Personalization services enable Internet users to benefit from tailored recommended content at their finger tips. Our interest in this contribution lies in video recommendation within the fitness domain to support an active lifestyle. We present 'Fitness that Fits', a preliminary platform for workout video recommendation, which benefits from the Youtube-8M labeled dataset and its rich variety of categorized video labels, thereby enabling fitness workout video recommendations predicated on the users' preferences and their recent viewing behavior. The proposed model also incorporates an approach to produce diversified recommendations and foster the practice of distinct fitness activities based on like-minded users' information. An initial experimental study shows the trade-ofs of the hybrid approach considered.
The Internet and its associated technologies have become an
indispensable tool to search products, services or frequently access
information needed in our daily lives, e.g. booking a hotel,
purchasing a new device or consulting the weather forecast. We are
presently reported to spend an average 6 hours per day connected
to the Internet. Amid this phenomenon, there is an increasing
interest in seeking aid in the Internet to embrace healthier lifestyles, e.g.
through the search and sharing of information related to fitness
exercises and wellness practices, or via smartphone apps [
Although gyms and leisure centers are a common choice for users who desire to adopt or maintain an active lifestyle, they are not always within the reach of every person, e.g. owing to ifnancial limitations, busy schedules, frequent traveling, etc. Taking HealthRecSys’18, October 6, 2018, Vancouver, BC, Canada © 2018 Copyright for the individual papers remains with the authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors. advantage of the growing demand for online resources to promote exercising, online workout videos have proliferated in recent years as an alternative means to keep users active from the comfort of home or beyond, with a number of advantageous characteristics: • They are convenient, providing 24/7 access to a wealth of fitness resources from anywhere with an Internet connection. • They do not require commitment to work out at an externally imposed day or time. • With a careful search and use of the resources available, they provide a wealth of workouts from a diversity of instructors. • They are cost-efective and can be undertaken in a more individual and private space.
Whether it is bodyweight workouts, aerobic exercises,
performance hacks or skill gaining tutorials, Youtube provides millions of
users with access to a wealth of video resources to support them in
practicing their preferred workouts anywhere and anytime. Despite
the potential benefits to Youtube users of receiving personalized
recommendations from the platform as a whole [
The Youtube-8M dataset is a clear example of large-scale dataset
containing comprehensive information about millions of videos
uploaded to Youtube. Despite having been primarily investigated for
tasks such as the classification and further understanding of video
data [
To overcome the challenges outlined above, this contribution presents ’Fitness that Fits’, a prototype platform for workout video recommendation, which relies on Youtube-8M video data describing fitness activities. Our main contribution is a recommendation model that extends principles from content-based and collaborative filtering by introducing mechanisms to provide end users with meaningful and diverse workout video recommendations. 2
Recommender systems applied to health-related domains are still
relatively scarce [
Ceron et al. [
To assist health professionals, patients and caregivers in the
process of finding relevant information amid a plethora of it,
SánchezBocanegra et al. [
Within the particular scope of the fitness domain, two studies on
recommendations for running were presented by Berndsen et al. [
and contextual information to produce socially enhanced fitness recommendations. 3
This section describes the currently used data in ’Fitness that Fits’ and briefly overviews its Web platform being developed.
YouTube-8M is a large-scale labeled video dataset which, as of June 2018, consists of over 6 million of YouTube video instances (which add up to 350,000 hours of video), namely video IDs with high-quality annotations generated by machine learning techniques, describing a highly diverse vocabulary of over 3.8K diferent entities (labels). Each video in the dataset contains an approximate average of three labels. These videos have been sampled uniformly with the aim of preserving the highly heterogeneous distribution of popular Youtube contents, whilst ensuring the dataset quality and stability by enforcing a series of restrictions. The dataset also incorporates precomputed audio-visual features from billions of frames and audio segments, which facilitates an eficient training of baseline models without the need for sophisticated computer settings, and (ii) enables a deep exploration of complex audio-visual models that otherwise would be impractical to train.
