Combining IGA and KG for Serendipitous Learning Contents Recommendation Emmanuel Ayedoun 1, Satoko Inoue 2, Hiroshi Takenouchi 3, and Masataka Tokumaru 1 1 Faculty of Engineering Science, Kansai University, 3-3-35 Yamate-cho Suita, Osaka 564-8680 Japan 2 Graduate School of Engineering, Kansai University, 3-3-35 Yamate-cho Suita, Osaka 564-8680 Japan 3 Department of System Management, Fukuoka Institute of Technology, 3-30-1 Wajiro-higashi, Higashi-ku, Fukuoka, 811-0295 Japan Abstract Although there have been few attempts to propose serendipity-oriented recommender systems in the field of education, such systems appear to lack of the essential ability to support learners’ agency, which is learners’ feeling of ownership and control over their own learning. In this paper, we propose an Interactive Evolutionary Computation driven recommender system that enables learners to take control and responsibility of their own learning while recommending learning resources that are novel and unexpected, yet still relevant to learners’ interests. The proposed system specifically employs Interactive Genetic Algorithm (IGA) and Knowledge Graphs (KG) for dynamic recommendation of learning contents related to the history of scientific discoveries. We conducted both numerical simulations that confirmed the effectiveness of the learning contents optimization algorithm and an experimental evaluation which hinted at the meaningfulness of the proposed approach towards inducing serendipity within learners. Keywords 1 Recommender Systems, Serendipitous Learning, Interactive Evolutionary Computation 1. Introduction overgeneralization of recommended information can impair the ability of learning support systems to provide learners with content that is interesting, The deployment of recommender systems in novel and more importantly unexpected [5]. As a the field of technology enhanced education has result, such approaches can lead to an overly attracted increased interest as a promising means narrow set of suggestions lacking in serendipity to help learners navigate through suitable learning and inadvertently placing the learner in what is resources, given the plethora of available digital known as a “filter bubble”, according to Pardos learning resources nowadays [1]. The principal [6]. That is, proposing recommender systems that and commonly used techniques to build also aim at helping learners make serendipitous recommender systems are collaborative filtering knowledge acquisition is necessary to tackle the [2] and content-based filtering [3]. However, in an filter bubble issue. educational context, both approaches present The term “Serendipitous learning” has been some shortcomings: risk of overgeneralization for used to refer to learning through gaining new collaborative recommenders and risk of insights, discovering interesting aspects and overspecialization as far as content-based recognizing new relations, which occurs by recommenders are concerned. Such issue has been chance or as by-product of other activities [7], [8]. framed as the “serendipity-problem” [4], to Serendipitous learning emphasizes the positive denote that the overspecialization or Proceedings Name, Month XX–XX, YYYY, City, Country EMAIL: email1@mail.com (A. 1); email2@mail.com (A. 2); email3@mail.com (A. 3) ORCID: XXXX-XXXX-XXXX-XXXX (A. 1); XXXX-XXXX- XXXX-XXXX (A. 2); XXXX-XXXX-XXXX-XXXX (A. 3) ©️ 2020 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR Workshop Proceedings (CEUR-WS.org) role of unexpected realization of hidden, informal learning, which depends to a large extent seemingly unrelated connections or analogies for on individual preferences or choices and is often learning and research [8], [9]. Although there self-directed [19], could be greatly enhanced by have been few attempts to propose serendipity- introducing in such learning environments oriented recommender systems in