Personalization of products is a popular aspect in various application domains, including videos. Enabling users to consume personalized learning videos that include only individually relevant content has the potential to deliver additional benefits in e-learning, for example, by making learning more eficient. In this paper, we present a practical approach to define configurable videos based on feature models, as well as an integrated solution to derive personalized videos by respecting given constraints. Additionally, the possibility to extend the configuration with interactive video elements is described to enable an improved user experience.
ing individual needs has been a popular topic of research
in recent years [
ent learning videos on video platforms, for example, YouTube1, covering various topics, explained to consumers with individual pre-knowledge. The large variety of options and poor possibilities to determine if videos are relevant before consumption make it complicated for users to find adequate videos [ 6]. Efective learning At the same time, they reduce learning time by excluding unrelated or already known information [5].
The usage of natural language queries to retrieve video summaries [7], and more recently the integration of chatbots to query and interact with videos2, have been published as a possibility to support users. While those approaches work well if users are able to specify what they are searching for, a knowledge-based configuration of videos [5] has been presented as a possibility to mitigate nEvelop-O https://www.youtube.com
this weakness, by giving the user more assistance. User requirements are collected and used in a Constraint Satisfaction Problem (CSP) to determine a video fulfilling the user requirements.
ized video using the Choco solver3. show an example instantiation that provides a personal
Previous work in the synthetic creation of videos has applied video processing techniques to change the visual content to generate artwork variants using variability management techniques [8], and to generate multiple video variants for algorithm test samples [9, 10, 11]. In our approach, we preserve the video content of existing videos and parts of videos while enhancing user experience by transforming it into a well-organized structure. This extends the work of an online video generator taking an initial selection of video clips as seed to create requirements and more complex constraints.
We extend our previous work on the problem definition of configurable videos [ 5], by demonstrating a practical implementation. A reusable and adaptable approach to defining the structure of configurable videos using feature model technologies is presented, including the integration of a solver to generate personalized videos with respect to specified user requirements. Furthermore, we explain how the solution can be extended to integrate decision points with interactive video elements [13] for an improved individual user experience.
plained in Section 2. In Section 3, we present an example CEUR Workshop Proce dings htp:/ceur-ws.org ISN1613-073
CEUR
Workshop Proceedings (CEUR-WS.org) resulting video. In Section 4, the reusability and limita
Developing feature models is a complex task for persons not experienced with the notation and technologies. In order to ease the problem definition, we enable users to define the structure of configurable videos using the JSON notation provided in [14]. The JSON format is heavily used in diferent applications, and many software developers have experience with it. For this reason, we expect that it will ease the future implementation of a GUI-based editor for configurable videos, such that they can be configured by everyday users.
Within this paper, we use the configuration of a learning video explaining the transformer model in machine learning [15] as a running example. A transformer model is a type of deep learning architecture designed to process sequential data, such as text or speech, by leveraging self-attention mechanisms. It is a popular and powerful model for a variety of natural language processing (NLP) applications, including, machine translation, language understanding, and text generation.
For our example, we use videos from the Hugging Face
tutorial on NLP4 hosted on YouTube. Using those tutorial
videos, we designed a configurable video with the
structure presented as feature model [
Following the structure, we see that each video includes at least an explanation of the individual components of the transformer model (encoder and decoder ), followed tion of this approach are discussed. Finally, open research by an explanation of the overall model. Examples of each issues are discussed in Section 5 before the paper is con- part can be included optionally, as well as a digression cluded with Section 6. on the carbon footprint of transformer models.
Using the formatting described in [14] the feature model can be described in a JSON format using a straight2. Specification of Configurable forward approach. The nested JSON structure enables the Videos user to define the model following a top-down approach, where for each node, an id, the type (mandatory, optional, root), child nodes as well as sibling-relation (alternative, or) can be defined. Furthermore, exclusions and required properties can be specified using their id. Following the algorithms described in [14], the JSON structure can be translated into a valid feature model. In Figure 2, a part of the configuration is shown as an example.
In addition to this generic approach to describing the structure of the configurable video, information on the included video segments is required. More specifically, an URL where the video is available is needed. Additionally, the duration of the segments in seconds is required to define constraints for the overall duration of the generated video. This is also defined using a JSON structure, where each key references an id of the model described as JSON. An example is shown in Figure 3.
The definition of the configured video is then used to instantiate a model for the CSP. We used PyCSP35 for this purpose, which is a Python framework, to describe models for CSPs in a declarative manner. It includes the possibility to choose between the solver of ACE (AbsCon Essence)6 and Choco. The generic code to use provided JSON files for the instantiation of the configurable video, as well as the complete example, are available in our repository7.
