Recommender systems help users to cope with information overload and have become one of the most powerful and popular tools in electronic commerce. In order to provide better recommendations and to be able to use recommender systems in arguably more complex types of applications, most of the typical used approaches need significant extensions. On the video recommendation domain, one of these extensions is based in Segments of Interest (SOI), i.e., video segments that the user liked more or is interested. For this work, our intention is to stress and discuss interaction design issues about SOI based video recommender systems and discuss the relation between SOI and collective intelligence. We present two approaches to marking SOI on a Web social environment and discuss their advantages and disadvantages, and we show why SOI can be seen as a source of collective intelligence and that information and knowledge emerged from a community that had marked SOIs can be used on consensus decision-making and to bring improvements to society.
RESUMO Sistemas de Recomendação auxiliam usuários a lidar com o problema da sobrecarga de informação e se tornaram uma das ferramentas mais populares e poderosas do comércio eletrônico. De modo a prover melhores recomendações e para que possam ser utilizados em tipos cada vez mais complexos de aplicação, grande parte das abordagens tipicamente utilizadas precisam de extensões significativas. Copyright © by the paper's authors. Copying permitted only for private and academic purposes. In: Proceedings of the V Workshop sobre Aspectos da Interação Humano-Computador na Web Social (WAIHCWS'13), Manaus, Brazil, 2013, published at http://ceur-ws.org. No domínio de vídeos, uma dessas extensões é baseada em Segmentos de Interesse, i.e., trechos de vídeo que o usuário mais gostou ou está interessado. Neste trabalho, nossa intenção consiste em analisar e discutir questões de design de interação em sistemas de recomendação de vídeo apoiados por segmentos de interesse, além de discutir a sua relação com inteligência coletiva. Para tanto, apresentamos duas abordagens para realizar a marcação de segmentos em um ambiente social Web e discutimos suas vantagens e desvantagens. Além disso, mostramos porque os segmentos podem ser vistos como uma fonte de inteligência coletiva e que a informação e o conhecimento que emerge de uma comunidade que marcou seguimentos pode ser usada na tomada de decisões consensuais e trazer melhorias para a sociedade.
Palavras-chave Sistemas de Recomendação; vídeo; segmento de interesse; inteligência coletiva.
INTRODUCTION Recommender Systems provide recommendations for items to be of use to a user. They help users to cope with information overload and have become one of the most powerful and popular tools in electronic commerce. In these systems, recommendations are generally made by two types of filtering: collaborative and content-based. Many systems use these types combined in a hybrid approach. Furthermore, in order to provide better recommendations and to be able to use recommender systems in arguably more complex types of applications, most of the typical approaches need significant extensions. On video recommendation domain, one of these extensions is based on Segments of Interest (SOI), i.e., video segments that the user liked more or is interested.
In a previews work [
For this work, our intention is to stress and discuss interaction design issues and the relation between SOI and collective intelligence. We present two approaches to marking SOI and discuss their advantages and disadvantages, and we show why SOI can be seen as a source of collective intelligence and discuss that information and knowledge emerged from a community that had marked SOIs can be used on consensus decisionmaking and to bring improvements to society.
RECOMMENDER SYSTEMS AND SEGMENTS OF
INTEREST
The recommendations provided by recommender systems
are aimed at supporting users in various decision-making
processes, such as what items to buy, what music to listen,
or what news to read. Recommender systems have proven
to be valuable means for online users to cope with the
information overload problem [
As already stated, recommender systems are generally
made by two types: collaborative filtering (CF) and
contentbased filtering (CBF). For them, various algorithms and
techniques have been proposed and successfully evaluated.
CBF takes the descriptions of the previously evaluated or
currently accessed items by the user to calculate the
similarity between items, and, then, recommend items of
interest to the user. This type of filtering enables
personalized recommendations for users. CF calculates the
similarity between users and recommends items that are
liked by similar users. This uses ratings given by users in
the past to find the best item to a similar user. Another
approach is to calculate the similarity between items to
produce recommendations [
These two approaches present some issues, if used alone.
For CBF, for instance, the two main disadvantages are: (i) it
depends on one objective description of the items; and (ii) it
tends to overspecialize recommendations [
Segments of Interest (SOI)
A SOI is a segment on video that the user liked more or is
interested. Users tend to like particular segments of the
video more than the rest [
In our previous work, we show that SOIs from a community of users can be used to find similar people, i.e. people who have similar interests about videos. When a pair of users has a quantity of intersections of SOIs above a threshold (and with a minimum acceptable size) in one video and this pattern occurs in a certain quantity above a threshold in a set of videos, they have similar interests about video. Figure 2, for instance, illustrates that users u2 and u4 will have similar interests about video if it is considered that similar users have at least 2 intersections of SOIs on a set of 3 or more videos.
