The suggestion of Points of Interest to people with Autism Spectrum Disorder (ASD) challenges recommender systems research because these users' perception of places is influenced by idiosyncratic sensory aversions which can mine their experience by causing stress and anxiety. Therefore, managing individual preferences is not enough to provide these people with suitable recommendations. In order to address this issue, we propose a Top-N recommendation model that combines the user's idiosyncratic aversions with her/his preferences in a personalized way to suggest the most compatible and likable Points of Interest for her/him. We are interested in finding a user-specific balance of compatibility and interest within a recommendation model that integrates heterogeneous evaluation criteria to appropriately take these aspects into account. We tested our model on both ASD and "neurotypical'' people. The evaluation results show that, on both groups, our model outperforms in accuracy and ranking capability the recommender systems based on item compatibility, on user preferences, or which integrate these two aspects by means of a uniform evaluation model.

1. Noemi Mauro, Liliana Ardissono, Federica Cena
noemi.mauro@unito.it, liliana.ardissono@unito.it, federica.cena@unito.it
Personalized Recommendation of PoIs
to People with Autism

2. INTRODUCTION
When suggesting PoIs to
people with Autism
Spectrum Disorder (ASD),
we must take into account
at least three aspects:
● Data scarcity (due to low number of
ASD people, interaction and attention
problems).
● Idiosyncratic sensory aversions to
avoid stress and anxiety.
● Interests to please the users.
In order to provide truly inclusive services, RS have to take into account
different factors which go farther than modeling user interests.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

3. SPATIAL NEEDS OF PEOPLE WITH AUTISM
● Atypical sensory perception in over
90% of ASD individuals: the brain
seems unable to appropriately
balance the senses.
● Reduced range of activities and
interests.
● Preference of mechanical,
deterministic situations and rigid,
repetitious routines.
● ASD people avoid places that
may negatively impact on their
senses.
● High sensory stimulation
negatively influences
individuals in their movements.
● Sensory aversions may result in
anxiety, fatigue, sense of
oppression or distraction.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

4. PROPOSAL AND PIUMA PROJECT
http://piuma.di.unito.it/
● The project has the aim to
develop digital solutions for
helping people with ASD in
their everyday movements.
● The final result of PIUMA will
be a mobile app showing maps
customized to ASD users.
PROPOSAL:
RS for ASD people that takes into
account user preferences and
aversions in feature-based user
profiles for the suggestion of safe
PoIs.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

5. USER MODEL
Preferences: elicited about
categories of PoIs such as
restaurants, parks, and so forth.
User aversions: short questionnaire
about aversions for features with
low or high levels.
Data will be elicited in the
registration phase of the PIUMA app.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

6. RATINGS AND VALUES OF FEATURES IN POIS
User Ratings:
Now by means of a questionnaire
for the evaluation of 50 specific
PoIs located in Torino city center in
a [1, 5] Likert scale with the “I don’t
know the place” option.
E.g., How much do you like Castle
Square?
soon using the PIUMA app
PoIs sensory features’ values:
● Now by means of a crowdsourcing
platform (Maps4all,
https://maps4all.firstlife.org/#/)
user can rate in the [1, 5] scale
the PoIs sensory features:
○ brightness
○ crowding
○ noise
○ smell
○ …
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

7. DATA COLLECTION
● 20 ASD adults (from 22 to 40
years-old, mean age: 26.3,
median 28; 11 men, 9 women)
patients of the Autistic Adult
center of the city of Torino,
medium- and high-
functioning.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.
● 128 neurotypical subjects
(from 19 to 71 years-old,
mean age: 28.1, median 23;
63 men, 65 women) from
university students and
authors’ contacts.
To collect ratings, preferences and idiosyncratic aversions we involved
two groups of users:

8. AVERSION TO ITEM’S SENSORY FEATURES
First case:
Given a feature f ∈ F of an item i and a user u
∈ U:
● The higher the value of f, the stronger is
its negative impact on the user; e.g.,
noise.
● vmax: maximum value of the Likert scale
(e.g.: 5).
● aufvmax: user aversion for the maximum
value of the feature.
The blue thick line shows u’s aversion
to the possible values of f.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

