DSCI553 Foundations and Applications of Data Mining

Fall 2024
Assignment 3
Deadline: Wednesday, October 30, 11:59 PM PST

1. Overview of the Assignment

In Assignment 3, you will complete two tasks. The goal is to familiarize you with Locality Sensitive
Hashing (LSH), and different types of collaborative-filtering recommendation systems. The dataset you
are going to use is a subset from the Yelp dataset used in the previous assignments.

2. Assignment Requirements
2.1 Programming Language and Library Requirements
a. You must use Python to implement all tasks. You can only use standard python libraries (i.e., external
libraries like numpy or pandas are not allowed). There will be a 10% bonus for each task (or case) if you
also submit a Scala implementation and both your Python and Scala implementations are correct.
b. You are required to only use the Spark RDD to understand Spark operations. You will not receive any
points if you use Spark DataFrame or DataSet.

2.2 Programming Environment

Python 3.6, JDK 1.8, Scala 2.12, and Spark 3.1.2
We will use these library versions to compile and test your code. There will be no point if we cannot run
your code on Vocareum. On Vocareum, you can call `spark-submit` located at /opt/spark/spark-3.1.2-
bin-hadoop3.2/bin/spark-submit`. *
*Do not use the one at `/home/local/spark/latest/bin/spark-submit (2.4.4)

2.3 Write your own code

Do not share your code with other students!!
We will combine all the code we can find from the Web (e.g., GitHub) as well as other students’ code
from this and other (previous) sections for plagiarism detection. We will report all the detected
plagiarism.

3. Yelp Data
In this assignment, the datasets you are going to use are from:
https://drive.google.com/drive/folders/1SufecRrgj1yWMOVdERmBBUnqz0EX7ARQ?usp=sharin
g
Make sure you are logged in with your USC gmail to view this folder.
We generated the following two datasets from the original Yelp review dataset with some filters. We
randomly took 60% of the data as the training dataset, 20% of the data as the validation dataset, and
20% of the data as the testing dataset.
● yelp_train.csv: the training data, which only include the columns: user_id, business_id, and
stars.
● yelp_val.csv: the validation data, which are in the same format as training data.
● other datasets: providing additional information (like the average star or location of a business)
Note: We are not sharing the test dataset. The goal of this assignment is to create a generalizable
recommender that works well on the unseen test data.

4. Tasks
Note: This Assignment has been divided into 2 parts on Vocareum. This has been done to
provide more computational resources.

4.1 Task 1: Jaccard based LSH (2 points)

In this task, you will implement the Locality Sensitive Hashing algorithm with Jaccard similarity using
yelp_train.csv.
In this task we focus on whether or not a user has reviewed a business, binary “0 or 1” ratings, rather
than the actual stars ratings. Specifically, if a user has rated a business, the user’s contribution in the
characteristic matrix is 1. If the user hasn’t rated the business, the contribution is 0. You need to
identify similar businesses whose similarity >= 0.5.
You can define any collection of hash functions that you think would result in a consistent permutation
of the row entries of the characteristic matrix. Some potential hash functions are:
f(x)= (ax + b) % m or f(x) = ((ax + b) % p) % m
where p is any prime number and m is the number of bins. Please carefully design your hash functions.
After you have defined all the hash functions, you will build the signature matrix. Then you will divide
the matrix into b bands with r rows each, where b x r = n (n is the number of hash functions). You should
carefully select a good combination of b and r in your implementation (b>1 and r>1). Remember that
two items are a candidate pair if their signatures are identical in at least one band.
Your final results will be the candidate pairs whose original Jaccard similarity is >= 0.5. You need to write
the final results into a CSV file according to the output format below in Table 1:

user1 user2 user3 user4

business1 0 1 1 1
business2 0 1 0 0
Jaccard Similarity (business1, business2) = #intersection / #union = 1/3

