Suficient physical activity is crucial for people's health and well-being. However, not enough people attain the weekly minimum of 150 minutes. Since current mobile health systems are not optimal to motivate and assist people to move more, this study investigates the efect of personalized suggestions generated by two types of recommender system algorithms: content-based (which provide more familiar recommendations, relevant to existing interests) and user-based collaborative ifltering (which deliver more diverse recommendations, allowing inspiration for new interests). By conducting a longitudinal between-subject user study over eight weeks, we will investigate how the two algorithms separately afect motivation and behavior change. We developed two versions of an Android smartphone application to deliver the recommendations, with the only diference being the implemented recommender algorithm. In all other aspects, the apps are identical: Both systems use the same datasets of physical activities and tips to break sedentary behavior, apply the user profile and contextual filter, and integrate the combination of star rating and momentary motivation feedback to provide personalization on preferences and well-being. We will analyze the diferences in people's star rating feedback, motivation to move, physical activity, and sedentary behavior. The main hypothesis is that inspiring recommendations from the collaborative algorithm will motivate people more for more physical activity and less sedentary behavior. The results of this study will provide insights for future mobile health recommenders in what type of algorithm and recommendations are most efective in the domain of increasing physical activity and motivating people to move more.
to break SB [6], more general healthy habits [7], or
reminders and tips [8]. Despite their great potential to
Insuficient physical activity (PA) is one of the modifiable motivate people, the interventions are often underused
underlying causes of chronic diseases, which cause most [9]. Furthermore, other research suggests that they
deaths worldwide [
Electronic health (eHealth) and mobile health and relevant interventions to the user [9, 11]. RSs gener(mHealth) interventions use technologies to promote ate personalized suggestions based on user preferences healthy behavior [4]. As such, they can also be used to help them with making decisions [12]. They can also to assist people in moving more by promoting PA be applied in the health domain as Health Recommender and prevent long periods of SB. In previous eHealth Systems (HRSs) to propose healthier suggestions, tailored and mHealth studies to increase PA, the content of to the user [13]. Previous work has applied RS techniques their interventions ranged from activities [5], ideas to provide personalized well-being recommendations for Ine Coppens, Luc Martens and Toon De Pessemier. 2023. Comparing food and PA [14], for personalized training sessions for Familiar with Inspiring Recommendations to Assist People in Moving marathon running [15], and for health activities [11]. AlMore. In Joint Proceedings of the ACM IUI 2023 Workshops, March though providing the most relevant health suggestion to 2023, Sydney, Australia, 8 pages. the user would optimize mHealth interventions, appli*$CoInrree.Cspoopnpdeinnsg@aUutGheonr.t.be (I. Coppens); Luc1.Martens@UGent.be cation of HRSs for behavior change is still in its infancy (L. Martens); Toon.DePessemier@UGent.be (T. De Pessemier) [9, 13].
0000-0002-3051-506X (I. Coppens); 0000-0001-9948-9157 To generate useful recommendations, the RS has to (L. Martens); 0000-0002-3920-7346 (T. De Pessemier) predict what the relevant items are for the user, for which © 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License several techniques exist [9, 12]. The content-based techCPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ACttEribUutRion W4.0oInrtekrnsahtioonpal (PCCroBYce4.0e).dings (CEUR-WS.org) nique suggests items similar to items liked by the user in behavior change by having more PA and less SB, as recomthe past, based on attributes that describe the items [12]. mended by the WHO [2]. Because the two RS techniques Alternatively, collaborative filtering uses other users’ rat- will provide diferent recommendations, we expect a difings and assumes that people with the same interests ferent efect on user motivation and behavior. To the will like the same items [9]. The item-based collaborative best of our knowledge, this efect of diferent types of RS ifltering technique recommends similar items based on algorithms has not been investigated. As such, we examthe ratings of other users, while the user-based collabora- ine which RS algorithm will perform best in motivating tive filtering technique focuses on recommending items users for more PA and less SB, responding to the demand that similar users with similar preferences liked in the of enhancing health interventions with the best personalpast [12]. While the item-based method provides more ization approach [9]. By developing two versions of the accurate recommendations as the user’s preferences are same Android app, we will conduct a between-subject modeled using similar items, the user-based approach user study with the following research question: can recommend more diverse and unexpected items [12]. Which recommender algorithm has the best efect on
