Project 2: Yelp Maps

voronoi

Let's go out to eat!
Show me places I would like
By learning my tastes.

Introduction

In this project, you will create a visualization of restaurant ratings using machine learning and the Yelp academic dataset. In this visualization, Berkeley is segmented into regions, where each region is shaded by the predicted rating of the closest restaurant (yellow is 5 stars, blue is 1 star). Specifically, the visualization you will be constructing is a Voronoi diagram.

In the map above, each dot represents a restaurant. The color of the dot is determined by the restaurant's location. For example, downtown restaurants are colored green. The user that generated this map has a strong preference for Southside restaurants, and so the southern regions are colored yellow.

This project uses concepts from Sections 2.1, 2.2, 2.3, and 2.4.3 of Composing Programs. It also introduces techniques and concepts from machine learning, a growing field at the intersection of computer science and statistics that analyzes data to find patterns and make predictions.

Download starter files

The maps.zip archive contains all the starter code and data sets. The project uses several files, but all of your changes will be made to utils.py, abstractions.py, and recommend.py.

  • abstractions.py: Data abstractions used in the project
  • recommend.py: Machine learning algorithms and data processing
  • utils.py: Utility functions for data processing
  • ucb.py: Utility functions for CS 61A
  • data: A directory of Yelp users, restaurants, and reviews
  • users: A directory of user files
  • ok: The autograder
  • proj2.ok: The ok configuration file
  • tests: A directory of tests used by ok
  • visualize: A directory of tools for drawing the final visualization

Logistics

This is a 1-week project, due on Friday, 07/13/18 at 11:59 PM. This is a solo project, so you will complete this project without a partner. You should not share your code with any other students, or copy from anyone else's solutions.

Remember that you can earn an additional bonus point by submitting the project at least 24 hours before the deadline.

The project is worth 22 points. 20 points are assigned for correctness, and 2 points for the overall composition of your program.

You will turn in the following files:

  • utils.py
  • abstractions.py
  • recommend.py

You do not need to modify or turn in any other files to complete the project. To submit the project, run the following command:

python3 ok --submit

You will be able to view your submissions on the Ok dashboard.

For the functions that we ask you to complete, there may be some initial code that we provide. If you would rather not use that code, feel free to delete it and start from scratch. You may also add new function definitions as you see fit.

However, please do not modify any other functions. Doing so may result in your code failing our autograder tests. Also, please do not change any function signatures (names, argument order, or number of arguments).

Testing

Throughout this project, you should be testing the correctness of your code. It is good practice to test often, so that it is easy to isolate any problems. However, you should not be testing too often, to allow yourself time to think through problems.

We have provided an autograder called ok to help you with testing your code and tracking your progress. The first time you run the autograder, you will be asked to log in with your Ok account using your web browser. Please do so. Each time you run ok, it will back up your work and progress on our servers.

The primary purpose of ok is to test your implementations, but there are two things you should be aware of.

First, some of the test cases are locked. To unlock tests, run the following command from your terminal:

python3 ok -u

This command will start an interactive prompt that looks like:

=====================================================================
Assignment: Yelp Maps
Ok, version ...
=====================================================================

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Unlocking tests

At each "? ", type what you would expect the output to be.
Type exit() to quit

---------------------------------------------------------------------
Question 0 > Suite 1 > Case 1
(cases remaining: 1)

>>> Code here
?

At the ?, you can type what you expect the output to be. If you are correct, then this test case will be available the next time you run the autograder.

The idea is to understand conceptually what your program should do first, before you start writing any code.

Once you have unlocked some tests and written some code, you can check the correctness of your program using the tests that you have unlocked:

python3 ok

Most of the time, you will want to focus on a particular question. Use the -q option as directed in the problems below.

We recommend that you submit after you finish each problem. Only your last submission will be graded. It is also useful for us to have more backups of your code in case you run into a submission issue.

The tests folder is used to store autograder tests, so do not modify it. You may lose all your unlocking progress if you do. If you need to get a fresh copy, you can download the zip archive and copy it over, but you will need to start unlocking from scratch.

If you do not want us to record a backup of your work or information about your progress, use the --local option when invoking ok. With this option, no information will be sent to our course servers.

Phase 0: Utilities

All changes in this phase will be made to utils.py.

Problem 0 (2 pt)

Before starting the core project, familiarize yourself with some Python features by implementing some functions in utils.py. Each function described below can be implemented in one line. These will be helpful in later parts of the project.