the field of serendipity-oriented recommender systems. As education [10], [11], [12], such systems do not evoked in the previous section, it should also be necessarily support learners’ agency, which is yet noted that most recommender systems dedicated an essential requirement, as serendipitous to learning support embed recommendation encounters also owe to the open-minded attitude techniques that could inadvertently place learners of the seekers, their curiosity, and their into “filter bubbles”, a type of swim-laning of perspicacity [13]. learners into a particular track based on a machine Interactive Evolutionary Computation (IEC) is learned stereotype [8]. Meanwhile, it has been a generic term which refers to a group of suggested that serendipitous experiences are optimization techniques or algorithms that uses valuable to learning at a personal level [10]. subjective human evaluation instead of a Therefore, to the extent of fostering learners’ numerical fitness function to solve optimization engagement in informal learning settings, the goal problems when the fitness function cannot be and major contribution of this study is to propose assumed or appropriately represented in the form a serendipity-oriented recommender system of a mathematical function [14]. Given such which fulfill the following requirements: characteristics, IEC techniques have been • Target support of learning in an informal successfully applied in many fields, such as face learning environment identification [15], fashion design [16], music • Facilitate learner’s agency by actively composition [17], hearing aid fitting [18]. In a supporting self-directed learning through typical scenario of IEC, a small number of exploratory interaction with the learning solutions (e.g., a population of ten solutions) are environment shown to a human user who is supposed to assign • Embed a resource recommendation one of a pre-specified set of ranks (e.g., 1: very algorithm that involves learners in the bad, 2: bad, 3: average, 4: good, 5: very good) to system recommendation refining process each solution in the population. by actively gathering their preferences In this paper, we propose an Interactive Evolutionary Computation (IEC) driven recommender system that enables learners to take 2.2. Approach control and responsibility of their own learning while exploring learning resources that are novel 2.2.1. Overview of proposed system and unexpected, yet still relevant to their interests. The proposed system specifically employs Figure 1 shows an overview of the proposed Interactive Genetic Algorithm (IGA) and system. In the system, the learning contents are Knowledge Graphs (KG) for dynamic generation represented in the form of “learning paths” of learning contents. covering related concepts and presented to learners via a dedicated interface, shown in Figure 2. Research Goal and Approach 2.1. Problem Statement and Research Goal In the domain of technology enhanced learning, a number of recommender systems have been proposed. Yet, a closer look to the current status of their development and evaluation reveals that such efforts present some limitations. For instance, available systems seem to target learning in formal settings, do not sufficiently support Figure 1: Schematic representation of learners’ agency and evaluate effectiveness only interactions between the learner and the from the standpoint of learners’ grade. However, system interesting to the user by leveraging IGA, and generates new paths based on those features. If the generated path exists in the path database, it is presented to the user as-is, and if it does not exist, it is replaced by the most similar path in the path database and presented to the user (Phase 2). By repeating the evaluation of the proposed paths, the system attempts to learn the learners’ taste and interests, and presents them with novel paths of greater interest, yet unexpected enough to achieve Figure 2: System Interface showing the paths recommendation of learning contents that could navigation window induce serendipity. 