1<!doctype html><!-- This site was created with Hugo Blox. https://hugoblox.com --><!-- Last Published: September 8, 2026 --><html lang=en-us><head><meta charset=utf-8><meta name=viewport content="width=device-width,initial-scale=1"><meta http-equiv=X-UA-Compatible content="IE=edge"><meta name=generator content="Hugo Blox Builder 5.9.7"><link rel=preconnect href=https://fonts.gstatic.com crossorigin><link rel=preload as=style href="https://fonts.googleapis.com/css2?family=Montserrat:wght@400;700&family=Roboto+Mono&family=Roboto:wght@400;700&display=swap"><link rel=stylesheet href="https://fonts.googleapis.com/css2?family=Montserrat:wght@400;700&family=Roboto+Mono&family=Roboto:wght@400;700&display=swap" media=print onload='this.media="all"'><link rel=stylesheet href=/css/vendor-bundle.min.26c458e6907dc03073573976b7f4044e.css media=print onload='this.media="all"'><link rel=stylesheet href=https://cdn.jsdelivr.net/gh/jpswalsh/[email protected]/css/academicons.min.css integrity="sha512-IW0nhlW5MgNydsXJO40En2EoCkTTjZhI3yuODrZIc8cQ4h1XcF53PsqDHa09NqnkXuIe0Oiyyj171BqZFwISBw==" crossorigin=anonymous media=print onload='this.media="all"'><link rel=stylesheet href=https://cdn.jsdelivr.net/npm/[email protected]/dist/leaflet.min.css integrity crossorigin=anonymous media=print onload='this.media="all"'><link rel=stylesheet href=/css/wowchemy.5691485ce8e88de823038b8e3d90ec13.css><link rel=stylesheet href=/css/libs/chroma/github-light.min.css title=hl-light media=print onload='this.media="all"'><link rel=stylesheet href=/css/libs/chroma/dracula.min.css title=hl-dark media=print onload='this.media="all"' disabled>
vendor: 63 bytes, line 1
1<script async src="https://www.googletagmanager.com/gtag/js?id=
1G-M1JBEB5F0H
vendor: 11 bytes, line 1
1"></script>
1<script>window.dataLayer=window.dataLayer||[];function gtag(){dataLayer.push(arguments)}function trackOutboundLink(e,t){gtag("event","click",{event_category:"outbound",event_label:e,transport_type:"beacon",event_callback:function(){t!=="_blank"&&(document.location=e)}}),console.debug("Outbound link clicked: "+e)}function onClickCallback(e){if(e.target.tagName!=="A"||e.target.host===window.location.host)return;trackOutboundLink(e.target,e.target.getAttribute("target"))}gtag("js",new Date),gtag("config","G-M1JBEB5F0H",{}),gtag("set",{cookie_flags:"SameSite=None;Secure"}),document.addEventListener("click",onClickCallback,!1)</script>
1<meta name=description content="Counterfactual explanations for recommender systems aim to identify minimal changes to a user's interaction history that would alter the model's recommendation, thereby providing actionable and user-centric explanations. However, existing CE methods often rely on local perturbation or mask-based strategies that require substantial profile modifications and exhibit unstable behavior under fine-grained ranking changes. In this work, we propose DiceRec, a diffusion-based, model-agnostic framework for generating interaction-level counterfactual explanations. DiceRec leverages denoising diffusion probabilistic models to learn the distribution of user interaction histories and integrates counterfactual objectives directly into the reverse diffusion process via guided denoising. By shaping the denoising trajectory at each step, DiceRec generates compact, semantically coherent counterfactuals that induce the desired ranking changes without relying on explicit sparsity regularization or heuristic perturbation budgets. Beyond the method itself, we introduce a fine-grained and constrained evaluation protocol that reveals important differences in explanation quality that are obscured by conventional coarse-grained, percentage-based perturbation schemes. Through extensive experiments, we show that DiceRec consistently outperforms state-of-the-art baselines, particularly in minimal-perturbation scenarios that reflect realistic counterfactual conditions. Our results highlight the advantages of diffusion-based counterfactual generation and underscore the necessity of fine-grained evaluation for faithful explanation assessment in recommender systems."