Predicting-Airbnb-Prices_files/figure-gfm/: Contains images to include in README.md.
Predicting Airbnb Prices.Rmd: Original .Rmd file where I created code
Predicting-Airbnb-Prices.md: GitHub friendly version
In October 2021, Melbourne, previously the worlds most livable city, and at the same time the most locked down city, came out of 262 days of lockdown for what is hopefully the final time. As a result, people are eager to travel and get out of the area they have been locked down in for so long.
Airbnb, a popular vacation rental company, connects people who are looking to travel, with people who are looking for accommodation in specific places. While Airbnb services in Melbourne have had some negative reviews recently due to COVID-19 restrictions and cancellation policies, with the end of lockdowns, hopefully this will no longer occur, and people can go back to enjoying their holidays.
Analysing Airbnb data in Melbourne from Inside Airbnb, we look to explore how well we can predict the prices of properties listed on Airbnb.
#Setting seed for reproducability
set.seed(30127)
#Loading packages
library(tidyverse)
library(geosphere)
library(caret)
library(knitr)
library(tidymodels)
#Loading the data
airbnb <- read.csv("listings.csv")
#Initial clean
airbnb_clean <- airbnb %>%
#Selecting only numerical columns that are useful for our analysis
select(latitude, longitude, bathrooms_text, bedrooms, reviews_per_month, review_scores_rating, host_total_listings_count, maximum_nights, price) %>%
mutate(
#Clean up price column and remove "$"
price = str_replace_all(price, "\\$", ""),
#Removing any "," in price column as well
price = str_replace_all(price, ",", ""),
#Creating new column with only number of bathrooms, removing the text
bathrooms = str_extract(bathrooms_text, "\\d+\\.?\\d*")
) %>%
#Ensuring all of our columns are in a numeric format
mutate_all(as.numeric)
#Calculating the distance to the city using latitude and longitude from
#LatLong.net (https://www.latlong.net/place/melbourne-vic-australia-27235.html)
dist_mat <- data.frame(airbnb_long = airbnb_clean$longitude,
airbnb_lat = airbnb_clean$latitude,
melb_long = rep(144.946457, length(airbnb_clean$longitude)),
melb_lat = rep(-37.840935, length(airbnb_clean$longitude))) %>%
rowwise %>%
mutate(distance_to_city = distm(x = c(airbnb_long, airbnb_lat),
y = c(melb_long, melb_lat),
#Dividng by 1000 to get distance in kilometres
fun = distGeo) / 1000)
#Final Clean
airbnb_clean <- airbnb_clean %>%
#Adding our calculated distances
bind_cols(distance_to_city = dist_mat$distance_to_city) %>%
#Only selecting our needed columns
select(bathrooms, bedrooms, reviews_per_month, review_scores_rating, host_total_listings_count, maximum_nights, distance_to_city, price) %>%
#Removing any NA or blank values
drop_na()
head(airbnb_clean) %>%
kable()| bathrooms | bedrooms | reviews_per_month | review_scores_rating | host_total_listings_count | maximum_nights | distance_to_city | price |
|---|---|---|---|---|---|---|---|
| 1.0 | 1 | 0.03 | 4.50 | 1 | 365 | 14.898270 | 60 |
| 1.0 | 1 | 0.30 | 4.68 | 13 | 14 | 3.394169 | 95 |
| 2.5 | 1 | 0.02 | 4.50 | 1 | 730 | 3.020933 | 1000 |
| 1.0 | 3 | 1.26 | 4.83 | 1 | 14 | 42.059901 | 110 |
| 1.0 | 1 | 1.17 | 4.71 | 3 | 365 | 16.647397 | 40 |
| 1.0 | 1 | 1.61 | 4.88 | 3 | 365 | 16.153476 | 99 |
K-Nearest Neighbours (k-NN) is a supervised machine learning technique that can be used for both regression and classification, however in this case we have used it for regression. Using k-NN for regression provides us an output based on the average value of the “nearest neighbours” of an object, where “near” is determined with the concept of distance. Distance, in a literal sense, measures how far one thing is from another, however in the world of machine learning, it refers to how similar (or dissimilar) two observations are. The larger the distance between two observations, the less similar they are, and vice versa.
By calculating the Euclidian distance between objects based on defined variables, k-NN looks to find the a specified number of the most similar objects, and in the case of regression, find the average of these labels. Using k-NN will allow us to take a number of important variables from our data, use this data to find Airbnb properties that have similar characteristics, then average these out to make our prediction of their price.