2. For instance, the knowledge database used for the study presented in this paper is a database in 2.2.2. Knowledge graph model which learning contents (i.e., scientific discoveries and inventions) are related to each In general, a knowledge graph G= {E, R, F} is other and such relationships can be quantitatively a collection of entities E, R, and facts F [22]. A expressed. To this extent, we built the learning fact is a triple (ℎ, 𝑟, 𝑡) ∈ F that denotes a link 𝑟 ∈ contents database of the system using the contents R between the head ℎ ∈E and the tail 𝑡 ∈ E of the of the book “Science: The Definitive Visual triple. In our proposed system, the relationship Guide, Adam Hart-Davis (Ed.)” [21]. It is a between nodes and edges is also represented using comprehensive book which tells the history of the common (ℎ, 𝑟, 𝑡) triples. Note that ℎ and 𝑡 science and technology from the earliest times to represent two different nodes in the knowledge the present day in chronological order by graph, while 𝑟 represents an edge linking these capturing every key moment of discovery, and nodes. In the following lines, we provide an showing how the concepts, the inventions, and the overview of how we define these triples in the individuals behind them have changed our world. context of this study. More interestingly, the book illustrates how one First of all, we expressed ℎ as a collection of discovery is connected to another by presenting the three parameters vectors ℋelement, ℋBefore and some pointers to events that preceded and ℋAfter. followed a current discovery or invention. Such ℋelement structure obviously holds the potential to make it ℎ = [ ℋBefore ] (1) easier for the reader to realize how scientific ℋAfter discoveries and inventions in a wide range of ℋelement represents the main contents of a page, scientific fields are interrelated to each other. In and is expressed as in equation (2), where ℎpage is the resulting knowledge graph, each piece of the page number of the node, ℎdiscipline is the information (i.e., major discovery or invention) is discipline (i.e., scientific field), and ℎera is the era represented by a node, and the relationship of the node contents. between related nodes is depicted by an edge. In 𝑒1 ℎpage other terms, each node holds the contents of each ℋelement = [𝑒2] = [ℎdiscipline] (2) page of the book, while an edge expresses the 𝑒3 ℎera relationship between two related pages. Therefore, what is called “learning path” in the ℋBefore represents the related pages labeled as context of this study is a collection of nodes and page BEFORE (=B) in the book, which refer to edges extracted from a knowledge graph. the related pages older than the current page. Generation and optimization of learning path to be ℋBefore is defined as in equation (3) according to presented to learners at a given time of the the number of older related pages NB, and each interaction are achieved by the means of an BEFORE page 𝑏i. interactive genetic algorithm (IGA), a kind of IEC 𝑏1 algorithm. 𝑏2 In the proposed system, users are first asked to ⋮ explore the knowledge graph and select paths of ℋBefore = 𝑏 = 𝑏 (1  i  NB) (3) i interest which they evaluate ((Phase 1). Then, the ⋮ system learns the features of the paths that are [𝑏𝑁𝐵 ] ℋAfter is defined similarly to ℋBefore and quantitatively express the degree of relevance or represents the related pages labeled as page divergence between two nodes (i.e., learning AFTER (=A) in the book, as shown in (4). Note contents). that NA stands for the number of related pages coming after the current page 𝑎j. 2.2.3. Learning path optimization 𝑎1 algorithm 𝑎2 ⋮ Path optimization here refers to the generation of ℋAfter = 𝑎 = 𝑎j (1  i  NA) (4) new