This is my first message here, and I would like to seek some assistance !
I have a technical test for a job that I really want, and I have 10 days to complete it. I've attempted to work on it, but I'm struggling to achieve the desired accuracy. The interviewer mentioned they achieved 0.95 accuracy in literature, but I can't surpass 0.52 on my validation set for a 10-class classification task.
Here's the challenge: I have no information about the origin of the data. There are no names on the features column, and no details on the classes. All I have is a training set of 50,000 examples with 1024 continuous features (already scaled between 0 and 1), and a test set of 10,000 examples.
I've experimented with various models, including xgboost, random forest, SVM, transformer, and more. I even considered the possibility that the data might be images of 32 x 32 pixels in grayscale.
The only hint provided is: "Tips: Tabular data are unstructured by definition. Finding the underlying structure will strongly enhance your results."
Up to now, I've employed data visualization to check for anomalies such as significant outliers or missing data, but I haven't found anything noteworthy. I've also attempted dimensionality reduction techniques like PCA, where 75 components explain 0.9 of the variance. Despite fine-tuning multiple models, my best model (a simple multi-layer perceptron with 75 components PCA) achieves no more than 0.52% accuracy.
I've also experimented with different scaling methods, such as standard scaling or min-max scaling.
I attempted to identify the problematic class, but upon reviewing the classification report, it appears that each class has an approximately 0.5 accuracy.
Does anyone have ideas on what kind of preprocessing steps I could take with the data ?
(I can provide the data and my notebooks !)
Thanks for your help !