Cognitive Learning Recommendation System in Indian Context Anurag Gulati Electrical Dept DTU, Delhi , India Anish Batra I.T Dept MSIT, Delhi, India Abstract—In a huge population like India, every student grows in a different psychological environment during formative years. Further formal education system leads to diversification of thought process and certain inclination towards a cognitive sense. Due to this it becomes too ambiguous for students to find the best suited methodology of learning. This paper presents machine learning based recommendation system to decipher previously hidden relationships between actual classroom teaching and optimal methods of learning for individual student. The proposed recommendation system provides a student with a series of methods of learning ranked from high to low based on the cognitive sense of the individual student and helps him to find the best learning methods for an individual student. The proposed recommendation system has used extensively surveyed data set compiled from multiple schools under the Indian education system for testing. The proposed system has increased the knowledge among students as a whole through recommended methods of learning, leading to greatly improved class room results. Index Terms— Recommendation system, collaborative filtering, methods of learning, machine learning I. INTRODUCTION I N India, there is a specific pattern to teach students wherein, they are categorized purely on the basis of marks obtained in merit based examinations. The only aids that are used to teach are black board and lecture along with visual based presentation tools which are restricted to the few elite schools that have adequate resources and infrastructure to support the technology and operate it. The problem here is that the students are collectively exposed to any random method of learning available disregarding the fact that each student has a different thought process and knowledge grasping channel. In this context, there is a requirement for recommending a number of learning methods and not just one to students. If just one learning method is recommended to a student, it might not work well due to certain circumstances the student might be in. Therefore, it would be appropriate to use a recommender system having large collection of learning methods out of which most suitable learning methods may be recommended to the students along with the ranking of the methods. Such a recommender system may work well as a whole . This type of recommender system might appear to be simplistic but its design is much more complex in a way that it has to handle a lot of unprocessed data. It also 978-1-5386-1922-3/17/$31.00 © 2017 IEEE Rohit Khurana CSE Dept JIIT, Noida, India Dr. MM Tripathi Electrical Dept DTU, Delhi, India gives a shape to the profuse data since there is a lot of data on the internet in the form of images, text and documents , . The quality of the recommendation system need to be taken care as poor results can cause high cost problem which do prevails in the world . There are millions of students in India studying in the schools and colleges, getting well structured education in a customary fashion. This paper proposes a cognitive learning recommending system based on machine learning for these students to help them to evolve and grow with a holistic perspective, leading to creation of leaders in diverse fields. By categorizing students on the basis of multiple intelligent questions and correlating them with a corresponding method of learning, an extremely rich and informative data set is generated. This data is fed to the proposed recommendation system and further the experiences of several students are also incorporated to predict the best category of learning method for a student. The recommendation system also gives other important outcome i.e names of students who are more likely to be similar in cognitive abilities to the one whose recommendations are being generated. The experience factor helps to increase quality of results as certain methods of learning which are common to a majority students (studying under the same environment), can be filtered by the recommendation system. During the present work presented in the paper, a data set of 3000 students was collected via extensive surveys in different schools (both government and private owned) under the ambit of the standardized central board of secondary education (CBSE) in India. These students were provided with a common questionnaire consisting of 31 questions, of which 30 were self-scoring and based on ‘multiple intelligence’. These questions were used to deduce the category of strengthened cognition in decreasing order of likeliness. The 31st question was a general question to rate financial status of parents from 1 to 4. This question helped us deduce impact of resources available in shaping the cognitive senses of an individual while growing up. The overview can be more easily understood by the following chart shown in fig. 1. The data set was fed into the program designed to generate recommendations and results. The program was written using python and employing analytical tools such as popularity model and personalized model to produce recommendations. The results from the program produced rank-wise recommendations for a particular student based on the complete data set for each of the stated models. Fig. 1: Weights given to Input Subsequently an examination was conducted based on the generated recommendations to test the validity of the experiment. The students were allowed to attempt an exam based on traditional techniques. In continuation to it another test was given, with the key difference that this time the student applied the best suited method of learning she/he thinks most suitable from a range of recommended options suggested to her/him. Further, with help of analytical tools, some relations were found within the data sets which were verified with observed cases. Additionally, the