To 100 (Non-existed) Boys I Dated Before
A Tinder clone contains machine-learning generated dating profiles and stories that I 'encountered' before
- Duration: 2 Weeks
- My role: Web designer and developer, Data collection, Ideation
- Skill set: #React #MaterialUI #Front-end #RunwayML
A Tinder clone tells my frustrating dating experiences with 100 non-existed men whose machine-learning generated dating profiles and our dating stories.
I started this project right after I ended my last long-term relationship. I was back on dating apps and had the familiar experience of swiping people, then got bored with it. I wondered how everyone could impress people with their profiles but had weird or boring conversations after getting matched. So we decided to do a project related to this experience and make fun of it with machine learning.
'To 100 (non-existed) Boys I've Dated Before' is a tinder clone that contains 100 profiles that are machine-learning generated, including the profile pictures and the bio. The user can swipe left or right on the website. If swiping right, there is a pop-up window that tells the bizarre dating story we had. The story is based on the training data and generated from context-free grammar.
The project process contains three parts: data collection, data training, and front-end development. It is a collaborative project with Tianxu Zhou.
- Profile photos are from StyleGan2 for the FFHQ dataset, one of NVIDIA Research Projects.
- Names are aggregated from the social security administration.
- Dating Bio is trained from Ok-Cupid profiles datasets from Kaggle, where contains around 60k users' dating profile information. From the open-ended questions, we picked the three most related ones, bio, fun fact, and "what are you looking for."
- Dating stories are trained from the 'what are you looking for' and fun fact categories of the Ok-Cupid profiles dataset. We put the training result into Tracey, a context-free grammar, and generated the story with particular grammar.
After exporting the data to text files, we used RunwayML for training the data. For training restraints, we used an initial medium-size model, line separator, and 1000 training steps. We had three different dataset models by the end and ready for organizing into a JSON file that we can fetch into our front-end.
The tech stack that I used for front-end development is 'React,' 'React-Router,' and 'Material-UI.'
Core interaction is built with React hooks ('useState' and 'useCallback'). Here is the code snippet of adding and removing users when swiping.
I used 'react-tinder-card' for building the tinder swiping animation and the interaction.
The interface is also mobile responsive, here is the mobile version:
You can find the code repo here.