring (Liu et al., 18 Jul 2025)	Texture prior $z_p$	FM-MSA and TM-FFN adaptive filtering
Depth super-resolution (Wang et al., 2024)	RGB, normal, semantic, depth priors	AiPP and MGF kernel construction
Combination forecasting (Luo et al., 10 Aug 2025)	Diversity signal $S_t$	Predictive-prior particle filtering
Graph denoising (Yan et al., 11 Sep 2025)	Partial prior information	JFRFT-domain Wiener-like filtering
LLM corpus curation (Seo et al., 23 Sep 2025)	Token priors from term frequency	Document ranking and removal
Capability shaping (Rathi et al., 29 Jan 2026)	Token-level medical labels	Loss masking or token removal
Filtered ANN (Gan et al., 20 Feb 2026)	Metadata/selectivity priors	Pre- vs post-filter query planning
Graph fairness (Krasanakis et al., 2021)	Edited graph signal priors	Posterior fairness control

2. Restoration and structured signal processing

In image restoration, prior-based filtering is instantiated as explicit prior-to-filter coupling. The deblurring framework in “Learning Deblurring Texture Prior from Unpaired Data with Diffusion Model” learns a spatially varying texture prior from unpaired sharp data by a Texture Prior Encoder with a memory bank $M \in \mathbb{R}^{N\times L}$, then trains a conditional diffusion model to generate that prior at inference, and finally injects the generated code into TTformer through Filter-Modulated Multi-head Self-Attention. In FM-MSA, the prior predicts offsets $\Delta p$ and mixing weights $\Delta m$ so that filtered features $p(x_t \mid x_{t-1}, S_t)$0 are formed before attention, and a wavelet-based adversarial loss preserves high-frequency details. The two-stage objective is

$p(x_t \mid x_{t-1}, S_t)$1

$p(x_t \mid x_{t-1}, S_t)$2

with $p(x_t \mid x_{t-1}, S_t)$3, $p(x_t \mid x_{t-1}, S_t)$4, $p(x_t \mid x_{t-1}, S_t)$5, $p(x_t \mid x_{t-1}, S_t)$6, and $p(x_t \mid x_{t-1}, S_t)$7. On unpaired training over GoPro, HIDE, RealBlur-R, and RealBlur-J, TP-Diff reports $p(x_t \mid x_{t-1}, S_t)$8, $p(x_t \mid x_{t-1}, S_t)$9, $p_{\mathrm{prior}}(t)$0, and $p_{\mathrm{prior}}(t)$1 respectively, while TP-Diff-se reports $p_{\mathrm{prior}}(t)$2, $p_{\mathrm{prior}}(t)$3, $p_{\mathrm{prior}}(t)$4, and $p_{\mathrm{prior}}(t)$5 using only $p_{\mathrm{prior}}(t)$6M parameters; ablating the diffusion model reduces performance to $p_{\mathrm{prior}}(t)$7 dB, ablating TPE to $p_{\mathrm{prior}}(t)$8 dB, ablating TTformer to $p_{\mathrm{prior}}(t)$9 dB, and ablating WaveLoss to $q$0 dB (Liu et al., 18 Jul 2025).

In depth super-resolution, SPFNet treats prior-based filtering as a remedy for modal gap, texture interference, and edge inaccuracy. It uses surface normals from Omnidata and semantic maps from SAM, encodes RGB, semantic, and normals at high resolution, then applies All-in-one Prior Propagation to compute patch-level similarities

$q$1

followed by One-to-one Prior Embedding with Mutual Guided Filtering in the empirically best order normal $q$2 semantic $q$3 RGB. The method reaches RMSE $q$4 on NYU-v2 $q$5 versus SGNet $q$6, $q$7 on Middlebury $q$8 versus SGNet $q$9, $r = H(W)q$0 on Lu $r = H(W)q$1 versus SGNet $r = H(W)q$2, and $r = H(W)q$3 on RGB-D-D versus SGNet $r = H(W)q$4 when trained on RGB-D-D. Joint use of normals and semantics gives the best performance, while AiPP + OPE reduces RMSE by about $r = H(W)q$5 cm on NYU-v2 and about $r = H(W)q$6 cm on Lu relative to a baseline without them (Wang et al., 2024).