Besides the description of a configurable video, user requirements need to be collected in order to personalize the video. While diferent possibilities, including the assessment of pre-knowledge, are possible, we restrict 4https://huggingface.co/learn/nlp-course 5http://pycsp.org 6https://github.com/xcsp3team/ace 7https://github.com/slubos/specifying-configurable-videos { ”id”:”TransformerVideo”, ”type”:”root”, ”parent”:””, ”relation”:””, ”requires”:[], ”excludes”:[], ”children”:[ { ”id”:”Encoder”, ”type”:”mandatory”, ”parent”:”TransformerVideo”, ”relation”:””, ”requires”:[], ”excludes”:[], ”children”:[ { ”id”:”EncoderExplanation”, ”type”:”mandatory”, ”parent”:”Encoder”, ”relation”:””, ”requires”:[], ”excludes”:[], ”children”:[ { ”id”:”EncoderExplanationShort”, ”type”:”optional”, ”parent”:”EncoderExplanation”, ”relation”:”alternative”, ”requires”:[], ”excludes”:[], ”children”:[] } those to the maximum video duration for our example, assuming that the video is suitable for a beginner level.
Especially in preparation for exams, students often follow the utility maximization problem [16], and try to learn as much as possible in a limited amount of time. To capture this requirement, the maximum acceptable video duration of a user is collected and translated to a maximum duration constraint. The duration is determined by summing the duration of each included video segment.
The personalized video is then generated following the configuration task described in [5], consisting of a feature { } ”EncoderExplanationShort”: { ”url”: ”https://youtu.be/H39Z...”, ”duration”: 45 }, ”EncoderExplanationLong”: { ”url”: ”https://youtu.be/MUqN...”, ”duration”: 141 }, ... model and a defined set of user requirements [ 17]. The solution of this task is a configuration , i.e., an assignment of variables in the CSP, such that the constraints of the model and user requirements are fulfilled [ 5].
In our presented approach, the variables state which of the individual video segments are included in the configured video. The respective variable domain is {true, false}, describing the inclusion or exclusion of a video segment. Nodes of the feature model described in the JSON file are defined as variables. Furthermore, for all video segments, a variable describing the duration is defined within the domain {0, videoduration}, where videoduration is the length of the video in seconds deifned in the JSON file. Using a constraint, the value of this variable is restricted to 0 or the defined videoduration depending on the inclusion of the segment in the conifgured video. Further knowledge base constraints are directly derived from the feature model described as JSON ifle, following the algorithms described in [ 14].
A configured video can be interpreted as a simple playlist, i.e., the included video segments will be played in an ordered fashion. Table 1 shows an example of the minimal and maximal video configuration in terms of video duration of the transformer model example. Depending on the maximum acceptable video duration of the user, diferent video segments are included or excluded.
Since multiple versions can be the result of a configuration task, the user has the choice to select one of the options. Considering, for example, 250 as the maximum acceptable duration, 15 configurations have been found. To enable the choice, the total duration could be shown, such that the user can select if they want to use most of their available time or not. Alternatively, a further explanation alternative might describe the content in a contrastive way, for example, video A contains more detailed explanations, while video B has more examples. Min.
Max.
A more advanced solution is based on interactive videos rather simple and the number of variables is manageable. which ofer an extension to classical videos by ofering Yet, we leave this experiment open for future work. several interactivity features [13]. This approach is used To simplify the process of specifying configurable to enable the user during the video consumption if parts videos, a GUI-based editor could be used instead of reof the video should be included, as long as they still fulfill lying on the video creator’s knowledge of JSON files the duration requirement. In terms of alternative videos, and feature models. The editor would allow the user this means a choice is presented which path is followed. to add videos by pasting links, then organize them into After each choice, the selection is included as a constraint, categories using a drag-and-drop approach. Constraints and new paths are configured on the fly. For ”or” video within categories could be specified using diferent group segments, the user can have the option to choose one or types, indicating whether they are alternatives or mulboth. In the case of optional segments, the user is asked tiple options to include. Additionally, the video creator if they want to watch it. could designate videos as mandatory or optional. Cross
Figure 4 sketches the possible path flow including the tree constraints could be added additionally to specify the decision points for the transformer video example de- requirement of videos from other categories. We expect scribed with the feature model in Figure 1, given the this user-friendly approach to be easily understandable, example requirement of 250 as the maximum accept- eliminating the need for understanding feature models. able duration. Decision points in the workflow diagram The translation of the GUI input to JSON is handled by are shown with the diamond symbol. For an interac- the application. tive video, this can be implemented as a question, with choices shown by the outgoing arrows, labeled with their description. The rectangle with rounded corners indi- 5. Open Issues for Future Work cates which video is played. A circle indicates the start, while a double-edged circle represents the end.
Each time a user takes a decision, the value is added as a constraint to the CSP, such that the remaining paths and options are computed dynamically while the user is consuming the video.
