Overview and Motivation

---

Project Information

Title: Dining with Remy: What is the nastiest restaurant in NYC?

Group Members:

• Hannah Anderson, hannahanderson@wustl.edu, 464841
• Nathan Springman, nspringman@wustl.edu, 464394

Repository Link: Click Here

Background and Motivation

New York City is known as the city that never sleeps, and its vast restaurant offerings support this. Many WashU students like us visit or get jobs in NYC and going out to eat is a big part of socializing. However, with so many restaurants to choose from, it's important to choose one where you can feel comfortable with the food you're eating.

When you go out to eat, you don’t want rats in the kitchen. A clean restaurant is important both for public health, but also for building customer trust. Our visualization would serve as a tool for both tourists and locals to find a restaurant that meets their expectations of cleanliness. While our tool is marketed as ‘find the worst restaurant’ to make it more engaging and memorable, the ultimate goal is for a user to explore the dataset and find a restaurant in which they would be comfortable dining.

Related Work

---

Our project, in part, was inspired by reports that rat sightings in NYC restaurants are up during the pandemic. The New York Times covered this phemonenon this November, which you can read about here.

The name of the project, Dining with Remy, is a silly reference to Pixar's Ratatouille, in which a rat is the chef for a fancy restaurant.

Questions

---

Initial

Below is a table of our initial questions we wanted to answer with our visualization and the benefits that answering them can provide.

#	Question	Benefit(s)
1	What are the most safe and unsafe places to eat in NYC?	Knowing a restaurant's food/building safety history can help lower one's risk of food poisoning or other health issues.
2	What are the most common health violations in NYC restaurants? Are they critical?	Even A-grade restaurants can be cited for violations, so contextualizing those violations as critical/noncritical and comparing them to how restaurants perform as a whole helps people realize how important/nonimportant different violations are.
3	Do certain boroughs or cuisine types experience more violations? Is there a specific type of violation they experience more of?	If one borough experiences more of a certain type of violation than others, we may be able to hypothesize about how its land/infrastructure may be causing a restaurant to deal with that violation. (e.g. If one borough of the city has more rat infestations than others, perhaps it has less alleyways to safely store food waste and is attracting rats.) The same goes for certain kinds of cuisines; We could hypothesize that rodents and bugs are more attracted to certain kinds of food.

Data

---

Source & Access Method

We used the Department of Health and Mental Hygiene (DOHMH) New York City Restaurant Inspection Results dataset, which we accessed through NYC OpenData by downloading a CSV file.

For each restaurant, the dataset has up to 26 features; Not every entry has all of the features due to the unreliable nature of data entry from different inspectors. However, we will only be using entries that have the following features:

• CAMIS: A unique number identifier for a given restaurant.
• DBA: The name that the restaurant does business as. (The public name can change easily.)
• BORO: The borough the restaurant is located in.
• BUILDING: The building number of the restaurant.
• STREET: The street of the restaurant.
• ZIPCODE: The zipcode of the restaurant.
• PHONE: The phone number of the restaurant.
• CUISINE DESCRIPTION: The type of food that the restaurant serves.
• INSPECTION DATE: The date of the last inspection.
• ACTION: The action associated with the inspection.
• VIOLATION CODE: The code of a restaurant's violation.
• VIOLATION DESCRIPTION: A description of what the violation code refers to.
• CRITICAL FLAG: Whether a violation was critical or not; Critical violations are most likely to contribute to food-borne illness.
• SCORE: The restaurant's inspection score.
• GRADE: The restaurant's inspection grade.
• LATITUDE: The restaurant's latitude.
• LONGITUDE: The restaurant's longitude.

Pre-Processing & Clean-Up

The original csv file is well-formatted, so we didn't have to do substantial clean-up; We just filtered the kind of data we're including and then modified some of the entries that weren't complete.

We pre-processed the data using Pandas, a Python-based data analysis tool, in Jupyter Notebook. We documented the detailed clean-up process in this notebook, which can be seen in our repository here. To summarize, we mainly did the following to the dataset:

• Removed incomplete/unreliable entries.
• Calculated grades for scored restaurants missing a grade based on the DOHMH grading system.
• Removed aspects of the dataset we weren't interested in.

