In this section you will use the techniques in the previous lessons to do some exploratory data analysis of a large dataset. Once you have a good understanding of the usefulness of the various columns, you will learn:
So far you've learned about how text data is quite unlike numerical types of data. If it's text that was written or spoken by a human, if can be analysed to find patterns and frequencies, sentiment and meaning. This lesson takes you into a real data set with a real challenge: **[515K Hotel Reviews Data in Europe](https://www.kaggle.com/jiashenliu/515k-hotel-reviews-data-in-europe)** and includes a [CC0: Public Domain license](https://creativecommons.org/publicdomain/zero/1.0/). It was scraped from Booking.com from public sources. The creator of the dataset was Jiashen Liu.
* The data set which is available on Kaggle [515K Hotel Reviews Data in Europe](https://www.kaggle.com/jiashenliu/515k-hotel-reviews-data-in-europe). It is around 230 MB unzipped. Download it to the root `/data` folder associated with these NLP lessons.
This challenge assumes that you are building a hotel recommendation bot using sentiment analysis and guest reviews scores. The dataset you will be using includes reviews of 1493 different hotels in 6 cities.
* What are the most frequently used words and phrases in reviews?
* Do the official *tags* describing a hotel correlate with review scores (e.g. are the more negative reviews for a particular hotel for *Family with young children* than by *Solo traveller*, perhaps indicating it is better for *Solo travellers*?)
* Do the NLTK sentiment scores 'agree' with the hotel reviewer's numerical score?
* According to the dataset creator, this column is the *Average Score of the hotel, calculated based on the latest comment in the last year*. This seems like an unusual way to calculate the score, but it is the data scraped so we may take it as face value for now.
- A freshness or staleness measure might be applied to a review (older reviews might not be as accurate as newer ones because hotel management changed, or renovations have been done, or a pool was added etc.)
-`Tags`
- These are short descriptors that a reviewer may select to describe the type of guest they were (e.g. solo or family), the type of room they had, the length of stay and how the review was submitted.
- Unfortunately, using these tags is problematic, check the section below which discusses their usefulness
- This might be a factor in a recommendation model, for instance, if you could determine that more prolific reviewers with hundreds of reviews were more likely to be negative rather than positive. However, the reviewer of any particular review is not identified with a unique code, and therefore cannot be linked to a set of reviews. There are 30 reviewers with 100 or more reviews, but it's hard to see how this can aid the recommendation model.
- Some people might think that certain nationalities are more likely to give a positive or negative review because of a national inclination. Be careful building such anecdotal views into your models. These are national (and sometimes racial) stereotypes, and each reviewer was an individual who wrote a review based on their experience. It may have been filtered through many lenses such as their previous hotel stays, the distance travelled, and their personal temperament. Thinking that their nationality was the reason for a review score is hard to justify.
| 7.8 | 1945 | 2.5 | This is currently not a hotel but a construction site I was terrorized from early morning and all day with unacceptable building noise while resting after a long trip and working in the room People were working all day i e with jackhammers in the adjacent rooms I asked for a room change but no silent room was available To make things worse I was overcharged I checked out in the evening since I had to leave very early flight and received an appropriate bill A day later the hotel made another charge without my consent in excess of booked price It's a terrible place Don't punish yourself by booking here | Nothing Terrible place Stay away | Business trip Couple Standard Double Room Stayed 2 nights |
As you can see, this guest did not have a happy stay at this hotel. The hotel has a good average score of 7.8 and 1945 reviews, but this reviewer gave it 2.5 and wrote 115 words about how negative their stay was. If they wrote nothing at all in the Positive_Review column, you might surmise there was nothing positive, but alas they wrote 7 words of warning. If we just counted words instead of the meaning, or sentiment of the words, we might have a skewed view of the reviewer's intent. Strangely, their score of 2.5 is confusing, because if that hotel stay was so bad, why give it any points at all? Investigating the dataset closely, you'll see that the lowest possible score is 2.5, not 0. The highest possible score is 10.
As mentioned above, at first glance, the idea to use `Tags` to categorize the data makes sense. Unfortunately these tags are not standardized, which means that in a given hotel, the options might be *Single room*, *Twin room*, and *Double room*, but in the next hotel, they are *Deluxe Single Room*, *Classic Queen Room*, and *Executive King Room*. These might be the same things, but there are so many variations that the choice becomes:
1. Attempt to change all terms to a single standard, which is very difficult, because it is not clear what the conversion path would be in each case (e.g. *Classic single room* maps to *Single room* but *Superior Queen Room with Courtyard Garden or City View* is much harder to map)
1. We can take an NLP approach and measure the frequency of certain terms like *Solo*, *Business Traveller*, or *Family with young kids* as they apply to each hotel, and factor that into the recommendation
Tags are usually (but not always) a single field containing a list of 5 to 6 comma separated values aligning to *Type of trip*, *Type of guests*, *Type of room*, *Number of nights*, and *Type of device review was submitted on*. However, because some reviewers don't fill in each field (they might leave one blank), the values are not always in the same order.
