Nowadays, more and more people are using the Web to search for health information. However, it is widely accepted that it is really hard for people to determine the quality of the presented information and to accurately judge on the relevance to their own condition. The FairGRecs system recommends to small groups of persons health documents selected by caregivers. The system exploits ontologies to model patient pro les and documents content, and then it uses a notion of semantic distance between patients in order to provide useful recommendations by incorporating the notion of fairness. In this paper, we describe the next step in this direction, namely adapting recommendations considering the educational level of the end-users and their psycho-emotional status.
During the last decade, the number of users who look for health and medical
information online has dramatically increased. However despite the increase in
those numbers, it is very hard for a patient to accurately judge the relevance
of some information to his/her own case and to identify the quality of the
provided information. On the other hand, existing health information services (e.g.
WebMD, MayoClinic Patient Care, Medicine Plus, HONSearch, PHIR [
Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
FairGRecs [
However recommendation algorithms so far ignore the fact that patients proles are multifaceted. For example, recommending the proper document should not only focus on the patients relevant problems but also on their health literacy (namely, the ability to obtain, read, understand, and use health care information in order to make appropriate health decisions and follow instructions for treatment), educational level and psychoemotional status, as emotions can greatly a ect the cognitive processes. In this paper, we explore those dimensions, as well paving the way for a new system incorporating all aforementioned aspects. 2
Ratings, Semantic Distance, Health Literacy and
Psychoemotional Status into the Mixer
The rst step in exploiting pro le information, is to be able to record it. To
this purpose, speci c short validated questionnaires [
Furthermore, for documents, we also need to have information regarding the target population concerning the 4 aforementioned dimensions. As such, all documents entered by the caregivers are annotated with numbers regarding target population health literacy, education level, cognitive closure and anxiety. In addition, the documents are automatically annotated using ICD-10 ontology3, and all annotations are stored into the document corpus.
Now, given a set of data items I and a set of patients U , we need to focus
rst on single user recommendations. A patient, or user, u might rate an item i
with a score r(u; i). The subset of items rated by a user u is denoted by I(u).
Typically, the cardinality of I is high and users rate only a few items. For the
items unrated by the users, recommender systems estimate a relevance score,
denoted as relevance(u; i). As we are using the collaborative ltering approach,
similar users should be located via a similarity function that will evaluate the
similarity between two users. Then items relevance scores should be computed for
3 http://www.icd10data.com/
users taking into account their most similar users. The novelty of our approach
lies in the fact that instead of using only classical similarity notions or based
only on their diseases as in [
Traditionally, two users are similar if they have rated data items in a similar way, i.e., they share the same interests. For calculating their similarity, we exploit the Pearson correlation metric:
RatS(u; u0) = pPi2X (r(u; i)
Pi2X (r(u; i) u)(r(u0; i)
u0 ) u)2pPi2X (r(u0; i) u0 )2 ; where X = I(u) \ I(u0), u is the mean of the ratings in I(u).
Pearson correlation actually measures the linear dependence between two users u and u0 : it has a value between +1 and 1, where +1 is total positive linear correlation, 0 is no linear correlation and 1 is total negative linear correlation.
Alternatively, in a content-like approach, users interests, or pro les, can be represented as structured, unstructured or semi-structured data. In structured pro les, there is a small number of attributes, each pro le is described by the same set of attributes, and there is a known set of values that the attributes may have. Unlike structured pro les, in unstructured pro les, there are no attribute names with well-de ned values. In between, in semi-structured pro les, there are some attributes with a set of restricted values and some free-text elds. A common approach to deal with free text ( elds) is to convert the text to a structured representation, in which each token may be viewed as an attribute with an integer value indicating the number of times the token appears in the text. In a more sophisticated approach, each token can be associated with a tf-idf value, v(t; d), that is, for a token t in a text d, a function of the frequency of t in d, the number of texts containing t, and the total number of texts. The intuition behind tf-idf is that the tokens with the highest values occur more often in that text than in other texts, and therefore are more important. In this scenario, RatS(u; u0) can be evaluated as the cosine similarity of the vectors representing the pro les of u and u0. 2.2
In the health domain, usually people have similar interest in health documents if
they have similar health problems. To identify similarities between health
problems and eventually between users, we exploit the ICD10 ontology. We represent
ICD10 as a tree, with health problems as its nodes. For a node A in the tree,
