Ismael Chang Ghalimi

A bar chart is not a single type of chart, instead it covers an entire family of charts, each having its own purpose and domain of applicability. - The Bar Plot visualizes raw numerical values. - The Histogram summarizes them to visualize distribution. - The Frequency Chart displays counts of categorical values. Some variables cannot be visualized with bar charts directly, and derivative variables must be used instead. For example, the lengths of texts can be visualized with a Histogram. Other categorical variables might have a very large number of distinct values, making the Frequency Chart impractical. Then, use a Frequency of Frequency Chart instead. It is not easy to interpret, but it is a very powerful tool for Data Quality.

All these pieces of advice are spot on. Great job LinkedIn AI! One more thing to consider is to design every single visual as if it was published on social media. "The medium is the message." —Marshall McLuhan Stakeholders have less and less time, and need every bit of help they can get to obtain the data they need for making the right decisions, as quickly as possible. Doing so will force you to focus on the most important data points, streamline your narrative, simplify your visuals, and highlight the key take-away points. Then, either give all your outputs a social media look-and-feel, or produce two versions of the same visuals, one for conventional media, and another for social media. Some tools can even help automate the process.

An alternative consists in rendering the visuals browser-side. This can be achieved by using a library like D3.js, which is downlaoded once and cached on the browser. Once minified, the latest version of D3.js is only 280 KB in size. If you need something a bit easier to use, Vega takes 511 KB. And Vega-Lite adds another 250 KB. Rendering visuals that way allows you to send only the data, while preserving animations. But be careful, this might actually use more bandwidth for visuals with many marks (shapes), like scatterplots for example. Finally, you could move the database itself on the browser, using in-process engines like DuckDB, then cache small datasets browser-side. This architecture will give you the best level of interactivity.

Try to answer the following questions: - Is the tool easy to learn? - Is the tool as easy to use as Excel for small datasets? - Can the tool handle large datasets (gigabytes to terabytes)? - Does the tool remain affordable with large datasets? - Does the tool offer no-code, low-code, high-code modes? - Does the tool support programming languages like Python? - Does the tool support query languages like SQL? - Does the tool support the most common data formats? - Does the tool support the definition of a data dictionary? - Does the tool support incremental transformations? - Does the tool support parametric transformations? - Does the tool offer data pre-visualization? - Does the tool intergate with source control systems like git?

I recently worked on a project to model the sales of every single Walmart store in the United States from 2013 to 2022. Here is how we found the data that we needed: - List of stores from a consulting firm. - Fiscal calendar from another consulting firm. - Store opening and closing dates from a Walmart fans site. - Missing dates from the online archives of local newspapers. - Country-wide revenues from Walmart press releases. - State-level metrics from Walmart website. - State and county-level metrics from government websites. - Store-level metrics for 45 stores from Kaggle. - Hourly store foot traffic from Google Maps. - Missing store sizes from Google Maps satellite images. - Products and product categories from Walmart online store.

Instead of using dedicated data quality dashboards, I like to embed data quality metrics and visuals directly within my data. For example, I usually add data quality bars at the top of every table column, with one stacked bar per type of value: - Green for valid - Red for invalid - Grey for missing I then display a summary data quality bar at table's bottom. These bars don't use a lot of screen real estate (10px × 140px), and provide a lot of value for very few pixels. For outliers, I like to display them on the summary charts that are added at the top of every table column. This works especially well for the histograms used to summarize numerical columns. Then, I make sure to provide visual explanations for these outliers (e.g. ±2.5σ).