paths of interest to the user by the system. Let ⋮ N be the number of paths generated from the [𝑎𝑁𝐴 ] knowledge graph G described in the previous section, and pathk (k ∈ N ), a path arbitrarily Next, 𝑡 which also represents a content node retrieved from the path database. In this study, each similarly to ℎ above is defined as follows. Let 𝑡page pathk has a fixed length and is composed of four denote the page number, 𝑡discipline denote the nodes h1, h2, h3, h4 (h1, h2, h3, h4 ∈ ℎ) and three discipline, and 𝑡era the era. t is expressed as in (5). edges 𝑟1, 𝑟2, and 𝑟3 (𝑟1, 𝑟2, 𝑟3 ∈ 𝑟). 𝑡1 𝑡page Considering that the edges 𝑟1, 𝑟2, 𝑟3 are 𝑡 = [𝑡2] = [𝑡discipline] (5) defined as in (6), 𝑟𝑝𝑎𝑡ℎ𝑘 which is the vector 𝑡3 𝑡era Finally, 𝑟 consists of the association of the representing the whole path (i.e., 𝑝𝑎𝑡ℎ𝑘 ) is following three vectors 𝒳element, 𝒳Before, and 𝒳After, expressed as the sum of 𝑟1, 𝑟2, and 𝑟3 as follows: as shown in equation (6). 𝑟𝑝𝑎𝑡ℎ𝑘 = [𝑟1, 𝑟2, 𝑟3] (10) 𝒳element In the present study, the process of path 𝑟 = [ 𝒳Before ] (6) optimization using IGA is based the gene 𝒳After information expressed by 𝑟𝑝𝑎𝑡ℎ𝑘 . To such extent, 𝒳element expresses the relation between the the learner first rates some paths presented to him main contents of node ℎ and the main content of by the system in terms of relevance with their node 𝑡 in terms of difference between discipline interests. Here, it seems important to bear in mind and era parameters, as shown in (7). |ℎ𝑑𝑖𝑠𝑐𝑖𝑝𝑙𝑖𝑛𝑒 − 𝑡𝑑𝑖𝑠𝑐𝑖𝑝𝑙𝑖𝑛𝑒 | that learners are not prompted to evaluate each 𝒳element = [ ] (7) edge or node, but the whole path with a focus on |ℎ𝑒𝑟𝑎 − 𝑡𝑒𝑟𝑎 | the connection between starting nodes and ending 𝒳Before is defined as the difference between nodes. The intention here, is to make the system node ℎ and 𝑡 in terms of three parameters: pages number, discipline, and era, as shown in equation capture how interesting the learner finds the (8). Note that here 𝒳𝐵𝑖 = 0 if 𝑡 = 𝑏𝑖 (i ∈ NB). connection between several related events across |𝑏𝑖1 − 𝑡1 | various scientific disciplines and eras. Based on the obtained evaluation values, the path is 𝒳Before = 𝒳Bi = [ 𝑖2 − 𝑡2 |] |𝑏 (8) |𝑏𝑖3 − 𝑡3 | optimized by genetic algorithm processing, and Similarly, 𝒳After is defined as the difference the next-generation solution candidate (i.e., between node ℎ and 𝑡 in terms of three learning path) is presented to the learner. The path parameters: pages number, discipline, and, and is optimized by repeating this process for a certain era, as shown in equations (9). Here as well, number of generations. Note that here, the path 𝒳𝐴𝑗 =0 if 𝑡 = 𝑎𝑗 (j ∈ NA). optimization differs from usual implementation of |𝑎𝑗1 − 𝑡1 | IGA as it requires an additional process that we call Path retrieval. When generating the next 𝒳After = 𝒳Aj = [|𝑎𝑗2 − 𝑡2 |] (9) generation of solutions, in most cases, Crossover |𝑎𝑗3 − 𝑡3 | or Mutation will cause the generation of candidate Based on the proposed knowledge graph solutions (i.e., paths) that do not exist in the path model, our key idea is to let an edge 𝑟 capture database RDB. Therefore, for example, a non- differences in terms of discipline, era and page existent path rpathA needs to be “replaced” by an number between two given nodes, ℎ and 𝑡. existing path rpathB with the constraint that both Besides, by expressing era and page number as time series parameters and adopting a similarity paths are similar enough (i.e., rpathA  rpathB). To the scale for the discipline parameter, we aim to extent of calculating the degree of similarity between two paths, we adopted the Dynamic Time contents (i.e., path) presented at the last Warping (DTW) algorithm [23], which is a well- generation was relatively high (M= 3.9, SD:0.92). known technique to find an optimal alignment Next, DTW values, which indicate similarity between two given (time-dependent) sequences degrees between the generated path by