weight(A) = w 2maxLevel level(A), where maxLevel is the maximum level of
the tree, level(A) returns the level of each node and w is a constant ([
For computing the semantic distance between two nodes A and B, we compute their distance from the lowest common ancestor C. The distance between A and C is calculated by accumulating the weight of each node in the path, as dist(A; C) = Pn2path(A;C) weight(n). In overall, the similarity between A and B is: simN (A; B) = 1 dist(A; C) + dist(B; C) maxLevel 2
Then, given two users u and u0, we calculate their overall similarity by taking into consideration all possible pairs of health problems between them. Speci cally, we take one by one all health problems of u, P roblems(u), and calculate the similarity with all the problems of u0, P roblems(u0), as follows: SemS(u; u0) =
Pi P roblems(u) ps(i; u0) jP roblems(u)j ; where ps(i; u0) = max(8j P roblems(u0)fsimN (i; j)g). 2.3
For documents, regarding the same information, people have similar interest in health documents that require the same educational and health literacy level to be comprehended. As such, the similarity between two users is calculated by the Euclidean distance between the corresponding values:
EducStatusS(u; u0) = p(HLiteracy(u)
2.4
Finally, anxiety and cognitive closure highly a ect the documents preferred by people in speci c periods of time - as anxiety and cognitive closure can uctuate over time. As such, we use the Euclidean distance between the values of those two properties. As psychoemotional questionnaires are being answered periodically, we consider each time only the latest measurements on these:
P sychStatusS(u; u0) = p(Anxiety(u)
2.5
To compute the similarity between two users u and u0, we use the function: S(u; u0) = AV G(RatS(u; u0); SemS(u; u0); EducStatusS(u; u0); P sychStatusS(u; u0)): Then, let Pu denote the most similar users to u. The overall relevance of i for u is estimated as: relevance(u; i) =
Pu02(Pu\U(i)) S(u; u0)r(u0; i)
Pu02(Pu\U(i)) S(u; u0) : After estimating the relevance scores of all unrated items for u, the items Au with the top-k relevance scores are suggested to u. 2.6
Since recommendations are typically personalized, di erent users are presented
with di erent suggestions. However, there are cases where a group of people
participates in a single activity. For this reason, recently, there are methods
for group recommendations, trying to satisfy the preferences of all the group
members. These methods can be classi ed into two approaches [
This way, our goal is to rst estimate the relevance scores of the unrated items for each user in the group, and then, aggregate these predictions to compute the suggestions for the group. That is, the relevance of an item i for a group of users G is:
relevanceG(G; i) = Aggru2G(relevance(u; i)):
As in [
u2G Alternatively, we focus on satisfying the majority of the group members and return the average relevance for each item: relevanceG(G; i) =
X relevance(u; i)=jGj: u2G
Targeting at increasing the fairness of the resulting set of recommendations, we also use the F air method. Here, we consider pairs of users in the group, in order to identify what to suggest. In particular, a data item i belongs to the top-k suggestions for a group G, if, for a pair of users u1; u2 2 G, i 2 Au1 T Au2 , and i is the item with the maximum rank in Au2 . For locating fair suggestions, initially, we consider an empty set D. Then, we incrementally construct D by selecting, for each pair of users ux and uy, the item in Aux with the maximum relevance score for uy. If k is greater than the items we found using the above method, then we construct the rest of D, by serially iterating the Au lists of the group members and adding the item with the maximum rank that does not exist in D. 3
In this paper, we argue that common problems and ratings are not enough for capturing similarity between users, and additional properties should be considered as well, such as educational and health literacy level, anxiety and cognitive closure. All these factors highly a ect the people's interest and understanding of information and especially in situations, where they are really stressed because of signi cant health problems.
The next step is to pilot and evaluate the system within the cancer domain.
We already have a corpus available for cancer patients through the
iManageCancer EU project [
Overall, we target at a general processing model that puts humans in the core, in order to produce recommendations for health-related documents that take into consideration additional perspectives like transparency and fairness. Transparency facilitates the understanding of data through, typically, exploration and explanation, used for assisting users identify the what, where, when and how of a data item. For example, exploration can support users by o ering sophisticated discovery capabilities. Di erently, explanations target at telling the story that the data has to say, by providing the reasons behind speci c recommendations. Fairness in data processing can be expressed as the lack of bias, where bias can come from data processing methods that re ect the preferences of the data scientists designing them. Regarding fairness in group recommendations, the goal is to locate, when possible or helpful, suggestions that include data items fair to the members of the group. That is, we should be able to recommend items that are both strongly related and fair to the majority of the group members
This work has been partially supported by the Virpa D project funded by Business Finland and by the BOUNCE project that has received funding from the European Unions Horizon 2020 Research and Innovation Programme.