What's the best format for your machine learning data?

Tabular data can contain missing values which have to be dealt with delicacy. The missing values can be missing at random or not thus omitting rows containing missing data or replacing them can be done according to the domain expertise. A simple example of missing values not at random is the field of number of births for male patients which shall always be 0. With enough domain knowledge the AI practitioner or the development team can handle “problems” in the data. In my experience, even the simple task of gathering data can solve some trivial problems. The quality of data is essential for all machine learning methods as they are data-driven approaches. Even a basic yet crucial error in the data can lead to machine learning model failures.

For Tabular data, storing them in Parquet file format is one of the best choice, as Parquet an open-source file format from Apache, is famous for its efficient data storage and retrieval. It employs data compression and encoding schemes to handle complex huge amount of data which is one of the primary need in Machine Learning

Choosing the suitable format for machine learning data is crucial for practical analysis and predictions. Tabular data is commonly used and stored in formats like CSV. Text data requires particular encoding and can be stored in JSON. Image and audio data need to be preprocessed and can be saved in JPEG or WAV formats, respectively. Graph data, helpful in understanding relationships, is stored in formats like GraphML. Your design should align with the data type, algorithm, and business goals, ensuring the data is flexible and efficient for your machine-learning tasks.

Tabular data is any data in the form of a table ( rows X columns). -its most common data form. -whenever you work in Excel, CSV, or databases, you get it -it's easiest to work with while coding (use Numpy for loading )

Tabular data is readily available for consumption because DB are maintained in that structure/ schema s historically. However as we grow the this is evolving and with GenAi technology the formats are also changing which will impact this also.

In the rapidly evolving world of machine learning and artificial intelligence, text data has emerged as a cornerstone of innovation. With the advent of LLMs like GPT series, the importance of text data has been underscored like never before. These models, trained on vast amounts of text, have showcased an unprecedented ability to understand, generate, and interact using natural language. LLMs have billions, or even trillions, of parameters. These parameters are fine-tuned during the training process, to allow the model to generate coherent and contextually relevant text. While there isn't a one-size-fits-all format for ML, the reliance on text data is expected to grow as LLMs become increasingly integrated across various industries.

Text data is any data we use for communication - Emails, reviews, posts, tweets, interactions -Its relatively difficult to load and analyze this data from an ML perspective -you need specialized NLP techniques to work on it -can give you high business returns once deployed

It is imperative to use data augmentation to improve model performance with image data. The nature of imaging techniques are prone to many different distortions such as brightness, blur, rotations and noise. These distortions will always be present in the real case even if they are not observed in the training image dataset. Data augmentation is artificially applying these distortions with static or dynamic variables that determine the probability or intensity of applying distortions. Through my experiments with generative data augmentation, I have observed that employing generative models to synthesize images that resembles training image dataset improved overall performance even with limited data.

Image data is any data with images. - photos, videos ( consider it as moving photos) -it needs complex methods to analyze, store and process -consumes highest amount of compute , reliable accuracy might be difficult depending on the use case

Audio data can be converted to 2 dimensional data (Mel-spectrogram) that can be used with one of the most widely used and proven image processing technique, Convolutional Neural Networks (CNN).

Audio data are the most less used among text and image data but audio data has huge potential like we can build our own ASR,TTS and voice cloning ,Voice Identification which can be used as voice biometric and many other cool use-cases are available .

any data from any sound - can have different formats based on player and recording method - complex data type as we need to split it in different dimensions to make sense of it - specialised computing & resources needed

Apart from the above, we also capture -location data ( Geospatial data) - Satellite data - Graph data ( connection between nodes and joins)

In building the next-gen tool, it's vital to embrace multimodal data for the future. This entails collecting diverse text, audio, image, and video data, centrally storing it, and then combining features from all modalities. A deep learning model, like a multimodal neural network, is trained to understand this rich dataset. It uses user feedback for continuous improvement, providing highly personalized content recommendations based on user preferences and interactions. This tool promises an advanced content discovery experience, emphasizing the importance of multimodal data in the future.

The best data format is the one that is fully understood, including its collection methods, update frequency, and the significance of each field. This is essential for data aggregates and user event data. Complete understanding is achievable through diligent documentation, ensuring its accuracy and regular updates. An example highlighting this need is in datasets with movie IDs liked by users. Assuming these IDs are in the order of user preference can be misleading. If they are not, it could cause significant issues in later model evaluations.