As you work through this phase, check your understanding of the problems by unlocking the test cases:

python3 ok -q 00 -u

Then, check your implementation by running the tests:

python3 ok -q 00

Problem 0.1: Using list comprehensions

A list comprehension constructs a new list from an existing sequence by first filtering the given sequence, and then computing an element of the result for each remaining element that is not filtered out. A list comprehension has the following syntax:

[<map expression> for <name> in <sequence expression> if <filter expression>]

For example, if we wanted to square every even integer in range(10), we could write:

>>> [x * x for x in range(10) if x % 2 == 0]
[0, 4, 16, 36, 64]

Implement map_and_filter, which takes in a sequence s, a one-argument function map_fn, and a one-argument function filter_fn. It returns a new list containing the result of calling map_fn on each element of s for which filter_fn returns a true value. Make sure your solution is only one line and uses a list comprehension.

Problem 0.2: Using min

The built-in min function takes a collection of elements (such as a list or a dictionary) and returns the collection's smallest element. The min function also takes in an optional keyword argument called key, which must be a one-argument function. The key function is called on each element of the collection, and the return values are used for comparison. For example:

>>> min([-1, 0, 1])                    # no key argument; return smallest input
-1
>>> min([-1, 0, 1], key=lambda x: x*x) # return input with the smallest square
0

Implement key_of_min_value, which takes in a dictionary d and returns the key that corresponds to the minimum value in d. This behavior differs from just calling min on a dictionary, which would return the smallest key. Make sure your solution is only one line and uses the min function.

Problem 0.3: Using zip

The zip function defined in utils.py takes multiple sequences as arguments and returns a list of lists, where the i-th list contains the i-th element of each original sequence. For example:

>>> zip([1, 2, 3], [4, 5, 6])
[[1, 4], [2, 5], [3, 6]]
>>> for triple in zip(['a', 'b', 'c'], [1, 2, 3], ['do', 're', 'mi']):
...     print(triple)
['a', 1, 'do']
['b', 2, 're']
['c', 3, 'mi']

Use the zip function to implement enumerate, which pairs the elements of a sequence with their indices, offset by a starting value. enumerate takes a sequence s and a starting value start. It returns a list of pairs, in which the i-th element is i + start paired with the i-th element of s. For example:

>>> enumerate(['maps', 21, 47], start=1)
>>> [[1, 'maps'], [2, 21], [3, 47]]

Make sure your solution is only one line and uses the zip function and a range.

Note: zip and enumerate are also built-in Python functions, but their behavior is slightly different than the versions provided in utils.py. The behavior of the built-in variants will be described later in the course.

Problem 1 (1 pt)

Implement the mean function which takes in a sequence of numbers, s, and returns the arithmetic mean, or average, of that sequence. The sequence cannot be empty: add an assert statement to ensure that empty sequences are not allowed.

Use Ok to unlock and test your code:

python3 ok -q 01 -u
python3 ok -q 01

Phase 1: Data Abstraction

All changes in this phase will be made to abstractions.py.

In this phase, we will implement data abstractions to represent a restaurant and its relevant features (such as name, location, and reviews).

Problem 2 (2 pt)

Complete the implementations of the constructor and selectors for the restaurant data abstraction in abstractions.py. Two of the data abstractions have already been completed for you: the review data abstraction and the user data abstraction. Make sure that you understand how they work.

You can use any implementation you choose, but the constructor and selectors must be defined together such that the restaurant selectors return the correct field from the constructed restaurant.

  • make_restaurant: return a restaurant constructed from five arguments:

    • name (a string)
    • location (a list containing latitude and longitude)
    • categories (a list of strings)
    • price (a number)
    • reviews (a list of review data abstractions created by make_review)
  • restaurant_name: return the name of a restaurant
  • restaurant_location: return the location of a restaurant
  • restaurant_categories: return the categories of a restaurant
  • restaurant_price: return the price of a restaurant
  • restaurant_ratings: return a list of ratings (numbers)

Use Ok to unlock and test your code:

python3 ok -q 02 -u
python3 ok -q 02

When you finish, you should be able to generate a visualization of all restaurants rated by a user. Use -u to select a user from the users directory. You can even create your own by copying one of the .dat files!

python3 recommend.py
python3 recommend.py -u one_cluster

Omitting the -u argument will default to user test_user.

You may have to refresh your browser to update the visualization.

Phase 2: Unsupervised Learning

All changes in this phase will be made to recommend.py.

Restaurants tend to appear in clusters (e.g. Southside restaurants, Downtown Berkeley restaurants, Gourmet Ghetto, etc.). In this phase, we will devise a way to group together restaurants that are close to each other into these clusters.

To do so, you will be implementing the k-means algorithm, a method for grouping data points into clusters by determining their center positions (which are called centroids).

K-means is considered an unsupervised learning method because the algorithm is not told what the correct clusters are; it must infer the clusters from the data alone.