Graph signal denoising provides a third restoration variant. JFRFFNet embeds a JFRFT-domain Wiener filter into a neural network and learns both the transform-order pair and the filter coefficients from data while requiring only partial prior information: graph topology and temporal window length. The core filtering step is

$r = H(W)q$7

where $r = H(W)q$8 is trained directly rather than computed from fully known PSDs. On eight real-world datasets with sample length $r = H(W)q$9, JFRFFNet ranks first on five datasets and second on three; examples include $z_p$0 on SST, $z_p$1 on BrestTemp, $z_p$2 on PEMS08, $z_p$3 on PEMS-BAY, and $z_p$4 on Quality, while consistently outperforming the graph-only GFRFFNet (Yan et al., 11 Sep 2025).

3. Sequential Bayesian filtering and predictive priors

In sequential inference, prior-based data filtering appears as prior shaping of state propagation. The DTVW framework for combination forecasting augments time-varying weights with a predictive prior driven by model diversity. If $z_p$5 is the latent state and $z_p$6 is the forward-looking diversity signal from $z_p$7-step-ahead forecasts, the state transition is

$z_p$8

with diversity defined by normalized forecast disagreement. Particle filtering then propagates $z_p$9 and updates weights by the observation likelihood. In simulation, DTVW improves on TVW from RMSFE $S_t$0 to $S_t$1, LS from $S_t$2 to $S_t$3, and CRPS from $S_t$4 to $S_t$5 in a complete model set, and from RMSFE $S_t$6 to $S_t$7, LS $S_t$8 to $S_t$9, and CRPS $M \in \mathbb{R}^{N\times L}$0 to $M \in \mathbb{R}^{N\times L}$1 in a misspecified nonlinear set. In oil-price forecasting it reduces RMSFE by $M \in \mathbb{R}^{N\times L}$2, $M \in \mathbb{R}^{N\times L}$3, and $M \in \mathbb{R}^{N\times L}$4 at horizons $M \in \mathbb{R}^{N\times L}$5, $M \in \mathbb{R}^{N\times L}$6, and $M \in \mathbb{R}^{N\times L}$7, and in U.S. PCE forecasting it achieves RMSFE $M \in \mathbb{R}^{N\times L}$8 versus TVW $M \in \mathbb{R}^{N\times L}$9 and CRPS $\Delta p$0 versus $\Delta p$1 (Luo et al., 10 Aug 2025).

A more classical use of priors appears in radar and multi-object tracking. For track-before-detect PHD filtering, the key result is that the Poisson prior cardinality distribution of TBD-PHD is conjugated to any likelihood function under the paper’s separable amplitude-echo assumptions. The updated intensity is

$\Delta p$2

and the posterior cardinality remains Poisson with rate $\Delta p$3. The paper emphasizes that importing DBT quantities such as $\Delta p$4 or clutter intensity $\Delta p$5 into TBD is incorrect, and introduces a capping operation to prevent divergence of target-number estimates as SNR increases. Monte Carlo experiments show good performance in Rayleigh noise and low-SNR scenarios. For extended objects, PMBM plays the analogous conjugate-prior role: a PPP models undetected targets, an MBM models detected targets, and both prediction and update preserve the PMBM form under the standard transition density and measurement likelihood (Mao et al., 2023, Granstrom et al., 2016).

The DEPF framework focuses on a different failure mode: prior support misalignment. It defines the Prior Boundary Phenomenon by recursive confinement of standard particle filtering support,

$\Delta p$6

and then breaks this confinement through exploratory particles sampled from $\Delta p$7, entropy-driven regularisation

$\Delta p$8

and kernel perturbations with covariance $\Delta p$9 and bandwidth $\Delta m$0. In the paper’s simulated Phase 1, DEPF reaches final distance mean about $\Delta m$1 versus TPF about $\Delta m$2 in $\Delta m$3D with $\Delta m$4, and about $\Delta m$5 versus $\Delta m$6 in $\Delta m$7D with $\Delta m$8; in Phase 2, success rates are $\Delta m$9 for Uniform, $p(x_t \mid x_{t-1}, S_t)$00 for Beta, $p(x_t \mid x_{t-1}, S_t)$01 for Gaussian, and $p(x_t \mid x_{t-1}, S_t)$02 for Dirichlet priors (Shi et al., 30 Jan 2025).