Additional Processing and Data Segmentation

There are too many restaurants and entires in the dataset to display at once, so we segmented the data by zip code.

When zoomed out, we want there to be a heatmap divided by zip code. In data/zip_processing.ipynb, we import a geojson, sourced here, and calculate an average grade for all the restaurants in that area.

Once zoomed in, the code will bring in data for zip codes that are visible. These are generated in data/zipcodeSorting.ipynb. The original dataset has entires by individual violations, so in these files, we have manipulated the data to have the restaurant as an entry with an array of violations.

In addition to segmenting the data by zipcode, we also wanted to visualize trends for specific violation types, so we segmented by violation code. This data is generated in data/violationSummaryData.ipynb and accessed via data/vioSummaryData.json.

Exploratory Data Analysis

---

To explore the data at a high level, we used NYC Open Data's own visualization tool. This tool was helpful because it allowed us to examine/identify potential areas of interest for our visualization. As a disclaimer, the online visualization tool uses the entire dataset, whereas we cleaned the dataset for our final visualization; Thus, some of the trends we visualize might be slightly different from these visualizations. Below are some of the charts we looked at.

Inspections by Inspection Date

One chart we made on the tool that was interesting was the above "Inspections by Inspection Date" line chart which visualizes how inspection counts varied across time. (We filtered it to exclude values before 2015, since 1900 was used as a placeholder in the dataset for restaurants that hadn't been inspected yet, and the graph has been panned to focus on mid 2018 to nearly the present.)

This line chart is interesting because it shows the sheer impact that the COVID-19 pandemic had on restaurant health inspections in NYC. The pandemic hit the U.S., especially NYC, hard in March 2020, which is where the line graph abruptly stops. There is some inspection data from later in 2020 (less than 10 inspections), but things didn't start to pick back up until summer 2021. Summer in general seems to be a big time of year for these inspections, since the data peaks for each year in those months.

We decided not to visualize inspection date trends, but this graph definitely was notable because of the pandemic.

Violations by Violation Type

We also made a "Violations by Type" bar chart using the violation codes in the data to see if there were any high-level trends. (The screenshot above just shows the most common codes.)

10F was the most common violation code, which represents "Non-food contact surface improperly constructed. Unacceptable material used. Non-food contact surface or equipment improperly maintained and/or not properly sealed, raised, spaced or movable to allow accessibility for cleaning on all sides, above and underneath the unit." This is a non-critical violation, so not super concerning.

However, it is extremely concerning that the 3rd most common violation is 04L, which stands for "Evidence of mice or live mice present in facility's food and/or non-food areas." These restaurants are quite literally Dining with Remy!

This visualization across violation codes showed us that it would be cool to contextualize restaurant violations based on how common they are in some way for our visualization.

Violations by Cuisine Type

We also visualized violation frequency by cuisine type. (As the sidebar in the screenshot shows, we were filtering by restaurants that had a specific violation code listed and an action that indicated violations were recorded at the time of the inspection.)

As the chart shows, there are definitely differences across cuisines types regarding violation frequency. However, much of this is just based on the fact that those cuisine types have more restaurants to be inspected to begin with. (We also looked at a graph that showed the total # of inspections for each cuisine type, and it was almost identical.) The takeaway from this graph isn't that American, Chinese, and Pizza restaurants have more violations for a reason; rather, we would say the takeaway is that no matter the cuisine type, restaurant violations in general seem to be common.

Though this chart didn't really show anything interesting, we still wanted to examine cuisine type in our visualization and filter the data down more by zipcode and violation code to see if there are any insights there.

Design Evolution

---

Initial Sketches

Revised Sketches

"Final" Design

We primarily want this site to make restaurant-goers more informed about restaurant cleanliness. As such, our primary navigation will be location-based. Drawing inspiration from sites such as AirBnB or Zillow, we want to have a split view with a map on one side and a table view on the other.

When zoomed out, the map will be a heatmap showing concentrations of violations. One can zoom in or use the filters on the table view to see individual restaurant pins. These pins will be color-coded to reflect their individual health scores and signal if serious violations have recently occurred.