We also show, in our previous work, that this similarity between pairs of users based on SOI can be used to enhance the accuracy of ratings predictions of video recommender systems with user based nearest neighbor collaborative recommendation.
Marking SOI One important aspect is how users can mark SOIs in video. The way users mark SOIs is directly related to the user interface of the system. This interface must provide specific components for marking SOIs. We have proposed two approaches for this purpose: • Buttons with predefined time slices: in this case, different buttons, with different time slices, are presented to the user. Each time the users want to indicate a SOI, they click on the button that corresponds better to the size of SOI they want (whose the end coincides with the current instant of time in the video). An example of this approach is presented in Figure 4. • Sliders that allow users to mark the beginning and the end of each SOI. In this case, users can mark videos while they are watching the video (like the previous approach) or after watching it. An example of this approach is presented in Figure 3.
The approach for marking SOI must be chosen accordingly
to the context and to the user tasks, because both have
advantages and disadvantages. Marking SOI with the use of
buttons on the user interface (or buttons in a remote control)
is indicated for video websites, video/movie on demand
services and personal video recorders, where the user's goal
is to enjoy entertainment or get information. The advantage
is that users do not lose so much attention when they want
to mark SOIs on the video, and this marking occurs quite
effortlessly, quickly and easily. The drawback is that the
beginning and the end of a segment may not be as precise
as intended (since the intervals are predefined). If SOI are
used to enhance the accuracy of the ratings predictions of
recommenders systems, this precision is relevant for the
accuracy [
The approach to mark SOI, that is part of the interaction
design, is extremely important. The designer of SOI based
recommender systems must be aware that users are
reluctant to give explicit feedback (in this case, marking
SOIs), that not everyone likes to participate more actively
and more interactively, and that, in our approach, there is a
significant level of user involvement [
THE DEVELOPED SYSTEM In our previews work we developed a video website with a recommender engine as a web application. The server hosts the application, and its interface is similar to many current video websites, i.e., it has a screen containing a video gallery where the user can browse videos, and a screen to watch videos. The client can be any device that contains a web browser, such as a desktop computer, a notebook, a smartphone, a tablet or a smart TV. This system only stores metadata to catalog videos and data generated during the use of the system. Videos are loaded directly from clients on demand from a video content provider (YouTube, in the specific case).
For performing an experimental evaluation of our recommender system approach, we have used educational videos of a given subject, but the system can deal with several domains, such as news, sports and movies. For this purpose, we have created a catalog containing 50 educational videos up 20 minutes of duration from YouTube.
Figure 3 shows part of the screen of the system where users
watch videos. Through it, users can rate videos (rating
options are, "Very bad", "Bad", "Ok", "Good" and "Very
good"), and the users can also mark their SOIs. This is
performed through a double slider below the progress bar of
the video player. Users drag and drop the sliders to mark
the beginning and the ending of each SOI. This approach to
marking segments was used previously in video remix tasks
[
If the context was other, like watching movies in a smart
TV, the segments of interest could be marked in another
way, using specific buttons on remote control or on a
smartphone application, for example, while the user
watches the video. Figure 4 illustrates the approach based
on buttons on the user interface of the video website. There,
three buttons can be seen: ‘10 seconds’, ‘20 seconds’, and
‘30 seconds’. If the user is watching a video and click on
the "20 seconds" button, a SOI is marked from the current
position on the video timeline until 20 seconds back. It is
easy and fast, and could be even done remotely, if the
buttons are in a remote device (remote control or
smartphone).
EXPERIMENTAL EVALUATION
As already stated, the system was described in a previous
work [
The focus of our previous work was to compare different
recommender system strategies. In this sense, we have
conducted an offline experimental evaluation based on the
leave-one-out strategy [
Questionnaire We have built a questionnaire to understand the user´s experience related to marking SOI on our system, to know about their collaboration habits on the Web, and about their habits related to watching video on the Web. In this sense, we have asked the subjects of the previous experiment to answer a questionnaire (anonymously) available on the Web. Results are described below. It is important to state that only 18 of 88 users have answered the questionnaire. Question #1: Did you mark segments of interest in what moments? 40% 30% 20% 10% 0% after while most while most after In this question, "after" means "I marked after I watched the video", "while" means "I marked while I was watching the video", "most while" means "I marked after and while I was watching the video, but in the most of times after", and, finally, "most after" means "I marked "after" and "while" I was watching the video, but in the most of times while". Question #2: How often do you watch videos on Web?
0%
17% 83% Question #3: Do you often collaborate with some information on the Web (such as 'add a comment', 'submit a link to a friend' or give a "like"')? no yes, rarely Question #4: Do you usually use somehow the information collaboratively added by others on websites? no yes, sometimes Question #5: Would you like to share your segments of interest with other people? yes, but with my permission The results showed that in the context of educational videos, considering the approach used to mark SOIs, most of users prefer to mark SOIs after watching videos. In another question, not viewed by graphs, 100% of the subjects answered that they enjoyed using the system to watch the educational videos offered.
