Seattleites pride themselves on their outdoor excursions. You'd be hard-pressed to find someone who didn't enjoy hiking. Washington Trails Association provides a website for people to find information on hikes and discover new ones.
With the hiking community so large in the Pacific Northwest, there's a strong market for outdoor sports gear. Imagine being a gear retailer, trying to find out how to target new customers, update products and create new ones.
With hiking being so popular, there are constants efforts made to build new trails, maintain current trails, and rescue estranged hikers who lose their way. Information on why a hike is so popular can help focus resources to the best locations. It can also be used for developing new hikes in places outside of Washington.
The following models, and overall study, are meant to provide insight into what makes a hike so popular. Feature importance can be extracted from models that make predictions on popularity quanitities to draw these insights.
The data for this project is collected from Washington Trails Association (WTA), a database for hikes and trip reports in Washington state.
A dataset of all of the hikes on the website, as well as urls to those hikes, can be found in a data.world database. This dataset provided a prelimiary set of data, as well as an easy way to access all of the hike web pages.
The data was collected by visiting each of the urls in the original dataframe and scraping the html. More details can be found in the data folder.
Data Features
A snippet of the final dataframe information can be seen here:<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3412 entries, 0 to 3411
Data columns (total 37 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 name 3412 non-null object
1 url 3412 non-null object
2 length 2193 non-null float64
3 highest point 1746 non-null float64
4 gain 2043 non-null float64
5 lat 2519 non-null float64
6 lon 2519 non-null float64
7 pass: Discover Pass 3412 non-null int64
8 pass: National Monument Fee 3412 non-null int64
9 pass: National Park Pass 3412 non-null int64
10 pass: None 3412 non-null int64
11 pass: Northwest Forest Pass 3412 non-null int64
12 pass: Sno-Parks Permit 3412 non-null int64
13 pass: Wilderness Permit 3412 non-null int64
14 Wildflowers/Meadows 3412 non-null float64
15 Dogs allowed on leash 3412 non-null float64
16 Good for kids 3412 non-null float64
17 Lakes 3412 non-null float64
18 Fall foliage 3412 non-null float64
19 Coast 3412 non-null float64
20 Mountain views 3412 non-null float64
21 Wildlife 3412 non-null float64
22 Old growth 3412 non-null float64
23 Summits 3412 non-null float64
24 Ridges/passes 3412 non-null float64
25 Established campsites 3412 non-null float64
26 Waterfalls 3412 non-null float64
27 Rivers 3412 non-null float64
28 rating 3251 non-null float64
29 votes 3251 non-null float64
30 reports 3251 non-null float64
31 description 3251 non-null object
32 drive distance 2031 non-null float64
33 drive time 2031 non-null float64
dtypes: float64(27), int64(7), object(3)
memory usage: 986.4+ KB
The meanings of the features are the following:
- length: hike length in miles
- highest point: max elevation of the hike in feet
- gain: elevation gain of the hike in feet
- lat: latitude of the hike
- lon: longitude of the hike
- pass: each of these features is binary (1 = True, 0 = False) whether this parking pass is required
- Dogs allowed on leash - Rivers: each of these features is binary (1 = True, 0 = False) whether this feature is present on this hike
- rating: 0-5 star rating of the hike
- votes: number of ratings for the hike
- reports: number of trip reports written for the hike
- description: text description of the hike
- drive distance: miles required to drive from Seattle to the hike
- drive time: minutes required to drive from Seattle to the hike
Two features of the data were identified as being related to popularity. These are reports and rating. Reports is a number that indicates how many trip reports have been written for a hike, related to how frequently a hike is visited. Rating is a 0-5 number that indicates hikers' favorability of the hike based on an aggregate of votes from the website's users. These two features are the focus of the machine learning predictions.
Among the data collected from WTA was the hike description, which provided a source for seeking out latent features of the data. Some natural language processing was performed to identify different features hidden in the text.
First, the text was vectorized, removing stopwords and creating term frequency matrices. Two processes approaches were taken to attempt to extract information from the text. A non-negative matrix factorization (NMF) was used to determine the most important latent features, and identify the associated words. Similarly, a K-Means clustering was done to group similar words and hikes together. Both produced similar results.
In just two clusters, and with a little bit of domain knowledge, one can clearly distinguish the different features in the text:
['required' 'trail' 'gear' 'experience' 'park' 'peak' 'area' 'scramble' 'summit' 'expertise' 'finding' 'route' 'trails' 'mountain' 'climbing']
['trail' 'lake' 'creek' 'miles' 'mile' 'road' 'hike' 'mountain' 'forest' 'views' 'river' 'ridge' 'feet' 'way' 'trailhead']
In this case, the first one refers to technical hiking, indicated by words like "climbing", "expertise", "gear", etc. The other cluster seems to lump together everything else.