As an example, take *Type of group*. There are 1025 unique possibilities in this field in the `Tags` column, and unfortunately only some of them refer to a group (some are the type of room etc.). If you filter only the ones that mention family, the results contain many *Family room* type results. If you include the term *with*, i.e. count the *Family with* values, the results are better, with over 80,000 of the 515,000 results containing the phrase "Family with young children" or "Family with older children".
There are a number of oddities or discrepancies with the data set that I can't figure out, but are illustrated here so you are aware of them when building your models. If you figure it out, please let us know in the discussion section!
The single hotel with the most reviews in this dataset is *Britannia International Hotel Canary Wharf* with 4789 reviews out of 515,000. But if we look at the `Total_Number_of_Reviews` value for this hotel, it is 9086. You might surmise that there are many more scores without reviews, so perhaps we should add in the `Additional_Number_of_Scoring` column value. That value is 2682, and adding it to 4789 gets us 7,471 which is still 1615 short of the `Total_Number_of_Reviews`.
If you take the `Average_Score` columns, you might surmise it is the average of the reviews in the dataset, but the description from Kaggle is "*Average Score of the hotel, calculated based on the latest comment in the last year*". That doesn't seem that useful, but we can calculate our own average based on the reviews scores in the data set. Using the same hotel as an example, the average hotel score is given as 7.1 but the calculated score (average reviewer score *in* the dataset) is 6.8. This is close, but not the same value, and we can only guess that the scores given in the `Additional_Number_of_Scoring` reviews increased the average to 7.1. Unfortunately with no way to test or prove that assertion, it is difficult to use or trust `Average_Score`, `Additional_Number_of_Scoring` and `Total_Number_of_Reviews` when they are based on, or refer to, data we do not have.
To complicate things further, the hotel with the second highest number of reviews has a calculated average score of 8.12 and the dataset `Average_Score` is 8.1. Is this correct score a coincidence or is the first hotel a discrepancy?
On the possibility that these hotel might be an outlier, and that maybe most of the values tally up (but some do not for some reason) we will write a short program next to explore the values in the dataset and determine the correct usage (or non-usage) of the values.
> When working with this dataset you will write code that calculates something from the text without having to read or analyse the text yourself. This is the essence of NLP, interpreting meaning or sentiment without having to have a human do it. However, it is possible that you will read some of the negative reviews. I would urge you not to, because you don't have to. Some of them are silly, or irrelevant negative hotel reviews, such as "The weather wasn't great", something beyond the control of the hotel, or indeed, anyone. But there is a dark side to some reviews too. Sometimes the negative reviews are racist, sexist, or ageist. This is unfortunate but to be expected in a dataset scraped off a public website. Some reviewers leave reviews that you would find distasteful, uncomfortable, or upsetting. Better to let the code measure the sentiment than read them yourself and be upset. That said, it is a minority that write such things, but they exist all the same.
That's enough examining the data visually, now you'll write some code and get some answers! This section uses the pandas library. Your very first task is to ensure you can load and read the CSV data. The pandas library has a fast CSV loader, and the result is placed in a dataframe, as in previous lessons. The CSV we are loading has over half a million rows, but only 17 columns. Pandas gives you lots of powerful ways to interact with a dataframe, including the ability to perform operations on every row.
From here on in this lesson, there will be code snippets and some explanations of the code and some discussion about what the results mean. Use the included _notebook.ipynb_ for your code.
In this case, the data is already *clean*, that means that it is ready to work with, and does not have characters in other languages that might trip up algorithms expecting only English characters.
✅ You might have to work with data that required some initial processing to format it before applying NLP techniques, but not this time. If you had to, how would you handle non-English characters?
Take a moment to ensure that once the data is loaded, you can explore it with code. It's very easy to want to focus on the `Negative_Review` and `Positive_Review` columns. They are filled with natural text for your NLP algorithms to process. But wait! Before you jump into the NLP and sentiment, you should follow the code below to ascertain if the values given in the dataset match the values you calculate with pandas.
The first task in this lesson is to check if the following assertions are correct by writing some code that examines the data frame (without changing it).
> Like many programming tasks, there are several ways to complete this, but good advice is to do it in the simplest, easiest way you can, especially if it will be easier to understand when you come back to this code in the future. With dataframes, there is a comprehensive API that will often have a way to do what you want efficiently.
1. Print out the *shape* of the data frame you have just loaded (the shape is the number of rows and columns)
2. Calculate the frequency count for reviewer nationalities:
1. How many distinct values are there for the column `Reviewer_Nationality` and what are they?
2. What reviewer nationality is the most common in the dataset (print country and number of reviews)?
3. What are the next top 10 most frequently found nationalities, and their frequency count?
3. What was the most frequently reviewed hotel for each of the top 10 most reviewer nationalities?
4. How many reviews are there per hotel (frequency count of hotel) in the dataset?
5. While there is an `Average_Score` column for each hotel in the dataset, you can also calculate an average score (getting the average of all reviewer scores in the dataset for each hotel). Add a new column to your dataframe with the column header `Calc_Average_Score` that contains that calculated average.