Table 1: Example of Jaccard Similarity

Input format: (we will use the following command to execute your code)
Python: /opt/spark/spark-3.1.2-bin-hadoop3.2/bin/spark-submit task1.py <input_file_name>
<output_file_name>
Scala: spark-submit --class task1 hw3.jar <input_file_name> <output_file_name>
Param: input_file_name: the name of the input file (yelp_train.csv), including the file path.
Param: output_file_name: the name of the output CSV file, including the file path.
Output format:
IMPORTANT: Please strictly follow the output format since your code will be graded automatically. We
will not regrade because of formatting issues.
a. The output file is a CSV file, containing all business pairs you have found. The header is
“business_id_1, business_id_2, similarity”. Each pair itself must be sorted in lexicographical order. This
means if there are 2 businesses with business IDs “abd” and “abc” respectively and their similarity is 0.7,
the output file should have
abc, abd, 0.7
And not
abd, abc, 0.7
The entire file also needs to be sorted in lexicographical order. There is no requirement for the number
of decimals for the similarity value. Please refer to the format in Figure 1.

Figure 1: a CSV output example for task1

Grading:
We will compare your output file against the ground truth file using precision and recall metrics.
Precision = true positives / (true positives + false positives)
Recall = true positives / (true positives + false negatives)
The ground truth file has been provided in the Google drive, named as “pure_jaccard_similarity.csv”.
You can use this file to compare your results to the ground truth as well.
The ground truth dataset only contains the business pairs (from the yelp_train.csv) whose Jaccard
similarity >=0.5. The business pair itself is sorted in the alphabetical order, so each pair only appears
once in the file (i.e., if pair (a, b) is in the dataset, (b, a) will not be there).
In order to get full credit for this task you should have precision >= 0.99 and recall >= 0.97. If not, then
you will get only partial credit based on the formula:
(Precision / 0.99) * 0.4 + (Recall / 0.97) * 0.4
Your runtime should be less than 100 seconds. If your runtime is more than or equal to 100 seconds,
you will not receive any point for this task.
4.2 Task 2: Recommendation Systems (5 points)
In task 2, you are going to build different types of recommendation systems using the yelp_train.csv to
predict the star ratings for given user ids and business ids. You can make any improvement to your
recommendation system in terms of speed and accuracy. You can use the validation dataset
(yelp_val.csv) to evaluate the accuracy of your recommendation systems, but please don’t include it as
your training data.
Here are some possible options to evaluate your recommendation systems (there are many more ways):
1. You can compare your results to the corresponding ground truth and compute the absolute
differences then take the mean across all data points. This is called Mean Absolute Error (MAE).
Example:
Predicted rating: 4.73 True rating: 3.5 abs(True - Predicted) = abs(3.5 - 4.73) = 1.23
2. You can divide the absolute differences into 5 levels and count the number for each level as
following example:
Rating diff. Count
>=0 and <1 n = 12,345
>=1 and <2 n = 3,506
>=2 and <3 n = 1,189
>=3 and <4 n = 36
>=4 n = 17
As calculated in example 2, there are 12,345 predictions with < 1 difference from the ground
truth. This is a good sign since most predictions are within the smallest possible error window.
Through method 2 you will be able to calculate the error distribution of your predictions and
improve the performance of your recommendation systems.
3. You can compute the Root Mean Squared Error (RMSE) by using following formula:

Where Predi is the prediction for business i and Ratei is the true rating for business i. n is the
total number of the business you are predicting. Squared error is similar to absolute error, but
will penalize large deviations from the predicted values harsher than absolute error.

In this task, you are required to implement:

Case 1: Item-based Collaborative Filtering (CF) recommendation system with Pearson similarity (2
points)
Case 2: Model-based recommendation system (RS) (1 point)
Case 3: Hybrid recommendation system (2 points)
4.2.1. Item-based CF recommendation system
Please strictly follow the slides to implement an item-based CF recommendation system with Pearson
similarity.
Note: Since it is a CF-based recommendation system, there are some inherent limitations to this
approach, like cold-start. You need to come up with a default rating mechanism for such cases. This
includes cases where the user or the business does not exist in the training dataset but is present in the
test dataset. This is a part of the assignment and you are supposed to come up with ways to handle such
issues on your own.
4.2.2. Model-based recommendation system
You should use XGBregressor (a regressor based on Decision Tree) to train a model. This is the
documentation for the API:
https://xgboost.readthedocs.io/en/release_0.72/index.html
XGBRegressor is inside the package xgboost.
Please use version 0.72 of the xgboost package on your local system to avoid any discrepancies you
might see between the results on your local system and Vocareum.
Please think like a real-world Data Scientist and select/develop your own features from the provided
extra datasets. For example, the average stars rated by a user and the number of reviews most likely
influence the prediction result. You need to select other features and train a model based on that. Use
the validation dataset to validate your result and remember don’t include it in your training data. This
will help you understand the generalizability of your model and improve performance on test data.