Providing diferent approaches to predict what a user people’s star rating feedback, motivation to move, physical might like, these RS techniques result in a diferent selec- activity, and sedentary behavior? tion of recommended items [9]. While content-based RSs succeed in recommending highly relevant items, they often sufer from overspecialization as they suggest items 2. Methods very similar to items the user already knows because the attributes are already defined in the user profile [ 16]. As We developed two HRSs that recommend personalized such, they fail at recommending more unexpected, sur- PAs and tips for breaking SB to assist users in their daily prising, and novel items that could still be relevant to the life in moving more. For these PA and tip items, our own user [12, 17]. Previous work has addressed this overspe- two datasets were created. The PA dataset was assembled cialization problem on the grounds that it leads to lower using 354 PAs from the Compendium of Physical Activiuser satisfaction [18, 17, 19, 16]. Collaborative systems ties [21]. The tip dataset contains ideas from the Belgian solve this problem because they can recommend items website for health (www.gezondleven.be/), resulting in with a very diferent content when it is liked by similar 81 items. The generated recommendations are delivered users [12, 16]. To summarize, there are content-based to the user in an Android app called MoveMoreApp, as algorithms that provide familiar recommendations which shown in Figure 1(a), with its interface similar to our are highly relevant to existing interests, and collaborative app from a previous study. This app shows three PA ifltering that can deliver more diverse and unexpected and three tip recommendations. When a user executes suggestions which allow new interests to be explored an item, manual feedback on the recommended items is [12, 19, 17]. Hybrid RS algorithms combine the advan- collected as a rating on five stars, as illustrated in Figure tages of the content-based and collaborative approach, 1(b) with the question “how do you rate the generated recproviding a balance between relevant and diverse recom- ommendation?”. Additionally, our system collects users’ mendations [12, 18]. momentary motivation to move with a slider measured
In this research, however, we do not want to balance on a 5-point Likert scale (from “not motivated” to “exthe characteristics of the algorithms by merging them tremely motivated” ), as depicted in Figure 1(c). in a hybrid RS. Rather, we want to study the algorithms and the impact of their advantages and disadvantages 2.1. The algorithms separately in the domain of PAs. For example, previous The PA and tip items are recommended to the users with research has shown that repetition of the same health two types of RS algorithms, as illustrated in Figure 2. behavior makes the behavior easier [20], suggesting that The initial filter based on the user’s profile (available overspecialization may not be a problem in the domain material and maximum impact level) and the contexof PA. Similarly, we chose to implement the user-based tual filter based on the current weather (obtained using version of collaborative filtering because it can recom- https://openweathermap.org/) and remaining daylight mend more diverse items than the item-based version are applied on the PA and tip datasets in both groups to [12], and because we want to emphasize the efect of remove unsuitable items. more diverse recommendations on people’s behavior. As In the next step, the RSs generate personalized recsuch, we investigate the content-based and user-based ommendations based on the users’ consumption hiscollaborative RS algorithm separately as two extremes tory. This history contains the PAs and tips the user (very relevant versus very diverse) to gain understanding engaged in, together with the provided star rating feedin how they each afect motivation and behavior change. back, momentary motivation, and the user’s mood. The
In this study, concrete PAs and tips to break SB are rec- star rating feedback and momentary motivation are both ommended with the goal to motivate people for healthy measured on a scale of five and are aggregated with sumption history also contains the situation history at equal weights in the formula: _ = the corresponding time. To re-rank the items, a value ( + )/2. In this way, our two RSs op- between 0 and 1 that represents how close in time the timize their recommendations on both the rating and item’s situation is to the situation’s occurrences in the motivation. The mood is asked at the beginning of every history is added to the preference estimation score. As a day and after every submit with several emoji, as shown result, items that match better with the estimated current in Figure 1(c). As such, the user’s current mood is used to situation appear higher in the list of recommendations. iflter the consumption history on previous consumptions Next, the recommended PA items go through the adapwith a similar mood, based on the mood micro-profile of tive step. Combined with the user’s current PA level and [22]. feedback on intensity and duration provided in the app,