Once someone has clicked on a restaurant, they can see more information about it including a history of violations. There, one can also see the context for the individual violation through an isotype graph. We will list the raw data below, stating, for example, "3,400 out of 26,000 restaurants have had this violation in the past 6 months." However, these numbers are somewhat difficult to comprehend. The isotype graph included will give an easier method of understanding "3 out of 10 restaurants." One will also be able to compare this to restaurants in different groupings: all restaurants, restuarants nearby, same cuisine, etc.. Although we have several years of data (back to 2014,) we don't believe serious violations from a long time ago are as important, and we will likely limit these sorts of comparisons to one year.

User Studies

Our "final" design continued to evolve to the exact end product available today through user studies conducted during the course. Our user study period had four stages:

1. Think-Aloud

During this stage, we gave users the opportunity to interact with and make sense of our visualization. We asked them to verbalize what they were doing and why they were doing it as they interacted.

Our main takeaway from the users during this section is that they wanted to be able to see more information/trends about the zipcodes while interacting with the heatmap layer. There was the assumption that hovering or clicking on a zipcode would produce some kind of response.

Users also seemed frustrated with how many clicks/zooms it took to get to the detailed circle layer.

We observed as well that even though we provide a modal of how to use the visualization and what everything means, users clicked out of it very quickly and instead just wanted to play with the visualization.

2. Task Completion

Next, we asked users to answer a series of questions using the visualization to see how well it accomplished our goals. Here is what we asked them:

1. Choose a random zipcode. Which restaurant would you most likely want to eat at? Why?
2. What violations were the most worrying to you from the ones you observed?

The users we worked with were able to easily click around the circle map of restaurants once zoomed in. Whatever zipcode they chose, they gravitated toward looking at different circles that were green (A grades), and then clicking around to examine their cuisine types and violations.

Users were most worried about vermin-related violations or violations related to food sanitation, e.g. the critical violations.

3. Feedback/Critique

Users were then given the opportunitiy to share ideas they had to improve our design.

• One user suggested we change the default heatmap focus to show all of the boroughs initially instead of focusing in on a specific area of the heatmap already, to give users a broader overview when they first load the page.
• Users critiqued how many clicks it took to zoom to the detailed map view, which was helpful because we assumed people would pan (which is much easier). One user suggested adding a button that zooms into the circle view.
• One user suggested that we add a tooltip with the restaurant name or label the circles in some way to make restaurants easier to identify on the map.
• Since there are so many green "Grade A" restaurants, users mentioned wanting them to be encoded differently by some other feature, like number of violations, to give more perspective.
• Users also wanted us to make it more obvious on both the table and the map when a restaurant is selected.
• Users said it was unclear that our table and violation sections could scroll.
• One user was confused that some restaurants only had data from 2019, even though restaurant inspections happen annually.(We originally thought much of that would be due to the pandemic, but we think it's also due to some of the data cleaning we did with missing values.)
• Users said the violation part was kind of small and didn't capture attention well.

4. Debrief

After the studies, we implemented the following changes...

• We plan to add more legends to the map, because even though we provide a useful "About" modal, many people click through it quickly and would like to be reminded.
• We plan to provide more summary data about each zipcode upon hovering/clicking instead of just whitespace, though we haven't decided how to visualize that yet. Additionally, upon clicking on the zipcode we also plan to zoom the user into the circle view.
• We plan to alter the circle layer to make the green less overwhelming by potentially adding other differentiating features.
• Based on instructor feedback, we have decided to get rid of the table since it doesn't do a great job of complementing the map. Instead we will probably use that space to display zipcode trends.
• We plan to highlight the violations more by providing them with their own visualization of trends and giving more space to them.

Final Design

Much of our final design is similar to our "final" design, but we did our best to take user and instructor feedback into account. The changes we made were aimed at making our visualization prettier and easier to understand.

The biggest change we made was the deletion of the table. As noted above, our instructor didn't think it was the best way to visualize information/trends, and we agreed. In place of the table, we want to implement a bar chart of violation frequencies that responds to borough selections and map refocusing. The bar chart will show the top 20 violation codes in a given area, and these codes will be colored depending on their critical flag.