4.2.3. Hybrid recommendation system.

Now that you have the results from previous models, you will need to choose a way (from the slides) to
combine them together and design a better hybrid recommendation system.
Here are two examples of hybrid systems:
Method 1:
You can combine them together as a weighted average, which means:
𝑓𝑖𝑛𝑎𝑙 𝑠𝑐𝑜𝑟𝑒 = 𝛼 × 𝑠𝑐𝑜𝑟𝑒𝑖𝑡𝑒𝑚_𝑏𝑎𝑠𝑒𝑑 + (1 − 𝛼) × 𝑠𝑐𝑜𝑟𝑒𝑚𝑜𝑑𝑒𝑙_𝑏𝑎𝑠𝑒𝑑
The key idea is: 𝛼 represents the weight of each component rating. You will need to determine the best
𝛼 value for the hybrid system. You can do that using domain knowledge about the system. For example,
item-based CF focuses on the neighbors of the item and the model-based RS focuses on features of the
users and items themselves. Specifically, if the item has a smaller number of neighbors, then the weight
of the CF should be smaller. Meanwhile, if two restaurants both have 4 star average ratings but the first
one has 10 reviews and the second one has 1000 reviews, the average star of the second one is more
trustworthy, so the model-based RS score should weigh more. You may need to find other features to
generate your own weight function to combine them together. You can also test different values of 𝛼
with the help of the error metrics listed above.
Method 2:
You can combine them together as a classification problem:
Again, the key idea is: the item-based CF focuses on the neighbors of the item and the model-based RS
focuses on the user and items themselves. As a result, in our dataset, some item-user pairs are more
suitable for the CF while the others are not. You need to choose some features to classify which model
you should choose for each item-user pair.
Some possible features (other features may also work):
● Average stars of a user, average stars of business, the variance of history review of a user or a
business.
● Number of reviews of a user or a business.
● Yelp account starting date, number of fans.
● The number of people who think a users’ review is useful/funny/cool. Number of compliments
(Be careful with these features. For example, sometimes when I visit a horrible restaurant, I will
give full stars because I hope I am not the only one who wasted money and time here.
Sometimes people are satirical. :-))

Input format: (we will use the following commands to execute your code)
Case 1:
spark-submit task2_1.py <train_file_name> <test_file_name> <output_file_name>
Param: train_file_name: the name of the training file (e.g., yelp_train.csv), including the file path
Param: test_file_name: the name of the testing file (e.g., yelp_val.csv), including the file path
Param: output_file_name: the name of the prediction result file, including the file path
Case 2:
spark-submit task2_2.py <folder_path> <test_file_name> <output_file_name>
Param: folder_path: the path of dataset folder, which contains exactly the same file as the google drive.
Param: test_file_name: the name of the testing file (e.g., yelp_val.csv), including the file path
Param: output_file_name: the name of the prediction result file, including the file path
Case 3:
spark-submit task2_3.py <folder_path> <test_file_name> <output_file_name>
Param: folder_path: the path of dataset folder, which contains exactly the same file as the google drive.
Param: test_file_name: the name of the testing file (e.g., yelp_val.csv), including the file path
Param: output_file_name: the name of the prediction result file, including the file path

Output format:
a. The output file is a CSV file, containing all the prediction results for each user and business pair in the
validation/testing data. The header is “user_id, business_id, prediction”. There is no requirement for the
order in this task. There is no requirement for the number of decimals for the similarity values. Please
refer to the format in Figure 2.