The content-based RS algorithm only needs the user’s as shown in Figure 1(b), the system provides a gradual inown consumption history. Calculating the similarity to crease in PA intensity and duration, following guidelines items consumed in the past relies on attributes that de- of the WHO [2] and the European Society of Cardiology scribe the items [12]. As such, our PA and tip dataset were [3]. In the final step, the recommended PAs and tips are extended with corresponding attributes to describe each shown to the user. item, such as aerobic, flexibility, or balance. The content- Right at the beginning, when the users did not submit based algorithm uses these to represent the user’s pref- and rate any PAs or tips yet, there is no consumption hiserences and match these with all the filtered PA and tip tory present to derive the user preferences from and base items using the cosine similarity [12]. In the other group, the recommendations on, resulting in the new user cold the collaborative filtering searches for similar users who start problem [12]. To provide an initial recommendation provided similar feedback to the same items using the with the available information, the two algorithms apply cosine similarity, and calculates a preference estimation the user profile filter and the contextual filter, and then score for all the filtered PAs and tips [12]. randomly select PAs and tips from this filtered set. As
At this point, both RS algorithms generated a list of more PAs and tips are chosen, the consumption history recommended items with corresponding preference es- will grow over time, resulting in better, more personaltimation scores. The contextual post-filter re-ranks the ized recommendations. It is however possible that users items based on the current estimated situation [12]. In do not like any of the (initial) recommendations and do our study, this situation can be: free time, during work, not select anything. In that case, the app allows users to household task, or active transport, and is assigned to select their own chosen PAs from the PA dataset with every item in the two datasets. In this way, the con- a search functionality when clicking on the “enter own
ContentBased Group
Filter
Context pre-filter
ContentBased RS
Context post-filter
Adaptive
Recommend PA
Recommend Tip activity” button, as shown in Figure 1(a). These PA con- the generated recommendations and they are stimulated sumptions are used in the RS algorithm for subsequent to only submit items when actually having engaged in recommendations. The new item cold start problem, on the them, rather than only rating them for more money. other hand, occurs when no ratings for the unexplored We designed the processing of data collected by our items are available yet [12], which is not a problem for app together with our ethical committee and data prothe content-based RS since this algorithm only uses at- tection oficers to be compliant with the General Data tributes to recommend other items. The collaborative Protection Regulation (GDPR) and our study received RS however, does depend on item ratings from other ethical approval. users [12]. We address this new item cold start problem by integrating an initial user-item consumption dataset 2.3. User study design from a previous study, in which items already received ratings from other users and to which no new items are added. Moreover, since our item datasets are relatively small compared to the amount of users (354 PAs and 81 tips) [12], and since we expect that users will engage in daily PA (which is any movement of the body, as defined by the WHO [2]), we estimate a suficient amount of consumptions after one week to alleviate the cold start problems. 2.2. Participants The target group of our study are adults who currently do not achieve the 150-minute weekly minimum of moderate PA. An initial screening with questions about age, weekly amount of PA [23], and a PA screening [24] in the app will decide whether or not the participant is eligible to join the study. Aimed at recruiting 50 participants, we promote our study via the Sona research participation system of Ghent University and several Facebook groups for paid studies. The study will run from March until June, 2023.
Participants will receive an incentive of 30 EUR when they used the app for eight weeks and answered all the questionnaires. They are not rewarded for having more PA or for the amount of PAs or tips they submit, because they can also use the app with “not now” and “enter own activity” submits. As such, they are free to choose from
A longitudinal user study will be conducted following a between-subject study design in which each user is assigned to either the content-based or collaborative filtering method. The advantage of between-subject user studies is the possibility to investigate the long-term effect of one system separately without having to switch between systems, but it also requires more users and more interactions [12]. As illustrated in Figure 2, the only diference between the two groups is the type of RS algorithm. The other steps (user profile filter, contextual pre-filter, contextual post-filter, and adaptive algorithm for PAs) are exactly the same.
