---
title: Weighted Importance Score (WIS)
url: https://www.emergentmind.com/topics/weighted-importance-score-wis
type: topic
---

# Weighted Importance Score (WIS)

Searching arXiv for the cited papers to ground the article and citations.
arxiv_search({"query":"id:2601.02440 OR title:\"Mitigating Long-Tailed Anomaly Score Distributions with Importance-Weighted Loss\"","max_results":5,"sort_by":"relevance"})
arxiv_search({"query":"id:1906.02126 OR title:\"Extractive Summarization via Weighted Dissimilarity and Importance Aligned Key Iterative Algorithm\"","max_results":5,"sort_by":"relevance"})
arxiv_search({"query":"id:2606.18575 OR title:\"Adaptive COVID-19 Trajectory Forecasting Using MAB-Inspired Ensemble Weighting\"","max_results":5,"sort_by":"relevance"})
Weighted Importance Score (WIS) is best understood, in the literature surveyed here, as an *Editor’s term* for scalar quantities that combine an importance notion with weighting, selection, or distributional correction. The term is not standardized. In some works it is only an interpretive label for a mathematically defined weight; in others the acronym WIS denotes a different object altogether, including **Watermarking with Importance Scoring**, **Weighted Importance Sampling**, or **Weighted Interval Score**. What unifies the genuinely importance-based usages is that a scalar derived from importance, rarity, difficulty, or utility is inserted into an objective, a sampler, or a selection rule to change how training instances, symbols, sentences, or tokens influence the final system [2601.02440] [1906.02126] [2311.09668] [2005.10718] [1905.04866] [1906.04328] [2606.18575].

## 1. Definition and conceptual scope

Across the surveyed works, WIS-like constructions fall into several recurrent patterns. In anomaly detection, the score is an importance-sampling ratio computed from anomaly-score densities. In extractive summarization, it is a weighted dissimilarity that combines sentence importance with diversity. In LLM watermarking, it is a token-importance gate used to protect semantically critical tokens. In source coding, variational inference, and off-policy prediction, related quantities appear as normalized importance weights or importance-aware costs. By contrast, epidemic forecasting uses the acronym WIS for a probabilistic scoring rule rather than an importance score.

| Context | Quantity | Role |
|---|---|---|
| Anomaly detection | \(w_i = p_b(b_i)/p_s(s_i)\) | Reweights reconstruction loss |
| Summarization | \(D(\vec{s}_i,\vec{s}_j)=\|\vec{s}_i\|\|\vec{s}_j\|\sin\theta_{ij}\) | Balances importance and non-redundancy |
| LLM watermarking | \(f(s,s_t)\in[0,1]\) | Gates whether watermarking may perturb a token |
| Source coding | \(p_i w_i\) or its normalized form | Encodes importance-aware coding cost |
| Variational inference | \(w_k=p_\theta(x,z_k)/q_\phi(z_k\mid x)\) | Defines multi-sample lower bounds |
| Off-policy RL | \(\tilde w_i=\rho_i/\sum_j \rho_j\) | Normalizes IS updates |
| Epidemic forecasting | Weighted Interval Score | Evaluates predictive distributions |

This suggests that “Weighted Importance Score” is most precise when it denotes a scalar that changes influence allocation, not merely any weighted metric. The literature also shows that the same acronym can name fundamentally different mathematical objects, so domain-specific definition is essential.

## 2. Score-space importance weighting in anomaly detection

The most explicit importance-weight construction appears in anomaly detection under long-tailed anomaly score distributions. For an input \(x_i\) and reconstruction \(f(x_i)\), the anomaly score is the squared residual norm,
\[
s_i=\|x_i-f(x_i)\|_2^2.
\]
A baseline autoencoder minimizes mean squared reconstruction error, but the proposed importance-weighted loss replaces plain MSE with
\[
L_{\text{IWL}}=\frac{1}{N}\sum_{i=1}^N w_i\,\|x_i-f(x_i)\|_2^2
=\frac{1}{N}\sum_{i=1}^N w(s_i)\,s_i.
\]
The weight is derived by importance sampling between two score distributions: the empirical anomaly-score distribution \(q(s)\), which may be long-tailed, and a target Gaussian-like distribution \(p(s)\). Scores are shifted positive, transformed by a monotone Box–Cox map \(b_i=f_b(s_i)\), Gaussian densities are fitted to original and transformed scores after removing outliers with a modified z-score rule, and the importance weight is defined as
\[
w_i=\frac{p_b(b_i)}{p_s(s_i)}.
\]
This quantity is not explicitly named “Weighted Importance Score” in the paper, but the paper states that its core idea is exactly to construct importance weights as a function of the anomaly score and use them inside the loss, which makes the interpretation natural [2601.02440].