The k-means algorithm begins by choosing k centroids at random. Then, it alternates between the following two steps:

  1. Update clusters. Group the restaurants into clusters, where each cluster contains all restaurants that are closest to the same centroid. In this step, centroid positions remain the same, but which cluster each restaurant belongs to can change.
  2. Update centroids. Compute a new centroid (average position) for each new cluster. In this step, restaurant clusters remain the same, but the centroid positions can move.

These steps are repeated to update the centroids until either an optimal list of centroids is found, or the centroid locations no longer change significantly each time (based on a maximum update threshold).

This visualization is a good way to understand how the algorithm works.

Glossary

As you complete the remaining questions, you will encounter the following terminology. Be sure to refer back here if you're ever confused about what a question is asking.

  • location: A pair containing latitude and longitude. Note that this is not a data abstraction, so we can assume its implementation is a two-element list.
  • centroid: A location (see above) that represents the center of a cluster.
  • restaurant: A restaurant data abstraction, as defined in abstractions.py.
  • cluster: A list of restaurants grouped around a centroid.
  • user: A user data abstraction, as defined in abstractions.py.
  • review: A review data abstraction, as defined in abstractions.py.
  • feature function: A single-argument function that takes a restaurant and returns a particular feature as a number, such as its mean rating or price.

Problem 3 (1 pt)

Implement find_closest, which takes a location and a sequence of centroids (locations). It returns the element of centroids closest to location.

You should use the distance function from utils.py to measure distance between locations. The distance function calculates the Euclidean distance between two locations. It has been imported for you.

If two centroids are equally close, return the one that occurs first in the sequence of centroids.

Hint: Use the min function to find the centroid with the minimum distance to location.

Use Ok to unlock and test your code:

python3 ok -q 03 -u
python3 ok -q 03

Problem 4 (2 pt)

Now, implement a helper function for the first step in the loop of the k-means algorithm, group_by_centroid, which takes a sequence of restaurants and a sequence of centroids (locations) and returns a list of clusters. Each cluster of the result is a list of restaurants that are closer to a specific centroid in centroids than any other centroid. The order of the list of clusters returned does not matter.

If a restaurant is equally close to two centroids, it is associated with the centroid that appears first in the sequence of centroids.

Hint: Use the provided group_by_first function to group together all values for the same key in a list of [key, value] pairs. You can look at the doctests to see how to use it.

Be sure not violate abstraction barriers! Test your implementation before moving on:

python3 ok -q 04 -u
python3 ok -q 04

Problem 5 (2 pt)

Implement find_centroid, which finds the centroid of a cluster (a list of restaurants) based on the locations of the restaurants. The centroid latitude is computed by averaging the latitudes of the restaurant locations. The centroid longitude is computed by averaging the longitudes.

Hint: Use the mean function from utils.py to compute the average value of a sequence of numbers.

Be sure not violate abstraction barriers! Test your implementation before moving on:

python3 ok -q 05 -u
python3 ok -q 05

Problem 6 (2 pt)

Finally, implement the full k_means algorithm. We've already filled out the first step of the algorithm, which was to randomly initialize a list of centroids. The rest of the algorithm consists of iteratively updating centroids by grouping the restaurants into clusters based on the current list of centroids and recomputing the new centroid for each cluster.

To complete the implementation, do the following in the body of the while statement:

  1. Group restaurants into clusters, where each cluster contains all restaurants closest to a particular centroid in centroids.
  2. Update centroids to contain the true centroid (i.e. average location) for each cluster.

Hint: Use the group_by_centroid and find_centroid helper functions.

These two steps repeat until an update doesn't change the list of centroids or after max_updates iterations.

Use Ok to unlock and test your code:

python3 ok -q 06 -u
python3 ok -q 06

Your visualization can indicate which restaurants are close to each other (e.g. Southside restaurants, Northside restaurants). Dots that have the same color on your map belong to the same cluster of restaurants. You can get more fine-grained groupings by increasing the number of clusters with the -k option.

python3 recommend.py -k 2
python3 recommend.py -u likes_everything -k 3

Congratulations! You've now implemented an unsupervised learning algorithm.

Phase 3: Supervised Learning

All changes in this phase will be made to recommend.py.

In this phase, you will predict what rating a user would give for a restaurant. You will implement a supervised learning algorithm that attempts to generalize from examples for which the correct rating is known, which are all of the restaurants that the user has already rated. By analyzing a user's past ratings, we can then try to predict what rating the user might give to a new restaurant. When you complete this phase, your visualization will include all restaurants, not just the restaurants that were rated by a user.

