Information Extraction for Inclusive Recommender Systems

Abstract

Inclusive recommender systems should take both user preferences and the compatibility of items with the user into account in order to generate suggestions that can be appreciated and smoothly experienced at the same time. For instance, considering people in the Autism Spectrum Disorder, the sensory features of a place that is potentially interesting to the user are important to predict whether it might make her/him uncomfortable when visiting it. However, information about users' experience with items can hardly be found in the metadata provided by online geographic sources. In order to address this issue, we suggest to retrieve it from the consumer feedback collected by location-based services that publish item reviews. This type of feedback represents a sustainable information source because it is supported by people through a continuous reviewing activity. Thus, it deserves special attention as a potential data source. In this paper, we outline how this type of information can be retrieved and we discuss its benefits to Top-N recommendation of Points of Interest.

Keywords

• eol>Recommender Systems
• Geographic Information Systems
• People with Autism

1. Introduction

In the development of inclusive recommender systems, multiple factors have to be taken into account to support a positive user experience with the suggested items. For instance, in Points of Interest (PoIs) recommendation, accessibility issues, such as architectonic barriers, should be considered to avoid imposing extra fatigue on the user, or making it hard to reach the place, if (s)he is physically impaired. Moreover, if the user is in the Autism Spectrum Disorder (ASD), the sensory features of places have to be taken into account to suggest PoIs that are compatible with her/his aversions. However, physical and sensory features of items are only a part of a broader scope of characteristics that can negatively influence user experience. For instance, cultural aspects might impose constraints on garment styles to be considered in order to avoid ofending the user with proposals that (s)he cannot accept.

These examples support the idea that both user preferences, and the compatibility of items with the user are key to the suggestion of items that (s)he can like and smoothly experience at the same time. Diferent compatibility aspects might be modeled within a recommender system, depending on its goals, such as impairments, sensory aversions, cultural principles, and so forth. The basic difference with respect to preference modeling 2. Background and related work is that the system should not assess whether the user likes more or less a property of an item, but if that property can cause any discomfort, or dificulties, to her/him. Thus, for As discussed by Ghose and Ipeirotis in [ 4 ], instance, a single, totally incompatible fea- online reviews are a precious source of inture might make an item unsuitable for the formation about products and services beuser, even though its other features would cause they describe previous consumers’ exsatisfy her/him very well. perience with items. Notice also that, as re

A few mobile guides propose models for views are voluntarily provided by people all the evaluation of items compatibility with over the world, and they are continuously users in personalized recommendation. For uploaded, they represent an ideally unlimited instance, INTRIGUE [ 1 ] focuses on physical source of up-to-date information about items accessibility of items, while PIUMA [ 2, 3 ] in- that can be used to feed a recommender sysvestigates their compliance with the sensory tem. aversions of people. In both works, the col- A lot of work has been carried out to exlection of information about items support- tract relevant data from online reviews with ing compatibility evaluation is a problematic the purpose of personalizing recommendatask. Specifically, while geographic informa- tion to the individual user [5, 6], or to astion sources provide some accessibility data sess review helpfulness [ 4, 7, 8, 9, 10 ]. More(e.g., wheelchair support), they typically of- over, a parallel research thread applies opinfer standard types of information which can ion mining to identify pros and cons of items, hardly represent the user experience with as observed by consumers, with the aim of items in a complete way. Moreover, metadata highlighting aspects that can be improved or do not always reflect real user experience. promoted. For instance, see [11], [12], and For instance, even though a hotel claims that [13]. However, those works rely on statistical it ofers Wi-Fi, the quality of internet ac- analyses of text and they focus on the most cess can only be measured when visiting the frequently reported aspects of items, such as place. the price and cleanliness of a hotel, or the

In order to address this issue, we want quality of the food served by a restaurant. to investigate the descriptive power of con- In contrast, compatibility evaluation is resumer feedback collected by location-based lated to individual idiosyncrasies. For this services that publish online item reviews. reason, it should be based on a deep invesSpecifically, we want to study the extrac- tigation of users’ perceptions of items, retion of data about sensory features from tex- gardless of how many people highlight the tual comments in order to check whether various issues in their reviews. For instance, leveraging this type of information in Top- even though a single person points out that a N recommendation improves the suggestion restaurant is challenging for somebody who of places. uses a wheelchair because the tables are too close to each other, this information should be recorded and taken into account by the system. For this reason, instead of identifying the main item properties that emerge from a statistical analysis of consumer feedback (bottom-up), we aim to start from the identification of the types of features that can Then, we plan to use Natural Language Prodetermine a compatibility problem. Then, cessing tools to extract the occurrences of we want to search for these features in the such words from the bulk of reviews associreviews (top-down). We hypothesize that ated with each individual item. In this way, this approach has the advantage that isolated we can build an item profile that specifies, for opinions can bring useful data to be used each feature, the mean value emerging from in the cautious type of recommendation we the complete set of occurrences of the assopursue. ciated words.

