Picture by Editor
# Introduction
In keeping with CrowdFlower’s survey, information scientists spend 60% of their time organizing and cleansing the info.
On this article, we’ll stroll via constructing an information cleansing pipeline utilizing a real-life dataset from DoorDash. It comprises almost 200,000 meals supply data, every of which incorporates dozens of options resembling supply time, complete gadgets, and retailer class (e.g., Mexican, Thai, or American delicacies).
# Predicting Meals Supply Occasions with DoorDash Knowledge
DoorDash goals to estimate the time it takes to ship meals precisely, from the second a buyer locations an order to the time it arrives at their door. In this information undertaking, we’re tasked with creating a mannequin that predicts the overall supply length primarily based on historic supply information.
Nonetheless, we gained’t do the entire undertaking—i.e., we gained’t construct a predictive mannequin. As a substitute, we’ll use the dataset supplied within the undertaking and create an information cleansing pipeline.
Our workflow consists of two main steps.
# Knowledge Exploration
Let’s begin by loading and viewing the primary few rows of the dataset.
// Load and Preview the Dataset
import pandas as pd
df = pd.read_csv("historical_data.csv")
df.head()
Right here is the output.
This dataset contains datetime columns that seize the order creation time and precise supply time, which can be utilized to calculate supply length. It additionally comprises different options resembling retailer class, complete merchandise depend, subtotal, and minimal merchandise worth, making it appropriate for varied kinds of information evaluation. We will already see that there are some NaN values, which we’ll discover extra carefully within the following step.
// Discover The Columns With data()
Let’s examine all column names with the data() technique. We’ll use this technique all through the article to see the modifications in column worth counts; it’s a great indicator of lacking information and general information well being.
Right here is the output.
As you’ll be able to see, we’ve got 15 columns, however the variety of non-null values differs throughout them. This implies some columns comprise lacking values, which might have an effect on our evaluation if not dealt with correctly. One last item: the created_at and actual_delivery_time information varieties are objects; these must be datetime.
# Constructing Knowledge Cleansing Pipeline
On this step, we construct a structured information cleansing pipeline to arrange the dataset for modeling. Every stage addresses frequent points resembling timestamp formatting, lacking values, and irrelevant options.
// Fixing the Date and Time Columns Knowledge Varieties
Earlier than doing information evaluation, we have to repair the columns that present the time. In any other case, the calculation that we talked about (actual_delivery_time – created_at) will go flawed.
What we’re fixing:
- created_at: when the order was positioned
- actual_delivery_time: when the meals arrived
These two columns are saved as objects, so to have the ability to do calculations accurately, we’ve got to transform them to the datetime format. To do this, we are able to use datetime capabilities in pandas. Right here is the code.
import pandas as pd
df = pd.read_csv("historical_data.csv")
# Convert timestamp strings to datetime objects
df["created_at"] = pd.to_datetime(df["created_at"], errors="coerce")
df["actual_delivery_time"] = pd.to_datetime(df["actual_delivery_time"], errors="coerce")
df.data()
Right here is the output.
As you’ll be able to see from the screenshot above, the created_at and actual_delivery_time are datetime objects now.
Among the many key columns, store_primary_category has the fewest non-null values (192,668), which implies it has essentially the most lacking information. That’s why we’ll deal with cleansing it first.
// Knowledge Imputation With mode()
One of many messiest columns within the dataset, evident from its excessive variety of lacking values, is store_primary_category. It tells us what sort of meals shops can be found, like Mexican, American, and Thai. Nonetheless, many rows are lacking this data, which is an issue. As an example, it will possibly restrict how we are able to group or analyze the info. So how can we repair it?
We’ll fill these rows as an alternative of dropping them. To do this, we are going to use smarter imputation.
We write a dictionary that maps every store_id to its most frequent class, after which use that mapping to fill in lacking values. Let’s see the dataset earlier than doing that.
Right here is the code.
import numpy as np
# World most-frequent class as a fallback
global_mode = df["store_primary_category"].mode().iloc[0]
# Construct store-level mapping to essentially the most frequent class (quick and strong)
store_mode = (
df.groupby("store_id")["store_primary_category"]
.agg(lambda s: s.mode().iloc[0] if not s.mode().empty else np.nan)
)
# Fill lacking classes utilizing the store-level mode, then fall again to world mode
df["store_primary_category"] = (
df["store_primary_category"]
.fillna(df["store_id"].map(store_mode))
.fillna(global_mode)
)
df.data()
Right here is the output.
As you’ll be able to see from the screenshot above, the store_primary_category column now has a better non-null depend. However let’s double-check with this code.
df["store_primary_category"].isna().sum()
Right here is the output displaying the variety of NaN values. It’s zero; we removed all of them.
And let’s see the dataset after the imputation.
// Dropping Remaining NaNs
Within the earlier step, we corrected the store_primary_category, however did you discover one thing? The non-null counts throughout the columns nonetheless don’t match!
It is a clear signal that we’re nonetheless coping with lacking values in some a part of the dataset. Now, on the subject of information cleansing, we’ve got two choices:
- Fill these lacking values
- Drop them
On condition that this dataset comprises almost 200,000 rows, we are able to afford to lose some. With smaller datasets, you’d should be extra cautious. In that case, it’s advisable to research every column, set up requirements (determine how lacking values will likely be stuffed—utilizing the imply, median, most frequent worth, or domain-specific defaults), after which fill them.
To take away the NaNs, we are going to use the dropna() technique from the pandas library. We’re setting inplace=True to use the modifications on to the DataFrame with no need to assign it once more. Let’s see the dataset at this level.
Right here is the code.
df.dropna(inplace=True)
df.data()
Right here is the output.
As you’ll be able to see from the screenshot above, every column now has the identical variety of non-null values.
Let’s see the dataset after all of the modifications.
// What Can You Do Subsequent?
Now that we’ve got a clear dataset, right here are some things you are able to do subsequent:
- Carry out EDA to know supply patterns.
- Engineer new options like supply hours or busy dashers ratio so as to add extra which means to your evaluation.
- Analyze correlations between variables to extend your mannequin’s efficiency.
- Construct completely different regression fashions and discover the best-performing mannequin.
- Predict the supply length with the best-performing mannequin.
# Remaining Ideas
On this article, we’ve got cleaned the real-life dataset from DoorDash by addressing frequent information high quality points, resembling fixing incorrect information varieties and dealing with lacking values. We constructed a easy information cleansing pipeline tailor-made to this information undertaking and explored potential subsequent steps.
Actual-world datasets could be messier than you assume, however there are additionally many strategies and methods to unravel these points. Thanks for studying!
Nate Rosidi is an information scientist and in product technique. He is additionally an adjunct professor educating analytics, and is the founding father of StrataScratch, a platform serving to information scientists put together for his or her interviews with actual interview questions from prime corporations. Nate writes on the most recent developments within the profession market, offers interview recommendation, shares information science initiatives, and covers every part SQL.
