Recommender systems are becoming very popular as they are able to predict the preferences of a user. This make recommendation based on the user pro le, past ratings or/and additional knowledge such as user contextual information. Applied to the health area, they can take advantage of context information to support health promotion and disease prevention. We present a recommender system for the promotion of physical activity called CoCARE. It recommends videos about physical activity based on a user pro le, his/her context. The main challenge of CoCARE is the small set of videos to be recommended, because the selection of the videos is done manually by of health experts. Several health recommender systems have this same problem. Although today there are a large number of videos available on the Internet related to physical activity. These could not be included in the data base of CoCARE; because these do not have enough information to be categorized and pro led. This article proposes a CBR system, this assigns a physical activity category to new video. In this way the new video will be added to the list of CoCARE recommendations. In this CBR process, basically consists on analyzing the description of the new video and compare it with the cases base of CoCARE, selecting the category of most similar cases.
Recommender systems in the health area have been proven as useful tools to
help patient-oriented decision making systems, promoting physical activity and
disease prevention, in general, to improve health conditions through healthier
habits. Health recommender systems (HRS) aim to promote health programs, to
provide patients with relevant information, products or services, using knowledge
about his/her personal health record systems [
In the literature, we nd few systems that recommend health educational
multimedia contents [2{6]. Users get recommendation of exercises (stretching,
Copyright © 2017 for this paper by its authors. Copying permitted for private and
academic purpose. In Proceedings of the ICCBR 2017 Workshops. Trondheim, Norway
strengthening,etc.), with outdoor or indoor sessions, based on the user
information taken from mobile devices, activity bracelets, sensors, and his/her personal
health records and risk factors [
We have developed \CoCARE" a platform for promotion of healthy lifestyle
on the basis of a context-aware recommendation system designed for mobile
smart devices [
CoCARE has a database with 80 videos. These have been tagged with its title, description, category and keywords (see example in Table 1). One video could be recommended to several users based on a decision model given by domain expert.
Concretely, the decision model is built from a dataset of 597 instances (rows),
6 attributes and 1 main class (see Table 2) created by the expert. CoCARE builds
a decision tree using a supervised learning algorithm. Then, the system classi es
the query with information about the user and his/her context and uses the tree
to recommend contents based on the current user situation [
In this paper we deal with the problem of video acquisition and tagging. Internet provides with a huge amount of videos, most of free use, related with physical activities: dancing, running, tness, GAP. Our goal is to use these videos as recommendation items in our system. To do that, we would need to annotate the videos with information about the potential users that would bene t from them. We propose a CBR system to automatically classify videos given its textual description. This CBR system also computes similarity between the CoCARE user pro les set and the new video categories, to found its categories.
The paper runs as follows. Section 2 describes the recommender system of CoCARE based on decision model. Section 3 explains the CBR process to automatically annotate new videos. Section 4 evaluates the CBR system. Section 6 concludes the paper and discusses some lines of future work. 2
CoCARE (see Figure 1) is a context aware recommender system that recommends videos on physical activity (PA) and healthy diet (HD) to patients for promotion of her healthy habits. CoCARE incorporates a context- adaptable interface based on decision trees.
Fig. 1: Mobile CoCARE
CoCARE recommends multimedia content of physical activity and healthy
diet based on user and contextual information. The basic user model includes
details on the user personal pro le (see Table 2). The system takes advantage
of additional contextual factors to provide with personalized recommendations
of multimedia content. The query includes static information like user pro le,
and dynamic features like geo-lacation or indoor location, date (day or season),
daily schedule of the user and it can detect when the user has company. [
Although the CoCARE system works well as a prototype, it relies on an initial video database of 80 videos. That means di erent problems: { Users get repeated contents after a while. { Lack of novelty contents provokes user desertion. { The task of including new videos is cumbersome. { New videos were included without expert supervision and they were misclassi ed and never recommended to the right users. We propose a CBR solution to solve these problems and assign tags (video category, keywords and user pro le tags) to new videos based on the comparison to existing ones.
Attribute BMI Life cycle Ethnicity Trauma Preference Cardiovascular disease Category { Description = prancing, tapping, dribbling, moving, braiding, waltz. { Solution = dancing.
This CBR1 module implements a k Nearest Neighbour algorithm to nd the most suitable categories for a given video description. Concretely, we use a 3-NN algorithm with a keyword based similarity measure to select the three categories with highest similarity values.
Fig. 2: CBR Process
Once the categories have been retrieved from the rst case base, the second CBR module(CBR2) estimates the most suitable user pro les for the video description.
This second module has a case base with 597 instances where every case is described by several categories and 6 user pro le attributes as the solution (see table 2) . By this way, the solution will be a new user pro le UP. The algorithm compares locally the similarity value for every attribute of the user pro le ( see table 2) using majority voting or weighted majority voting to select the t attribute to the solution.