The method is motivated by the observation that, under Gaussian reconstruction noise and independent residual components, anomaly scores for homogeneous normal data should be approximately Gaussian. Real normal data often contain multiple subgroups, producing a right-skewed, long-tailed score distribution: many “easy” normal instances lie near low scores, while minority normal patterns occupy the tail. The importance ratio \(w(s)\) counteracts this imbalance. Regions that are overrepresented relative to the target Gaussian are down-weighted; underrepresented regions are up-weighted. The gradient contribution of sample \(i\) becomes
\[
\nabla_\theta\!\left(w_i\|x_i-f_\theta(x_i)\|_2^2\right)
=
w_i\,\nabla_\theta \|x_i-f_\theta(x_i)\|_2^2,
\]
so the weight is a direct scalar multiplier on training influence.

To stabilize optimization, the method clips weights by a skewness-dependent cap:
\[
T=\max(|\text{skewness}(s)|,\epsilon),\qquad
T=\min(\alpha T,T_0),\qquad
w_i\leftarrow \mathrm{clip}(w_i,0,T).
\]
The paper reports \(\alpha=4\) and \(T_0=20\). Empirically, with DSVDD on MNIST, the skewness of log anomaly scores decreases from \(0.623\) to \(0.176\), indicating that the long right tail is substantially reduced. On long-tailed MNIST, DSVDD average AUROC improves from \(0.725\) to \(0.768\); on three real-world hyperspectral datasets, DSVDD average AUPR improves from \(0.763\) to \(0.794\); and the reported overall average improvement across settings is about \(0.043\) [2601.02440].

A central feature of this formulation is that it operates in **score space** rather than label space. Only the label “normal” is available, and the method does not require prior knowledge of normal subclasses. The model becomes implicitly subgroup-aware by reshaping the anomaly-score distribution itself.

## 3. Sentence-level WIS in extractive summarization

In extractive summarization, a WIS-like construction appears through sentence importance and weighted dissimilarity. Each sentence \(s\) is represented by
\[
\vec{s}=\frac{1}{N(s)}\sum_{w\in s} N_w \vec e_w,
\]
where \(N(s)\) is the number of words in the sentence, \(N_w\) is the number of occurrences of word \(w\) in the document, and \(\vec e_w\) is the unit basis vector. The paper defines sentence importance as the squared norm,
\[
\mathrm{imp}(s)=\|\vec{s}\|^2.
\]
Its key pairwise quantity is the weighted dissimilarity
\[
D(\vec{s}_1,\vec{s}_2)=\|\vec{s}_1\|\,\|\vec{s}_2\|\sin\theta_{12},
\]
with \(\theta_{12}\) the angle between the sentence vectors. The term “Weighted Importance Score” is not used in the paper, but this pairwise scalar can be interpreted as one: it is large only when both sentences are important and mutually dissimilar [1906.02126].

The Importance Aligned Key Iterative Algorithm orders sentences by descending importance, then performs thresholded greedy selection. Given a current summary set \(S_{\text{sel}}\), a candidate sentence is accepted when its minimum weighted dissimilarity to already selected sentences exceeds a threshold \(T\):
\[
\min_{s'\in S_{\text{sel}}} D(\vec{s},\vec{s}') \ge T.
\]
A sentence-level WIS interpretation is therefore
\[
\mathrm{WIS}(s\mid S_{\text{sel}})
=
\min_{s'\in S_{\text{sel}}} D(\vec{s},\vec{s}').
\]
Binary search over \(T\) yields the desired summary size \(S\), and the reported computational complexity is \(O(SN\log N)\).

The algorithm is evaluated on the Opinosis dataset with two-sentence summaries. Reported ROUGE scores are \(R\!-\!1=32.34\) and \(R\!-\!2=9.05\), compared with human \(R\!-\!1=30.88\), \(R\!-\!2=10.69\), biclique \(R\!-\!1=32.6\), \(R\!-\!2=8.4\), TextRank \(R\!-\!1=27.56\), \(R\!-\!2=6.12\), and ParaFuse \(R\!-\!1=33.86\), \(R\!-\!2=9.74\). The paper also reports that, for *A Christmas Carol* with \(1420\) sentences, runtime is approximately \(1.74\) s for IMAKITA, \(3.71\) s for biclique, and \(106\) s for TextRank. In this setting, WIS-like behavior is tied to the classical summarization trade-off between relevance and redundancy, but the combination is multiplicative rather than additive.