Firstly, we will focus on sensory features to use them in compatibility evaluation of 3. Item recommendation PoIs in relation to people with autism. For this purpose, we are investigating online reWe plan to exploit the feature values ex- view repositories providing consumer feedtracted from consumer feedback to evaluate back about PoIs, such as Yelp [14], TripAdvithe compatibility of each feature of an item sor [15], and Google Maps [16]. However, we with a user , taking her/his aversions into will extend our analysis to other types of feaaccount. Following the approach presented tures, such as those related to physical accesin [ 2 ], we will combine the compatibility of sibility, in order to complement the standard the features of with ’s preference for the type of information provided by data-sources category of to obtain the final score of the such as OpenStreetMap [17] with the percepitem. Compatibility and preference informa- tions of previous visitors. tion can be integrated in diferent ways. For We also plan to combine the extracted instance, in [ 2 ] we proposed to compute an information with other data sources. For overall compatibility value for the item instance, we will consider OpenStreetMap and to combine it with ’s preference for metadata provision and possibly crowdfor the category of (e.g., cinema, park, etc.) sourcing platforms such as Maps4All [18], in order to estimate the rating ̂ of : which support a flexible type of geo-data ̂ = ∗ + (1 − ) ∗ (1) rmicahpeprinigte. mIn pthroisfilewsatyo, wbee wuisleldpofsosribplyerosbotna-in where ∈ [ 0, 1 ] tunes the influence of com- alized recommendation. We are aware that patibility and preference information in rat- is hard to collect real, objective sensory feaing estimation. However, other methods can tures of PoIs, since the same place can be perbe applied, which we plan to investigate. ceived diferently from person to person, especially with notable diferences between individuals with autism or not. However, we 4. Extraction of think that, by merging a large amount of difcompatibility features ferent points of view on the same place, as provided by online reviews, we will be able from online reviews to obtain an image as similar as possible to its real characteristics.

In order to support a top-down search of compatibility features in item reviews, we plan to identify the words that refer to such features and to map words to feature values.

Acknowledgments

We thank Claudio Mattutino and the Adult Autism Center of the City of Torino for their support to this work.

actions on on Knowledge and Data Engineering 23 (2011) 1498–1512. URL: https://doi.org/10.1109/TKDE.2010.188. doi:10.1109/TKDE.2010.188. [5] M. Hernández-Rubio, I. Cantador, A. Bellogín, A comparative analysis of recommender systems based on item aspect opinions extracted from user reviews, User Modeling and User-Adapted Interaction 29 (2019) 381–441. URL: https: //doi.org/10.1007/s11257-018-9214-9. doi:10.1007/s11257-018-9214-9. [6] L. Chen, G. Chen, F. Wang, Recommender systems based on user reviews: the state of the art, User Modeling and User-Adapted Interaction 25 (2015) 99–154. URL: https: //doi.org/10.1007/s11257-015-9155-5. doi:10.1007/s11257-015-9155-5. [7] R. Dong, M. Schaal, M. P. O’Mahony, B. Smyth, Topic extraction from online reviews for classification and recommendation, in: Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence, IJCAI ’13, AAAI Press, 2013, p. 1310–1316. [8] H. Hong, D. Xu, G. Wang, W. Fan, Understanding the determinants of online review helpfulness, Decision Support Systems 102 (2017) 1–11. URL: https: //doi.org/10.1016/j.dss.2017.06.007.

doi:10.1016/j.dss.2017.06.007. [9] M. Siering, J. Muntermann, B. Rajagopalan, Explaining and predicting online review helpfulness: the role of content and reviewerrelated signals, Decision Support Systems 108 (2018) 1 – 12. URL: http://www.sciencedirect.com/ science/article/pii/S0167923618300149. doi:10.1016/j.dss.2018.01.004. [10] N. Mauro, L. Ardissono, G. Petrone, User and item-aware estimation of review helpfulness, Information Process

References

[1] L. Ardissono , A. Goy , G. Petrone, M. Segnan , P. Torasso, INTRIGUE: personalized recommendation of tourist attractions for desktop and handset devices , Applied Artificial Intelligence, Special Issue on Artificial Intelligence for Cultural Heritage and Digital Libraries 17 ( 2003 ) 687 - 714 . URL: https://www.tandfonline.com/ doi/abs/10.1080/713827254. doi:https: //doi.org/10.1080/713827254.

[2] N. Mauro , L. Ardissono , F. Cena , Personalized recommendation of PoIs to people with autism , in: Proceedings of the 28th ACM Conference on User Modeling, Adaptation and Personalization , UMAP'20 , ACM , New York, NY, USA, 2020 , pp. 163 - 172 . URL: https://dl. acm.org/doi/10.1145/3340631.3394845. doi: 10 .1145/3340631.3394845.

[3] A. Rapp , F. Cena , C. Mattutino , G. Boella, C. Schifanella , R. Keller, S. Brighenti, Designing an urban support for autism , in: Proceedings of the 21st International Conference on Human-Computer Interaction with Mobile Devices and Services , MobileHCI 2019 , Taipei, Taiwan, October 1- 4 , 2019 , ACM, 2019 , pp. 43 : 1 - 43 : 6 . URL: https: //doi.org/10.1145/3338286.3344390. doi: 10 .1145/3338286.3344390.

[4] A. Ghose , P. G. Ipeirotis, Estimating the helpfulness and economic impact of product reviews: mining text and reviewer characteristics , IEEE Trans-