impact of available resources in shaping the cognitive abilities of a child during the growing stages was studied. Based on the results it may be concluded that the proposed recommendation system can yield higher order benefits. Rest of the paper is organized in following fashion. Section II presents the theoretical aspect of the proposed recommender system and section III discusses the modeling and prediction algorithm for recommender system. Data preparation is presented in section IV whereas results are presented in section V. The conclusion is presented in section VI. II. RECOMMENDER SYSTEM In last decade, recommender system using content-based filtering has got attention of many researchers , . In this paper, the collaborative filtering has been employed as this approach looks promising. Collaborative filtering is also used in other areas such as e-commerce, where each user is recommended the products according to her/his own prior choices, as well as the taste choices of other users who bought some common product as the one in question . Since it is proposed to recommend the learning methods to users using the data of the considerable population, the concept of item based collaborative filtering as shown in fig. 2 is employed in present work. One similarity measure is used to find the similarity between different items in the dataset which is employed to predict ratings of user-item pairs that is not given in the dataset . Fig. 2: Weights given to Input in item based collaborative filtering Different types of similarity concepts are used while implementing item based collaborative filtering . One type is cosine based similarity formulation that views items and their ratings as vectors. It also considers the angle between given vectors as similarity between them. Two similarity measure methods are discussed below. A. Correlation based similarity The deviation of ratings of common users from average ratings for a perticular item is used to predict the similarity as shown in eq. 1 below. (1) B. Adjusted cosine similarity In this theory, it is considered that some users may give higher ratings to items, whereas others may rate low as a preference . Hence for a pair of item average rating of a user are subtracted from each user's rating under consideration as given in eq. 2 (2) The rating of any user-item pair may be predicated, using weighted sum as described in eq. 3, after model is ready using any similarity measure. (3) From identified items similar to the target, the items rated by active user is picked. The similarity between identified and target item is used to assign weight to every user's rating for each item. At the end, scaling of prediction is done using the sum of similarities to achieve the closest predicted rating. III. MODELLING AND PREDICTION ALGORITHM In this paper, the personalized algorithm of Python has been used for prediction. The flow chart of proposed algorithm and the method we have used is shown in fig. 3 which is self-explanatory. The internal processing is done with a combination of nearest neighbor searching, dense tables for tracking item-item similarities, and sparse item-item tables. A model is created that makes recommendations using item popularity. When no target column is provided, the popularity is determined by the number of observations involving each item. When a target is provided, popularity is computed using the item’s mean target value. When the target column contains ratings, for example, the model computes the mean rating for each item and uses this to rank items for recommendation. In order to understand the connectivity between the theory of the recommender system and our implementation, consider the item based filtering at basic wherein the item is the Learning Scheme described in the Table 2. Now, imagine a student inclined towards interpersonal – collaborative learning. The student is used to studying with that technique and he is not focused upon according to his mind, psychology and surroundings. Now, there are several other students in the same/other schools with similar matching circumstances who might be using other learning techniques as that student. As a matter of fact, that student is recommended with the other student’s learning technique which might work well for him. In a similar fashion, various techniques are recommended to him in the decreasing order of the probabilities of the match. The beauty of the recommendation relies in the fact that the items recommended to the students are in the dataset itself and used or pursued by other students/users with similar psychological trends who might be producing different results for the same task. In this manner, the users/students learn from the other users; who are similar to the one in consideration. This helps the user choose the best out of the options amalgamated by both environmental factors as well as the ideal scenario. Refer to fig.2 and the formulas used in the section II. IV. DATA PREPARATION Data was collected from different schools of Delhi, India (both private and government) for the research. Collected data was molded into the required datasets so as to analyze it in the required manner. The learning schemes were categorized into 30 categories, hereafter named as LSCAT, which could tell about the present learning methods of the students and recommend them the best suited learning methods using the collaborative filtering . Table 1: Parent’s Income Data Parent’s Income Category Parent’s income Category 1 0-4 lakhs per year Category 2 4-7 lakhs per year Category 3 Category 4 7-10 lakhs per year Above 10 lakhs per year The details of the students and the schools is not presented due to privacy reasons. The income of the parents of the students was also taken into account so as to find out the relation between financial condition, type of the society and different learning methods. The parent’s income has been categorized as given in table 1. Fig. 3: Flow chart of prediction algorithm Table 2: Learning method data set sample LSCAT Learning scheme 1 Interpersonal Collaborative learning Interpersonal Connected learning Interpersonal Cooperative learning Interpersonal - Inside outside circle learning Interpersonal - Jigsaw technique Kinesthetic – Experimental learning Kinesthetic - Interpretive dance Kinesthetic - Role playing Kinesthetic - Serious play Kinesthetic - Tactical 2 3 4 5 6 7 8 9 10 Parent’s User ID Score Income Category 301 5 2 301 3 2 301 5 2 301 3 2 301 5 2 301 5 2 301 5 2 301 4 2 301 5 2 301 5 2 A sample learning methods data set for a student (user ID 301) is shown in table 2 which contains 10 categories out of 31 categories. The score in table 2 is a self-rated entity given by student for corresponding learning scheme. It ranges from 1 to 5, wherein 1 is strongly disinterested and 5 is strongly interested. The learning methods of the students were analyzed and rated using the psychometric exam (which was a 31 question exam) and each question corresponded to a particular learning method . The questions were rated by the students in the form of marks from 1 to 5. There were 3000 students in totality in the sample space. The best suited learning methods were recommended to all the students using popularity and personalized methods. Table 3 shows the recommendations made to a particular user using popularity method whereas table 4 shows the recommendations made to the same user using personalized method. It is evident from both tables that both the models recommend the learning methods of user id 15 but with different learning schemes. The recommendations using personalized method recommends the category 23 as rank 1 but the popularity method recommends the category 29 to the same user as preferred rank. Table 3: Recommendation to user using popularity method V. RESULTS AND DISCUSSIONS Data analysis was performed on a data set experimented upon aaprox. 3000 students. The students were provided with a common questionnaire consisting of 31 questions, of which 30 were self-scoring and based on ‘multiple intelligence’ to deduce the category of strengthened cognition in decreasing order of likeliness. The 31st question was a general question to rate financial status of parents from 1 to 4. Based on the experimental results, the graph in fig. 4 is constructed for a sample of 100 students, where x-axis depicts category id and y-axis depicts sum of scores of all respective category ids of all students in a sample set. Fig. 4: User ID vs. learning scheme category ID The fig. 4 indicates the inclination of students towards to method of learning of category id 5 which is Jigsaw Technique. Jigsaw technique is a classroom activity wherein students are dependent on each other to succeed. Students with different abilities work in together on small problems, each topic at a time to finish over the entire syllabus or course. A total of 30 different samples were considered and a significant majority of them shows the same result as shown in fig. 4. Result indicates that the Jigsaw technique has a significant overlap to the Indian education system. This shows that due to growing up under this system, the students have been moulded in a way through which students have got used to this type of learning method even if their cognitive abilities prove otherwise. User ID 15 15 15 15 15 15 15 15 15 15 LSCAT 29 28 11 8 20 10 5 30 23 21 Score 85.0 85.0 85.0 85.0 84.0 84.0 84.0 83.0 82.0 82.0 Rank 1 2 3 4 5 6 7 8 9 10 Table 4: Recommendation to user using personalized method User ID 15 15 15 15 15 15 15 15 15 15 Code 23 19 21 14 1 13 18 2 17 27 Score 0.709072499828 0.705763818942 0.705517884365 0.703353118817 0.69693447047 0.696670768696 0.68471544573 0.68346608484 0.679943936084 0.671205678618 Rank 1 2 3 4 5 6 7 8 9 10 The 31st question of the surveyed data set was the parent’s income on a rated scale of 1 to 4. Each student was made to answer this question, this helped to deduce a basic relation between available resources and methods of learning acquired during the early growth stages of life. An assumption was made that resources for some categories such as graphical visual techniques or audio books etc require more resources over some other categories which required relatively lower resources such as read aloud or collaborative learning. The scatter plots which relates category ID to mean score of all students relative to each category ID were plotted. Fig. 5 shows that the interpersonal jigsaw technique is the most preferred by the students lying in parent’s income category 1. Since, these students have scarce availability of resources and knowledge acquisition sources through other means, therefore, these students are highly dependent on the classroom study of the Indian education system. Hence, they have evolved to accept this as a common learning scheme. Linguistic cramming is the most preferred learning method by the students lying in parent’s income category 2 as indicated by fig. 6. They have the availability of sufficient reading material and text books, and since, books are an easy source of knowledge, the thorough reading and cramming is the common practice in this category. were made for both popularity and personalized models and comparison of results is as shown in table 5 Fig. 8: Scatter plot corresponding to parents’ income category = 4 . The test with conventional learning scheme is categorized as A and test with recommended learning scheme is categorized as B. The maximum mark was taken as 50. Fig. 5: Scatter plot corresponding to parents’ income category = 1 Fig. 6: Scatter plot corresponding to parents’ income id category = 2 Fig. 7: Scatter plot corresponding to parents’ income category = 3 No. of student 1 2 3 4 5 6 7 8 9 10 Total Table 5: Comparison of test results Popularity Model Personalized A B A 40 44 40 26 27 26 45 44 45 45 47 45 34 39 34 29 41 29 17 16 17 22 19 22 31 38 31 42 44 42 331 359 331 Model B 45 28 43 49 49 45 20 17 45 50 391 These students were given a subjective test and