4. Corpus selection, token masking, and query planning

In large-scale text curation, prior-based data filtering becomes an inference-free proxy for quality selection. The prior-based noisy text filter estimates token priors by

$p(x_t \mid x_{t-1}, S_t)$03

then scores each document $p(x_t \mid x_{t-1}, S_t)$04 by

$p(x_t \mid x_{t-1}, S_t)$05

$p(x_t \mid x_{t-1}, S_t)$06

$p(x_t \mid x_{t-1}, S_t)$07

and removes extremes relative to corpus medians $p(x_t \mid x_{t-1}, S_t)$08 and $p(x_t \mid x_{t-1}, S_t)$09. On Dolma with a GPT-2 1.5B model, prior-based filtering reaches average normalized accuracy $p(x_t \mid x_{t-1}, S_t)$10 across $p(x_t \mid x_{t-1}, S_t)$11 downstream benchmarks, compared with $p(x_t \mid x_{t-1}, S_t)$12 for PPL-based filtering, $p(x_t \mid x_{t-1}, S_t)$13 for DSIR, and $p(x_t \mid x_{t-1}, S_t)$14 for no filtering, while reducing wall-clock from $p(x_t \mid x_{t-1}, S_t)$15 GPU hours for PPL-based filtering to about $p(x_t \mid x_{t-1}, S_t)$16 hours. On Pile-github, it reaches $p(x_t \mid x_{t-1}, S_t)$17 versus $p(x_t \mid x_{t-1}, S_t)$18 for PPL-based filtering, and the method is reported to be over $p(x_t \mid x_{t-1}, S_t)$19 faster than PPL-based filtering. The same paper reports multilingual adaptability: when Chinese Wiki-ch is mixed into English Dolma at ratio $p(x_t \mid x_{t-1}, S_t)$20, nearly all Chinese documents are flagged as outliers, while beyond about $p(x_t \mid x_{t-1}, S_t)$21 the outlier rate falls to about $p(x_t \mid x_{t-1}, S_t)$22 (Seo et al., 23 Sep 2025).

Capability shaping in language-model pretraining uses a more targeted form of prior manipulation. “Shaping capabilities with token-level data filtering” treats prior-based filtering as deliberate modification of the pretraining data distribution for a forget domain, here medicine. Token-level labels are bootstrapped from Gemma Scope sparse autoencoders at layer $p(x_t \mid x_{t-1}, S_t)$23 with width $p(x_t \mid x_{t-1}, S_t)$24k, reduced to $p(x_t \mid x_{t-1}, S_t)$25 medical-related latents after autointerp and embedding-score filtering, and distilled into bidirectional token probes. The best $p(x_t \mid x_{t-1}, S_t)$26M bidirectional probe reaches F1 $p(x_t \mid x_{t-1}, S_t)$27 on validation and $p(x_t \mid x_{t-1}, S_t)$28 on test, while a document-level probe reaches $p(x_t \mid x_{t-1}, S_t)$29 and $p(x_t \mid x_{t-1}, S_t)$30. During pretraining, medical tokens are either loss-masked or replaced by a special <|hidden|> token. The paper’s central empirical result is that token filtering Pareto dominates document filtering, and for $p(x_t \mid x_{t-1}, S_t)$31B-parameter models token removal yields at least a $p(x_t \mid x_{t-1}, S_t)$32 effective compute slowdown on the forget domain, whereas document filtering gives about $p(x_t \mid x_{t-1}, S_t)$33. A delayed-filtering ablation shows that starting filtering late in training drastically reduces effectiveness, and delaying onset by $p(x_t \mid x_{t-1}, S_t)$34 makes filtering about an order of magnitude less effective (Rathi et al., 29 Jan 2026).