One topic for future work is the implementation of the interactive video approach described in this paper. Frameworks for this purpose, e.g., FrameTrail8 or H5P9, ofer the possibility to define the interactive elements and use them for playout. As we expect that this kind of video consumption improves learning efectiveness, conducting a user study to examine this assumption is planned. 4. Discussion A between-subject study could be conducted, where one group uses interactive videos, while the other views This paper presents a reusable approach for creating con- the complete video without interaction. Questionnaires ifgurable videos adaptable to any topic. It requires the about the video topic immediately after the video and availability of manually structured videos by the creator, after a few weeks could be used to analyze the shortand the specification of video sources must be updated ac- and long-term learning efectiveness of the interactive cordingly (see Figure 3). While YouTube videos were used approach. as an example, any video source could be utilized. The Further topics include the support of users in the defivideo creator is responsible for adapting the model for the nition of configurable videos. While our approach ofers configurable video to represent their desired structure the possibility to describe the structure of those, it is still and constraints (see Figure 2). This approach is versatile a lot of manual annotation work to describe the video, and can represent any video structure, also including more videos and constraints. We expect the approach to 8https://frametrail.org scale well for more complex videos, as the constraints are 9https://h5p.org/ for this purpose, options to automatize parts of this work should be considered. This includes the automated indexing of video content to ease recognition of what is included [18, 19], as well as their semantic segmentation defining the individual video segments that can be included ([20, 21]). Also, the possible inclusion of recommendation technologies [22, 23] to support the definition of those videos, e.g., by recommending which options could further be included.
Finally, the inclusion of diagnoses [24, 25] should be considered to relax situations where no solution can be found for given user requirements. Those can help to ifnd a configuration that takes into account as much as possible of the original user requirements.
With this paper, we present an initial implementation to define and generate personalized videos using feature models. Using an easy-to-use JSON notation, a solution was presented that is able to add additional benefit to knowledge transfer with videos by reusing already existing material. Following a practical example learning video, we showed how the approach can be used, and further extended to enable its usage with interactive videos, which is part of our future work.
The presented work has been developed within the research project Streamdiver which is funded by the Austrian Research Promotion Agency (FFG) under the project number 886205. org/10.1145/3109729.3109748. doi:10.1145/3109729. tor, in: Proceedings of the Ninth International 3109748. Workshop on Variability Modelling of Software[5] S. Lubos, M. Tautschnig, A. Felfernig, V.-M. Le, Intensive Systems, VaMoS ’15, Association for Knowledge-based configuration of videos using fea- Computing Machinery, New York, NY, USA, 2015, ture models, in: Proceedings of the 26th ACM p. 96–102. URL: https://doi.org/10.1145/2701319. International Systems and Software Product Line 2701328. doi:10.1145/2701319.2701328. Conference - Volume B, SPLC ’22, Association for [13] A. Palaigeorgiou, George and Papadopoulou, Computing Machinery, New York, NY, USA, 2022, I. Kazanidis, Interactive video for learning: A p. 188–192. URL: https://doi.org/10.1145/3503229. review of interaction types, commercial platforms, 3547052. doi:10.1145/3503229.3547052. and design guidelines, in: M. Tsitouridou, J. A. Di[6] A. Imran, F. Alaya Cheikh, S. Kowalski, Automatic niz, T. A. Mikropoulos (Eds.), Technology and annotation of lecture videos for multimedia driven Innovation in Learning, Teaching and Education, pedagogical platforms, Knowledge Management Springer International Publishing, Cham, 2019, pp. and E-Learning 7 (2015). 503–518. [7] M. Vahedi, M. M. Rahman, F. Khomh, G. Uddin, [14] H. Shatnawi, H. C. Cunningham, Encoding G. Antoniol, Summarizing relevant parts from tech- feature models using mainstream json technolonical videos, in: 2021 IEEE International Confer- gies, in: Proceedings of the 2021 ACM Southence on Software Analysis, Evolution and Reengi- east Conference, ACM SE ’21, Association for neering (SANER), 2021, pp. 434–445. doi:10.1109/ Computing Machinery, New York, NY, USA, 2021, SANER50967.2021.00047. p. 146–153. URL: https://doi.org/10.1145/3409334. [8] J. Martinez, G. Rossi, T. Ziadi, T. F. D. A. Bissyandé, 3452048. doi:10.1145/3409334.3452048.