><link rel=alternate hreflang=en-us href=https://evazangerle.at/publication/mohammadi-sigir-2026/><link rel=canonical href=https://evazangerle.at/publication/mohammadi-sigir-2026/><link rel=manifest href=/manifest.webmanifest><link rel=icon type=image/png href=/media/icon_hu15160663797846276485.png><link rel=apple-touch-icon type=image/png href=/media/icon_hu12043829570639752632.png><meta name=theme-color content="#1565c0"><meta property="twitter:card" content="summary"><meta property="twitter:image" content="https://evazangerle.at/media/icon_hu9002677103972937553.png"><meta property="og:type" content="article"><meta property="og:site_name" content="Eva Zangerle"><meta property="og:url" content="https://evazangerle.at/publication/mohammadi-sigir-2026/"><meta property="og:title" content="Minimal-Perturbation Counterfactuals through Guided Denoising Diffusion for Recommender Systems Explanation | Eva Zangerle"><meta property="og:description" content="Counterfactual explanations for recommender systems aim to identify minimal changes to a user's interaction history that would alter the model's recommendation, thereby providing actionable and user-centric explanations. However, existing CE methods often rely on local perturbation or mask-based strategies that require substantial profile modifications and exhibit unstable behavior under fine-grained ranking changes. In this work, we propose DiceRec, a diffusion-based, model-agnostic framework for generating interaction-level counterfactual explanations. DiceRec leverages denoising diffusion probabilistic models to learn the distribution of user interaction histories and integrates counterfactual objectives directly into the reverse diffusion process via guided denoising. By shaping the denoising trajectory at each step, DiceRec generates compact, semantically coherent counterfactuals that induce the desired ranking changes without relying on explicit sparsity regularization or heuristic perturbation budgets. Beyond the method itself, we introduce a fine-grained and constrained evaluation protocol that reveals important differences in explanation quality that are obscured by conventional coarse-grained, percentage-based perturbation schemes. Through extensive experiments, we show that DiceRec consistently outperforms state-of-the-art baselines, particularly in minimal-perturbation scenarios that reflect realistic counterfactual conditions. Our results highlight the advantages of diffusion-based counterfactual generation and underscore the necessity of fine-grained evaluation for faithful explanation asse
1ssment in recommender systems."><meta property="og:image" content="https://evazangerle.at/media/icon_hu9002677103972937553.png"><meta property="og:locale" content="en-us"><meta property="article:published_time" content="2026-07-14T06:07:55+00:00"><meta property="article:modified_time" content="2026-07-19T00:00:00+00:00">
1<script type=application/ld+json>{"@context":"https://schema.org","@type":"Article","mainEntityOfPage":{"@type":"WebPage","@id":"https://evazangerle.at/publication/mohammadi-sigir-2026/"},"headline":"Minimal-Perturbation Counterfactuals through Guided Denoising Diffusion for Recommender Systems Explanation","datePublished":"2026-07-14T06:07:55Z","dateModified":"2026-07-19T00:00:00Z","author":{"@type":"Person","name":"Amir Reza Mohammadi"},"publisher":{"@type":"Organization","name":"Eva Zangerle","logo":{"@type":"ImageObject","url":"https://evazangerle.at/media/icon_hu18335020596408119049.png"}},"description":"Counterfactual explanations for recommender systems aim to identify minimal changes to a user's interaction history that would alter the model's recommendation, thereby providing actionable and user-centric explanations. However, existing CE methods often rely on local perturbation or mask-based strategies that require substantial profile modifications and exhibit unstable behavior under fine-grained ranking changes. In this work, we propose DiceRec, a diffusion-based, model-agnostic framework for generating interaction-level counterfactual explanations. DiceRec leverages denoising diffusion probabilistic models to learn the distribution of user interaction histories and integrates counterfactual objectives directly into the reverse diffusion process via guided denoising. By shaping the denoising trajectory at each step, DiceRec generates compact, semantically