## 4. Token-level importance gating in LLM watermarking

In LLM watermarking, WIS is an official acronym, but it stands for **Watermarking with Importance Scoring** rather than Weighted Importance Score. The baseline watermarking scheme partitions the vocabulary into a greenlist \(G\) and redlist \(R\) at each generation step and adds a logit bias \(\delta\) to greenlist tokens:
\[
\mathbf{l}'_t[i]=
\begin{cases}
\mathbf{l}_t[i]+\delta,& v_i\in G,\\
\mathbf{l}_t[i],& \text{otherwise}.
\end{cases}
\]
The proposed method introduces an importance function
\[
f(s=[s_1,\dots,s_{t-1}],s_t):V^*\times V\to[0,1],
\]
where importance is defined as “the degree to which it contributes to the correctness of output text given an input \(s\).” The base model’s preferred token is
\[
v_t^*=M(s_1,\dots,s_{t-1}),
\]
and WIS modifies the logits so that, if this preferred token is important, it is also boosted even when it lies in the redlist:
\[
\mathbf{l}'_t[i]=
\begin{cases}
\mathbf{l}_t[i]+\delta,& v_i\in G,\\
\mathbf{l}_t[i]+\delta,& v_i=v_t^* \text{ and } f([s_1,\dots,s_{t-1}],v_i)\ge r_0,\\
\mathbf{l}_t[i],& \text{otherwise}.
\end{cases}
\]
The score therefore acts as a binary gate rather than a continuous rescaling of \(\delta\) [2311.09668].

Three importance predictors are reported. **WIS-Perturbation** uses a BERTScore-style cosine similarity,
\[
f([s_1,\dots,s_{t-1}],s_t)=
\cos\!\big(\phi([s_1,\dots,s_{t-1}]),\phi([s_1,\dots,s_t])\big),
\]
with a sliding window of length \(w\). **WIS-Regression** trains a BERT-base model to predict the paraphrase retention frequency \(n_t/N\) for each token. **WIS-Classification** trains a BERT-base token classifier on the binary label \(\mathbb{I}(n_t>0)\). Reported thresholds include \(r_0=0.02\) for WIS-Perturbation and \(r_0=0.9\) for WIS-Regression; sliding windows are \(w=16\) on FIB and \(w=10\) on ELI5.

Detection remains unchanged and uses the same green-token \(z\)-statistic,
\[
z=\frac{|s|_G-\gamma T}{\sqrt{T\gamma(1-\gamma)}}.
\]
Experiments use LLaMA-2-13B, beam search width \(2\), maximum generated length \(T=100\), and the FIB and ELI5 datasets. The reported qualitative finding is that WIS improves ROUGE-1 at comparable detection rates. A concrete example on ELI5 with \(\gamma=0.25\) reports baseline watermarking at \(\delta=2.5\) with \(z=5.05\), and WIS-Perturbation at \(\delta=2.75\) with \(z=5.47\), while producing a more complete answer. A limitation is also reported: if too many tokens are judged important, especially at larger \(\gamma\), the achievable detection rate can decrease because the green-token bias is partially neutralized.

## 5. Importance weighting in coding, variational inference, and off-policy prediction

Importance-aware source coding provides a formal information-theoretic version of WIS. For a discrete source \(X\) with symbol probabilities \(p_i\), symbol-specific importance weights \(w_i\), and codeword lengths \(l_i\), the importance-aware expected codeword length is
\[
\bar L_w=\sum_{i=1}^N p_i w_i l_i.
\]
The associated information measure is
\[
\mathcal{L}(w,X)=
-\sum_{i=1}^N p_i w_i \log_D \frac{p_i w_i}{H_w(X)},
\qquad
H_w(X)=\sum_{j=1}^N p_j w_j.
\]
The optimal weighted code satisfies
\[
\mathcal{L}(w,X)\le \bar L_w^* < \mathcal{L}(w,X)+H_w(X).
\]
A natural per-symbol WIS is the normalized importance-weighted mass
\[
WIS_i=\frac{p_i w_i}{\sum_j p_j w_j}.
\]
The paper also studies Message Importance Measure,
\[
MIM(X;\omega)=\sum_{i=1}^N p_i e^{\omega(1-p_i)},
\]
and its normalized form
\[
MIM_N(x_i;\omega)=
\frac{p_i e^{\omega(1-p_i)}}{\sum_j p_j e^{\omega(1-p_j)}},
\]
which acts as an importance-dependent weighting of symbol influence [2005.10718].

