Which features are correlated with rating review score?

From the moment that someone views a listing on the Airbnb website, to the moment they hand back the keys as they return home, there are many factors that can influence the rating score that a guest may give for their stay.

How then, might we get an idea of the ratings for a particular listing?

The dataset contains descriptions and review scores for each listing in Boston (3585 listings) and all review comments written by guests (68.3k reviews) during that time period.

To guide me on this journey, I asked myself the following questions:

The listings dataset contains almost 100 columns, and as such, I narrowed my investigation down to a small subset of columns.

Interestingly, the deskewed review score implies that an inverse relationship with total number of reviews may also exist.

This may perhaps be because:

On the other hand, a larger proportion of the listings which are not instant-bookable have higher ratings than the listings which are instant-bookable. This may be because being instant-bookable naturally results in more bookings, and as we may have identified earlier, having more bookings tends to lead to lower overall ratings.

Density curves for the neighbourhoods with the top 5 ratings (the first 5 from the left) tend to be sarcophagus-shaped, while the distributions for lower-rated neighbourhoods tend to be more pear-shaped.

Interestingly, the neighbourhoods with the highest average listing rating scores are not necessarily the ones with the most expensive listings.

However, neighbourhoods with lower average reviews have lower listing prices.

As previously explained, review rating scores are strongly biased towards higher values.

The average rating score in the Boston Airbnb listing dataset is 92, and the median score is 94. It is as such somewhat challenging to define what a ‘good’ or ‘bad’ rating score is.

From the perspective of a data scientist working at Airbnb, it may be more useful to know which listings are performing significantly worse than their counterparts.

For this reason, I defined ‘bad’ listings as the bottom 25% of all properties in Boston. This corresponded to listings with a rating score of 89 or less.

From this result, we can conclude that review comments alone can be used to predict if a listing is bad or good with a fairly high degree of accuracy.

A natural follow up question to this finding was: What words are associated with good or bad reviews?

Stemming of the words results in a loss of interpretability, but looking at the top 20 words used by the model used to discriminate good and bad listings, we can see that words like positive words like home, love, great, beauti and comfort (home, love, beauti and comfort were perhaps originally homely, lovely, beautiful and comfortable) would likely have been used to identify good listings.

On the other hand, words like dirti, howev, noisi (dirty, however and noisy prior to stemming) would likely have been attributed with listings with low rating scores.

What will YOU be looking out for in a listing or host when you book your next stay?

Related posts:

When our daughter was a little toddler we started fostering dogs. Each foster dog bounded into our lives damaged in their own way. When they left they wagged their tails with such force it knocked…

My husband Chris said he needed to talk and sat me down in the living room. I had our 18-month-old son on my lap and our second daughter in my womb. At that moment, my husband of seven years…

Our daily routines would not be complete without a shower, which can also be extremely calming and energizing. But if you don’t take care, you could unintentionally harm your hair while taking a…