About the habits of the subjects, they frequently watch videos on the Web and use collaboration mechanisms to obtain information. However, very few subjects effectively collaborated with others (to provide information). Finally, most of them stated that they would like to share their SOIs. It could be done, for instance, in a social network among their friends.
Segments of Interest on Video and Collective
Intelligence
In this work, another relevant aspect to discuss is the
relation between SOI and collective intelligence.
Collaboration among people has been gaining attention due
to the popularity of social media tools such as forums,
blogs, wikis and social networks. These tools are
collaborative systems that enable users to share their
knowledge, skills and other information. As a result, they
are becoming important sources of collective intelligence
[
Collective intelligence led to the rise of a new business
model known as crowdsourcing. This term describes a
model that takes advantage of several creative solutions that
people can propose [
The other approach is known as implicit crowdsourcing,
which can take two forms: standalone and piggyback [
In this sense, SOI, which is collected in an explicit way and is used to enhance the accuracy of rating predictions of video recommender systems, can be used implicitly as source of information and knowledge about a user community. For instance, depending of the context, a cluster of SOIs can sign something relevant in one video. For instance, in the entertainment video domain, a cluster of SOIs can represent the most popular segments of a particular movie for the community; similarly, in the educational environment, a cluster of SOIs can be the starting point for a teacher to discover the more relevant segments for a community of students and, based on this information, figure out why that point is important for learning. In this sense, SOI can be seen as a source of collective intelligence.
SOI Data Visualization
To discover clusters of SOIs we have implemented a data
visualization display in the developed system. This display
presents the set of SOIs marked by each user in a video. As
we had only two dimensions (SOI vs. users), we have used
a simple data visualization technique that uses rectangles on
orthogonal axes to represent intervals [
The display illustrated in the Figure 5 corresponds to the aforementioned visualization technique. There, users are represented on the vertical axis and the horizontal axis is the length of the video. For instance, in this figure, most users marked just one SOI, one marked two and another marked three. In the display, each rectangle represent a SOI. The size of each rectangle indicates the size of the interval in relation to the video timeline. The position of the rectangle (and it size) corresponds to where it was marked in the video timeline.
To aid the analysis, we included a combo box to allow browsing videos. Once a video is selected, one key frame is shown to illustrate the video selected. With this tool, we were able to discover the parts that were most marked by the community of users. For instance, Figure 5 shows the data visualization display for video "Visualization of Quicksort", with the 13 users who had marked SOIs in it. The video is an animation about sorting algorithms. As we can see, many users had marked SOI in a special part of the video (on the right). Such part is the moment where the animation starts to compare the Quicksort and the Bubblesort algorithms, running side-by-side in a dispute. When we pass the mouse pointer over a rectangle representing a SOI, a tooltip presents its exact beginning and end, in seconds.
It was another example that shows why SOI can be seen as a source of collective intelligence. Information and knowledge that emerged from a community that had marked SOI can be used on consensus decision-making and to bring improvements to society. For instance, using information and knowledge originated from SOI clusters a video on demand provider can improve its recommender engine, or a TV channel can discover groups, segment viewers and send personalized ads.
CONCLUSION AND FUTURE WORK Segments of interest, which can be used to enhance the accuracy of rating predictions of video recommender system, can be marked by different ways. Each way is directly related to the user interface of the system that contains the recommender system. In this work, we showed that each way have advantages and disadvantages. Therefore, the designer of SOI based recommender systems must be aware that the approach chosen to allow users to mark SOIs may be directly related of the success or failure of the system and its recommendations.
Furthermore, we have evaluated the use of SOI concerning collaboration and user's habits about consuming video on the Web. The results showed that, in the context of educational videos and considering the approach used to mark SOIs, most of the subjects prefer to mark SOI after watching videos, and all subjects enjoyed to use the developed system to watch videos. About their habits, they stated that the frequently watch videos on the Web and use advices obtained through collaboration. However, very few of them admitted to collaborate with others on the Web. In relation to sharing their SOIs, most of them stated that they would like to share them with friends, for instance, in a social network.
We also discussed why SOI could be seen as a source of collective intelligence. Information and knowledge emerged from a community that had marked SOIs can be used on consensus decision-making and to bring improvements to society. We implement data visualization over the developed system to illustrate how SOI can be such a source of collective intelligence.
In future works, we will make more experiments to compare even further the presented approaches. The goal is to verify which one is better suited for different contexts and scenarios.
ACKNOWLEDGMENTS This work was partially supported by CNPq and CAPES.