Dynamic Weighted Sliding-Window Strategy

• Dynamic weighted sliding-window strategy is a methodological framework that restricts computation to recent data segments while assigning unequal importance to elements through adaptive weighting.
• It is applied in diverse fields such as frequency estimation, portfolio optimization, reinforcement learning, and clustering to improve responsiveness while managing computational cost.
• The approach effectively balances trade-offs among adaptation speed, memory usage, and statistical accuracy, as demonstrated in experiments across networking, finance, and dynamic scene reconstruction.

Searching arXiv for the cited works and closely related papers to ground the article. Dynamic weighted sliding-window strategy denotes a family of recency-restricted computational schemes in which only a recent portion of a stream, sequence, or trajectory is actively used, while the elements inside or around that window are treated unequally through explicit weights, discounting, adaptive selection, or learned prioritization. The literature does not present a single canonical algorithm under this exact name; rather, it presents related designs for approximate frequency estimation, log-optimal portfolio selection, dynamic 3D Gaussian Splatting, non-stationary reinforcement learning and bandits, submodular subset selection, clustering, summarization, and density tracking. Across these settings, the common objective is to balance responsiveness to change against statistical stability, memory, or computational cost (Shahout et al., 2024, Wang et al., 2022, Shaw et al., 2023, Gajane et al., 2018, Wang et al., 2024).

1. Conceptual scope

In this literature, a sliding window is the restriction of computation to the latest $W$, $w$, or $M$ observations, events, edges, or frames. The “dynamic” component may refer to online re-estimation at each step, motion-adaptive or agenda-adaptive window movement, changing window size, or a time-dependent selection region. The “weighted” component may refer to portfolio weights, per-object blending coefficients, exponentially discounted samples, convex weights in density estimation, knapsack costs, or biased parent selection within a moving cost interval (Wang et al., 2022, Shaw et al., 2023, Wang et al., 2023, Wang et al., 2024, Wang et al., 2017, Neumann et al., 2023).

The resulting strategies are heterogeneous but structurally comparable.

Setting	Window mechanism	Weighting or prioritization mechanism
Frequency estimation	last $W$ arrivals, frames and blocks	prediction-based filtering of non-essential items
Portfolio optimization	rolling window of length $M$	time-varying portfolio weights $K^*(k)$
Dynamic 3DGS	motion-adaptive overlapping windows	learnable per-Gaussian $\alpha_i$
Non-stationary bandits	recent data or discounted history	per-arm windows or exponential discounting
Dynamic KDE	last $T$ batches	optimal convex weight vector $\alpha$

This suggests that the term is best understood as a methodological umbrella rather than a standardized theorem-level object. Some works emphasize dynamic window motion or resizing; others keep the window fixed and vary the weighting; still others combine both.

2. Sliding-window dynamics

A first axis of variation is how the active window is defined and updated. In approximate frequency estimation, Window Compact Space-Saving partitions the stream into frames of size $W$, each frame into $w$0 equal blocks of size $w$1, and exploits the fact that the query window overlaps with at most two frames. The learned variant LWCSS retains this architecture and modifies only the update logic (Shahout et al., 2024).

In non-stationary reinforcement learning, SW-UCRL discards all but the last $w$2 time steps. At episode $w$3, it recomputes $w$4, $w$5, and $w$6 from $w$7, builds confidence sets, and terminates the episode when $w$8. Because of that stopping rule, no episode can last longer than $w$9 steps (Gajane et al., 2018).

Other works make the window itself adaptive. In meeting summarization, the source transcript is processed chunk by chunk, but the stride is dynamic: given window size $M$0, the next start index is chosen as $M$1, and the “Retrospective” mechanism predicts the last supporting utterance of the current context segment to determine the next window boundary (Liu et al., 2021). In dynamic 3D Gaussian Splatting, the sequence is partitioned into windows of varying length according to 2D optical flow magnitude estimated with RAFT: high motion yields shorter windows sampled more frequently, low motion yields longer windows sampled less frequently, and neighboring windows share one overlapping frame for later temporal consistency fine-tuning (Shaw et al., 2023).