At the same time, we will use hyperparameter optimisation to find the optimal value of k for our k-NN model. Hyperparameter optimaisation will take our training set, divide it into a set number of folds, and will use k-fold cross-validation to calculate the error metrics for each potential value of k for our k-NN model. This will become our validation error metric - specifically root mean square error (RMSE) - and the value of k that gives us our lowest RMSE will be our considered our optimal value of k.
#Splitting the data
airbnb_split <- initial_split(airbnb_clean)
#Using the split data to create our training set
airbnb_train <- training(airbnb_split)
#Using the split data to create our test set
airbnb_test <- testing(airbnb_split)
#Defining that we want to use cross validation with 5 folds when evaluating our k-nearest neighbours model
train_control <- trainControl(method = "cv",
number = 5)
#Creating a grid of possible tuning values
knn_grid <- expand.grid(k = 1:100)
#Creating k-nearest neighbours model
knn_model <- train(price ~.,
data = airbnb_train,
method = "knn",
trControl = train_control,
#Standardising our numerical values to ensure no single variable overpowers the formula
preProcess = c("center", "scale"),
tuneGrid = knn_grid
)
knn_model## k-Nearest Neighbors
##
## 9659 samples
## 7 predictor
##
## Pre-processing: centered (7), scaled (7)
## Resampling: Cross-Validated (5 fold)
## Summary of sample sizes: 7726, 7729, 7726, 7728, 7727
## Resampling results across tuning parameters:
##
## k RMSE Rsquared MAE
## 1 461.7767 0.02266360 114.65170
## 2 399.7350 0.03644223 104.10335
## 3 380.3024 0.04423620 100.47418
## 4 372.6794 0.04962224 98.69786
## 5 363.6873 0.05792913 97.43847
## 6 362.9174 0.05787541 97.44310
## 7 357.7486 0.06455719 96.59421
## 8 356.4038 0.06709347 96.51860
## 9 355.0781 0.06863879 96.28321
## 10 352.3814 0.07252608 95.80300
## 11 351.6872 0.07251317 95.93213
## 12 350.6351 0.07448632 95.96371
## 13 349.1888 0.07704391 95.79552
## 14 348.4796 0.07773070 95.68525
## 15 347.3496 0.08018353 95.35490
## 16 346.5130 0.08174387 95.21843
## 17 344.9686 0.08669312 94.62149
## 18 344.1920 0.08925251 94.32749
## 19 343.6927 0.09030606 94.28614
## 20 343.0477 0.09233387 94.08664
## 21 342.2949 0.09481826 93.71308
## 22 342.1452 0.09483092 93.67136
## 23 341.6057 0.09661353 93.51357
## 24 341.0414 0.09864331 93.35787
## 25 340.8358 0.09962435 93.51770
## 26 340.4806 0.10052152 93.41384
## 27 340.1011 0.10194081 93.24096
## 28 339.8340 0.10314779 93.25555
## 29 339.6591 0.10368554 93.28006
## 30 339.4765 0.10427292 93.18097
## 31 339.3611 0.10479302 93.27900
## 32 339.1218 0.10558658 93.03957
## 33 338.9793 0.10612834 93.14913
## 34 338.6866 0.10732658 93.06157
## 35 338.6465 0.10741712 93.11371
## 36 338.5537 0.10772396 93.08434
## 37 338.4671 0.10800913 93.06607
## 38 338.3122 0.10862395 92.97393
## 39 338.2259 0.10898718 93.01399
## 40 338.1128 0.10948477 92.99419
## 41 338.1420 0.10916695 93.00781
## 42 338.1186 0.10924141 93.00450
## 43 338.0690 0.10956181 93.05368
## 44 338.0314 0.10958508 93.04599
## 45 337.5550 0.11300168 92.99154
## 46 337.4757 0.11341892 93.00779
## 47 337.2070 0.11481148 92.92187
## 48 337.2244 0.11470866 92.93962
## 49 337.1809 0.11489170 92.87062
## 50 337.0793 0.11536773 92.77549
## 51 337.1048 0.11520938 92.88265
## 52 337.1763 0.11455252 92.96257
## 53 337.1631 0.11456983 92.94214
## 54 337.1547 0.11462454 92.96760
## 55 337.0922 0.11512604 92.93026
## 56 336.9560 0.11601098 92.82235
## 57 336.8654 0.11657056 92.70146
## 58 336.8122 0.11697100 92.68124
## 59 336.8594 0.11662161 92.75839
## 60 336.7318 0.11746627 92.61898
## 61 336.7112 0.11764483 92.59586
## 62 336.7147 0.11765860 92.66799
## 63 336.6202 0.11829810 92.57573
## 64 336.5973 0.11847921 92.58687
## 65 336.5999 0.11850056 