Participants are asked to install the Android application on their own smartphone. Immediately after installation, the app randomizes the participants in the contentbased or collaborative filtering group. Then, participants are asked to answer the pre-test questionnaire, followed by an eight-week study. During these eight weeks, they can use the app in their daily life to look at the recommendations and choose an item to execute. When an item is selected, as shown in Figure 1(a), this is saved in the app even when the app is closed during the execution of the activity. When the activity or tip is executed, the user goes back to the app to submit and rate it, as depicted in Figure 1(b), in which the eventual duration of the executed PA is also asked. As such, participants are requested to only submit PAs and tips after engaging in them to provide proper feedback on the eventual rating, European Health Interview Survey - Physical Activity motivation, and duration. After eight weeks, the app Questionnaire (EHIS-PAQ) [23], and SB, surveyed with shows a final post-test questionnaire. the Sedentary Behavior Questionnaire (SBQ) [29] because
Since the goal of our study is to investigate the difer- they both allow participants to reflect on their average ences of receiving personalized recommendations from weekly PA and SB behavior, and they both distinguish either the content-based or the collaborative RS algo- between diferent situations, such as PA or SB at work rithm, the study duration is dependent on the time it or as transport. Repeated Measures ANOVA tests will takes for the RSs to succeed in generating personalized be conducted to investigate the evolution in motivation recommendations. By providing solutions for the cold regulation style and behavior change between the prestart problems as discussed earlier, we expect that the RSs and post-test measurements and between the two groups will be able to provide personalization after one week. In [30]. total, we decided on a study duration of eight weeks, rea- A manipulation check will validate whether the masoning that longer durations would result in more user nipulation succeeded. The manipulation in our study dropout [9]. We expect that users will have submitted is generating either familiar recommendations with the suficient consumptions, and that suficient PAs and tips content-based RS, or diverse recommendations with the will have been recommended in eight weeks to answer collaborative RS. The user’s experience of these recomour research question. mendations can be measured with the questionnaires of [31]. In these questionnaires, diferent RS properties are 2.4. Measures and analyses surveyed, such as perceived recommendation accuracy and quality (e.g., “The recommended items fitted my prefWhen the study is finished, statistical analyses will be erence” ), and additional properties that measure beyond conducted using IBM SPSS Statistics Version 28 to answer accuracy, such as perceived recommendation diversity our research question. The research question is divided and variety (e.g., “The list of recommendations was varinto four main dependent variables: star rating feedback, ied” ) [31, 12]. To keep the app user friendly, the app will motivation to move, amount of PA, and SB. These vari- not ask these questionnaires every time the user receives ables are all measured using the Android app at diferent a recommendation. Instead, the app will randomly show points in time. Depending on the timing of measurement these questionnaires in 20% of the time after the user of the dependent variable, diferent types of statistical chose and submitted a PA or tip recommendation. As a tests will be conducted on the longitudinal dataset and result, these data will also be longitudinal with repeated the pre-post dataset. measures over eight weeks, and Generalized Estimating
Firstly, measurements per individual are repeated over Equations [25] will be conducted for the analysis of the the eight-week study resulting in a longitudinal dataset. manipulation check.
The repeated measurements include: star rating feedback As the success and usefulness of an RS algorithm is on a recommended item, momentary motivation to move, based on how well it can predict the user’s preferences and the daily executed PAs and tips. Because of this [12], the stability of the preferences determines which longitudinal data, in which the data can be unbalanced algorithm will provide the best recommendations [17]. (e.g., not every user engages in the same amount of PAs), In some domains, such as movies, user preferences are analyses will be conducted with Generalized Estimating mostly stable over time, thus eliminating the need for Equations [25] to investigate diferences between the diverse recommendations [17]. On the other hand, some groups. people seek variety in their behavior, indicating the need
Secondly, motivation and behavior change are also for novelty and diversity in the recommendations [12]. measured in both the pre- and post-test questionnaires to In this case, RSs should take into account the diferences investigate their evolution after the eight-week study. To in user preferences, which can be depended on their measure motivation, we chose to utilize the regulation personality [12] or change over time [32]. For this reason, types of motivation as defined by the self-determination we also survey the user’s preference for variety in the theory (SDT), a theory of motivation that distinguishes pre-test questionnaire with our own questions, rated on between autonomous and controlled motivation [26]. a 5-point Likert scale from “Disagree strongly” to “Agree Based on the SDT, the motivation for PA (RM4-FM) ques- strongly”: “I like variety in my daily physical activities” tionnaire [27] and the Behavioral Regulations for Exer- and “I prefer routine in my daily physical activities”. This cise Questionnaire (BREQ) [28] measure the motivation independent variable will serve as a control variable in types for PA and exercise, respectively. By using separate the aforementioned analyses. questionnaires, we diferentiate between PA, which the To evaluate the overall performance of all the steps WHO defines as any movement of the body [ 2], and exer- of the algorithms, the “not now” button allows users to cise, which is a subset of PA. To measure behavior change, provide a reason why now is not a good time for PA. We we chose to analyze changes in PA, surveyed with the provided our own feedback sentences to check whether or not the recommendations fit with the weather (e.g., “It is raining too much” ) or with the current mood (e.g., “I do not feel good” ), whether or not they are adapted to the user’s PA level (e.g., “The recommendations are too intense” ), and whether or not the situation is suited for the recommendation (e.g., “I’m still at work/school” ).