coherent counterfactuals that induce the desired ranking changes without relying on explicit sparsity regularization or heuristic perturbation budgets. Beyond the method itself, we introduce a fine-grained and constrained evaluation protocol that reveals important differences in explanation quality that are obscured by conventional coarse-grained, percentage-based perturbation schemes. Through extensive experiments, we show that DiceRec consistently outperforms state-of-the-art baselines, particularly in minimal-perturbation scenarios that reflect realistic counterfactual conditions. Our results highlight the advantages of diffusion-based counterfactual generation and underscore the necessity of fine-grained evaluation for faithful explanation assessment in recommender systems."}</script>
1<title>Minimal-Perturbation Counterfactuals through Guided Denoising Diffusion for Recommender Systems Explanation | Eva Zangerle</title></head><body id=top data-spy=scroll data-offset=70 data-target=#TableOfContents class=page-wrapper data-wc-page-id=d3beefdc4bfc07ba74f46a0d5637c6ec>
1<script src=/js/wowchemy-init.min.9e4214442a7711d35691acd58f6f6361.js></script>
1<aside class=search-modal id=search><div class=container><section class=search-header><div class="row no-gutters justify-content-between mb-3"><div class=col-6><h1>Search</h1></div><div class="col-6 col-search-close"><a class=js-search href=# aria-label=Close><i class="fas fa-times-circle text-muted" aria-hidden=true></i></a></div></div><div id=search-box><input name=q id=search-query placeholder=Search... autocapitalize=off autocomplete=off autocorrect=off spellcheck=false type=search class=form-control aria-label=Search...></div></section><section class=section-search-results><div id=search-hits></div></section></div></aside><div class="page-header header--fixed"><header><nav class="navbar navbar-expand-lg navbar-light compensate-for-scrollbar" id=navbar-main><div class=container-xl><div class="d-none d-lg-inline-flex"><a class=navbar-brand href=/>Eva Zangerle</a></div><button type=button class=navbar-toggler data-toggle=collapse data-target=#navbar-content aria-controls=navbar-content aria-expanded=false aria-label="Toggle navigation"> 2<span><i class="fas fa-bars"></i></span></button><div class="navbar-brand-mobile-wrapper d-inline-flex d-lg-none"><a class=navbar-brand href=/>Eva Zangerle</a></div><div class="navbar-collapse main-menu-item collapse justify-content-start" id=navbar-content><ul class="navbar-nav d-md-inline-flex"><li class=nav-item><a class=nav-link href=/#about><span>Home</span></a></li><li class=nav-item><a class=nav-link href=/#news><span>News</span></a></li><li class=nav-item><a class=nav-link href=/#publications><span>Publications</span></a></li><li class=nav-item><a class=nav-link href=/#projects><span>Projects</span></a></li><li class=nav-item><a class=nav-link href=/#service><span>Service</span></a></li><li class=nav-item><a class=nav-link href=/#contact><span>Contact</span></a></li></ul></div><ul class="nav-icons navbar-nav flex-row ml-auto d-flex pl-md-2"><li class=nav-item><a class="nav-link js-search" href=# aria-label=Search><i class="fas fa-search" aria-hidden=true></i></a></li><li class="nav-item dropdown theme-dropdown"><a href=# class=nav-link data-toggle=dropdown aria-haspopup=true aria-label="Display preferences"><i class="fas fa-moon" aria-hidden=true></i></a><div class=dropdown-menu><a href=# class="dropdown-item js-set-theme-light"><span>Light</span> 3</a><a href=# class="dropdown-item js-set-theme-dark"><span>Dark</span> 4</a><a href=# class="dropdown-item js-set-theme-auto"><span>Automatic</span></a></div></li></ul></div></nav></header></div><div class=page-body><div class=pub><div class="article-container pt-3"><h1>Minimal-Perturbation Counterfactuals through Guided Denoising Diffusion for Recommender Systems Explanation</h1><div class=article-metadata><div><span>Amir Reza Mohammadi</span>, <span>Andreas Peintner</span>, <span>Michael M. Müller</span>, <span>Eva Zangerle</span></div><span class=article-date>July, 2026</span></div><div class="btn-links mb-3"><a href=# class="btn btn-outline-primary btn-page-header js-cite-modal" data-filename=/publication/mohammadi-sigir-2026/cite.bib>Cite 