Personalized Window Size Mechanism

• PWSM is a family of techniques that replace fixed windows with data-conditioned rules to determine optimal aggregation scales across various domains.
• It leverages methods such as interpolation, reinforcement learning, and statistical stopping to balance trade-offs in accuracy, resource usage, and privacy.
• Applications include adaptive sizing in online advertising, process mining, stream mining, vision transformers, and personalized privacy-preserving estimation.

Searching arXiv for papers on Personalized Window Size Mechanism and closely related adaptive window-sizing formulations. Personalized Window Size Mechanism (PWSM) denotes a class of methods in which the effective window is not fixed globally but is selected, adapted, or personalized according to local evidence. The term is used explicitly in private infinite-stream estimation under personalized $w$-event privacy, where users may have distinct $(w_i,\epsilon_i)$ constraints, and closely related mechanisms appear in online advertising, streaming process mining, frequent itemset mining over streams, multivariate sliding-window control, scan statistics, and windowed attention in vision transformers (Du et al., 9 May 2026, Du et al., 10 Sep 2025, Zhang et al., 23 Jan 2025, Imenkamp et al., 25 Oct 2025, Reddy et al., 2014, Zarghani et al., 9 Jul 2025, Reiner-Benaim, 2013, Zhang et al., 2022). A plausible unifying interpretation is that PWSM replaces a single hand-crafted temporal or spatial scale with a data-conditioned rule that determines how much history, neighborhood, or support should be aggregated for each decision.

1. Conceptual scope and window semantics

The meaning of a “window” in PWSM is domain-dependent. In online advertising, the window is a conversion horizon such as 1 day or 7 days. In streaming process mining, it is the current subset of recent events used for discovery or conformance. In continual private release, it is the temporal interval over which a user’s cumulative privacy loss must remain bounded. In scan statistics, it is the moving-sum support used to search for a local peak. In vision transformers, it is the attention area from which keys and values are sampled (Zhang et al., 23 Jan 2025, Imenkamp et al., 25 Oct 2025, Du et al., 9 May 2026, Reiner-Benaim, 2013, Zhang et al., 2022).

Domain	Window object	Adaptation basis
Online advertising	Conversion-delay horizon	Interpolation between $T_s$ and $T_l$
Process mining	Event window over stream	Coverage and dynamic threshold
Private stream estimation	Privacy window per user	Per-user $(w_i,\epsilon_i)$
Frequent itemset mining	Transaction mining window	Context similarity between windows
Vision transformers	Attention region	Regressed scale and offset

This diversity of semantics is important because PWSM is not a single algorithmic template. Some instantiations choose a window by thresholding representativeness, some by explicit optimization, some by post-hoc interpolation, and some by learned regression or reinforcement learning. What they share is that the window is an operational variable rather than a static hyperparameter.

2. Why fixed windows are insufficient

Across application areas, fixed windows fail because the relevant scale is heterogeneous. In online advertising, conversion delay differs significantly across advertisers and products: gaming apps may have long cycles from install to in-app event, e-commerce may care about purchases within promotional periods, and a single fixed window such as 1 day or 7 days is therefore not universally optimal. A short window is fresher but incomplete; a long window is more complete but suffers from stale labels and delayed feedback. The same paper notes that supporting daily windows within 7 days would require at least 7 models, plus separate label pipelines, serving logic, and storage/computation overhead (Zhang et al., 23 Jan 2025).

In streaming process mining, a fixed window is described as brittle under concept drift. If the window is too small, it may miss important activities, trace variants, or directly-follows relations; if too large, it includes outdated behavior and increases resource usage. In frequent itemset mining over streams, a fixed window can mix transactions from different contexts, making itemsets appear infrequent or noisy. In multivariate stream analytics, fixed-size windows struggle with concept drift and bursty patterns. In scan statistics, the window size is “proportional to the length of the peak searched for,” and the null tail probability depends on both the covariance structure and the size of the searching window. In windowed vision transformers, fixed windows limit long-range dependency modeling and object-size adaptivity (Imenkamp et al., 25 Oct 2025, Reddy et al., 2014, Zarghani et al., 9 Jul 2025, Reiner-Benaim, 2013, Zhang et al., 2022).