The most explicit window-size control appears in RL-Window, where the action space is a discrete set of candidate sizes. The methodology uses $M$2, while the experiments expand this to $M$3. The agent observes state features such as variance, pairwise correlations, rate of change, entropy, spectral features, and concept drift signals, and learns a policy maximizing discounted return with $M$4 (Zarghani et al., 9 Jul 2025).

A different form of dynamics appears in evolutionary multi-objective optimization. SW-GSEMO does not change the window over time in the data-stream sense; instead, it slides a narrow selection interval over the cost axis by setting $M$5 and restricting parent selection to $M$6, with fallback to the full population if the filtered set is empty (Neumann et al., 2023).

3. Weighting, discounting, and selective retention

The second axis is how importance is assigned within the active window. In LWCSS, the key prediction is whether $M$7. If the next arrival is predicted to be within $M$8, the item is inserted into $M$9; otherwise the algorithm uses a Bloom filter $W$0 to avoid repeated unverified suppression. The update rule is: if next arrival of $W$1 $W$2, then $W$3; else, if $W$4, then $W$5; else $W$6 and return. Query answers add a $W$7 correction because the paper identifies two sources of underestimation and notes that a query window overlaps at most two frames (Shahout et al., 2024).

In portfolio selection, the weighting is literal. The portfolio vector is

$W$8

and the sliding-window method computes

$W$9

The defining feature is that $M$0 is time-varying rather than constant, so weighting and windowing are inseparable (Wang et al., 2022).

In SWinGS, weighting is implemented by a learnable per-Gaussian blending parameter $M$1. For the $M$2-th Gaussian,

$M$3

High $M$4 emphasizes the MLP for that Gaussian, whereas low $M$5 leaves it closer to the canonical representation. Initialization uses a rough dynamic/static mask produced by projecting Gaussians into each view, computing L1 image differences, averaging labels over views and frames, and setting $M$6 for dynamic Gaussians and $M$7 for static Gaussians (Shaw et al., 2023).

Bandit and MDP papers use weighting as discounting. In non-stationary parametric bandits, past observations receive weight $M$8, with weighted ridge or quasi-likelihood estimators and confidence sets built from

$M$9

The stated motivation is that gradual drifting is often better matched by soft forgetting than by hard truncation or restart (Wang et al., 2023, Wang et al., 3 Jan 2026). In dynamic density tracking, weights are a convex vector $K^*(k)$0 on batchwise KDEs,

$K^*(k)$1

and the exact MISE reduces to a constrained quadratic program,

$K^*(k)$2

Here the weighting sequence itself is the object of optimization (Wang et al., 2024).

A further interpretation of weighting is cost-sensitive selection. In KnapStream and KnapWindow, the “weights” are costs $K^*(k)$3 in $K^*(k)$4-knapsack constraints, and selection uses the threshold

$K^*(k)$5

In SW-GSEMO and the dynamic chance-constrained knapsack work, priority is given to solutions whose costs or weights fall in a moving interval rather than to the whole Pareto population (Wang et al., 2017, Perera et al., 2024).

4. Theoretical properties and algorithmic trade-offs

Theoretical analyses in this area typically formalize a bias-variance, memory-accuracy, or adaptation-exploration trade-off. In LWCSS, the core guarantee is robustness relative to the baseline sliding-window estimator: the proof sketch states that LWCSS can underestimate an item by at most $K^*(k)$6 in the worst case, and the query correction restores the desired one-sided additive guarantee. The paper explicitly frames the contribution as an improvement in the practical memory-accuracy trade-off rather than an asymptotic space improvement of WCSS itself (Shahout et al., 2024).

SW-UCRL makes the trade-off explicit through the regret bound

$K^*(k)$7

and the optimized window size

$K^*(k)$8

The first term quantifies the regret cost of change-points; the second quantifies the learning cost within stationary segments (Gajane et al., 2018).

For dynamic representative subset selection, KW and KW$K^*(k)$9 obtain approximation guarantees of $\alpha_i$0 and $\alpha_i$1, respectively. KW achieves this by maintaining checkpoints and KnapStream instances over the window, while KW$\alpha_i$2 uses SubKnapChk to delete a checkpoint whenever it can be approximated by its successors, improving efficiency at the cost of a weaker factor (Wang et al., 2017). In sliding-window $\alpha_i$3-clustering, the objective is approximated by a compact weighted summary that supports expiration, suffix extraction, and replacement of expired centers; the theorem gives an $\alpha_i$4-approximation with polylogarithmic dependence on window-related parameters in the reported regime (Borassi et al., 2020).