In variational inference, a different but related WIS-like object is the importance weight
\[
w_k=\frac{p_\theta(x,z_k)}{q_\phi(z_k\mid x)},
\]
which underlies the IWAE lower bound
\[
\mathcal{L}_K(q_\phi)
=
\mathbb{E}\!\left[
\log\!\left(\frac{1}{K}\sum_{k=1}^K w_k\right)
\right].
\]
“Hierarchical Importance Weighted Autoencoders” replace i.i.d. proposals with a hierarchical joint proposal involving a shared meta-latent \(z_0\), with the stated aim of inducing negative correlation among importance weights and reducing the variance of the estimator. The paper reports both theoretically and empirically that maximization of the lower bound implicitly minimizes variance, and attributes this to negative correlation induced by the hierarchical meta-sampling scheme [1905.04866].

In off-policy reinforcement learning, WIS usually denotes **Weighted Importance Sampling**, not Weighted Importance Score. For transitions with importance ratios \(\rho_i=\pi(a_i\mid s_i)/\mu(a_i\mid s_i)\), the canonical normalized weights are
\[
\tilde w_i=\frac{\rho_i}{\sum_j \rho_j}.
\]
The paper studies several WIS baselines, including
\[
\Delta\theta=
\frac{\sum_{i=1}^n \rho_i \delta_i \nabla_\theta V_\theta(s_i)}{\sum_{j=1}^n \rho_j}
\]
for full-buffer WIS-Optimal. Its main alternative is Importance Resampling, which samples replay-buffer transitions with probabilities proportional to \(\rho_i\) and then applies on-policy TD updates. The paper proves that plain IR has the same bias as WIS-Optimal, while a bias-corrected version BC-IR is unbiased and consistent. Empirically, IR is reported to have improved sample efficiency and lower variance updates than IS, WIS variants, and V-trace in several microworlds and in a TORCS image-based value-learning experiment [1906.04328].

Taken together, these three literatures show distinct but mathematically adjacent uses of importance weighting: reweighting code costs, tightening variational lower bounds, and normalizing off-policy estimators. In each case, the central scalar changes how probability mass is treated, but the semantic meaning of “importance” differs.

## 6. Distinction from Weighted Interval Score and the problem of acronym ambiguity

A major source of confusion is that WIS is already a standard acronym for **Weighted Interval Score** in epidemic forecasting. In that literature, WIS is a proper scoring rule for probabilistic forecasts represented by quantiles or central prediction intervals. For interval level \(1-\alpha\), the interval score is
\[
\mathrm{IS}_{\alpha,t}
=
U_t^\alpha-L_t^\alpha
+\frac{2}{\alpha}(L_t^\alpha-D_t)\mathbf{1}\{D_t<L_t^\alpha\}
+\frac{2}{\alpha}(D_t-U_t^\alpha)\mathbf{1}\{D_t>U_t^\alpha\},
\]
and the per-time weighted interval score is
\[
\mathrm{WIS}_t
=
\frac{1}{J+\tfrac12}
\left(
\tfrac12 |D_t-\hat D_t|
+
\sum_{\alpha\in\mathcal A}\frac{\alpha}{2}\,\mathrm{IS}_{\alpha,t}
\right),
\]
with \(\mathcal A=\{0.02,0.05,0.10,0.20,\ldots,0.90\}\) and \(J=11\). Mean WIS is then the average of \(\mathrm{WIS}_t\) over the evaluation set. This object evaluates calibration and sharpness; it is not an importance score [2606.18575].

The same paper also defines an **InverseWIS** ensemble benchmark with weights
\[
v_k=\frac{1/\mathrm{WIS}_k}{\sum_{j=1}^K 1/\mathrm{WIS}_j},
\]
again reinforcing that WIS there is a forecasting metric. Reported mean forecast WIS values include \(24{,}438.7\) for EXP3Stoch, \(24{,}772.5\) for EXP3Det, \(24{,}897.6\) for EPSStoch, \(24{,}915.4\) for InverseWIS, and \(25{,}381.2\) for the unweighted ensemble. The main gains are reported in probabilistic forecast quality, especially WIS and interval coverage, rather than uniformly lower RMSE.

The surveyed literature therefore suggests that “Weighted Importance Score” is not a universally accepted term but a context-dependent description. In anomaly detection, summarization, watermarking, source coding, variational inference, and off-policy learning, WIS-like objects are scalar mechanisms for reallocating influence. In epidemic forecasting, by contrast, WIS is a quantile-based proper scoring rule. A plausible implication is that any technical use of “WIS” should be accompanied by its defining equation, because the acronym alone does not determine the underlying mathematics.

Source: https://www.emergentmind.com/topics/weighted-importance-score-wis