These observations establish the central design pressure behind PWSM: the statistically or operationally correct aggregation scale varies across advertisers, products, users, objects, regions, or time steps. A fixed window therefore hard-codes a mismatch between the model’s support and the signal’s natural extent.

3. Mathematical design patterns

One recurrent pattern is interpolation between two known endpoints. In the advertising formulation, flexible optimization window estimation is treated as approximation of the CDF of conversion time. For $T_f \in (T_s,T_l)$,

$$\mathbb{P}(\tau \le T_f|e) = (1-\alpha(T_s,T_f,T_l))\cdot \mathbb{P}(\tau \le T_s|e) + \alpha(T_s,T_f,T_l)\cdot \mathbb{P}(\tau \le T_l|e),$$

with

$$\alpha(T_s,T_f,T_l) = \mathbb{P}(\tau \in (T_s,T_f]|e)/\mathbb{P}(\tau \in (T_s,T_l]|e) \in [0,1].$$

When a model outputs short-window and conditional long-window predictions, the estimated flexible-window probability becomes

$$\widehat{\mathbb{P}(\tau \le T_f|e)} = f_{T_s}(e) + (1-f_{T_s}(e))\cdot(\alpha \cdot f_{T_l|T_s}(e)).$$

The paper proposes linear, rational, and exponential designs for $\alpha$ (Zhang et al., 23 Jan 2025).

A second pattern is representativeness-based stopping. In streaming process mining, the method adapts species estimation from biodiversity research. For a window sample $(w_i,\epsilon_i)$0 of size $(w_i,\epsilon_i)$1, Chao1 richness is estimated by

$(w_i,\epsilon_i)$2

and coverage is

$(w_i,\epsilon_i)$3

The window remains open until coverage exceeds a dynamically chosen threshold, with the Dynamic Threshold Heuristic locating an elbow-like point through

$(w_i,\epsilon_i)$4

This makes window closure a function of estimated sample representativeness rather than elapsed time or event count (Imenkamp et al., 25 Oct 2025).

A third pattern is context-similarity-driven growth. In variable-window frequent itemset mining, the similarity between the current transaction window and a context-change analysis window is computed as

$(w_i,\epsilon_i)$5

equivalently described as a Jaccard similarity over attribute sets. If $(w_i,\epsilon_i)$6, the two windows are merged; otherwise the current mining window is finalized and the analysis window seeds the next one (Reddy et al., 2014).

A fourth pattern is explicit sequential decision-making. In RL-Window, the adaptive window-sizing problem is formulated as a reinforcement-learning objective

$(w_i,\epsilon_i)$7

where the action is a discrete window size $(w_i,\epsilon_i)$8. The state encodes variance, pairwise correlations, temporal trend, entropy, out-of-order indicators, spectral features, and concept drift signals, and the agent is implemented with a Dueling DQN with prioritized experience replay (Zarghani et al., 9 Jul 2025).

A fifth pattern is learned geometric regression. In Varied-Size Window Attention, default windows $(w_i,\epsilon_i)$9 are summarized and a lightweight regression module predicts target window scale and offset through

$T_s$0

with per-head predictions used to sample keys and values from a learned attention area rather than a hand-crafted fixed block (Zhang et al., 2022).

4. Advertising realization: Personalized Interpolation

The advertising paper "Personalized Interpolation" operationalizes the PWSM idea as a lightweight mechanism attached to existing conversion models. Its workflow is deliberately simple: train or obtain predictions for two windows, $T_s$1 and $T_s$2; add a post-hoc interpolation layer; and produce a flexible-window estimate as a convex combination of the two window predictions. The interpolation layer is not used in training and is applied only at inference/evaluation time. The paper emphasizes that this can be attached to an existing black-box production model with essentially no additional training cost (Zhang et al., 23 Jan 2025).

The experimental setup uses a multi-task multi-label (MTML) consolidation model with two tasks: SOW as 1-day conversion prediction and LOW as 7-day conversion prediction. The data are impression data from 2023-12-27 to 2024-02-04, and flexible windows are evaluated for $T_s$3. The empirical conversion CDF is reported as concave, rising from $T_s$4 to $T_s$5. Calibration scores are near 1 for all interpolation methods, with rational interpolation being the closest.