vation, we hypothesize that the increase of PA and the decrease of SB will be stronger in the collaborative filtering group because autonomous motivation results in more efective healthy behavior change [26].
Lastly, we hypothesize that the collaborative RS will perform better (e.g., higher star ratings, momentary motivation, and amount of PA) when combined with a user who needs more variety in their behavior because it generates more diverse recommendations [12] and allows exploration of new items and interests [17]. Similarly, we hypothesize that the content-based RS will perform better when combined with a user who prefers routine because it generates recommendations similar to items the user already engaged in and already knows [12, 17, 16].
Moreover, repeating the same behaviors can make them easier [20], mitigating the overspecialization problem of the content-based RS. Since this research examines whether RSs should focus on existing interests or on discovering new interests in the domain of PAs in an eight-week period, we will not investigate whether or not these interests persist as habits, as previous research has indicated that habit formation may take up to 254 days [20].
We will first check whether our manipulation succeeded by analyzing the users’ experience with the generated recommendations. We expect that participants in the content-based group will assign larger scores for perceived recommendation accuracy and quality [31] because the content-based algorithm will generate recommendations that fit better with the user preferences [ 12].
Furthermore, we check whether the collaborative algorithm provided more diverse recommendations, as we expect larger scores for perceived recommendation diversity and variety [12, 31].
As content-based RSs generate recommendations that are similar to previously consumed items, and thus, fit better with user preferences [12], we hypothesize that the assigned star rating feedback will be higher in the content-based group. However, content-based RSs do not 4. Conclusions and future work provide an exploration of new items and expansion of their knowledge [17], and they ignore items with little This research investigates whether content-based or colsimilarities [18]. Moreover, we expect that integrating laborative filtering recommendations have a better efect more variety and unexpected items in the recommenda- on people’s motivation and behavior change for PA when tions with collaborative filtering will enhance their en- implemented in an HRS that assists people in moving joyment [9], inspire them with new interests, and expand more. The efectiveness of the HRSs will be evaluated their horizon [12, 17]. We hypothesize that increasing with a between-subject eight-week user study and an Aninspiration for new ways to move will motivate peo- droid application that randomly assigns each participant ple more because varied content is important to keep to either the content-based or the user-based collaborathe users engaged [33]. As such, we hypothesize that tive filtering RS algorithm. Expecting diferent efects on momentary motivation to move, and thereby also the motivation and behavior, we hypothesize that collaboraamount of executed PAs and tips, will be higher in the tive filtering will provide inspiration with new ways to collaborative filtering group. move, and motivate users more than the familiar items
Since both groups of participants receive an app aimed suggested by the content-based algorithm. at increasing PA, we expect that both groups will have To the best of our knowledge, the most optimal type of more PA and less SB in the post-test compared to the pre- algorithm for an HRS in the domain of PA has not been test. By following a between-subject study design, the investigated. Understanding how the algorithms sepalong-term efect of the applied system can be assessed as rately afect motivation and behavior change is impora whole [12], allowing us to compare the evolution in mo- tant before combining them in a hybrid system. As such, tivation regulation style and behavior change between this study will contribute to new insights in efective the two groups. Following the SDT, the autonomous algorithms for developers of future HRSs. For example, motivation regulation types are associated with people’s future hybrid RS algorithms can assign diferent weights own willingness to engage in the behavior and with more to content-based and collaborative filtering recommenpsychological health, while controlled motivation is as- dation outcomes, depending on the degree to which the sociated with pressure to behave in a certain way [26]. user prefers a familiar routine or varied inspiration in Because we expect more enjoyment with the inspiring daily activities. recommendations of the collaborative filtering group [ 9], we hypothesize that their autonomous motivation for PA will increase. As a result of higher autonomous moti