Weighted matching in the sliding-window model exhibits a different trade-off. One result gives a $\alpha_i$5-approximation using $\alpha_i$6 space, and another gives a $\alpha_i$7-approximation in $\alpha_i$8 space. The stronger approximation relies on running multiple instances of the Paz–Schwartzman algorithm on carefully selected substreams and, crucially, processing some blocks in reverse order (Alexandru et al., 2022).

A more abstract structural contribution comes from fully-dynamic-to-incremental reductions in the deletions-look-ahead model. Because sliding window is FIFO, it falls into that model. The reduction preserves query time $\alpha_i$9 and yields worst-case update time $T$0, or amortized $T$1 in the simpler version, provided the incremental algorithm already has worst-case update time and the deletion order is known (Peng et al., 2022).

Finally, some papers optimize the weighting rule itself. Dynamic KDE derives the exact MISE for evolving Gaussian densities and shows that weight selection is a constrained quadratic program, while weighted non-stationary bandit and MDP papers derive dynamic regret bounds under exponential discounting and emphasize that a refined local-norm analysis removes the need for extra covariance matrices in earlier weighted approaches (Wang et al., 2024, Wang et al., 2023, Wang et al., 3 Jan 2026).

5. Representative applications and empirical behavior

In streaming analytics, LWCSS is evaluated on CAIDA Internet traces from Chicago, NY, and SJ, and on synthetic Zipf streams. The LSTM-based next-arrival predictor is trained as a binary classifier and reports F1 scores of $T$2, $T$3, and $T$4 on Chicago, NY, and SJ, respectively. LWCSS outperforms WCSS in RMSE on all real datasets; query performance is nearly the same for WCSS and LWCSS; update performance of LWCSS is slightly worse because of Bloom filter operations; and, for Chicago, RMSE increases as window size increases, with better performance for smaller windows (Shahout et al., 2024).

In finance, the sliding-window log-optimal portfolio is tested on a three-asset portfolio consisting of VT, BND, and BNDX, using daily data from February 14, 2018 to February 14, 2020, with $T$5. The classical benchmark computed from one year of data gives $T$6. For $T$7, the best reported setting is $T$8, with cumulative return $T$9, realized log-growth $\alpha$0, max drawdown $\alpha$1, and annualized Sharpe ratio $\alpha$2, compared with $\alpha$3, $\alpha$4, $\alpha$5, and $\alpha$6 for the classical fixed-weight strategy. In the flipped bear-market scenario, the classical strategy gives cumulative return $\alpha$7, whereas $\alpha$8 gives $\alpha$9 cumulative return and $W$0 Sharpe ratio (Wang et al., 2022).

In dynamic scene reconstruction, SWinGS reports that training separate dynamic 3D Gaussian models per motion-adaptive window supports arbitrary-length sequences, large motion, scene content changes, new objects appearing, and topological changes. The temporal consistency step uses one overlapping frame and the loss

$W$1

The ablation shows a deliberate trade-off: without temporal consistency there can be slightly better per-frame metrics in some cases, but with temporal consistency there is a substantial improvement in perceptual temporal coherence despite a minor degradation in some frame-wise scores (Shaw et al., 2023).

In meeting summarization, the AMI corpus analysis reports average source length $W$2 sentences / $W$3 words and average reference summary length $W$4 sentences / $W$5 words. The extractive alignment study reports that $W$6 of sentences in reference summaries match the same agenda order as in the source content and that local supporting segments provide $W$7 word-level coverage. On the AMI test set, BART-SW-Dynamic achieves ROUGE-1 $W$8, ROUGE-2 $W$9, and ROUGE-L $w$00, improving over BART-Truncate and BART-SW-Fixed, while BART-SW-GoldSeg reaches $w$01, $w$02, and $w$03. The retrospective boundary prediction has average utterance-level distance $w$04, equivalent to about $w$05 characters (Liu et al., 2021).