5</a><a class="btn btn-outline-primary btn-page-header" href=https://doi.org/10.1145/3805712.3809584 target=_blank rel=noopener>DOI 6</a><a class="btn btn-outline-primary btn-page-header" href=https://doi.org/10.1145/3805712.3809584 target=_blank rel=noopener>URL</a></div></div><div class=article-container><h3>Abstract</h3><p class=pub-abstract>Counterfactual explanations for recommender systems aim to identify minimal changes to a user’s interaction history that would alter the model’s recommendation, thereby providing actionable and user-centric explanations. However, existing CE methods often rely on local perturbation or mask-based strategies that require substantial profile modifications and exhibit unstable behavior under fine-grained ranking changes. In this work, we propose DiceRec, a diffusion-based, model-agnostic framework for generating interaction-level counterfactual explanations. DiceRec leverages denoising diffusion probabilistic models to learn the distribution of user interaction histories and integrates counterfactual objectives directly into the reverse diffusion process via guided denoising. By shaping the denoising trajectory at each step, DiceRec generates compact, semantically coherent counterfactuals that induce the desired ranking changes without relying on explicit sparsity regularization or heuristic perturbation budgets. Beyond the method itself, we introduce a fine-grained and constrained evaluation protocol that reveals important differences in explanation quality that are obscured by conventional coarse-grained, percentage-based perturbation schemes. Through extensive experiments, we show that DiceRec consistently outperforms state-of-the-art baselines, particularly in minimal-perturbation scenarios that reflect realistic counterfactual conditions. Our results highlight the advantages of diffusion-based counterfactual generation and underscore the necessity of fine-grained evaluation for faithful explanation asse
6ssment in recommender systems.</p><div class=row><div class=col-md-1></div><div class=col-md-10><div class=row><div class="col-12 col-md-3 pub-row-heading">Type</div><div class="col-12 col-md-9"><a href=/publication/#paper-conference>Conference paper</a></div></div></div><div class=col-md-1></div></div><div class="d-md-none space-below"></div><div class=row><div class=col-md-1></div><div class=col-md-10><div class=row><div class="col-12 col-md-3 pub-row-heading">Publication</div><div class="col-12 col-md-9"><em>Proceedings of the 49th International ACM SIGIR Conference on Research and Development in Information Retrieval</em></div></div></div><div class=col-md-1></div></div><div class="d-md-none space-below"></div><div class=space-below></div><div class=article-style></div><div class=article-tags><a class="badge badge-light" href=/tag/recommender-systems/>Recommender Systems</a> 7<a class="badge badge-light" href=/tag/counterfactual-explanation/>Counterfactual Explanation</a> 8<a class="badge badge-light" href=/tag/denoising-diffusion-probabilistic-models/>Denoising Diffusion Probabilistic Models</a> 9<a class="badge badge-light" href=/tag/evaluation/>Evaluation</a></div><div class=share-box><ul class=share><li><a href="https://twitter.com/intent/tweet?url=https%3A%2F%2Fevazangerle.at%2Fpublication%2Fmohammadi-sigir-2026%2F&text=Minimal-Perturbation+Counterfactuals+through+Guided+Denoising+Diffusion+for+Recommender+Systems+Explanation" target=_blank rel=noopener class=share-btn-twitter aria-label=twitter><i class="fab fa-twitter"></i></a></li><li><a href="https://www.facebook.com/sharer.php?u=https%3A%2F%2Fevazangerle.at%2Fpublication%2Fmohammadi-sigir-2026%2F&t=Minimal-Perturbation+Counterfactuals+through+Guided+Denoising+Diffusion+for+Recommender+Systems+Explanation" target=_blank rel=noopener class=share-btn-facebook aria-label=facebook><i class="fab fa-facebook"></i></a></li><li><a href="mailto:?subject=Minimal-Perturbation%20Counterfactuals%20through%20Guided%20Denoising%20Diffusion%20for%20Recommender%20Systems%20Explanation&body=https%3A%2F%2Fevazangerle.at%2Fpublication%2Fmohammadi-sigir-2026%2F" target=_blank rel=noopener class=share-btn-email aria-label=envelope><i class="fas fa-envelope"></i></a></li><li><a href="https://www.linkedin.com/shareArticle?url=https%3A%2F%2Fevazangerle.at%2Fpublication%2Fmohammadi-sigir-2026%2F&title=Minimal-Perturbation+Counterfactuals+through+Guided+Denoising+Diffusion+for+Recommender+Systems+Explanation" target=_blank rel=noopener class=share-btn-l