In vector retrieval, the same logic appears as prior-informed query planning. A filtered ANN query $p(x_t \mid x_{t-1}, S_t)$35 requires top-$p(x_t \mid x_{t-1}, S_t)$36 nearest neighbors satisfying predicate $p(x_t \mid x_{t-1}, S_t)$37, and the core trade-off is between pre-filtering and post-filtering. The learning-based planner estimates selectivity $p(x_t \mid x_{t-1}, S_t)$38 from metadata priors—frequency dictionaries, pairwise co-occurrence matrices, $p(x_t \mid x_{t-1}, S_t)$39-bin histograms, and gradient-boosting corrections for multi-label or mixed predicates—then feeds dataset and predicate features into a two-layer MLP with hidden widths $p(x_t \mid x_{t-1}, S_t)$40 and $p(x_t \mid x_{t-1}, S_t)$41. Labels are defined by the utility

$p(x_t \mid x_{t-1}, S_t)$42

measured for both execution plans. The paper reports build-time reductions of $p(x_t \mid x_{t-1}, S_t)$43 to $p(x_t \mid x_{t-1}, S_t)$44 compared with ACORN-1, and on ArXiv the planner reaches Recall@10 about $p(x_t \mid x_{t-1}, S_t)$45 with up to $p(x_t \mid x_{t-1}, S_t)$46 acceleration over ACORN-1 and about $p(x_t \mid x_{t-1}, S_t)$47 over pre-filtering; the evaluation emphasizes recall at or above $p(x_t \mid x_{t-1}, S_t)$48 (Gan et al., 20 Feb 2026).

5. Priors as fairness controls and embedded hardware filters

Prior-based filtering is also used to control graph-filter outputs under fairness constraints. In “Prior Signal Editing for Graph Filter Posterior Fairness Constraints,” the base relation is $p(x_t \mid x_{t-1}, S_t)$49, where $p(x_t \mid x_{t-1}, S_t)$50 is the prior graph signal and $p(x_t \mid x_{t-1}, S_t)$51 the posterior score. FairEdit edits priors by

$p(x_t \mid x_{t-1}, S_t)$52

then re-filters to obtain fairer posteriors while minimizing

$p(x_t \mid x_{t-1}, S_t)$53

Across $p(x_t \mid x_{t-1}, S_t)$54 graphs and $p(x_t \mid x_{t-1}, S_t)$55 base filters, FairEdit with sweep ratio attains AUC about $p(x_t \mid x_{t-1}, S_t)$56, pRule about $p(x_t \mid x_{t-1}, S_t)$57, and pRule $p(x_t \mid x_{t-1}, S_t)$58 in about $p(x_t \mid x_{t-1}, S_t)$59 of experiments; FairEdit-C reaches AUC about $p(x_t \mid x_{t-1}, S_t)$60, pRule about $p(x_t \mid x_{t-1}, S_t)$61, and pRule $p(x_t \mid x_{t-1}, S_t)$62 in about $p(x_t \mid x_{t-1}, S_t)$63 of experiments. The theoretical argument is that positive-definite graph filters are robust enough that coarse prior editing can still drive the posterior to a local optimum (Krasanakis et al., 2021).

At the hardware level, prior-based filtering appears as physics-motivated edge classification embedded inside the detector. The prototype in-pixel readout chip for particle tracking detectors integrates a $p(x_t \mid x_{t-1}, S_t)$64-bit flash ADC in each $p(x_t \mid x_{t-1}, S_t)$65m pixel, sums row-wise outputs into a $p(x_t \mid x_{t-1}, S_t)$66-bit $p(x_t \mid x_{t-1}, S_t)$67-profile, and feeds that profile into a quantized two-layer fully connected NN implemented as combinational logic. The class output $p(x_t \mid x_{t-1}, S_t)$68 is $p(x_t \mid x_{t-1}, S_t)$69 for high-$p(x_t \mid x_{t-1}, S_t)$70, $p(x_t \mid x_{t-1}, S_t)$71 for low-$p(x_t \mid x_{t-1}, S_t)$72 negative, $p(x_t \mid x_{t-1}, S_t)$73 for low-$p(x_t \mid x_{t-1}, S_t)$74 positive, and $p(x_t \mid x_{t-1}, S_t)$75 for invalid. The underlying prior is physical rather than statistical: cluster shape in a magnetic field, including the paper’s empirical observation that negatively charged particles produce broader charge profiles than positively charged ones. Charge injection is modeled as