the under certain restrictions. The DTW distance algorithm and the one retrieved from the database, 𝐷(𝑝𝑎𝑡ℎ𝑘 , 𝑝𝑎𝑡ℎ𝑙 ) which indicates the degree of tend to converge to a value near 0 around the last similarity of two different paths 𝑝𝑎𝑡ℎ𝑘 and generation. This suggests that the proposed 𝑝𝑎𝑡ℎ𝑙 is recursively calculated using the system was able to generate paths that are close to paths which exist in the database. This is a good following equations: indication that the proposed DTW-based path 𝐷(𝑝𝑎𝑡ℎ𝑘 , 𝑝𝑎𝑡ℎ𝑙 ) similarity calculation method performed well. = 𝛿(𝑟𝑝𝑎𝑡ℎ𝑘 , 𝑟𝑝𝑎𝑡ℎ𝑙 ) However, when analyzing the transition of 𝐷(𝑝𝑎𝑡ℎ𝑘 , 𝑝𝑎𝑡ℎ𝑙−1 ) DTW values for some subjects, there were cases + min { 𝐷(𝑝𝑎𝑡ℎ𝑘−1 , 𝑝𝑎𝑡ℎ𝑙 ) (11) in which DTW values rose rapidly even near the 𝐷(𝑝𝑎𝑡ℎ𝑘−1 , 𝑝𝑎𝑡ℎ𝑙−1 ) last generation or did not show a decreasing trend despite the number of generations increased, such where 𝛿(𝑟𝑝𝑎𝑡ℎ𝑘 , 𝑟𝑝𝑎𝑡ℎ𝑙 ) denotes the distance as in the case of Subject B (Figure 4). Therefore, we cannot rule the hypothesis that using a method between respective edges of 𝑝𝑎𝑡ℎ𝑘 and 𝑝𝑎𝑡ℎ𝑙 other than DTW distance as a method for calculated as: calculating path similarity may lead to higher 𝛿(𝑟𝑝𝑎𝑡ℎ𝑘 , 𝑟𝑝𝑎𝑡ℎ𝑙 ) = |𝑟𝑝𝑎𝑡ℎ𝑘 − 𝑟𝑝𝑎𝑡ℎ𝑙 | (12) performance for path optimization. From the results of the questionnaire survey, 3. Pilot Evaluation we note that the proposed system was able to present interesting and surprising learning contents to two out of three subjects. Moreover, We conducted an experimental pilot two subjects also declared that they were able to evaluation to investigate whether the proposed experience serendipity through their interaction system could present information of interest but with the system. Such results seem to suggest the yet unexpected enough to the extent to induce meaningfulness of the proposed approach. serendipity within participants. The subjects were 3 university students majoring in science-related fields. Subjects were asked to visit and then evaluate the paths proposed by the system in terms of preference level on the scale of 0 to 5. Based on their ratings, the system generated new paths and the same operation was repeated until the ending condition (i.e., 10 generation rounds) was met. At the end of the interactions, we administrated a questionnaire survey, to collect participants’ subjective opinions on the meaningfulness of their interaction with the system. Figure 3~5 show the transition of the DTW values and evaluation scores of the most highly rated paths by each of the three participants (Subjects A~C) between the first and last generation rounds. First, from these results, it can be noted that the proposed system was able to optimize the paths according to each user since the highest evaluation scores from participants seem to stabilize around the last generations. In other terms, the system was able to gradually present subjects with learning contents that were highly Figure 3: Evaluation scores of the best paths and rated. The average evaluation score of the learning corresponding DTW values (Top to Bottom Subject A, B and C). [11] M. J. Ibarra, C. Serrano, and A. F. Navarro, [1] H. Drachsler, K. Verbert, O. C. Santos, and “Recommender system to identify students N. Manouselis, “Panorama of recommender with learning deficiencies in assessments” In systems to support learning,” In 2016 International Symposium on Recommender systems handbook, pp. 421– Computers in Education (SIIE), IEEE, pp. 1– 451, Springer, Boston, MA, 2015, doi: 6, 2016. 