To predict ratings, you will implement simple least-squares linear regression, a widely used statistical method that approximates a relationship between some input feature (such as price) and an output value (the rating) with a line. The algorithm takes a sequence of input-output pairs and computes the slope and intercept of the line that minimizes the mean of the squared difference between the line and the outputs.

least-squares-gif

Problem 7 (3 pt)

First, let's use least-squares linear regression to write a function find_predictor that computes a predictor function for a user based on their existing restaurant ratings. A predictor function predicts a restaurant's rating based on a given feature, such as price or location. find_predictor takes in a user, a list of restaurants that have been reviewed by the user, and a feature function called feature_fn and returns two values: a predictor function and an r_squared value.

The predictor function is represented as the line y = a + bx, where y is the predicted rating for a restaurant given x, a feature of the restaurant. The r_squared value measures how accurately this line describes the original data.

To compute a, b, and r_squared, start by calculating the sums of squares S_xx, S_yy, and S_xy of the existing data (i.e. the feature values and corresponding user ratings for each restaurant in restaurants).

  • Sxx = Σi (xi - mean(x))2
  • Syy = Σi (yi - mean(y))2
  • Sxy = Σi (xi - mean(x)) (yi - mean(y))

After calculating the sums of squares, the regression coefficients (a and b) and r_squared are defined as follows:

  • b = Sxy / Sxx
  • a = mean(y) - b * mean(x)
  • R2 = Sxy2 / (Sxx Syy)

Hint: The mean and zip functions can be helpful here.

Use Ok to unlock and test your code:

python3 ok -q 07 -u
python3 ok -q 07

Problem 8 (2 pt)

Now we need a way to decide which feature is the best predictor of restaurant ratings. This can differ from user to user; for example, some users may base ratings mostly on location, while others may base them more on price.

Implement best_predictor, which takes a user, a list of restaurants, and a sequence of feature_fns. It computes a predictor function for each feature function and returns the predictor that has the highest r_squared value. All predictors are learned from the subset of restaurants reviewed by the user (called reviewed in the starter implementation).

Hint: The max function can also take a key argument, just like min.

Use Ok to unlock and test your code:

python3 ok -q 08 -u
python3 ok -q 08

Problem 9 (2 pt)

Now that we are able to find an optimal predictor function for a given user, we can compile a full collection of restaurant ratings for the user, including ratings for restaurants they haven't actually rated yet!

Implement rate_all, which takes a user, a list of restaurants, and a sequence of feature_fns. It returns a dictionary where the keys are the names of each restaurant in restaurants and the values are the corresponding ratings (numbers).

If a restaurant has already been rated by the user, rate_all will assign the restaurant the user's rating. Otherwise, rate_all will assign the restaurant the rating computed by the best predictor for the user. The best predictor is chosen using a list of feature_fns.

Hint: user_rating, implemented in abstractions.py, returns a user's rating for a restaurant given the user and restaurant_name.

Be sure not violate abstraction barriers! Test your implementation before moving on:

python3 ok -q 09 -u
python3 ok -q 09

In your visualization, you can now predict what rating a user would give a restaurant, even if they haven't rated the restaurant before. To do this, add the -p option:

python3 recommend.py -u likes_southside -k 5 -p

If you hover over each dot (a restaurant) in the visualization, you'll see a rating in parentheses next to the restaurant name.

Problem 10 (1 pt)

As a final addition to our visualization, let's add the ability to focus the visualization on a particular restaurant category by implemening search. The search function takes a category query and a sequence of restaurants. It returns all restaurants that have query as a category.

Be sure not violate abstraction barriers! Test your implementation:

python3 ok -q 10 -u
python3 ok -q 10

Congratulations, you've completed the project! The -q option allows you to filter based on a category. For example, the following command visualizes all sandwich restaurants and their predicted ratings for the user who likes_expensive restaurants:

python3 recommend.py -u likes_expensive -k 2 -p -q Sandwiches

Predicting your own ratings

Once you're done, you can use your project to predict your own ratings too! Here's how:

  1. In the users directory, you'll see a couple of .dat files. Copy one of them and rename the new file to yourname.dat (for example, john.dat).
  2. In the new file (e.g. john.dat), you'll see something like the following:

    make_user(
        'John DoeNero',     # name
        [                   # reviews
            make_review('Jasmine Thai', 4.0),
            ...
        ]

    Replace the second line with your name (as a string).

  3. Replace the existing reviews with reviews of your own! You can get a list of Berkeley restaurants with the following command:

    python3 recommend.py -r

    Rate a couple of your favorite (or least favorite) restaurants.

  4. Use recommend.py to predict ratings for you:

    python3 recommend.py -u john -k 2 -p -q Sandwiches

    (Replace john with your name.) Play around with the number of clusters (the -k option) and try different queries (with the -q option)!

How accurate is your predictor?