6. Do any hotels have the same (rounded to 1 decimal place) `Average_Score` and `Calc_Average_Score`?
1. Try writing a Python function that takes a Series (row) as an argument and compares the values, printing out a message when the values are not equal. Then use the `.apply()` method to process every row with the function.
7. Calculate and print out how many rows have column `Negative_Review` values of "No Negative"
8. Calculate and print out how many rows have column `Positive_Review` values of "No Positive"
9. Calculate and print out how many rows have column `Positive_Review` values of "No Positive" **and**`Negative_Review` values of "No Negative"
3. What are the next top 10 most frequently found nationalities, and their frequency count?
```python
print("The highest frequency reviewer nationality is " + str(nationality_freq.index[0]).strip() + " with " + str(nationality_freq[0]) + " reviews.")
# Notice there is a leading space on the values, strip() removes that for printing
# What is the top 10 most common nationalities and their frequencies?
print("The next 10 highest frequency reviewer nationalities are:")
print(nationality_freq[1:11].to_string())
The highest frequency reviewer nationality is United Kingdom with 245246 reviews.
The next 10 highest frequency reviewer nationalities are:
United States of America 35437
Australia 21686
Ireland 14827
United Arab Emirates 10235
Saudi Arabia 8951
Netherlands 8772
Switzerland 8678
Germany 7941
Canada 7894
France 7296
```
3. What was the most frequently reviewed hotel for each of the top 10 most reviewer nationalities?
```python
# What was the most frequently reviewed hotel for the top 10 nationalities
# Normally with pandas you will avoid an explicit loop, but wanted to show creating a new dataframe using criteria (don't do this with large amounts of data because it could be very slow)
for nat in nationality_freq[:10].index:
# First, extract all the rows that match the criteria into a new dataframe
nat_df = df[df["Reviewer_Nationality"] == nat]
# Now get the hotel freq
freq = nat_df["Hotel_Name"].value_counts()
print("The most reviewed hotel for " + str(nat).strip() + " was " + str(freq.index[0]) + " with " + str(freq[0]) + " reviews.")
The most reviewed hotel for United Kingdom was Britannia International Hotel Canary Wharf with 3833 reviews.
The most reviewed hotel for United States of America was Hotel Esther a with 423 reviews.
The most reviewed hotel for Australia was Park Plaza Westminster Bridge London with 167 reviews.
The most reviewed hotel for Ireland was Copthorne Tara Hotel London Kensington with 239 reviews.
The most reviewed hotel for United Arab Emirates was Millennium Hotel London Knightsbridge with 129 reviews.
The most reviewed hotel for Saudi Arabia was The Cumberland A Guoman Hotel with 142 reviews.
The most reviewed hotel for Netherlands was Jaz Amsterdam with 97 reviews.
The most reviewed hotel for Switzerland was Hotel Da Vinci with 97 reviews.
The most reviewed hotel for Germany was Hotel Da Vinci with 86 reviews.
The most reviewed hotel for Canada was St James Court A Taj Hotel London with 61 reviews.
```
4. How many reviews are there per hotel (frequency count of hotel) in the dataset?
```python
# First create a new dataframe based on the old one, removing the uneeded columns
You may notice that the *counted in the dataset* results do not match the value in `Total_Number_of_Reviews`. It is unclear if this value in the dataset represented the total number of reviews the hotel had, but not all were scraped, or some other calculation. `Total_Number_of_Reviews` is not used in the model because of this unclarity.
5. While there is an `Average_Score` column for each hotel in the dataset, you can also calculate an average score (getting the average of all reviewer scores in the dataset for each hotel). Add a new column to your dataframe with the column header `Calc_Average_Score` that contains that calculated average. Print out the columns `Hotel_Name`, `Average_Score`, and `Calc_Average_Score`.
```python
# define a function that takes a row and performs some calculation with it
You may also wonder about the `Average_Score` value and why it is sometimes different from the calculated average score. As we can't know why some of the values match, but others have a difference, it's safest in this case to use the review scores that we have to calculate the average ourselves. That said, the differences are usually very small, here are the hotels with the greatest deviation from the dataset average and the calculated average:
You may have noticed that there are 127 rows that have both "No Negative" and "No Positive" values for the columns `Negative_Review` and `Positive_Review` respectively. That means that the reviewer gave the hotel a numerical score, but declined to write either a positive or negative review. Luckily this is a small amount of rows (127 out of 515738, or 0.02%), so it probably won't skew our model or results in any particular direction, but you might not have expected a data set of reviews to have rows with no reviews, so it's worth exploring the data to discover rows like this.
This lesson demonstrates, as we saw in previous lessons, how critically important it is to understand your data and its foibles before performing operations on it. Text-based data, in particular, bears careful scrutiny. Dig through various text-heavy datasets and see if you can discover areas that could introduce bias or skewed sentiment into a model.
Take [this Learning Path on NLP](https://docs.microsoft.com/learn/paths/explore-natural-language-processing/?WT.mc_id=academic-15963-cxa) to discover tools to try when building speech and text-heavy models.