10.1007/978-1-4899-7637-6_12. [12] R. Rismanto, A. Rachmad Syulistyo, and B. [2] M. D. Ekstrand, “Collaborative Filtering P. Citra Agusta, “Research Supervisor Recommender Systems,” Foundations and Recommendation System Based on Topic Trends® in Human–Computer Interaction, Conformity,” International Journal of vol. 4, no. 2, pp. 81–173, 2011, doi: Modern Education and Computer Science, 10.1561/1100000009. vol. 12, no. 1, pp. 26–34, Feb. 2020, doi: [3] P. Lops, M. D. Gemmis, and G. Semeraro, 10.5815/ijmecs.2020.01.04. “Content-based recommender systems: State [13] O. Yaqub, “Serendipity: Towards a of the art and trends,” In Recommender taxonomy and a theory,” Research Policy, systems handbook, pp. 73–105, Springer, vol. 47, no. 1, pp. 169–179, Feb. 2018, doi: Boston, MA, 2011, doi: 10.1007/978-0-387- 10.1016/j.respol.2017.10.007. 85820-3_3 [14] H. Takagi, “Interactive evolutionary [4] L. Iaquinta, M. de Gemmis, P. Lops, G. computation: fusion of the capabilities of EC Semergro, M. Filannino, and P. Molino, optimization and human evaluation,” “Introducing serendipity in a content-based Proceedings of the IEEE, vol. 89, no. 9, pp. recommender system”, In Proceedings of the 1275–1296, 2001, doi: 10.1109/5.949485. Eighth International Conference on Hybrid [15] C. Caldwell and V. S. Johnston, “Tracking a Intelligent Systems, IEEE, pp. 168–173, criminal suspect through” face-space “with a 2008. genetic algorithm”, In Proceedings of [5] T. Gup, “Technology and the end of International Conference on Genetic serendipity”, The Chronicle of Higher Algorithms, 1992, pp. 416–421. Education, vol. 44, p. A52, 1997. [16] J. H. Lee, H. S. Kim, and S. B. Cho, [6] Z. A. Pardos, and W. Jiang, “Designing for “Accelerating evolution by direct serendipity in a university course manipulation for interactive fashion design,” recommendation system” In Proceedings of In Proceedings of the fourth International the tenth international conference on learning Conference on Computational Intelligence analytics & knowledge, pp. 350–359, 2020. and Multimedia Applications. ICCIMA, [7] I. Buchem, “Serendipitous learning: 2001, pp. 343–347. Recognizing and fostering the potential of [17] N. Tokui, and H. Iba, “Music composition microblogging”, Form@re, vol. 11, no. 74, with interactive evolutionary computation.” pp. 7–16, Mar. 2013, doi: 0.13128/formare- In Proceedings of the third international 12559. conference on generative art, 2000, vol. 17, [8] G. A. Fine and J. G. Deegan, “Three no. 2, pp. 215–226. principles of Serendip: insight, chance, and [18] H. Takagi and M. Ohsaki, “Interactive discovery in qualitative research,” Evolutionary Computation-Based Hearing International Journal of Qualitative Studies Aid Fitting,” IEEE Transactions on in Education, vol. 9, no. 4, pp. 434–447, Oct. Evolutionary Computation, vol. 11, no. 3, pp. 1996, doi: 10.1080/0951839960090405. 414–427, Jun. 2007, doi: [9] J. Gritton, “Can serendipitous browsing lead 10.1109/tevc.2006.883465. to serendipitous [19] R. G. Brockett and R. Hiemstra, Self- learning?,”http://www.futurelab.org.uk/reso direction in Adult Learning: Perspectives on urces/publications-reports-articles/web- Theory, Research and Practice. Routledge, articles/Web-Article795 (accessed June 1994. 30th, 2022). [20] E. Olshannikova, T. Olsson, J. Huhtamäki, S. [10] O. R. Zaiane, “Building a recommender Paasovaara, and H. Kärkkäinen, “From agent for e-learning systems,” In Chance to Serendipity: Knowledge Workers’ Proceedings of International Conference on Experiences of Serendipitous Social Computers in Education, IEEE, pp. 55-59, Encounters,” Advances in Human-Computer 2002, doi: Interaction, vol. 2020, pp. 1–18, Feb. 2020, doi: 10.1155/2020/1827107. [21] J. Gribbin, A. Hart-Davis, D. Palmer, J. Cherfas, and Dorling Kindersley Limited, Science, The definitive visual guide. London: Dorling Kindersley Publishers Ltd. [22] A. Rossi, D. Barbosa, D. Firmani, A. Matinata, and P. Merialdo, “Knowledge Graph Embedding for Link Prediction,” ACM Transactions on Knowledge Discovery from Data, vol. 15, no. 2, pp. 1–49, Apr. 2021, doi: 10.1145/3424672. [23] T. Giorgino, “Computing and Visualizing Dynamic Time Warping Alignments inR: ThedtwPackage,” Journal of Statistical Software, vol. 31, no. 7, 2009, doi: 10.18637/jss.v031.i07.