The baselines are: Use $T_s$6, Use $T_s$7, 7-TaskHead Model (7THM), Dedicated Consolidations (DC), and No Delay Upper Bound (NDUB). The reported day-7 normalized entropy values are as follows.

Method	Day-7 NE
NDUB	0.6156
DC	0.6175
Linear-INTRPL	0.6175
Rational-INTRPL	0.6175
Exp-INTRPL	0.6175
7THM	0.6200
Use $T_s$8	0.6391
Use $T_s$9	0.6175

The main interpretation offered in the paper is that interpolation methods match DC very closely while using only one model, outperform naive fixed-window shortcuts, and do better than the 7-TaskHead model, likely because multi-head training introduces task balancing issues. The gap to NDUB is described as very small, around $T_l$0 NE loss, and over a week of recurring training the interpolation methods outperform the 7-TaskHead baseline by over $T_l$1 NE gain. The exponential interpolation is reported as least sensitive to $T_l$2, while performance degrades as the target window approaches 1 day if the interpolation factor is misspecified.

5. Adaptive windowing in streaming analytics, mining, and statistical detection

In streaming process mining, adaptive window size is framed as a representativeness problem rather than a mere buffering problem. The mechanism continuously updates species counts for activities $T_l$3, trace variants $T_l$4, or directly-follows relations $T_l$5, recomputes Chao1-derived completeness and coverage in $T_l$6, and closes the current window when coverage reaches the dynamically updated threshold. The implementation uses Python, Apache Kafka, and Faust. On synthetic drift logs, window size changed by about $T_l$7 under sudden drift, with coefficient of variation $T_l$8. Under gradual drift, average window size increased from $T_l$9 pre-drift to $(w_i,\epsilon_i)$0 during drift to $(w_i,\epsilon_i)$1 post-drift, corresponding to a $(w_i,\epsilon_i)$2 relative increase during the drift period and coefficient of variation $(w_i,\epsilon_i)$3. On Sepsis Cases, the approach achieved an average F1-score of $(w_i,\epsilon_i)$4, compared with $(w_i,\epsilon_i)$5 for a landmark window and $(w_i,\epsilon_i)$6 for a fixed count-based tumbling window of size 20. Average throughput is reported as $(w_i,\epsilon_i)$7 events/s with standard deviation $(w_i,\epsilon_i)$8 events/s, while latency grows linearly with window size (Imenkamp et al., 25 Oct 2025).

In frequent itemset mining over streams, the proposed mechanism uses a two-level structure: a transaction window Wtran and a context change analysis window Wcca. Initially, WSmin transactions are moved into Wtran; then s(Wcca) transactions are moved into Wcca; context variation analysis compares their attribute sets; and the main window either grows until WSmax or is finalized when similarity falls below threshold. The finalized window is mined by TIFIM using bushes whose coverage exceeds cov. The implementation is in Java 7, with streaming data emulated by flat files, parallelism simulated by Java multithreading, and stream behavior emulated using Java RMI. Against the matrix-based MFIM baseline, the paper reports average memory increment $(w_i,\epsilon_i)$9 for MFI-VWS-CVA versus $T_f \in (T_s,T_l)$0 for matrix-based FIM, and execution-time increment $T_f \in (T_s,T_l)$1 versus $T_f \in (T_s,T_l)$2 under divergent coverage conditions (Reddy et al., 2014).

