KMITL

Permanent URI for this communityhttps://dspace.kmitl.ac.th/handle/123456789/1

Browse

Settings

Subject: search.filters.subject.Cosine similarity

Search Results

Now showing 1 - 3 of 3
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Cyberbullying detection on Tweets
    (2021-05-19)
    Phanomtip, Aekachai
    ;
    Sueb-In, Thaiyathorn
    ;
    Vittayakorn, Sirion
    Nowadays the young generation spends a median of 3 hours a day on social media sites such as Twitter, Facebook, Instagram, Snapchat or TikTok. As a consequence, these platforms have gradually become part of their daily life as well as an influence on their attitude and behaviour. Although the social media platforms were originally designed to share information/news or connect families and friends, they also spread fake news, conspiracy theories, hate speech and cyberbullying. Cyberbullying is bullying with the use of digital technologies which is repeated, aimed at scaring, angering or shaming those who are targeted. Feeling vulnerable, powerless, humiliated, isolated, depressed or suicidal are examples of the negative effects from cyberbullying. Unfortunately, study has shown that one-third of the students have experienced cyberbullying in their lifetime. To mitigate this problem, we propose 1) a novel dataset of 67K tweets collected from Twitter, 2) automatic annotation methods for a large-scale dataset and 3) a detection system that to identify these toxic behaviours. The experimental results demonstrate that our method outperforms the baseline by 4%.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Comparison of gesture in Thai boxing framework using angular dynamic time warping
    (2019-07-01)
    Chantaprasert, Benjarat
    ;
    Chumchuen, Phadermpong
    ;
    Wangsiripitak, Somkiat
    We developed an algorithm that compared the movements of a Thai boxing trainer and those of a trainee and produced a video clip that a trainer can use for training novice boxers in a training camp or that a trainee could use for self-training. In the developed system, a Microsoft Kinect sensor was used to capture 3D joint positions of the body of a trainer or trainee. A cosine similarity and an angular dynamic time warping were used to determine a similarity score between the postures or joint positions of the trainer and the trainee for static and dynamic poses respectively. Based on these similarity scores, suggestions to trainees on how to improve their boxing postures were overlaid on the recorded video clip, helping the trainee to improve his/her posture and movement. The proposed algorithm and developed framework could be easily used for other training such as yoga, dance, cardio workouts because the system allows to track and compare the whole body of users.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    A novel technique for feature subset selection based on Cosine similarity
    (2012-11-27)
    Suebsing, Anirut
    ;
    Hiransakolwong, Nualsawat
    Nowadays, data mining has been playing an important role in the various disciplines of sciences and technologies. Data mining is composed of many tasks but one of the essential procedures of data mining is feature selection, which is the technique mostly based on the machine learning for selecting a subset consisted of significant features, building a stronger learning model, and enhancing the efficiency of prediction rate. Normally, a processing of building a learning model from the huge amount of collected data needs high computation cost. Therefore, with feature selection, the computation cost can be reduced by selecting relevant features. In the previous researches on feature selection, the criteria and algorithms for selecting the features from the raw data are mostly complicated and difficult to implement. Therefore, this paper presents a novel method by applied Cosine similarity to feature selection method. The proposed algorithm begins with selecting a robust feature subset using the Cosine similarity. This method is the simple algorithm using smallerstorage space, reducing computation time and gaining higher predictive performance. During the evaluation phase, the tendata sets from UCI benchmark data sets are used to evaluate the performance of proposed approach by using the C5.0, CARTand Neural Networks classifiers. Experimental results show that the method based on the Cosine similarity can improve the performance of accuracy detection rate with less error rate.