➩ A function that returns an image response can be used for @app.route("/plot.png") or @app.route("/plot.svg"). (SVG is Scalable Vector Graphics and vector graphics is represented by a list of geometric shapes such as points, lines, curves, Link, PNG is Portable Network Graphics and raster graphics is represented by a matrix of pixel color values, Link)

➩ An image response can be created using flask.Response(image, headers={"Content-Type": "image/png"}) or flask.Response(image, headers={"Content-Type": "image/svg+xml"}), where the image is a bytes object and can be obtained using io.BytesIO.getvalue() where io.BytesIO creates a "fake" file to store the image.

➩ On the web page, display the image as <img src="plot.png"> or <img src="plot.svg">.

➩ One to three dimensional data items can be directly plotted as points (positions) in a 2D or 3D plot.

➩ For data items with more than three dimensions ("dimensions" will be called "features" in machine learning), visualizing them effectively could help with exploring patterns in the dataset.

➩ Summary statistics such as mean, variance, covariance etc might not be sufficient to find patterns in the dataset.

➩ For item dimensions that are discrete (categorical), plotting them as positions may not be ideal since positions imply ordering, and the categories may not have an ordering.

➩ Position, size, shape (style), value (light to dark), color (hue), orientation, and texture can be used to present data points in different dimensions. These are called visual encodings.

➩ Some of these encodings are better for different types of features (data dimensions).

➩ seaborn is one of the data visualization libraries that can make plots for exploring the datasets with a few dimensions (features): Link

➩ Suppose the columns are indexed by c1, c2, ..., then seaborn.relplot(data = ..., x = "c1", y = "c2", hue = "c3", size = "c4", style = "c5") visualizes the relationship between the columns by encoding c1 by x-position, c2 by y-position, c3 by color hue if the feature is discrete, and by color value if it is continuous, c4 by size, c5 by shape (for example, o's and x's for points, solid and dotted for lines) if the feature is discrete.

➩ For discrete dimensions with a small numbers of categories, multiple plots can be made, one for each category.

➩ seaborn.relplot(data = ..., ..., col = "c6", row = "c7") produces multiple columns of plots one for each category of c6, and multiple rows of plots one for each category of c7.

➩ seaborn.pairplot produces a scatter plot for each pair of columns (features) which could be useful for exploring relationships between pairs of continuous features too.

➩ If there are large numbers of dimensions and data points, plotting them directly is inappropriate.

➩ To figure out the most important dimensions, which are not necessarily one of the original dimensions, unsupervised machine learning techniques can be used.

➩ One example of such dimensionality reduction algorithms is called Principal Component Analysis (PCA): Link, Link.