In multivariate stream analytics, RL-Window recasts the selection of sliding-window size as a discrete-action control problem. The agent selects from $T_f \in (T_s,T_l)$3, uses a Dueling DQN with hidden layers of 256, 128, and 64 neurons, and is trained with replay buffer size 200,000, minibatch size 128, target network updates every 2000 steps, learning rate decaying from $T_f \in (T_s,T_l)$4 to $T_f \in (T_s,T_l)$5 over 50,000 steps, and $T_f \in (T_s,T_l)$6-greedy exploration from $T_f \in (T_s,T_l)$7 to $T_f \in (T_s,T_l)$8. Reported accuracies are $T_f \in (T_s,T_l)$9 on UCI HAR, $$\mathbb{P}(\tau \le T_f|e) = (1-\alpha(T_s,T_f,T_l))\cdot \mathbb{P}(\tau \le T_s|e) + \alpha(T_s,T_f,T_l)\cdot \mathbb{P}(\tau \le T_l|e),$$0 on PAMAP2, and $$\mathbb{P}(\tau \le T_f|e) = (1-\alpha(T_s,T_f,T_l))\cdot \mathbb{P}(\tau \le T_s|e) + \alpha(T_s,T_f,T_l)\cdot \mathbb{P}(\tau \le T_l|e),$$1 on Yahoo! Finance Stream, with drift robustness losses of $$\mathbb{P}(\tau \le T_f|e) = (1-\alpha(T_s,T_f,T_l))\cdot \mathbb{P}(\tau \le T_s|e) + \alpha(T_s,T_f,T_l)\cdot \mathbb{P}(\tau \le T_l|e),$$2, $$\mathbb{P}(\tau \le T_f|e) = (1-\alpha(T_s,T_f,T_l))\cdot \mathbb{P}(\tau \le T_s|e) + \alpha(T_s,T_f,T_l)\cdot \mathbb{P}(\tau \le T_l|e),$$3, and $$\mathbb{P}(\tau \le T_f|e) = (1-\alpha(T_s,T_f,T_l))\cdot \mathbb{P}(\tau \le T_s|e) + \alpha(T_s,T_f,T_l)\cdot \mathbb{P}(\tau \le T_l|e),$$4, respectively. The ablation results attribute performance degradation to removal of Dueling DQN, prioritized replay, spectral features, or the stability penalty (Zarghani et al., 9 Jul 2025).

The scan-statistic literature supplies an important theoretical complement. For

$$\mathbb{P}(\tau \le T_f|e) = (1-\alpha(T_s,T_f,T_l))\cdot \mathbb{P}(\tau \le T_s|e) + \alpha(T_s,T_f,T_l)\cdot \mathbb{P}(\tau \le T_l|e),$$5

the null tail probability

$$\mathbb{P}(\tau \le T_f|e) = (1-\alpha(T_s,T_f,T_l))\cdot \mathbb{P}(\tau \le T_s|e) + \alpha(T_s,T_f,T_l)\cdot \mathbb{P}(\tau \le T_l|e),$$6

is governed by the covariance matrix $$\mathbb{P}(\tau \le T_f|e) = (1-\alpha(T_s,T_f,T_l))\cdot \mathbb{P}(\tau \le T_s|e) + \alpha(T_s,T_f,T_l)\cdot \mathbb{P}(\tau \le T_l|e),$$7 of the moving sums, and exact covariance formulas are given under general, common, and auto-covariance structures. This does not constitute an explicit personalized algorithm, but it shows that the statistical behavior of the detector is inseparable from window size, overlap, and dependence structure. A plausible implication is that any PWSM for detection should adapt $$\mathbb{P}(\tau \le T_f|e) = (1-\alpha(T_s,T_f,T_l))\cdot \mathbb{P}(\tau \le T_s|e) + \alpha(T_s,T_f,T_l)\cdot \mathbb{P}(\tau \le T_l|e),$$8 to expected peak length and local covariance rather than treat it as a generic smoothing parameter (Reiner-Benaim, 2013).

6. PWSM in personalized $$\mathbb{P}(\tau \le T_f|e) = (1-\alpha(T_s,T_f,T_l))\cdot \mathbb{P}(\tau \le T_s|e) + \alpha(T_s,T_f,T_l)\cdot \mathbb{P}(\tau \le T_l|e),$$9-event privacy

