Housematch ft. Restb.ai

Inspo

Finding a good place to live can be a boring and generic task, we intend to solve this problem with the development of an app which provides a high quality, engaging and custom experience for its users.

Our idea is based on obtaining high quality data from rental or on-sale properties using the restb.ai models which provide a large amount of clean and powerful data.

We noticed that swiping apps are trendy, especially amongst youngsters, as they provide a very engaging experience which keeps the user involved and they also can provide valuable data in understanding the user’s preferences.

So, with this in consideration, we came up with the idea to combine the data obtained from the restb.ai models with the data obtained from the user’s interaction in swipes to describe an algorithm that characterizes a learning process. This allows us to understand the user’s taste and preferences in the real estate sector to make high quality predictions which we convert into recommendations.

How it works

A user enters a swiping app where properties with multiple pics appear allowing him to swipe right if he likes it or left if he does not, while swiping our algorithm learns it's taste and allows us to provide predictions on which properties will be a good fit for the user.

Results

We end up obtaining a rich learning process composed by +7 computer vision restb.ai models, a proper machine learning model we created and our custom learning algorithm. For more information about the algorithm feel free to check the github repo linked in this post.

What we developed it with

The frontend was built with react and the backend was built with python.

The learning algorithm:

We built an AI meta-model ending in a boosted-random-forest classifier fed with 14 numerical values computed from the outputs of 7 models of the restb API. The restb API models provided us with categorical data that we converted into numerical inputs using different techniques such as taking significant statistical measures selected after careful reasoning, profiling with normalized frequency vectors and a defined correlation measure and lastly by defining score functions after discussing what could be significant.

While the user swipes the dataset we have for this user grows so we can train the model and it ends up providing better results for the predictions section.

Other

The project was awarded with a selection to the top 5 best projects in HackUPC 2022

Built With

• javascript
• json
• python
• react
• restb-api
• scraping
• sklearn