In this paper, we summarize our previous contribution to the research area of Explainable Recommender Systems in the healthcare domain, called ICARE (Intuitive Context-Aware Recommender with Explanations), which is a framework based on data-mining algorithms that can provide personalized recommendations with contextual and intuitive explanations. In particular, we consider the scenario related to physical activities to improve sleep quality, and we now describe how the system satisfies the four principles of Explainable Artificial Intelligence.
Collecting data on people’s behaviors and habits has become easier since sensors and mobile devices are widespread and low-cost elements of modern daily life. Recommendation Systems are commonly used to exploit data, often collected through IoT devices, and support users in decision-making processes that cover diferent domains. The proposed suggestions are based on the users’ past habits or their profiles. tems (HRS) provide recommendations in the healthcare context [1]. Possible applications range from healthy lifestyle or balanced nutrition plans to healthcare information and medical treatment suggestions. Collecting data related to people’s behaviors and well-being has also become easier thanks to wearable devices; indeed, many people own them and can monitor with simple apps their movements, record their sleep quality and heartbeats while also being surrounded by IoT devices embedded in common appliances in homes, ofices and means of transport. This situation ensures a constant stream of new data that can be integrated with external data sources. It ofers increasing opportunities for data analytics in the nEvelop-O Published in the Proceedings of the Workshops of the EDBT/ICDT 2024 ⋆Summary of the bookchapter ICARE: An Intuitive Context-Aware Recommender with Explanations, in Advances in AI-Enhanced ∗Corresponding author. †These authors contributed equally. [3].
The temporal aspect of data collected from sensors, as wearable devices, is very important and can be considered another contextual dimension for obtaining useful knowledge. Sensors collect information about events happening in succession, and thus (stream) data are temporal. When considering the temporal aspect, detecting events’ periodicity or temporal correlations is possible and allows the prediction of future situations by improv
To be sure that users will accept the provided recommendations, an HRS should not only focus on eficiency and accuracy because clear explanations accompanying suggestions can help users understand and trust the system [5]. Providing clear, timely recommendations and intuitive and simple explanations is very important, especially in healthcare.
collecting data can also help people gain more awareness healthcare domain to produce useful insights. Moreover, ing the provided suggestion [4]. CEUR
ceur-ws.org τ w D
Augmentation
DTw of frequent sequences , used to generate a set of totally ordered sequential rules , formalized as implications → , where and are two sets of ordered data items, such that ∩ = ∅ , according to specific thresholds to carry out to reach and follow a healthy lifestyle and im- for confidence and support. Support is the frequency prove sleep quality. To support the user in the choice of the next activities to perform, we propose the ICARE (Intuitive Context-Aware Recommender with Explanations) of the set ∪
in the dataset, while confidence is the conditional probability of finding , having found and is given by ( ∪ )/( ) . Based on the confidence framework based on data mining algorithms for personaland completeness of rules, the recommender system orized recommendations. The supplied recommendations, in the form of totally ordered sequential rules, include ders to produce a set of totally ordered sequential rules that can be queried to extract a positive recommendation the context and perform well in providing explanations + and a negative one − .
In this Section, we describe the overall approach of ICARE, which is represented in Figure 1 and the ALBA (Aged LookBackApriori) algorithm used to infer totally ordered sequential rules that are then used to provide contextual recommendations [10]. nario, the log of physical activity levels and sleep scores collected by Fitbit), enriched with contextual information. We leverage the temporal dimension to capture the user’s habits on some specific days, i.e., we distinguish whether each day is a weekday or part of the weekend. Whenever it is possible, we also integrate weather conditions to establish if the user’s preferences are related to specific situations. The dataset D is fed into ALBA, which then considers a temporal window to construct an augmented data set . This augmented dataset is then fed to a version of the Apriori algorithm [11] that calls an aging mechanism at each iteration to calculate the following: 1 ∶ { ∶ 3, ∶ 2} 2 ∶ { , ∶ 3} −1 ∧ { ∶ 3, ∶ 2}
The increasing complexity of decision-making processes and intelligent systems has made them more challenging to understand. This has motivated the research community’s attention regarding AI systems’ transparency, ethics, and accountability. Explainable AI focuses on creating and deploying AI systems that explicitly explain
ICARE needs as input a temporal dataset (in our sce- the weather conditions are good, and the user performed their decision-making processes and the outcomes gen- explains the suggestion (i.e., the consequent of mined erated by machine learning algorithms. rules) in their antecedent. For example, consider the
Four principles for explainable AI have been identified following two rules: [8] and are: 1 ∶ {, , , ∶ 1, ∶ 2} Explanation This principle asserts that a system can explain, provide evidence, and explicitly support every decision made. Our choice to use ALBA, an algorithm based on Apriori, ofers the possibility of fulfilling this requirement because the form of mined rules directly
Now, we will show how these principles are naturally integrated into ICARE, since the recommendation process is based on mined sequential rules that contain their explanation.
