/milestones/competitions/naisc2024

First slide of my

National AI Student Challenge 2024 (Track 1)

The

is a competition organised by

AISG

for students to showcase their skills in AI. For 2024, there are 3 tracks. Track 1, by the AI Professionals Association requires participants to create a model to predict the adoption of rates of pets listed on

PetFinder.my

, while the other 2 tracks involves developing or tuning an existing large-language model. Due to the eligibility criteria (NSF), I was only able to join the competition under track 1. The challenge consisted of 3 phases. The first phase is the Technical Ability Test that involves participants creating a model based on the problem statement. This is followed by the second phase is the interview, where participants present their findings and models to the judges. Then, if they're lucky enough, participants would compete in the final judging round. I managed to pass all three rounds to clinch 1st prize.

This page is just a summary of my project. For more info, please have a look at the

repo

Me the weekend after IFC.

Link to article

Before I could create my model, I had to conduct Exploratory Data Analysis

(EDA)

and present the data to the judges. There were roughly 15 thousand pets (or data points, in this case) which had to cleaned up for mistakes like typos and duplicates. In addition, I inferred additional data using the descriptions that came with each pet listing with

TextBlob

for sentiment analysis, and

Polyglot

to detect the language. After cleaning the data, I used

ydata-profiling

to generate a simple report for me. However, most of the analysis is contained in the notebook eda.ipynb where I used

Seaborn

and

Pandas

to produce graphs and to ultimate draw the judges to important correlations between certain variables and the time it took for a pet to go from being listed to being adopted. These variables include the pet's age, health and number of photos in the listing.

With a clearer picture of the data, I proceeded to create my prediction models. I opted to make models using both

scikit-learn

and

tensorflow

as I wanted to try both (to be fair, I had loads of time after A-levels). In the end, I had 6 models. For each of scikit-learn and tensorflow, I created the following 3 models, a binary classifier (adopted or not), a speed classifier (how fast a pet was adopted) and a regressor (how fast a pet was adopted, but as a scale). Due to the low correlation between most variables and adoption speed, the models could not learn many patterns from the data, and thus accuracy was relatively low.

For this challenge, it was not sufficient to just have a working prediction model, ethics also seem to play a big part in the overall judging. From the problem statement:

Please note that the development and use of such a prediction model might involve several ethical and safety concerns that may need to be carefully considered so that the system may be trusted.

Hence, I placed a greater emphasis on the ethics of my models. For example, I suggested that my model could highlight such disadvantaged pets that my models predict would not be adopted, and recommend them to users.

While I was very happy with the outcome of the competition, I felt that there could be some improvements to my project. For example, when trying to infer data from the description of each pet listing, I could employ the use of LLMs to pull keywords, the presence or absence of which can be used to train the model. I could even find the actual listing for each pet on PetFinder.my and pull the images and videos included in the listing and train a deep learning model on them to predict adoption outcomes. Moreover, we only got data about the pets. If we had information on the users, we can get an idea of what kind of users are more likely to adopt what kind of pets, making recommendations more personal.

/milestones/competitions/naisc2024

National AI Student Challenge 2024 (Track 1)

Links