All number of clusters for K-Means and factors for NMF provided similar results, where one category clearly identified techncial hiking, and the others were not distinguishable. Silhouette graphs of two different clusterings show this effect (with silhouette score being a measure of how appropriately grouped a data point is within a cluster). One cluster (the "technical" grouping) has a high positive score, meaning the data point is close to the center of the cluster, while all other clustered data points are negative. Five clusters for the second graph was arbitrarily chosen to highlight the affect.
This latent feature in the text was added to the data frame as a new feature called technical, hopefully to add additional insights to the predictions.
Here is a plot of hikes on a map of Washington by location, distinguished by whether they are "technical" or not, as determined by the natural language processing.
Three models were considered in making predictions of reports and rating: an ordinary least squares linear regression, a ridge regression, and a gradient boosting regression.
As a baseline point of comparison, the average and standard deviation for the reports and rating data was recorded. A model that has a "root mean squared error" that is lower than those listed here would be an imporvement in predictive capability.
| feature | mean | standard deviation |
|---|---|---|
| reports | 111 | 204 |
| rating | 3.20 | 1.19 |
An ordinary least squares linear regression model was used to attempt to predict hike reports and rating features. The model was built with the statsmodels python library. The goal of the model was to extrect feature coefficients to gain some insight into which are important for the model. Most importanly, identifying the sign and p-value of the coefficients would identify which features were effective in the model and for which direction.
This plot summarizes the coefficients with p-values less tahn 0.05 from the models for the prediction of reports and rating.
This model identifies that some features are more important than others via the magnitude of the coefficients. More importantly, the sign of the coefficients indicates the direction of their effect. For instance, waterfalls on a hike have a large positive affect, while certian parking passes and the hike being technical make a hiker less likely to travel there.
The ridge regression model was built to normalize the scale of the coefficients in the linear regression to get a better sens of their relative magnitude. This regression used a standard scaler in a pipeline, so the output coefficients are listed relative to each other.
By optimizing the alpha hyperparameter for the model of reports and rating we can glean a better understanding of feature importance with regularized coefficients.
The ridge regression adds a correction factor called alpha to the linear regression to avoid one coefficient overpowering the other. The features must be normalized to remove the effect of scale. Here are the alpha optimizations for each model.
The ridge regression models provided more insight into what features of a hike lead to the hike being more or less popular. This plot shows that the feature coefficients from modeling rating and reports generally produce the same results. The scale of the coefficients in this plot are arbitrarily scaled for visibility.
The ridge regression model error results are the following:
| feature | baseline standard deviation | average root mean squared error |
|---|---|---|
| reports | 204 | 152 |
| rating | 1.19 | 1.09 |
The idea behind the gradient boosting model is to see if there could be improvement by accounting for the nonlinearity of the data. The pitfall of this model is that while it can identify feature importance, it can't identify whether a feature has a positive or negative outcome on the prediction.
The model hyperparameters were optimized by doing a simple grid search of varying parameters.
The gradient boosting feature importances provide similar insight as the ridge regression model, however, the misalignment of features between the reports and rating model stand out more here.
The gradient boosting regression model error results are the following:
| feature | baseline standard deviation | average root mean squared error |
|---|---|---|
| reports | 204 | 158 |
| rating | 1.19 | 1.10 |
To draw some reasonable conclusions, we'll assume that consistency between the models highlights the actual importance of features in determining the populatiry of a hike.
To understand what people like and don't like about hikes, we can separate features into the ones that had a significant positive or negative impact on the models. This is done essentailly by eyeballing, in close to descending order of importance.
| Positive Impacts | Negative Impacts |
|---|---|
| Waterfalls | drive time |
| Good for Kids | length |
| Lakes | technical |
| Wildflowers/Meadows | pass: Sno-Parks Permit |
| Mountain Views | |
| Summits | |
| pass: Northwest Forest Pass |
In essence, people like to see pretty things and epic views on their hikes, but they don't like to drive very far or hike very long. A kid-friendly hike is important, and waterfalls are the jackpot. The "Northwest Forest Pass" parking permit is required for all national forest lands, which indicates that hikes in those protected lands also tend to be more popular (or maybe just more abundant).
The "technical" feature was unique in that it had a strong negative impact for "rating" but minimal impact for "reports". This could indicate that people don't have as much fun on hikes that are more challenging. The "Sno-Parks Permit" feature had a similar impact on "rating" and not on "reports". This pass is generally used for snowshoeing in winter.
With updates to the data and soem additional analysis, a next step for this project would be to create a hiking recommender. The recommender would act in real time to determine the best hike for a person at a certain time. It would have the following attributes:
- draw overall insights from all previous trip reports to include user sentiment about a hike and determine the best times of year
- pull the latest trip reports and perform sentiment analysis to determine if hiking conditions are currently good
- draw on the current weather forecast to determine if the weather will be appropriate at a given hike
