DeepInsight

The DeepInsight method transforms non-image tabular data into synthetic images by spatially arranging features based on their similarity. The process begins by applying dimensionality reduction techniques (such as t-SNE, PCA, or Kernel PCA) to project the feature space into a 2-dimensional plane. A convex hull algorithm is then employed to identify the smallest rectangle encompassing the feature points, which is subsequently rotated to align with the image axes. Finally, the feature values are mapped to the corresponding pixel locations within this grid. This structured arrangement ensures that correlated features are placed in proximity, enabling Convolutional Neural Networks (CNNs) to exploit local spatial dependencies effectively. This implementation extends the original method by adding techniques for mapping features to pixels. Furthemore, a new method to compute pixels values when using PCA is added

Import DeepInsight

To import DeepInsight model use:

>>> from TINTOlib.deepInsight import DeepInsight
>>> model = DeepInsight(image_dim=30)

Inherited base functionality

DeepInsight inherits from ParamImageMethod (see Base classes), so it also provides:

• Feature coordinates: a feature-to-pixel mapping exported to features_positions.csv after fit/fit_transform.

• Programmatic access: _get_features_mapping() returns the feature coordinates as a DataFrame once the model is fitted.

• Shared utilities: saveHyperparameters / loadHyperparameters and the standard fit / transform / fit_transform workflow.

Hyperparameters & Configuration

When creating the DeepInsight class, some parameters can be modified. The parameters are:

Parameters	Description	Default value	Valid values
image_dim	Order size for a square matrix image.	Required	integer
problem	The type of problem, defining how the images are grouped.	‘classification’	[‘classification’, ‘unsupervised’, ‘regression’]
transformer	Preprocessing transformations like scaling, normalization, etc.	MinMaxScaler()	Scikit Learn transformers or custom implementation inheriting CustomTransformer.
verbose	Show execution details in the terminal.	False	[True, False]
algorithm_rd	Dimensionality reduction algorithm to determine feature coordinates.	‘PCA’	[‘PCA’, ‘t-SNE’, ‘kPCA’]
assignment_method	Technique to map features to pixels.	‘bin’	str (e.g., ‘bin’)
relocate	Relocate features so that each pixel can represent a single feature (only if algorithm_rd is ‘PCA’).	False	[True, False]
algorithm_opt	Optimization algorithm applied in the pixel assignment stage.	‘linear_sum’	str (e.g., ‘linear_sum’)
group_method	Technique to calculate pixel values sharing multiple features.	‘avg’	‘avg’
zoom	Multiplication factor determining the size of the saved image relative to the original size.	1	integer > 0
format	Output format using images with matplotlib with [0,255] range for pixel or using npy format.	‘png’	[‘png’, ‘npy’]
cmap	Color map to use with matplotlib.	‘gray’	‘viridis’, ‘plasma’, ‘inferno’, ‘magma’, ‘cividis’, ‘Greys’, etc.
random_seed	Seed for reproducibility.	23	integer

Code example:

>>> model = DeepInsight(image_dim=50, algorithm_rd='t-SNE', cmap='viridis', random_seed=42)

All the parameters that aren’t specifically set will have their default values.

Functions

DeepInsight has the following functions:

Function	Description	Output
saveHyperparameters(filename)	Allows to save the defined parameters.	.pkl file with the configuration
loadHyperparameters(filename)	Load DeepInsight configuration previously saved with saveHyperparameters(filename) filename: .pkl file path
fit(data)	Trains the model on the tabular data. This step performs dimensionality reduction to map features to the 2D grid. data: A path to a CSV file or a Pandas DataFrame containing the features and targets. The target column must be the last column.
transform(data, folder)	Generates and saves synthetic images in a specified folder. Requires the model to be fitted first. data: A path to a CSV file or a Pandas DataFrame containing the features and targets. The target column must be the last column. folder: Path to the folder where the synthetic images will be saved.	Folders with synthetic images
fit_transform(data, folder)	Combines the training and image generation steps. Fits the model to the data and generates synthetic images in one step. data: A path to a CSV file or a Pandas DataFrame containing the features and targets. The target column must be the last column. folder: Path to the folder where the synthetic images will be saved.	Folders with synthetic images

• The model must be fitted before using the transform method. If the model isn’t fitted, a RuntimeError will be raised.

Citation

Paper: https://doi.org/10.1038/s41598-019-47765-6

Code Repository: https://github.com/alok-ai-lab/pyDeepInsight