In evolutionary optimization, SW-GSEMO improves the expected time to a $w$06-approximation in monotone submodular maximization under a uniform constraint to $w$07 with probability $w$08, removing dependence on the maximum population size in the analyzed setting. Experiments on maximum coverage show that SW-GSEMO usually outperforms GSEMO, especially when many trade-offs arise, and that in some large instances it produces up to about four times as many trade-off solutions as GSEMO (Neumann et al., 2023). In the dynamic chance-constrained knapsack problem, three-objective methods outperform bi-objective methods overall, and sliding-window selection is especially beneficial in the 3D case; the reported qualitative conclusion is that SW-3D achieves the highest mean profit values in the final dynamic state (Perera et al., 2024).

In non-stationary bandits, the fixed-size per-arm SW-UCB baseline is reported as robust, whereas D-UCB is described as suffering from a fundamental learning failure leading to linear regret. In the MovieLens experiments, FDSW-UCB with optimistic max aggregation achieves cumulative regret $w$09 under abrupt drift, $w$10 under gradual drift, and $w$11 under stationary conditions, compared with $w$12 for SW-UCB in abrupt drift (Chen, 24 Nov 2025). In multi-dimensional stream classification, RL-Window reports accuracies of $w$13 on UCI HAR, $w$14 on PAMAP2, and $w$15 on Yahoo! Finance, with post-drift degradation $w$16, $w$17, and $w$18, computational cost around $w$19–$w$20 ms, and average chosen window sizes around $w$21–$w$22 (Zarghani et al., 9 Jul 2025).

6. Conceptual distinctions, limitations, and open directions

Several recurring misconceptions are corrected by the literature itself. First, not every sliding-window method is “weighted” in the same sense. In dynamic RSS, weighting largely means knapsack costs rather than recency coefficients (Wang et al., 2017). In weighted matching, “weighted” refers to edge weights in the objective, while the sliding-window mechanism is realized through substream selection and reverse or forward processing (Alexandru et al., 2022). In asynchronous event streams, the method is explicitly not about weighted windows; it studies a distributed sliding window $w$23 and the maintenance of the corresponding lattice Lat-Win (Yang et al., 2011).

Second, not every method labeled dynamic actually learns or optimizes the window online. The non-stationary UCB paper explicitly states that it does not propose a fully dynamic weighted sliding-window rule; SW-UCB uses a fixed window size, uniform weighting within the retained window, and per-arm queues, while the dynamic element in FDSW-UCB comes from ensemble aggregation rather than adaptive window-length learning (Chen, 24 Nov 2025). Conversely, RL-Window is explicitly an adaptive window-size controller, but its flexibility is limited by a discrete action space and by the need for reward shaping, replay, target-network updates, and periodic classifier retraining (Zarghani et al., 9 Jul 2025).

Third, the costs of adaptivity are recurrent. LWCSS pays slightly worse update throughput because of Bloom filter operations (Shahout et al., 2024). SWinGS introduces overhead from multiple model trainings, adaptive partitioning, overlapping windows, and sequential fine-tuning (Shaw et al., 2023). Fully-dynamic-to-incremental reductions require worst-case incremental update time and known deletion order; without those conditions, the black-box reduction is not available (Peng et al., 2022). Dynamic KDE derives optimal weights only under the evolving-Gaussian model with Gaussian kernels, and the paper explicitly identifies extension to Gaussian mixtures as future work (Wang et al., 2024).

The literature also leaves several open design questions. A plausible implication is that adaptive control of forgetting remains one of the central unresolved problems. Weighted parametric bandit papers argue that soft forgetting is especially natural under gradual drift, yet also state that an adaptive weighted method with time-varying $w$24, analogous to optimal black-box restart methods, remains open (Wang et al., 2023, Wang et al., 3 Jan 2026). In portfolio optimization, the window length $w$25 is treated as a design variable and the possibility of an “optimal” $w$26 is explicitly suggested rather than resolved (Wang et al., 2022). More generally, this body of work suggests that dynamic weighted sliding-window strategy is less a single method than a design principle: preserve recency, bias computation toward informative or high-value content, and control the error introduced by forgetting through verification, correction, confidence bounds, or post hoc consistency mechanisms.