9. AVERSION TO ITEM’S SENSORY FEATURES
First case:
Given a feature f ∈ F of an item i and a user u
∈ U:
● vmax: maximum value of the Likert scale
(e.g.: 5).
● aufvmax: user aversion for the maximum
value of the feature.
● eaufi: u’s aversion to feature f of item i.
E.g. in the figure:
Feature value of item i: 3
eaufi: 2.5
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

10. AVERSION TO ITEM’S SENSORY FEATURES
Second case:
Given a feature f ∈ F of an item i and a user u
∈ U:
● Extreme values make users
uncomfortable, while the middle ones
are less problematic; e.g., brightness.
● vmax: maximum value of the Likert scale
(e.g.: 5).
● aufvmax and aufvmin: user aversion for the
maximum-minimum value of the
feature.
The blue-green thick line shows u’s
aversion to the possible values of f.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

11. AVERSION TO ITEM’S SENSORY FEATURES
Second case:
● eaufi: u’s aversion to feature f of item i.
● eaufi: max( , )
Example:
Feature value: 3
: 2.5 : 2
eaufi: max (2.5, 2) = 2.5
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

12. COMPATIBILITY OF FEATURES WITH THE USER
eaufi: higher values mean that the feature
generates more discomfort to u.
Compatibility is the opposite of discomfort.
User’s aversion: 2
User’s compatibility: 4
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

13. FEATURES AGGREGATION MEASURES
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

14. FEATURES AGGREGATION MEASURES
RMSD: complement of the Root Mean
Square Deviation between the features’
values of i and those of an ideal item.
Cos: distance between the features’
values of i and those of an ideal item
which best matches u’s
idiosyncrasies.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.
= ideal item which best matches u’s idiosyncrasies.
= features values of item i.

15. RATINGS PREDICTION (Ind recommender)
Ratings prediction:
Ind: the model identifies a specific
α value for each user to optimize
item recommendation to her/him.
● puci: preference in the item
category.
● compiu: compatibility between
item i and user u.
α parameter: used to balance item
compatibility and preferences for the
individual user.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

16. TEST METHODOLOGY
Methodology:
● Comparison of the results on
autistic users dataset (AUT) and on
neurotypical users one (NOR).
● 5-fold cross validation.
● For Ind RS, optimization of the α
parameter with respect to MAP.
● Metrics: Precision, Recall, F1, MAP,
MRR, MAE, RMSE.
Baseline recommenders:
● Multi-Criteria (MC): α = 0.5 →
uniformly treats items
compatibility and preferences.
● C-only: α = 1 → evaluates items
exclusively on the basis of their
compatibility with the user.
● Pref-only: α = 0 → evaluates
items on the exclusive basis of
the user’s preferences.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

17. RESULTS - ASD PEOPLE
● IndCos excels in accuracy and
ranking capabilities.
● IndMin excels in error minimization.
● Pref-only is the best baseline
regarding MAP.
● The configurations of MC have
middle to low performance.
● Similar results for neurotypical
users - see paper.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

18. DISCUSSION
● Compatibility information or
preference information alone are
not enough to generate relevant
recommendations for user.
● Personalized balance between
compatibility and user’s
preference increase the
recommendation performance.
● Personalized balance improves
item suggestion both for ASD and
neurotypical people.
The integration of possibly
heterogeneous evaluation criteria
concerning user interests and
idiosyncratic aversions is
promising for inclusive
recommender systems.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

19. FUTURE WORK
Extension of information about PoIs:
● VGI campaign with people with ASD
and their caregivers, as well as with
the general population, to acquire a
larger amount of data.
● Extraction of sensory information
from consumer reviews available in
online platforms such as
TripAdvisor.
Extension of the mobile app:
● Integration of the Ind
recommendation algorithm in the
PIUMA app.
● Learn detailed information about
user interests and aversions
based on an analysis of user
behavior.
● Users test in the city of Torino
with ASD people.
N. Mauro, L. Ardissono and F. Cena. Personalized Recommendation of PoIs to People with Autism. UMAP 2020. Genoa, Italy.

20. THANKS!
Noemi Mauro, Liliana Ardissono, Federica Cena
noemi.mauro@unito.it, liliana.ardissono@unito.it, federica.cena@unito.it