The use of Youtube-8M has been illustrated in recent research
eforts including large-scale video classification [
We built an initial dataset by filtering the original Youtube-8M labeled video dataset 2 predicated on the following filtering criteria: (1) Highly-viewed: Only videos with a minimum of 50K views in Youtube are selected for the scope of the preliminary research presented in this contribution. Whilst important, tackling the popular cold-start problem associated to newly added content lies outside the scope of the present study. (2) Fitness-related: Videos having machine-generated annotations of ’Beauty and Fitness’ narrowed down to 16 labels, associated with popular and highly-viewed types of fitness activities in accordance with criterion (1).
The resulting video dataset comprises 1.7K fitness workout videos
with over 1K user views each, which supposes an elevated number
of videos in contrast to other related works [
Figure 1 provides an overview of the user interface for the ’Fitness that Fits’ Web platform. Besides the interface for exploring videos or viewing the recommendation list, the user can establish and update anytime a profile by selecting at least two labels as her
favorite types of workout. In the following, our interest focuses on presenting the recommender model implemented in the platform. (B) Measuring diversity of recommendations. (C) Diversity-aware replacement process. 4
The model for the recommendation process underlying our platform is illustrated in Figure 2. It consists on a hybrid approach incorporating basic principles from content-based and neighborbased collaborative filtering, on top of which three features are introduced: (A) Identifying user preferences.
These three novel features are labeled using (A), (B), (C) in Figure 2. The content-based and collaborative filtering steps are implemented upon basic approaches in the current prototype version of this work, and they can be seamlessly replaced by other existing methods with similar aim. We therefore concentrate the subsequent discussion on the three distinctive features listed previously. User behavior (recently viewed video history) (A) Identifying user preferences
User preferences
Video features
with M = {1, 2, . . . , m} the label index set. As shown in Eq. (1),
pˆj relies on a weighted average of (i) the relative frequency of
i
workout videos containing topic j in the recent viewing history,
and (ii) the (binary) profile information given by pij . The weighting
parameter ωi ∈ [
is dynamically assigned to each user as follows: ωi = min (
hl (i) 2 · n1 Ík hl (k) , 1 ) (2) with hl (i) ∈ N the history length or number of recently viewed videos by ui . This dynamic adjustment of ωi adapts to the degree to which the user utilizes the system: the more engaged ui has recently been (larger recent viewing history), the higher ωi and the more relevance is assigned to the behavior information. If ui recent engagement equals the average recent engagement of all users, i.e. if hl (i) = n1 Ík hl (k), then ωi = 0.5.
As a result, a user’s current preference vector Pˆi = (pˆi1, pˆi2, . . . pˆim ) is yielded for each ui ∈ U . Based on the cosine similarity between a user preference and a video representation Pv = (pv1 , . . . , pvm ), pvj ∈ {0, 1}, a content-based filtering process is subsequently applied, leading to a preliminary video recommendation list of size N . This list might still be adjusted before being delivered to the end user, as explained in the following two subsections. 4.2 (B) Measuring diversity of recommendations One of the aims of the proposed system is to provide users with recommended videos that are both relevant (in accordance with their current preferences) and diverse. Diversity in workout recommendations may not only help exploring “new” types of workout the user might potentially like, but also fosters variety of workouts in such recommendations to prevent an eventual sense of boredom. Let Ri = (r vi j )N ×M be a matrix representation of ui ’s recommendation list, where the vth row contains the M-dimensional vector representation of the vth recommended video, hence r vi j = 1 if the vth recommended video is labeled with topic j, and r vi j = 0 otherwise. By introducing a diversity threshold δ ∈]0, 1], the diversity level of Ri , denoted D(Ri ), is measured and compared against δ , predicated on the number of topics appearing in at least one video in Ri (Ô denotes the logical disjunction ’OR’ operation): Íj ÔN i
v=1 rv j D(Ri ) =
M If D(Ri ) ≥ δ (i.e. the ratio D(Ri )/δ ≥ 1 as shown later on in experiments) then the recommended videos for ui are suficiently diverse and they are supplied to the user. Conversely, if D(Ri ) < δ then a neighborhood-based collaborative filtering approach is adopted to further diversify Ri based on the information extracted from similar users’ recommendation lists. Importantly, we adopt a variant of classical user-user collaborative filtering in which, once the K most similar users to the target user have been identified (predicated on the similarity between their current preference vectors Pˆi ), we analyze those neighbor users’ recommendation lists, as opposed to existing approaches that apply a rating prediction function. (3) 4.3