their results were evaluated by a teacher, and then the same students were given access to our software which made them suitable recommendations. Then another test of similar level and subject was given to them and results obtained are compared. It is clear that in both the models the results of the students have improved using recommended learning scheme. Fig. 7 shows that the music with ideas is the most preferred by the students lying in the parent’s income category 3. Access to audio and visual techniques due to the availability of smart phones, internet, and computers prompts them to use videos and audio techniques and hence they use the musical: music with ideas techniques for their learning. Experimental learning is the best suited for the students lying in this parent’s income category 4 as shown in fig. 8. Learning in the private schools and availability of enough resources in the schools and their home allows them to learn through equipment and physical activities rather than mere text books. Hence, this learning for the category 4 is completely justified. An examination was conducted on 10 samples from the sample space both before and after the recommendations Fig. 9: Efficiency of popularity model The difference between the conventional and the recommended results were also plotted as shown in fig. 9 and fig. 10. In Popularity model there was 8.4592 percent increase in the overall result of the students using as shown by fig. 9. In the Personalized model there was 18.1268 percent increase in the overall result of the students using the results of personalized model as evident in fig. 10.       Fig. 10: Efficiency of the personalized model As it is seen from the graphs, the recommended models undoubtedly work much better than the conventional models. Moreover, the personalized model works better than the popularity model in the Indian education system. VI. CONCLUSION Indian education system is very complex due to large variation in terms of social, economic and geographical conditions. Students may develop a learning pattern which may not be best method for them. In this paper it is proposed to develop a recommender system based on machine learning which can recommend them better learning schemes. Data set of 3000 students was collected and experiments were performed on sample data sets. Results indicate that Indian education system is more inclined to the Jigsaw technique which is related to the observable classroom based study techniques where topic wise syllabus is covered in groups of mixed students. The recommender system produced multiple suggestions for the students. Students who appeared for the tests chose themselves, the most suited methodology, among the ones available. The algorithms help filter inclinations of cognitive senses due to environmental factors and embrace the strengthened cognitive ability of the student. The tests results shows better performance via popularity model of 8.4592 and 18.1268 via personalized model. Further a relation between financial status and preferred method of learning was observed. It may be concluded that the classrooms can be divided based on the recommendations of the algorithm presented in this paper to yield better results and optimize the learning time. Instead of dividing the students into class sections in random order or by merit (as traditionally done in our Indian education system), students can be grouped by learning abilities and have a teacher specialized in that particular learning technique using the required tools to teach them. VII.  REFERENCES G. Walsh and J. Golbeck, “Curator: A game with a purpose for       collection recommendation,” in ACM CHI ’10, pp. 2079–2082, ACM, ACM ID: 1753643, 2010. Cosley, S. K. Lam, I. Albert, J. A. Konstan, and J. Riedl, “Is seeing believing? How recommender system interfaces affect users’ opinions,” in Proceedings of the SIGCHI conference on Human factors in computing systems, 2003, pp. 585– 592. Oard and J. Kim, “Implicit feedback for recommender systems,” in AAAI Workshop on Recommender Systems, Madison, Wisconsin, 1998. L. Pizzato, T. Rej, K. Yacef, I. Koprinska, and J. Kay, “Finding someone you will like and who won’t reject you,” User Modeling, Adaption and Personalization, pp. 269–280, 2011. https://link.springer.com/article/10.1007/s00799-015-0156-0 JoeranBeel, BelaGipp, Stefan Langer, and CorinnaBreitinger, “Research Paper Recommender Systems: A Literature Survey”, International Journal on Digital Libraries · July 2015, DOI: 10.1007/s00799-015-01560 http://www.cs.carleton.edu/cs_comps/0607/recommend/recomm ender/itembased.html Ghuli, Poonam, AtanuGhosh, and RajashreeShettar. "A collaborative filtering recommendation engine in a distributed environment", 2014 International Conference on Contemporary Computing and Informatics (IC3I), 2014. Verbert, Katrien, HendrikDrachsler, Nikos Manouselis, Martin Wolpers, RiinaVuorikari, and Erik Duval. "Dataset-driven research for improving recommender systems for learning", Proceedings of the 1st International Conference on Learning Analytics and Knowledge - LAK 11 LAK 11, 2011. Song, Guohui, Shutao Sun, and Wen Fan, "Applying User Interest on Item-based Recommender System", 2012 Fifth International Joint Conference on Computational Sciences and Optimization, 2012. Robert J. Durkin, AakashVerma. "Experiential Learning in Engineering Technology: A Case Study on Problem Solving in Project-Based Learning at the Undergraduate Level", Journal of Engineering Technology, 2016 S. Lam, D. Frankowski, and J. Riedl, “Do you trust your recommendations? An exploration of security and privacy issues in recommender systems,” Emerging Trends in Information and Communication Security, pp. 14–29, 2006. Gardner, H. (2011). The Theory of Multiple Intelligences: The Battle-Scarred Journey (An excerpt from The theory of multiple intelligences: As psychology, as education, as social science. Address delivered at José Cela University on October 29, 2011. Madrid, Spain.) The Daily Riff.