The most explicit formalization of PWSM appears in private infinite-stream estimation under personalized $$\alpha(T_s,T_f,T_l) = \mathbb{P}(\tau \in (T_s,T_f]|e)/\mathbb{P}(\tau \in (T_s,T_l]|e) \in [0,1].$$0-event privacy. Here the stream is $$\alpha(T_s,T_f,T_l) = \mathbb{P}(\tau \in (T_s,T_f]|e)/\mathbb{P}(\tau \in (T_s,T_l]|e) \in [0,1].$$1, the released statistic at time $$\alpha(T_s,T_f,T_l) = \mathbb{P}(\tau \in (T_s,T_f]|e)/\mathbb{P}(\tau \in (T_s,T_l]|e) \in [0,1].$$2 is an estimate $$\alpha(T_s,T_f,T_l) = \mathbb{P}(\tau \in (T_s,T_f]|e)/\mathbb{P}(\tau \in (T_s,T_l]|e) \in [0,1].$$3 of $$\alpha(T_s,T_f,T_l) = \mathbb{P}(\tau \in (T_s,T_f]|e)/\mathbb{P}(\tau \in (T_s,T_l]|e) \in [0,1].$$4, and each user $$\alpha(T_s,T_f,T_l) = \mathbb{P}(\tau \in (T_s,T_f]|e)/\mathbb{P}(\tau \in (T_s,T_l]|e) \in [0,1].$$5 specifies a privacy window $$\alpha(T_s,T_f,T_l) = \mathbb{P}(\tau \in (T_s,T_f]|e)/\mathbb{P}(\tau \in (T_s,T_l]|e) \in [0,1].$$6 and budget $$\alpha(T_s,T_f,T_l) = \mathbb{P}(\tau \in (T_s,T_f]|e)/\mathbb{P}(\tau \in (T_s,T_l]|e) \in [0,1].$$7. The optimization constraint is

$$\alpha(T_s,T_f,T_l) = \mathbb{P}(\tau \in (T_s,T_f]|e)/\mathbb{P}(\tau \in (T_s,T_l]|e) \in [0,1].$$8

so every user’s privacy loss across any sliding window of length $$\alpha(T_s,T_f,T_l) = \mathbb{P}(\tau \in (T_s,T_f]|e)/\mathbb{P}(\tau \in (T_s,T_l]|e) \in [0,1].$$9 must remain within $$\widehat{\mathbb{P}(\tau \le T_f|e)} = f_{T_s}(e) + (1-f_{T_s}(e))\cdot(\alpha \cdot f_{T_l|T_s}(e)).$$0. The privacy guarantee is formulated as $$\widehat{\mathbb{P}(\tau \le T_f|e)} = f_{T_s}(e) + (1-f_{T_s}(e))\cdot(\alpha \cdot f_{T_l|T_s}(e)).$$1-EPDP, and the later dynamic extension generalizes this to dynamic backward and forward constraints (Du et al., 10 Sep 2025, Du et al., 9 May 2026).

At each time slot, PWSM computes current user budgets, splits the budget into a dissimilarity-testing part and a publication part, uses the first part to estimate dissimilarity, uses the second part to estimate reporting error through Optimal Budget Selection (OBS), and publishes a fresh noisy result only if

$$\widehat{\mathbb{P}(\tau \le T_f|e)} = f_{T_s}(e) + (1-f_{T_s}(e))\cdot(\alpha \cdot f_{T_l|T_s}(e)).$$2

Otherwise, the last publication is reused. The private dissimilarity test is based on

$$\widehat{\mathbb{P}(\tau \le T_f|e)} = f_{T_s}(e) + (1-f_{T_s}(e))\cdot(\alpha \cdot f_{T_l|T_s}(e)).$$3

while OBS selects the threshold minimizing

$$\widehat{\mathbb{P}(\tau \le T_f|e)} = f_{T_s}(e) + (1-f_{T_s}(e))\cdot(\alpha \cdot f_{T_l|T_s}(e)).$$4

Two fixed-setting instantiations are proposed. Personalized Budget Distribution (PBD) distributes each user’s privacy budget more regularly across the window and guarantees that the budget reserved for the next time step is no smaller than the budget consumed in the previous release. Personalized Budget Absorption (PBA) fully absorbs the privacy budget from the previous $$\widehat{\mathbb{P}(\tau \le T_f|e)} = f_{T_s}(e) + (1-f_{T_s}(e))\cdot(\alpha \cdot f_{T_l|T_s}(e)).$$5 time slots while also borrowing from the next $$\widehat{\mathbb{P}(\tau \le T_f|e)} = f_{T_s}(e) + (1-f_{T_s}(e))\cdot(\alpha \cdot f_{T_l|T_s}(e)).$$6 time slots, thereby increasing the budget available for the current time slot. The paper characterizes PBD as preferable when the stream changes relatively often or steadily, and PBA as preferable when the stream changes less often or more abruptly (Du et al., 10 Sep 2025).