(C) Diversity-aware replacement process An iterative procedure is introduced to diversify the recommendation list Ri . The procedure is characterized by replacing - at each iteration - one of the videos recommended to the target user ui with another video stemming from one of her neighbor users’ recommendation list, based on the following steps: (1) Sample one of the K neighbor users of ui , denoted ui′ , by normalizing similarities sim(u, u ′) into probabilities for each neighbor to be picked. Retrieve from the database the matrix Ri′ containing his latest list of recommendations. (2) Check the rows (recommended videos) in Ri′ in descending order. Choose the first occurrence of video containing at least one topic j that does not appear in Ri . If no videos in Ri′ hold this condition, return to step (1). (3) Assume that the selected video in Ri′ corresponds to the vth row of the neighbor’s recommendation matrix. Then, in the target user matrix Ri , replace the existing video in the vth position with the selected video from ui′ recommendations. As a result of an iteration, a single-video replacement is made on Ri , after which its diversity level is measured again. The overall iterative process described in Section 4.2 and 4.3 is repeated until the recommendations are diverse enough, or an a priori fixed maximum number of iterations is exceeded. In either case, the final adjusted recommendation list is provided to ui . 4.4
This subsection briefly outlines a preliminary evaluation conducted on the current version of ’Fitness that Fits’ underlying model. We remark that despite the considerable volume of real labeled video data available, the present study relies exclusively on a small and partly synthetic dataset of user information (profiles and viewing histories). Deploying the platform into a Web environment and gathering larger amounts of user data for a comprehensive evaluation, constitutes our most immediate direction of future work.
We consider a sample of 10 users, a recommendation list size of
N = 30, a number of neighbor users K = 3 for the diversification
strategy, and a diversity threshold δ = 0.375 (at least 6 out of 16
topics to appear in Ri ). This preliminary experiment focuses on
measuring the diversity ratio D(Ri )/δ and the average similarity
or relevance S(Ri ) in the user’s recommendation matrix, before and
after applying some replacements under the proposed
diversification strategy: until δ is achieved or at most N /2 replacements are
made on Ri . The similarity score S(Ri ) ∈ [
This initial evaluation shows that, although the model can effectively increase the diversity of recommended workout videos, this normally comes at the expense of a decline in the relevance (similarity) or such recommendations with respect to the target user preferences. This is not surprising, since the initial content-based ifltering stage (before diversifying) relies exclusively on the users’ preferences, and the current use of a cosine similarity may make the resulting relevance sensitive to any zeros in either user preferences of video representations. Moreover, data containing recent viewing 0.89 ; 1 0.81 ; 0.77 0.80 ; 0.71 0.73 ; 0.62 histories is rather scarce in the present prototype, therefore the weighting parameter ωi tends to be low for most users and their static profile information (favorite topics) is prioritized. We argue that an online evaluation of users’ experience with the system, for instance by tracking clicks on recommended videos and analyzing whether such clicks have been predominantly on “replacement” videos or not, will be an interesting direction to extend the proposed approach into a more adaptive one, where depending on the user’s response towards diversity, a tailored trade-of between diversity and relevance is sought for her/him. Another interesting ifnding is the dramatic increase in the diversity of R1 and R3 with respect to other users. This is largely due to some videos in the system having associated more topic labels than others. 5
This contribution presented ’Fitness that Fits’, a prototype platform for recommending physical workout videos upon labeled video data from the Youtube-8M dataset. Besides integrating basic content-based and collaborative filtering mechanisms, the proposed recommender model incorporates novel features for the flexible modeling of user preferences based on their profile and recent viewing behavior. Furthermore, an iterative replacement strategy inspired by neighborhood-collaborative filtering is introduced to promote diversified recommendation lists for users to enhance with diferent types of fitness activities.
Besides the platform deployment, subsequent acquisition of more
real user data and the elaboration of a complete experimental study,
we also postulate the following directions for future research:
• Recent eforts have been put in personalization services for
promoting healthy habits, e.g. via food recommendations for
positive nourishment practices [
The authors would like to thank anonymous reviewers for their insightful and constructive comments on the initial stage of this research, some of which would lay the foundations for future work.