Analyzing the Rise of TikTok as a News Platform| Informative Essay

The Rise of TikTok as a News Platform

By [Student’s Name]

Course Code and Name

Professor’s Name

University

City and State

Introduction:

Since its release in 2017, TikTok has become an important tool for news organizations–including the Times of Israel. During the July 2018 Gaza border protests, Palestinian protesters streamed their activity on TikTok. This has the potential to inform news organizations on what is happening and where events are going. The app, which is available in over 50 languages and is widely used in the U.S., has created a new genre of activism where protests, hashtags, and the app itself have become the main ways of disseminating information.

The first protest was organized by a group called the Bekaa Youth Movement, an organization of the Shiite minority in Lebanon that works to protect the rights of Christians, Druze, and Sunnis. While most protesters in Lebanon are Sunni, Christian protesters make up the majority of the Bekaa valley. On Monday, July 17, in the village of Khiam, a group of protesters gathered and called for three things: a complete end to the Israeli occupation, the release of the 12 Lebanese who were kidnapped by Israel, and the formation of a local council that is run by and for the people. The protest came on the heels of a similar sit-in by the group on July 4, 2018, in Jounieh, Lebanon.

Here we are a total of the team of 3 people where we have divided all the respective tasks equally throughout the complete project and model building

Literature review:

This report provides a systematic review of the literature on various MSs. commonly used

Here, applications, recommendation methods, and machine learning algorithms are described. these RSS provides personalized recommendations to online users.

Time and effort to select a product or service according to one’s own taste. This section addresses the first research question and presents relevant research in this area.

This method can measure the number of users and the number of items independently. when to use Element-based collaborative filtering compares what users have bought and rated.

on the related topic. A similar item is then added to the list of recommendations. According to the study, this The main problem with element-based RSS is the impact of context awareness, loss of neighbors, metastatic and rarity.

Approach to solving the problem:

Deep Learning is a subfield of machine learning that relies on creating artificial neural networks to make predictions. It is incredibly powerful and is used in everything from auto-subtitling movies to powering self-driving cars. The structure of the neural networks also makes it especially good at recognizing complex features like faces in images. For this project, I chose Deep Learning to base my models on because TikTok videos can be broken down into component images and these images can be used to train a powerful Deep Learning model.

Content-based filtering takes into account your ratings and provides future recommendations when browsing Internet services. Users tend to rate what they like or dislike. His rating reflects his reaction to this item. If he likes any subject, he gives a high score, and a low score means he doesn’t care much. These rated elements act as “content” in content-based filtering. Based on this content, users make recommendations for future items that they can approve. Here, users recommend recommended movies related to specific genres that they like.

Team responsibility distribution:

1. The one person should study in getting to know about the how background and literature survey work that how actually these sought of thing works and likewise
2. And the next things come with data collection and analysis and performing eda on it where another person will full perform work related to it for building the respective model and then will continue with the data labelling work
3. And coming to the last person where he will run the through a base model and then train the images with model and finally will save the respective model in h5 file

Project phases and expected time frames:

So, in the complete project where we divide into 2 different phases respective with 3 different persons

On toal it has been taken total 9 hours to complete the work

Phase 1: 2 and half hours of work

Phase 2: 3 and half hours of work

Phase 3: 3 hours of work

TikTok scraper:

TikTok does not currently have an accessible API, so an open-source scraper was used for this project. I choose to use TikTok-scraper user drawrowfly, but you are welcome to use any of the TikTok scrapers available to you, you only need to adjust that part in the shell script accordingly. This TikTok scraper also does not need a login, so you can be sure you are not scraping a custom trending page but rather the generic trending page for all users on the app

Data_deeplearning

This is a dataset of 700+ images and a csv that labels all the images. I personally hand-labeled these images from tiktok videos with Trump, random trending tiktok videos, and Google images. You will need to upload this to Google Drive first and then mount the colab notebook onto it.