J. Klein, Y. Le Traon, Estimating and predicting [15] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, average likability on computer-generated artwork L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Atvariants, in: Proceedings of the Companion Publi- tention is all you need, Advances in neural inforcation of the 2015 Annual Conference on Genetic mation processing systems 30 (2017). and Evolutionary Computation, GECCO Compan- [16] A. Mas-Colell, M. D. Whinston, J. R. Green, et al., ion ’15, Association for Computing Machinery, New The Utility Maximization Problem, volume 1, OxYork, NY, USA, 2015, p. 1431–1432. URL: https://doi. ford university press New York, 1995. org/10.1145/2739482.2764681. doi:10.1145/2739482. [17] L. Hotz, A. Felfernig, M. Stumptner, A. Ryabokon, 2764681. C. Bagley, K. Wolter, Configuration Knowledge [9] M. Acher, M. Alférez, J. A. Galindo, P. Romenteau, Representation and Reasoning, 1 ed., Elsevier B.V., B. Baudry, Vivid: A variability-based tool for syn- Netherlands, 2014, pp. 41–72. thesizing video sequences, in: Proceedings of the [18] Y. Deldjoo, Enhancing Video Recommendation Us18th International Software Product Line Confer- ing Multimedia Content, Springer International ence: Companion Volume for Workshops, Demon- Publishing, Cham, 2020, pp. 77–89. URL: https:// strations and Tools - Volume 2, SPLC ’14, Asso- doi.org/10.1007/978-3-030-32094-2_6. doi:10.1007/ ciation for Computing Machinery, New York, NY, 978- 3- 030- 32094- 2_6.
USA, 2014, p. 143–147. URL: https://doi.org/10.1145/ [19] M. Elahi, F. Bakhshandegan Moghaddam, R. Hos2647908.2655981. doi:10.1145/2647908.2655981. seini, M. H. Rimaz, N. El Ioini, M. Tkalcic, [10] J. A. Galindo, M. Alférez, M. Acher, B. Baudry, D. Be- C. Trattner, T. Tillo, Recommending Videos in navides, A variability-based testing approach for Cold Start With Automatic Visual Tags, Associsynthesizing video sequences, in: Proceedings ation for Computing Machinery, New York, NY, of the 2014 International Symposium on Software USA, 2021, p. 54–60. URL: https://doi.org/10.1145/ Testing and Analysis, ISSTA 2014, Association for 3450614.3461687.
Computing Machinery, New York, NY, USA, 2014, [20] T. Tuna, M. Joshi, V. Varghese, R. Deshpande, p. 293–303. URL: https://doi.org/10.1145/2610384. J. Subhlok, R. Verma, Topic based segmentation 2610411. doi:10.1145/2610384.2610411. of classroom videos, in: 2015 IEEE Frontiers in [11] M. Alférez, M. Acher, J. A. Galindo, B. Baudry, Education Conference (FIE), 2015, pp. 1–9. doi:10.
D. Benavides, Modeling variability in the video 1109/FIE.2015.7344336. domain: Language and experience report, Soft- [21] P. A. Co, W. R. Dacuyan, J. G. Kandt, S.-C. ware Quality Journal 27 (2019) 307–347. URL: https: Cheng, C. L. Sta. Romana, Automatic topic//doi.org/10.1007/s11219-017-9400-8. doi:10.1007/ based lecture video segmentation, in: Innos11219- 017- 9400- 8. vative Technologies and Learning: 5th Interna[12] G. Bécan, M. Acher, J.-M. Jézéquel, T. Menguy, On tional Conference, ICITL 2022, Virtual Event, Authe variability secrets of an online video genera- gust 29–31, 2022, Proceedings, Springer-Verlag,
doi.org/10.1007/978-3-031-15273-3_4. doi:1 0 . 1 0 0 7 / 9 7 8 - 3 - 0 3 1 - 1 5 2 7 3 - 3 _ 4 . [22] A. Falkner, A. Felfernig, A. Haag, Recommendation Technologies for Configurable Products, AI Magazine 32 (2011) 99–108. [23] A. Felfernig, V.-M. Le, A. Popescu, M. Uta, T. N. T.
Tran, M. Atas, An overview of recommender systems and machine learning in feature modeling and configuration, in: 15th International Working Conference on Variability Modelling of Software-Intensive Systems, VaMoS’21, Association for Computing Machinery, New York, NY, USA, 2021. URL: https://doi.org/10.1145/3442391.3442408. doi:1 0 . 1 1 4 5 / 3 4 4 2 3 9 1 . 3 4 4 2 4 0 8 . [24] A. Felfernig, M. Schubert, C. Zehentner, An eficient diagnosis algorithm for inconsistent constraint sets, AI for Engineering Design, Analysis, and Manufacturing (AIEDAM) 26 (2012) 53–62. [25] A. Felfernig, R. Walter, J. Galindo, D. Benavides, M. Atas, S. Polat-Erdeniz, S. Reiterer, Anytime Diagnosis for Reconfiguration, Journal of Intelligent Information Systems 51 (2018) 161–182.