CB1 = < C11; C21; : : : ; Cm1 >
Ci1 = < keywords; categories > CB2 = < C12; C22; : : : ; Cn2 >
Ci2 = < categories; U P > U P = < BM I; age; et; tr; pr; mc; category > (1) (2) (3) (4) (5) 4
We evaluate our system using leave-1-out cross-validation.
We used the video description of each case on CB as a query. We proposed
the category and pro le for U Pr, that is compared to the stored solution U Pq.
We compute the similarity between attributes of U Pq and U Pr, using a binary
function [
= where = = = = = = = =
Sim(BM Iq; BM Ir) [0; 1] Sim(ageq; ager) [0; 1] Sim(etq; etr) [0; 1] Sim(trq; trr) [0; 1] Sim(prq; prr) [0; 1] Sim(mcq; mcr) [0; 1] Sim(categoryq; categoryr) [0; 1] 0:1 (a + b + c + d + e + f ) + 0:4 g (6) (7)
The table 3 shows an example. Our query in this example is U Pq and D= \Steve and Jackie take you through how to get the most out of power walking and show you how bene cial it truly is. Yes it is an Olympic sport!".
First the CBR1 module got 3NN categories as: Walk, Exercises and HIIT. So the CBR2 module compared the UP associated to these categories and found the most similarity U Pr. Next the system uses cross validation and retrieves a success solution only if the similarity measure 0:7. This process is the comparison between attributes U Pq and U Pr. For example the table 3 shows a test case, we obtained a score of greater than 0:7, so U Pr was added to CB.
Test Case
U Pq
U Pr Test Case Test Case
BMI age et tr pr cv V normal weight Adult other without trauma beauty diabetes Walk overweight Young Other without trauma beauty diabetes Walk a b c d e f g 0 0 1 1 1 1 1
Table 3: Example cross validation
In each leave-1-out step, we obtained 3 values: the similarity of the best cases returned by the CBR1 module (1-NN), the 2 best cases (2-NN) and the 3 best cases (3-NN). Next we made 2 tests with Majority Voting (MV) and Weighted Majority Voting (WMV) in CBR2 module. Figure 3 shows results from our experiment as: { Case 1, it is represented by the blue bar. We found the user pro le with majority voting (MV) for 1NN, 2NN and 3NN. { Case 2, it is represented by the red bar. We found the user pro le with weighted majority voting (WMV) only for 2NN and 3NN.
We conclude that the better result from our CBR system was 3NN with WMV. In this case, we obtained a value of greater than 0.85 surpassing the results achieved of the other tests. Our experiment shows that greater than 90% of the cases are correctly classi ed. The literature to continuation are about design and implementation of CoCARE. We approached works of health recommendation systems.
Related works such as [
The previously works mention the use of user models, data modeling and use of recommendation techniques, but they do not describe the selection of recommendation techniques and how performance the validation them. On the other hand, these works didn`t make mention about how many recommendations have their systems, apparently the systems has few contents to recommender.
The next related works relate with CBR topic. [
Our CBR systems retrieves a user pro le from the video description. This could be useful in others areas such as education, commerce and / or advertising. For example one recommendation systems could nd a user pro le t for learning content or advertising videos with just the description of item. 6
We have described our CoCARE recommender system. CoCARE recommends videos of physical activity categorized by health experts. But the problem is that they are very few, to include a new video must be properly categorized for a user pro le. Our CBR allows you to categorize the video and nd an appropriate pro le from the description of a video. In this work we proposed a system composed of 2 CBR system, the rst categorizes the new video and the second delivers the appropriate pro le. We have evaluated that the CBR system delivers a better response if the rst CBR is 3NN and CBR2 is with similarity.
Our CBR system uses little input knowledge to get an adequate solution. It o ers a simpler alternative to associate videos to the needs and preferences of di erent users.
Our system bene ts the user and the health expert, with the possibility of having new recommendations that help the adherence of the physical activity program.
The work presented in this paper opens several lines of future work.
When you have very short video descriptions the CBR system loses precision in nding the right category, although the results we obtained are very promising we have considered that they can be improved if we extend the description from synonyms using an ontology of synonyms and algorithms matching of learning.
We plan to take information about the most viewed videos on the Internet (YouTube) and use their description to classify them, assign to the new video an appropriate user pro le and add to CoCARE case base CB automatically using collaborative ltering and CBR.
This work was performed under the doctoral thesis \Context-Aware Recommender System to Physical Activity Promotion" nanced by Colciencias, under call \Convocatoria No 6172 (Doctorados Nacionales)". Supported by UCM (Group 921330) and Spanish Committee of Economy and Competitiveness (TIN201455006-R).