
Inspiration:
Going into SelfieHacks II, we had no concepts on what to make, however we knew that we wished to create a challenge that empowered content material creators. Throughout our brainstorming session, we puzzled “What’s one thing that each one content material creators battle with?,” the place we got here up with the thought to assist content material creators to develop their communities. We then narrowed the scope to serving to YouTube creators, and serving to their content material attain a wider viewers.
If we wished to trace how many individuals are actively partaking with a channel, the most effective indicators is the view rely of their movies. As views and subscribers are primary indicators of the success of a video or channel, we wished to make a device that will increase these numbers. This then will increase the publicity of their movies to new customers, permitting the channel to develop. Considering on this means, we lastly got here up with our challenge thought, which we known as YouTube Creator Assistant.
Breaking down the issue:
We began with wanting on the YouTube homepage and figuring out what parts would persuade a person to click on on a sure video over one other, such because the title and thumbnail. In our program, we wished to take these elements of the video to generate a predicted view rely. The person can check varied combos of elements comparable to thumbnails, titles, video period and classes to maximise the variety of views. Whereas modifying could be completed with trial and error after the video has been printed, our answer permits it to be completed beforehand. Views can then be gained extra simply throughout the time that’s most vital: proper originally.
How did we implement it?:

To create our mannequin, we first seemed on the kinds of information we had, which included pictures for the thumbnails, language information for the titles, and tabular information for the remainder of the data. For the thumbnails, we used a convolutional neural community (CNN) to establish eye-catching parts of the picture (AKA clickbait). Subsequent for the titles, we extracted options that we thought have been helpful, such because the size and the proportion of capital letters. Lastly for the tabular information, we used a totally linked neural community to foretell how every variable pertains to the ensuing variety of viewers. Then, we mixed the outputs of the 2 networks, giving us the expected viewer rely.

eng = matlab.engine.start_matlab()
We have been then in a position to take inputs from our Gradio internet app in Python, feed them into our MATLAB mannequin, and output the expected view rely as a Python integer.
Outcomes:
We examined our mannequin by making a pretend thumbnail and filling in some particulars about our hypothetical video and channel. We then tried altering the thumbnail and title to at least one that we thought would entice extra viewers and as anticipated, the expected variety of views elevated!
Total, our completed mannequin carried out properly on new enter++s, the place a extra “clickbait-y” thumbnail or title is predicted to have a better variety of views. Regardless of this, we discovered throughout testing that it has some difficulties with outputting an correct prediction of the viewer rely for
+ channels with a small variety of subscribers. That is high-quality for the meant objective of the mannequin, however we really feel that it will profit from some extra information, as the present information solely options probably the most trending and in style creators. Sooner or later, we plan to present the mannequin extra information particularly containing YouTube channels with fewer subscribers, in order that the mannequin can higher establish how a particular function impacts the ensuing viewer rely. Watch this video to see how our code works

Key Takeaways:
In comparison with Python, we discovered that MATLAB was simpler to make use of for prototyping, as there have been many built-in features to make coding fast and straightforward. The massive quantity of documentation diminished the issue of making an attempt new issues, permitting us to discover extra of MATLAB’s many options. YouTube Creator Assistant was a enjoyable challenge to work on and we realized a ton about MATLAB’s options for information science and machine studying, in addition to its Deep Community Designer and MATLAB Engine.