• Explanation: AI systems should be able to provide
clear explanations for their actions 2 ∶ { ∶ 3, ∶ 2} −1 ∧ { ∶ 3, ∶ 2} 0 → { ∶ 1} 0 • Meaningful: explanations ofered by AI systems 1 states that on a cold day with good weather condimust be comprehensible and relevant to humans, tions during the week (WD), if the user performs a low particularly those who are not experts in the field. level of heavy physical activity ( ∶ 1 ) and a medium • Explanation accuracy: explanations need to be level of light activity ( ∶ 2 ), the resulting sleep score precise and accurate. Thus, it is possible to iden- will have a low value ( ∶ 1 ). tify the most important variables/features in a 2 states that independently from the context, if the decision-making process. user yesterday performed a high level of heavy physical • Knowledge limits: AI systems should recognize activity ( ∶ 3 ) and a medium level of light activity their limitations and uncertainties, operating ex- ( ∶ 2 ), and today performs the same level of physical clusively within their designed conditions. activity, the resulting sleep score will have a low value ( ∶ 1 ).
In Fig. 2, we show two screenshots of the ICARE mobile application with information provided to the user.
ICARE is an Android app with a Python backend. The left screenshot shows the predicted sleep score and the best positive (green) and negative (red) recommendations.
The right screenshot shows the contextual explanation (the antecedents of mined rules) of the sleep score, reported in the lower part ”Because you have been doing”.
when the final user understands the explanations provided by the system, which should be generated together with each recommendation. The development of the ICARE app has been performed by considering this principle; indeed, both positive and negative suggestions should help the user better understand what influences his/her sleep quality. In Fig. 2, the left-hand side screenshot reports positive and negative suggestions and the sleep quality explanation. In this example, the current context is Sunny Holiday, and the goal is a score greater than or equal to 11 for sleep quality. Recommendations for activities to perform are represented in part ”Next step” on a green background, and recommendations for activities to avoid in part ”Avoid” on a red background.
ating trustworthy in suggestions proposed by automated systems, and thus, it characterizes trust in AI systems.
Trust is fundamental, especially in the medical domain.
A method to define desired properties is needed to characterize a good explanation from an AI system. The four principles of Explainable Artificial Intelligence have been proposed in [8] for defining the factors to consider for producing a suitable explanation and possibly evaluating its quality. The proposed principles cover diferent perspectives and consider the importance of explaining the outcome of the process, but also the style, the aim, and the recipient of the explanation itself. In [12], the four principles proposed in [8] have been applied to BioExplanation Accuracy The Explanation and Meaning- metrics and Facial Forensic Algorithms in order to conful principles require a system to generate explanations sider trust and societal norm of AI systems and provide a that are comprehensible to humans. These principles do foundation of explainability. not impose that a system provides precise explanations. In [13], the authors consider problems related to exThe Explanation Accuracy principle introduces a chal- plaining black-box algorithms and classify them concernlenge in motivating accuracy for a system’s explanations ing the notion of explanation. Starting from the descripthrough metrics. tion of a problem definition, a type of black box, and
During the implementation of ICARE, we have col- a preferred explanation, the survey should support the lected data from 4 users, specifically for this study. For researcher in finding the proposals that best fit their reour approach’s experimental evaluation, we verified that quirements. relevant features influencing sleep quality may difer for In [14], the authors provide an analytical review of the users. Thus, after a training phase, we have considered papers in the literature focusing on the explainability of only relevant features for each user. We have performed artificial intelligence in the context of machine learning interviews for our evaluation to understand if the mined and deep learning. The work briefly describes the difinsights are accurate. ferent terms used to indicate the understandability of a
For example, one user has sleep quality that is highly system and maps out the main challenges to deal with to correlated to weather conditions, as shown by the follow- fulfill the explainability issues. ing rules: In [15], the notion of explainable AI has been explored in the biomedical contexts for proposing a functional 1 1 ∶ { , , ∶ 2, ∶ 3} 0 → { ∶ 3} 0 definition and a conceptual framework that can be used 1 2 ∶ { , , , ∶ 2, ∶ 1} 0 → { ∶ 1} 0 when considering explainable AI.
In this work, we show that data mining algorithms still represent an alternative that fits well for producing explainable recommendations.
Another user has sleep quality correlated to the part of the week (i.e., weekend or weekday), as shown by the following rules: 2 2 1 ∶ { , ∶ 1, ∶ 1} 1 ∶ { , ∶ 1, ∶ 3} 0 → { ∶ 3} 0 → { ∶ 1}
In this paper, we have described an Intuitive ContextKnowledge limits This principle requires identifying Aware Recommender with Explanations (ICARE), which and pointing out cases the system could not manage. is a framework for collecting and enriching wearable ICARE provides suggestions based on collected data, and device data with contextual information to produce relin selecting rules to consider for producing recommen- evant insights. Indeed, sequential rules correlating sedations, it considers the thresholds set for confidence quences of past events with a specified future goal are and support. This means that the ALBA algorithm may discovered by analyzing temporal logs. Finally, ICARE not produce suggestions when it does not mine suitable provides explainable recommendations using an intuitive rules, but it cannot be found in a situation out of its scope. application.
When users do not wear their Fitbit constantly, the pre- As for future work, we plan to extend the ICARE app cision of rules decreases. In the future, we should add a to collect the user’s feedback on the received predictions module to delete outliers from data. and recommendations. Moreover, we are applying the framework in other scenarios, particularly for maintaining glucose levels in the normal range, by suggesting how to organize physical activity and meals. Another possible domain is a recommendation system that suggests balanced meals based on what they recently ate, also considering where meals are consumed (e.g., home, school canteen, …).