92.55709
## 66 336.6594 0.11802207 92.72655
## 67 336.6438 0.11820633 92.72487
## 68 336.6672 0.11802600 92.74795
## 69 336.6844 0.11786621 92.82076
## 70 336.6139 0.11835047 92.73105
## 71 336.6498 0.11805533 92.73442
## 72 336.6835 0.11775667 92.81577
## 73 336.6285 0.11809340 92.77570
## 74 336.6198 0.11817932 92.76272
## 75 336.6456 0.11793041 92.79982
## 76 336.6084 0.11822792 92.79621
## 77 336.6188 0.11820877 92.80810
## 78 336.5674 0.11861341 92.84763
## 79 336.6029 0.11838234 92.82770
## 80 336.6198 0.11826487 92.76239
## 81 336.6220 0.11825914 92.75727
## 82 336.5939 0.11848347 92.73046
## 83 336.5378 0.11893383 92.69921
## 84 336.4958 0.11917810 92.66600
## 85 336.4385 0.11953031 92.61774
## 86 336.4291 0.11958392 92.62762
## 87 336.3587 0.12016244 92.63855
## 88 336.2816 0.12075728 92.52643
## 89 336.2626 0.12096731 92.51771
## 90 336.2159 0.12129728 92.47242
## 91 336.2010 0.12136107 92.43030
## 92 336.2473 0.12108639 92.46837
## 93 336.2390 0.12120947 92.48346
## 94 336.1359 0.12173442 92.43634
## 95 336.0959 0.12203893 92.44687
## 96 336.1060 0.12192804 92.39700
## 97 336.1217 0.12175872 92.38973
## 98 336.1386 0.12164314 92.41770
## 99 336.1655 0.12144357 92.40092
## 100 336.1527 0.12156517 92.32283
##
## RMSE was used to select the optimal model using the smallest value.
## The final value used for the model was k = 95.
plot(knn_model)
Looking at the above plot, it is difficult to see the optimal value of k. After conducting hyperparameter optimsation, the optimal value of k = 95. This is more clear in the above output from our k-NN model, where we can see that the RMSE continues to decrease until k reaches 95, then following 95 the RMSE begins to decrease.
#Creating our predictions on our testing set
predictions <- predict(knn_model, airbnb_test) %>%
as_tibble()
#Printing our results for our model
airbnb_test %>%
bind_cols(predictions) %>%
metrics(truth = price, estimate = value) %>%
kable()| .metric | .estimator | .estimate |
|---|---|---|
| rmse | standard | 317.1482665 |
| rsq | standard | 0.0963688 |
| mae | standard | 92.8375615 |
Since we are measuring how well we can predict the price of an Airbnb in Melbourne, and since the RMSE can be used to penalise large errors, we can look at the Mean Absolute Error (MAE) to evaluate the performance of our model. The MAE tells us that on average, our model will predict the price of the Airbnb with a distance of $92.84, or in other words, we will predict the price with an average error of $92.84, either more or less.
#Calculating the range of the prices of our Airbnb's
tibble(min = range(airbnb_clean$price)[1],
max = range(airbnb_clean$price)[2]) %>%
kable()| min | max |
|---|---|
| 14 | 13379 |
Since we have quite a large range of different prices for Airbnb’s in Melbourne, it appears our model does a fairly good job of predicting the prices.
#Getting our five number summary to calculate outliers
summary_stats <- summary(airbnb_clean$price)
#(1.5 * IQR) + 3rd Quartile
outlier <- (1.5 * (summary_stats[[5]] - summary_stats[[2]])) + summary_stats[[5]]
airbnb_clean %>%
ggplot(aes(x = price)) +
geom_histogram(bins = 50) +
labs(x = "Price",
y = "Count",
title = "Distribution of the different prices of Airbnb properties in Melbourne") +
theme_bw()
airbnb_clean %>%
filter(price > outlier) %>%
ggplot(aes(x = price)) +
geom_histogram() +
ylim(0, 50) +
labs(x = "Price",
y = "Count",
title = "Distribution of the different prices of Airbnb properties in Melbourne considered \n outliers") +
theme_bw()
As we can see in the above plots, there are some heavy outliers, that may be potentially skewing the performance of our model. It is not beneficial for us to remove these as if we do not have them, however, it would be interesting to see how our model performs after removing the outliers.