Another big change we want to make to the design is the addition of a violation visualization, which shows up when one of a restaurant's violations (in bottom right of the main webpage) is clicked. This visualization is a simple pie chart where users can see the frequency of a given violation in the context of all of NYC, the restaurant's zipcode, and the restaurant's cuisine type.

We also reimagined our map designs to have legends and tooltips/hover information. For the circle view, we want to add an outline to the circles to distinguish their critical status. For the heatmap view, we want to change the colors so that it's more clear what they mean. We also plan to implement functionality that allows users to click on a zipcode and immediately zoom into the circle view.

We are confident that these changes address the concerns/feedback from our users as noted above.

Justification

We believe that our visualization design is a successful exploration of NYC health inspection data because it encodes information at a macro level - of big trends across the city, its boroughs, and different zipcodes with our heatmap and bar chart - and a micro level - of small trends within each zipcode and neighborhood with our circle map, modified bar chart, and individual restaurant view. Our addition of the bar chart allowed us to target the frequency of violations in a way that's more suitable to identifying trends, and our addition of the restaurant violation pop-up and pie chart allows users to contextualize violations of each restaurant that they select among other restaurants in the same zipcode, of the same cuisine type, and across NYC. Though we ultimately didn't address widespread cuisine type trends in a formal way, we think that was for the best because during our exploratory data analysis, it appeared that certain cuisines had more violations simply because there were more restaurants of those cuisine types.

Implementation

---

Milestone 1

We foresaw that implementing the map would be the most challenging part of the assignment as it was unfamiliar to us. We thus decided to tackle its implementation first, starting with a heatmap for each zip code. As discussed above, there is too much data to load in all at once, so we had to segment the data by zip code. Individual restaurant data will be loaded in once the user has zoomed in enough.

Starting with a geojson dataset of each individual county in the five New York City boroughs, we calculated and assigned an average restaurant grade for the area. A score of A is a 1, B a 2, and C a 3, so zip codes with a higher average score will have more violations per restaurant on average.

We used leaflet.js as our mapping library, and adapted code from the leaflet docs, Mike Bostock, and the observable forum to initialize and project the geojson.

Milestone 2

Implementing more of the map was again our first priority. We broke up the data by zip code, so once you have zoomed in a certain amount, the data for the areas in view will load. This data includes information about each restaurant like its name, address, and cuisine type in addition to data about its health grade and inspection results.

Milestone 2 Resubmission

Based on feedback from submission two in addition to feedback from user studies, we implemented the following changes.

Evaluation

---

We learned several things about our data. Most notably, we noticed in what way restaurants with lower or higher scores are clustered and how violations that seem serious don't always result in a poor grade.

Clusters of data

One of the most interesting discoveries was that restaurant grades are clustered on the macro/neightboorhood level, but not necessarily on the street level. For instance, zip codes in the Bronx tend to have more grades of B and C compared to that of lower Manhattan. On our map, one can see how the Bronx has zones colored more orange, and lower Manhattan has zones colored more yellow.

Above, the map on the left shows income inequality throughout New York City. Our map on the right, showing health code data, is colored similarly, implying there might be some relationship between neighborhood wealth and cleanliness of restaurants.

We had initially hypothesized that restaurants on the street level would share similar scores or violations. For example, we thought several restaurants next to one another would all have rats or not have rats at all. This, however, turned out not to be the case. This could imply that while there is potentially a more macro level answer to health data, preventing violations still lies with the individual restaurant.

Critical violations

The most concerning discovery was with how scores are calculated. A restaurant can have critical violations, such as evidence of rats or roaches, and still have a grade of A. This is the most important takeaway from our data. Restaurant grades must be considered in tandem with the actual violations that contributed to that grade in order to fully understand the safety of a restaurant. Yelp or Google Reviews can provide one perspective, but those don’t give you the full picture of how safe it is to eat there.

Reflection

Overall, we think our visualization works well to answer the main questions we asked. To further improve our design, we think adding the ability to search for a restaurant by name would be helpful for someone who wants to find information quickly about a specific place. Similarly, adding filters on the restaurant level such as by cuisine or by violation could also help someone have more control over the navigation.