A colab notebook was used to generate the model. This file has a jupyter notebook version that has been added to this repo. To build the model, first, save the Data deeplearning repository to your Google Drive, then open the notebook in Google Colab (which is free) and run the cells. This is due to the fact that you will have access to free GPUs, which will allow you to complete your training much more quickly. This notebook will then generate a h5 file of the completed, trained model, which you can save to your computer. There are four broad steps in the jupyter notebook that will lead you through training the model.

Labeling the data: For this research, we’re using photos from TikTok videos as well as some photographs from Google Images. I choose to hand-label tiktok video clips. This is because deep learning networks are particularly sensitive to picture attributes, and I wanted to train the model on TikTok-specific photos because I was constructing a model for TikTok. (For example, most of the time Trump appears on TikTok, it’s a recording of a television or computer screen, as people rarely submit first-hand experiences of witnessing Trump on the platform.) As a result, the labelling process takes the longest in this model to complete, As a result, the labelling process takes the longest in this model to complete. For this project, I manually classified 700+ photos and then enhanced them with data augmentation (detailed in the jupyter notebook). Simply replace this step with photographs of the desired feature you wish to recognise and label accordingly if you want to substitute a different image for the model to recognise besides Trump.

Running the photos through a base model: I chose to run the images through the VGG16 basis model first in this experiment. VGG16 is a trained model that translates (224,224) RGB pictures into features. It’s pre-trained on millions of photos from ImageNet and comes out of the box with the keras library.

Model development: Then, using many Dense layers and dropout layers, we build a sequential model (to prevent overfitting). We train this sequential model with our data-augmented photos. You only need to make sure the input and output layers are consistent as stated in the notebook for this phase; the remainder of the model can be customised as desired.

The model is then saved as a h5 file and downloaded to our local machine. The model should then be placed in the same directory as the github repository. When your shell script runs, it will call the model and expect it to be in the same parallel directory as the model.

Implementation of model:

The model’s prediction on the video is generated by this python file. This file will load the pretrained model from the jupyter notebook using the h5 file, and then it will take an input video filename (provided by the shell script) and convert it to images at a frame rate of 5 frames per second. It will then predict each of those photos by first extracting features from the image using the VGG16 model from the Keras library, and then running those features through our pretrained model. We output that Trump has been spotted in a video if any 1 label is anticipated.

RSS (Recommender System) is being used in various fields such as e-commerce, tourism, healthcare, and e-learning. Such a system is widely used to recommend products to users in the field of e-commerce their interest. Products are recommended based on the customer’s purchase history. based Analysis and prediction based on user preferences are also possible. This technique can be viewed as part of: Because personalization helps each customer buy the product they care about. This system can be installed If you have a good relationship with your customers and your recommendations are relevant, you can increase your sales.

Remaining all the work, outputs and implantation is in the respective .ipynb file

summary notes:

On the completing of the hoel project, we get to know about a lot of interesting things and also planned some future work to be done such as:

One of the project’s major flaws right now is that the TikTok scraper isn’t always reliable, which is likely due to TikTok’s own anti-scraping methods on its app. As a result, you may need to run the run.sh script several times before a trending video download. However, because that component is somewhat separated from the rest of the project, it may be quickly replaced or enhanced without requiring a complete reworking of the project’s architecture.

From the evaluation of RSS, it’s far discovered that there may be an exponential increase of studies on this field. Collaborative filtering is the maximum famous method used for advice structures however the major downside with this method is its incapability to remedy the bloodless begin problem. The usually used method subsequent to CF is content-primarily based totally filtering, however, this method additionally suffers from the identical problem.

Hence the hybrid method, combining the alternative strategies appear to provide a higher-end result and it’s far from turning into the maximum famous method followed in RSS. With the ever-growing online utilization in all elements of human lifestyles because of the existing pandemic situation, novel synthetic intelligence-primarily based totally structures are evolving on this area. In this paper, we’ve mentioned the present era of RSS. There are feasible extensions to those algorithms that may result in greater fruitful studies on this area

Another drawback is that this model can only be trained to recognise a single object, such as a single person’s face or a single sign. As a result, if you wanted it to recognise more presidential candidates, you’d have to relabel all of the photos and restructure the model, because the output layer would need to be larger to accommodate the new labels.