The dynamic extension introduces Dynamic Personalized Budget Distribution (DPBD) and Dynamic Personalized Budget Absorption (DPBA), which permit users to adjust privacy requirements over time while satisfying $$\widehat{\mathbb{P}(\tau \le T_f|e)} = f_{T_s}(e) + (1-f_{T_s}(e))\cdot(\alpha \cdot f_{T_l|T_s}(e)).$$7-Event $$\widehat{\mathbb{P}(\tau \le T_f|e)} = f_{T_s}(e) + (1-f_{T_s}(e))\cdot(\alpha \cdot f_{T_l|T_s}(e)).$$8-Personalized Differential Privacy. The dynamic methods enforce both backward feasibility and forward feasibility, so that current budget allocation does not violate either past consumption or future commitments (Du et al., 9 May 2026).

The empirical results are reported in terms of AMRE and AJSD on Taxi trajectories, Foursquare check-ins, and synthetic TLNS, Sin, and Log streams. In the 2025 paper, PBD achieves 68% less error than BD on average on real data sets, while PBA achieves 24.9% less error than BA on average on synthetic data sets. In the 2026 extension, the proposed methods reduce estimation error by at least $$\widehat{\mathbb{P}(\tau \le T_f|e)} = f_{T_s}(e) + (1-f_{T_s}(e))\cdot(\alpha \cdot f_{T_l|T_s}(e)).$$9 compared with state-of-the-art algorithms, with PBD and DPBD strongest on volatile real streams and PBA and DPBA strongest on smoother synthetic streams (Du et al., 10 Sep 2025, Du et al., 9 May 2026).

7. Related mechanisms, misconceptions, and limitations

A common misconception is that PWSM refers to a single canonical procedure. The literature does not support that reading. In privacy-preserving stream release, PWSM is an explicit framework with OBS, dissimilarity testing, and budget splitting. In advertising, it appears as post-hoc CDF interpolation between two endpoint windows. In process mining, it is a coverage-driven window closure rule. In stream mining, it is context-sensitive growth and finalization. In reinforcement learning, it is a policy over discrete window sizes. In vision transformers, it is a learned per-head regression of window scale and offset (Du et al., 9 May 2026, Zhang et al., 23 Jan 2025, Imenkamp et al., 25 Oct 2025, Reddy et al., 2014, Zarghani et al., 9 Jul 2025, Zhang et al., 2022).

A second misconception is that “personalization” must always mean per-user customization. The sources show several distinct granularities. Advertising interpolation is personalized at the event level, where different advertisers, products, and events can induce different effective window behavior. RL-Window adapts the window to recent stream state rather than to identity. VSA applies adaptation independently for each attention head, so different heads obtain different target windows even within the same default window. Scan-statistic theory suggests adaptation to local covariance pattern and expected peak width. Thus, personalization may target users, events, objects, regions, or time steps depending on the domain (Zhang et al., 23 Jan 2025, Zarghani et al., 9 Jul 2025, Zhang et al., 2022, Reiner-Benaim, 2013).

The limitations are likewise domain-specific. In process mining, many erroneous or spurious singleton events can prevent exact completeness from being reached, and very small samples can make coverage or completeness misleadingly high if no singletons are observed; the method also introduces parameters such as smoothing factor, decay rate, minimum threshold, stagnation threshold $\alpha$0, and stagnation window $\alpha$1. In variable-window frequent itemset mining, WSmin, WSmax, and ss are pre-set, and the evaluation uses synthetic datasets. In VSA, runtime overhead is modest but not zero: Swin-T throughput drops from 1557 fps to 1297 fps on A100, while GPU memory increases from 15.8G to 16.1G. In private stream estimation, PBD is generally more expensive than simpler homogeneous baselines because of optimal budget selection and more frequent comparisons and publications (Imenkamp et al., 25 Oct 2025, Reddy et al., 2014, Zhang et al., 2022, Du et al., 10 Sep 2025).

The broader significance of PWSM is therefore methodological rather than terminological. It organizes a family of techniques that treat window size as a first-class adaptive variable. A plausible implication is that future work will continue to hybridize the existing design patterns—interpolation, statistical stopping, optimization, control, and privacy accounting—rather than converge on a single universal mechanism.