$p(x_t \mid x_{t-1}, S_t)$76

and comparator behavior is tied to conversion gain by $p(x_t \mid x_{t-1}, S_t)$77. The chip achieves $p(x_t \mid x_{t-1}, S_t)$78 agreement with offline RTL for $p(x_t \mid x_{t-1}, S_t)$79 test vectors at $p(x_t \mid x_{t-1}, S_t)$80 mV thresholds. At threshold triplet $p(x_t \mid x_{t-1}, S_t)$81 electrons, on-chip signal efficiency is $p(x_t \mid x_{t-1}, S_t)$82 with data reduction $p(x_t \mid x_{t-1}, S_t)$83; at $p(x_t \mid x_{t-1}, S_t)$84, signal efficiency is $p(x_t \mid x_{t-1}, S_t)$85 with data reduction $p(x_t \mid x_{t-1}, S_t)$86. The paper describes this as the first physical realization of in-pixel signal processing with integrated AI-based data filtering for particle tracking detectors (Parpillon et al., 8 Oct 2025).

6. Limitations, misconceptions, and directions of extension

The literature does not present prior-based data filtering as universally robust. In SPFNet, performance depends on prior quality: normals and semantics from large-scale models may be inaccurate in out-of-domain scenes, highly textured scenes with weak geometric cues can still induce interference, and extremely thin structures can be oversmoothed. In TP-Diff, prior resolution scales with input resolution, performance depends on reblurring quality during training, and texture-sparse scenes offer weaker prior signals. In JFRFFNet, the graph topology is fixed, eigendecomposition costs $p(x_t \mid x_{t-1}, S_t)$87, and supervised training assumes access to clean targets. In DEPF, dense covariance handling becomes costly in very high dimension. In the filtered-ANN planner, exact higher-order dependencies are not modeled explicitly and numeric multi-range predicates across different attributes are not yet covered. In token-level capability filtering, borderline or dual-use content remains difficult, adversaries may reintroduce capabilities via finetuning, and filtering must be applied from the beginning of training to remain effective (Wang et al., 2024, Liu et al., 18 Jul 2025, Yan et al., 11 Sep 2025, Shi et al., 30 Jan 2025, Gan et al., 20 Feb 2026, Rathi et al., 29 Jan 2026).

Corpus-frequency priors raise a separate representational issue. Frequency-based filtering can underrepresent minority languages or rare high-quality technical content when those domains occupy a small share of the corpus; the same paper therefore recommends monitoring distributions, blending targeted corpora into prior estimation, or using $p(x_t \mid x_{t-1}, S_t)$88-only filtering when minority retention is important. Fairness-oriented prior editing likewise requires sensitive attributes for all nodes and can overcorrect if the fairness weight is too large. These cases indicate that prior-based filtering is often a precision–coverage trade-off rather than a monotone improvement (Seo et al., 23 Sep 2025, Krasanakis et al., 2021).

The proposed extensions are correspondingly diverse. The vision literature suggests adaptive prior weighting, uncertainty modeling, self-supervised priors, causal consistency, and prompted or adapter-based integration of large-scale models for depth super-resolution, while TP-Diff suggests a fixed-number set of priors with learned upsampling or placement to reduce diffusion cost. The forecasting literature proposes alternative diversity measures, hierarchical predictive priors, smoothing and Rao-Blackwellization, and alternative nonlinear filters. The ANN planner points toward joint optimization of index/search parameters with plan selection. JFRFFNet suggests adaptive or learned graphs and fast spectral approximations. Text-corpus work proposes combining priors with cheap structural or semantic signals and curriculum-like widening of admissible $p(x_t \mid x_{t-1}, S_t)$89 and $p(x_t \mid x_{t-1}, S_t)$90 envelopes over time. Token-level capability shaping suggests future combinations of representation-aware filtering and data attribution to target influential tokens directly (Wang et al., 2024, Liu et al., 18 Jul 2025, Luo et al., 10 Aug 2025, Gan et al., 20 Feb 2026, Yan et al., 11 Sep 2025, Seo et al., 23 Sep 2025, Rathi et al., 29 Jan 2026).

Across these lines of work, prior-based data filtering consistently serves the same technical purpose: it inserts an intermediate object between raw data and the filtering action. Whether that object is a texture prior, a scene prior, a predictive state prior, a token-frequency prior, a selectivity prior, or an edited graph signal, the effect is to make filtering data-dependent before the final estimator acts. This suggests that the central unifying criterion is not the source of the prior but the fact that the prior explicitly parameterizes what the filter is allowed to pass, reject, or reconstruct.