#Removing outliers
airbnb_clean1 <- airbnb_clean %>%
filter(price < outlier)
#Splitting the data
airbnb_split1 <- initial_split(airbnb_clean1)
#Using the split data to create our training set
airbnb_train1 <- training(airbnb_split1)
#Using the split data to create our test set
airbnb_test1 <- testing(airbnb_split1)
#Defining that we want to use cross validation with 5 folds when evaluating our k-nearest neighbours model
train_control1 <- trainControl(method = "cv",
number = 5)
#Creating a grid of possible tuning values
knn_grid1 <- expand.grid(k = 1:40)
#Creating k-nearest neighbours model
knn_model1 <- train(price ~.,
data = airbnb_train1,
method = "knn",
trControl = train_control1,
#Standardising our numerical values to ensure no single variable overpowers the formula
preProcess = c("center", "scale"),
tuneGrid = knn_grid1
)
knn_model1## k-Nearest Neighbors
##
## 8908 samples
## 7 predictor
##
## Pre-processing: centered (7), scaled (7)
## Resampling: Cross-Validated (5 fold)
## Summary of sample sizes: 7126, 7127, 7125, 7127, 7127
## Resampling results across tuning parameters:
##
## k RMSE Rsquared MAE
## 1 79.16563 0.2520463 56.29225
## 2 68.71546 0.3311861 49.95490
## 3 65.17857 0.3664675 47.65490
## 4 63.53086 0.3841872 46.63859
## 5 62.43200 0.3972593 45.89566
## 6 61.79044 0.4051373 45.49702
## 7 61.53385 0.4079895 45.39458
## 8 61.26140 0.4115750 45.23018
## 9 61.11225 0.4132456 45.08093
## 10 60.83970 0.4173213 44.95116
## 11 60.78025 0.4178634 44.98929
## 12 60.62130 0.4200795 44.87491
## 13 60.53756 0.4211457 44.84985
## 14 60.44518 0.4225771 44.80366
## 15 60.40039 0.4230976 44.76659
## 16 60.23946 0.4259123 44.71126
## 17 60.23200 0.4259133 44.71481
## 18 60.24994 0.4254167 44.73686
## 19 60.27217 0.4248590 44.73274
## 20 60.21634 0.4257513 44.69792
## 21 60.10742 0.4276534 44.60538
## 22 60.08584 0.4279608 44.57969
## 23 60.05727 0.4284252 44.56771
## 24 60.01568 0.4291959 44.55678
## 25 60.00442 0.4293278 44.54700
## 26 60.01582 0.4290873 44.58534
## 27 60.05929 0.4282236 44.63602
## 28 60.06146 0.4281724 44.62644
## 29 60.06922 0.4280059 44.65460
## 30 60.10599 0.4273086 44.67889
## 31 60.14494 0.4265700 44.70795
## 32 60.12248 0.4269964 44.73052
## 33 60.14440 0.4265795 44.74629
## 34 60.17410 0.4260144 44.78338
## 35 60.13487 0.4267565 44.74448
## 36 60.10144 0.4274082 44.70727
## 37 60.13469 0.4267941 44.74908
## 38 60.15319 0.4264533 44.75752
## 39 60.13716 0.4267687 44.74978
## 40 60.15518 0.4264274 44.76737
##
## RMSE was used to select the optimal model using the smallest value.
## The final value used for the model was k = 25.
plot(knn_model1)
Looking at the above plot we can see much clearer the optimal value for k. Using k = 25 is where our RMSE is lowest, and after this number the RMSE increases, implying that our average error increases.
#Creating our predictions on our testing set
predictions1 <- predict(knn_model1, airbnb_test1) %>%
as_tibble()
#Printing our results for our model
airbnb_test1 %>%
bind_cols(predictions1) %>%
metrics(truth = price, estimate = value) %>%
kable()| .metric | .estimator | .estimate |
|---|---|---|
| rmse | standard | 58.1527814 |
| rsq | standard | 0.4410414 |
| mae | standard | 43.3387293 |
Similar to our first model, MAE tells us that on average, our model will predict the price of the Airbnb with a distance of $43.34, or in other words, we will predict the price with an average error of $43.34, either more or less, for those properties that are less than $380.
At the same time R2 has significantly increased, with approximately 44.10% of the variation in price of Airbnb’s in Melbourne can be explained by the variables we have chosen.
After our analysis, we were able to not only gain an insight into the different prices of Airbnb properties in Melbourne, but we were able to create a model that attempts to predict the prices of these properties. As Melbourne continues to open back up by easing restrictions, we hope this model can be useful to those who may be looking to take a holiday somewhere in Melbourne, allowing them to see what the price should be based on already active listings.
One improvement that could be made to the model is a slight variation on the distance to the city variable created. For those looking to stay closer to the Central Business District (CBD) of Melbourne, this is an important variable to consider, however a number of Airbnb properties may be further from the CBD, and instead closer to other popular areas, like the beaches on the Mornington Peninsula, or the Dandenong Ranges. Just because these properties are further from the city, does not mean that this should affect price, as they are closer to other desirable areas that aren’t the CBD.