Kevin > Recent Status Updates

Choose an aspect ratio that does not exaggerate or minimize the change (in the data).
A general rule-of-thumb*: a 1/x increase or decrease should have a roughly x:1 aspect ratio (width:height)
- *This produces distorted charts when x is relatively large or small

"Chart design is similar to writing"
"Decoding a chart is similar to reading text"

C2: Charts may lie by being poorly designed. When the mapping between values and symbols is not proportional. There can be visual distortions (scale, proportions).
The choice of baseline plays a large role in how a chart is perceived. The data ultimately govern this choice. E.g. Fahrenheit and Celsius don't have a minimum of zero, so a 0° baseline is often a bad choice.
Choose sensible baselines, not possible ones.

This chapter explores three methods useful for recommendation systems:
1) Linear Regression
2) Singular Value Decomposition
3) Principal Component Analysis

SVD decomposes the observation matrix into 3 matrices so that less important features can be identified and dropped from the model.
PCA identifies "latent features" that can be used to create a simpler model. Both involve lots of computation for large datasets.

Vectorized String Operations

Pandas excels at data munging with vectorized operations on strings.

"A chart may lie because:
- It's poorly designed.
- It uses the wrong data.
- It shows an inappropriate amount of data—either too little or too much.
- It conceals or confuses uncertainty.
- It suggests misleading patterns.
- It panders to our expectations or prejudices."

"To read a chart correctly, we must pay attention to it and never take anything for granted."
"...if you just skim them, you won't understand them, and well-designed charts aren't just informative but also graceful..."
Elements:
- Title, intro, caption, source
- Measurements, units, scales, legends
- Visual encoding methods
- Annotations
- Patterns, trends, relationships (using bird's-eye view)

C1:How Charts Work
Charts encode information using: height/length, width, position, size/area, angle, color hue, and color shade.
Effective communication between a designer and an audience requires a shared understanding of what the chart is about and how the data is symbolized. Mental models save us time and effort.

Combining Datasets: Concat and Append

Concatenation of Pandas Series and DataFrames works much the same as NumPy arrays, but indices are preserved in Pandas whereas they are not in NumPy.

Multi-indexing can be useful for computing tables of aggregated statistics. It is very similar to performing the groupby operation.

Time series analysis is tricky and involves making a justifiable assumption about what the time is modeling. This choice that you make for one research question might be different for another question.