9inkedin aria-label=linkedin-in><i class="fab fa-linkedin-in"></i></a></li><li><a href="whatsapp://send?text=Minimal-Perturbation+Counterfactuals+through+Guided+Denoising+Diffusion+for+Recommender+Systems+Explanation%20https%3A%2F%2Fevazangerle.at%2Fpublication%2Fmohammadi-sigir-2026%2F" target=_blank rel=noopener class=share-btn-whatsapp aria-label=whatsapp><i class="fab fa-whatsapp"></i></a></li><li><a href="https://service.weibo.com/share/share.php?url=https%3A%2F%2Fevazangerle.at%2Fpublication%2Fmohammadi-sigir-2026%2F&title=Minimal-Perturbation+Counterfactuals+through+Guided+Denoising+Diffusion+for+Recommender+Systems+Explanation" target=_blank rel=noopener class=share-btn-weibo aria-label=weibo><i class="fab fa-weibo"></i></a></li></ul></div></div></div></div><div class=page-footer><div class=container><footer class=site-footer><p class="powered-by copyright-license-text">© 2025 Eva Zangerle. This work is licensed under <a href=https://creativecommons.org/licenses/by-nc-nd/4.0 rel="noopener noreferrer" target=_blank>CC BY NC ND 4.0</a></p><p class="powered-by footer-license-icons"><a href=https://creativecommons.org/licenses/by-nc-nd/4.0 rel="noopener noreferrer" target=_blank aria-label="Creative Commons"><i class="fab fa-creative-commons fa-2x" aria-hidden=true></i> 10<i class="fab fa-creative-commons-by fa-2x" aria-hidden=true></i> 11<i class="fab fa-creative-commons-nc fa-2x" aria-hidden=true></i> 12<i class="fab fa-creative-commons-nd fa-2x" aria-hidden=true></i></a></p><p class=powered-by>Published with <a href="https://hugoblox.com/?utm_campaign=poweredby" target=_blank rel=noopener>Hugo Blox Builder</a> â the free, <a href=https://github.com/HugoBlox/hugo-blox-builder target=_blank rel=noopener>open source</a> website builder that empowers creators.</p></footer></div></div>
12<script src=/js/vendor-bundle.min.50933d940896e49f984a778650d5f7f5.js></script>
12<script src=https://cdn.jsdelivr.net/npm/[email protected]/dist/leaflet.min.js integrity crossorigin=anonymous></script>
12<script id=search-hit-fuse-template type=text/x-template> 13 <div class="search-hit" id="summary-{{key}}"> 14 <div class="search-hit-content"> 15 <div class="search-hit-name"> 16 <a href="{{relpermalink}}">{{title}}</a> 17 <div class="article-metadata search-hit-type">{{type}}</div> 18 <p class="search-hit-description">{{snippet}}</p> 19 </div> 20 </div> 21 </div> 22 </script>
22<script src=https://cdn.jsdelivr.net/gh/krisk/[email protected]/dist/fuse.min.js integrity="sha512-o38bmzBGX+hD3JHWUFCDA09btWaqrNmoJ3RXLlrysA7PP01Kgs4UlE4MhelE1v5dJR3+cxlR4qQlotsW7jKsnw==" crossorigin=anonymous></script>
22<script src=https://cdn.jsdelivr.net/gh/julmot/[email protected]/dist/jquery.mark.min.js integrity="sha512-mhbv5DqBMgrWL+32MmsDOt/OAvqr/cHimk6B8y/bx/xS88MVkYGPiVv2ixKVrkywF2qHplNRUvFsAHUdxZ3Krg==" crossorigin=anonymous></script>
22<script id=page-data type=application/json>{"use_headroom":true}</script>
22<script src=/js/wowchemy-headroom.db4755770454eb63685f8de785c0a172.js type=module></script>
22<script src=/en/js/wowchemy.min.abfe36939207af9cb82df4fe2604ba17.js></script>
22<script src=/js/wowchemy-map.a26e9d2f7238ba5b868384f1c5bc6477.js type=module></script>
22<div id=modal class="modal fade" role=dialog><div class=modal-dialog><div class=modal-content><div class=modal-header><h5 class=modal-title>Cite</h5><button type=button class=close data-dismiss=modal aria-label=Close> 23<span aria-hidden=true>×</span></button></div><div class=modal-body><pre><code></code></pre></div><div class=modal-footer><a class="btn btn-outline-primary my-1 js-copy-cite" href=# target=_blank><i class="fas fa-copy"></i> Copy 24</a><a class="btn btn-outline-primary my-1 js-download-cite" href=# target=_blank><i class="fas fa-download"></i> Download</a><div id=modal-error></div></div></div></div></div>
24<script src=/js/wowchemy-publication.9c0e895144aef5a693008b5c5d450147.js type=module></script>
24</body></html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.