Figure 2: Output example in CSV for task2

Grading:
We will compare your prediction results against the ground truth. We will grade all the cases in Task2
based on your accuracy using RMSE. For your reference, the table below shows the RMSE baselines and
running time for predicting the validation data. The time limit of case3 is set to 30 minutes because we
hope you consider this factor and try to improve on it as much as possible (hint: this will help you a lot in
the competition project at the end of the semester).
Case 1 Case 2 Case 3
RMSE 1.09 1.00 0.99
Running Time 130s 400s 1800s
For grading, we will use the test data to evaluate your recommendation systems. If you can pass the
RMSE baselines in the above table, you should be able to pass the RMSE baselines for the test data.
However, if your recommendation system only passes the RMSE baselines for the validation data, you
will receive 50% of the points for each case.

5. Submission
You need to submit following files on Vocareum with exactly the same name:
a. Four Python scripts:
● task1.py
● task2_1.py
● task2_2.py
● task2_3.py
b. [OPTIONAL] hw3.jar and Four Scala scripts:
● task1.scala
● task2_1.scala
● task2_2.scala
● task2_3.scala
6. Grading Criteria
(% penalty = % penalty of possible points you get)
1. You can use your free 5-day extension separately or together.
a. Google Forms Link for Extension: https://forms.gle/S7nsS1QXKe2bysvC6
2. There will be a 10% bonus if you use both Scala and Python.
3. We will combine all the code we can find from the web (e.g., Github) as well as other students’ code
from this and other (previous) sections for plagiarism detection. If plagiarism is detected, you will
receive no points for the entire assignment and we will report all detected plagiarism.
4. All submissions will be graded on Vocareum. Please strictly follow the format provided, otherwise
you won’t receive points even though the answer is correct.
5. If the outputs of your program are unsorted or partially sorted, there will be a 50% penalty.
6. Do NOT use Spark DataFrame, DataSet, sparksql.
7. We can regrade your assignments within seven days once the scores are released. We will not
accept any regrading requests after a week. There will be a 20% penalty if our grading is correct.
8. There will be a 20% penalty for late submissions within a week and no points after a week.
9. Only if your results from Python are correct will the bonus of using Scala be calculated. There are no
partial points awarded for Scala. See the example below:

Example situations
Score for Scala
Task Score for Python Total
(10% of previous column if correct)
Task
Correct: 3 points Correct: 3 * 10% 3.3
1
Task
Wrong: 0 point Correct: 0 * 10% 0.0
1
Task
Partially correct: 1.5 points Correct: 1.5 * 10% 1.65
1
Task
Partially correct: 1.5 points Wrong: 0 1.5
1

7. Common problems causing fail submission on Vocareum/FAQ

(If your program runs seem successfully on your local machine but fail on Vocareum, please check these)
1. Try your program on Vocareum terminal. Remember to set python version as python3.6,

And use the latest Spark

/opt/spark/spark-3.1.2-bin-hadoop3.2/bin/spark-submit
2. Check the input command line format.
3. Check the output format, for example, the header, tag, typos.
4. Check the requirements of sorting the results.
5. Your program scripts should be named as task1.py task2.py etc.
6. Check whether your local environment fits the assignment description, i.e. version, configuration.
7. If you implement the core part in Python instead of Spark, or implement it in a high time complexity
way (e.g. search an element in a list instead of a set), your program may be killed on Vocareum because
it runs too slowly.
8. You are required to only use Spark RDD in order to understand Spark operations more deeply. You will
not get any points if you use Spark DataFrame or DataSet. Don’t import sparksql.
9. Do not use Vocareum for debugging purposes, please debug on your local machine. Vocareum can
be very slow if you use it for debugging.
10. Vocareum is reliable in helping you to check the input and output formats, but its function on
checking the code correctness is limited. It can not guarantee the correctness of the code even with
a full score in the submission report.
11. Some students encounter an error like: the output rate …. has exceeded the allowed
value ….bytes/s; attempting to kill the process.
To resolve this, please remove all print statements and set the Spark logging level such that
it limits the logs generated - that can be done using sc.setLogLevel . Preferably, set the log level to
either WARN or ERROR when submitting your code. (e.g. add sc.setLogLevel('ERROR') to your code)