In point-of-interest recommendation for people with autism, standard preference-driven recommenders often misalign with sensory sensitivities and severe data scarcity, which can yield suggestions that are hard to trust and potentially harmful for everyday exploration. Knowledge-based reasoning and explanation mechanisms can enable more data-efficient, safety-aware personalization in this setting by explicitly modeling user–place sensory compatibility.
Supporting spatial exploration for autistic people is a recommender-systems problem with unusually high stakes. Visiting unfamiliar places can involve unpredictable stimuli, and for many individuals, sensory overload is a trigger for anxiety and distress. In this context, a recommendation is a decision aid that can either reduce uncertainty or amplify it.
This changes what “good personalization” means. First, the system must account for both sensory preferences and sensory aversions, because avoiding problematic stimuli can be as important as seeking appealing ones. Second, the system must be robust under very limited interaction data, because recruiting and engaging autistic participants for research and longitudinal logging is inherently difficult. Third, to be usable, recommendations need to be justified in a way that supports structured decision-making and is communicated in forms that fit individual cognitive and linguistic profiles.
In a study, in cooperation with Federica Cena, Mirko Marras, Noemi Mauro, and Giacomo Medda, and published in the Proceedings of ACM SIGIR 2025, we introduce an agenda and initial design direction for point-of-interest recommendation tailored to autistic individuals under severe resource constraints.
Rather than presenting a fully finalized system with extensive benchmarking, the work is positioned around identifying the core challenges that make this domain distinctive and outlining a coherent approach to progress despite small samples, sparse sensory signals, and the need for explanation and accessibility from the outset.
High-level solution overview
Our central idea is to treat sensory-aware POI recommendation as a structured reasoning problem grounded in explicit domain knowledge, not just as a pattern-mining task over clicks or ratings. Concretely, we aim to: (i) collect limited but high-value data through a reproducible involvement protocol co-designed with specialists, (ii) encode that data into a knowledge graph that captures relations between users, sensory characteristics, and places, and (iii) generate recommendations through path-based reasoning that can be surfaced as explanations. We then focus on translating those explanations into accessible natural language and interfaces appropriate for autistic users.
This framing emphasizes three capabilities that matter in low-resource, high-impact settings: sample efficiency (because the data will remain small), controllability (because sensory aversions must be handled explicitly), and transparency (because the user’s decision process benefits from clear reasoning pathways).
Approach
Structured user involvement as a data-quality mechanism
The first problem we address is that the main bottleneck is not only “insufficient data,” but “insufficient reliable and safe-to-collect data.” Standard recruitment and evaluation procedures can fail when participants have limited engagement windows, strong routine preferences, or discomfort with unfamiliar interactions.
Our methodological move is to treat user involvement as a designed protocol rather than an ad hoc phase. We introduce a structured, reproducible process shaped with clinicians and ASD specialists to support recruitment, ethical handling, and elicitation of both preferences and aversions. The key abstraction is that each collected datum should be interpretable and actionable for later modeling (e.g., it should map onto sensory concepts and POI attributes). In small-data regimes, the quality and structure of signals dominate the downstream viability of modeling choices.
Ontology-guided representation of sensory compatibility
The second problem is that sensory characteristics of places are rarely available in a form that supports systematic personalization. Even when some proxy information exists, it tends to be unstructured and does not directly expose relations among a user’s sensitivities, environmental stimuli, and the kinds of places that may be suitable.
We address this by defining an ontology that captures the relevant entities and relations: autistic users, sensory aversions, POI categories, and sensory features of POIs. The assumption is that making these concepts explicit enables a coherent representation even when observations are sparse. We then instantiate this ontology as a knowledge graph constructed from the limited dataset available to the project. This choice is important because the graph becomes a scaffold for generalizing from a few observations: it allows the system to reason over relations (e.g., how a user’s aversion connects to a place’s sensory properties via intermediate concepts) instead of relying on dense behavioral histories.
Path-based reasoning to unify recommendation and explanation
The third problem is transparency under personalization. For autistic users, “why” can be as important as “what,” and explanations should reduce cognitive load rather than introduce additional complexity. At the same time, explanations should be faithful to the decision logic used by the recommender system; otherwise, they risk undermining trust.
Our approach uses path reasoning over the knowledge graph, where a recommendation is supported by a structured pathway connecting the user to a candidate POI through sensory and categorical relations. The introduced signal is the path itself: it functions as a compact representation of the rationale. We explore two complementary reasoning paradigms: one that treats traversal as a sequential decision process over the graph, and one that treats graph elements as symbols that can be composed into meaningful sequences. In this scenario, path-based reasoning provides an intrinsic bridge between decision and justification: the same structure that supports ranking can also support explanation, and the explanation can be made more cognitively aligned by controlling the complexity of the path.
Accessibility-aware delivery and evaluation as key constraints
The fourth problem is that even correct recommendations with faithful rationales can fail if they are not delivered in an accessible format. Information must be presented in ways that fit perceptual and cognitive patterns, and evaluation must be adapted to engagement constraints and conducted with appropriate support.
We therefore integrate psychologists into the design loop to shape explanation templates and tailor the user interface, including controlling explanation complexity to match individual capacities. The underlying assumption is that explanation usefulness is not monotonic in detail: more information can increase overload, so the system should enable calibrated, structured delivery. This matters for evaluation as well: success criteria and study designs must reflect the lived constraints of the target population, requiring close coordination with caregivers and healthcare professionals rather than standard large-scale A/B testing paradigms.
Conclusions
This paper contributes a research-oriented direction for sensory-aware POI recommendation for autistic people under severe resource limitations, emphasizing that progress depends on aligning data collection, representation, reasoning, and delivery with the realities of the user population. By grounding personalization in an ontology-driven knowledge graph and using path reasoning to connect recommendation and explanation, we aim to make transparent, controllable decision support feasible even when conventional large-scale learning is not.
Several extensions naturally follow from this framing. We can expand how sensory information about places is obtained, including careful use of proxy sources while preserving interpretability and minimizing noise. We can develop risk-aware recommendation strategies that explicitly manage uncertainty when evidence is thin. We can study how explanation complexity should adapt over time and across user profiles, treating accessibility as a measurable objective rather than a static design choice. Finally, we can design longitudinal, clinician-supported evaluations that capture not only acceptance but also well-being and confidence during real-world exploration.