In this repository, we provide implementations of several data mining algorithms, including:

• 1. MapReduce for finding nodes with many mutual neighbours
• 2. A-Priori algorithm for frequent itemset mining
• 3. MinHash-based LSH for finding similar documents
• 4. k-Means algorithm using Spark
• 5. Collaborative filtering
• 6. PageRank algorithm using Spark
• 7. Efficient triangle counting
• 8. Gradient Descent SVM
• 9. DGIM algorithm

The implementations are written from scratch in Python and Spark.

Notes

Due to privacy issues, this repository only includes the final code and documentation without any commit history.

Some of the implementations in this repository require Spark, which can be installed with the following steps:

1. Download Spark

wget https://archive.apache.org/dist/spark/spark-3.1.2/spark-3.1.2-bin-hadoop2.7.tgz

2. Unzip the downloaded file

tar -zxvf spark-3.1.2-bin-hadoop2.7.tgz

3. Check that the Spark binary /spark-3.1.2-bin-hadoop2.7/bin/spark-submit exists
4. Install Open JDK 8

$ sudo apt-get update
$ sudo apt-get install openjdk-8-jdk

In this section, we describe the implemented algorithms and provide example usage.

Given a graph, we want to find pairs of nodes that have the highest number of mutual neighbours. We implement this algorithm using MapReduce.

The input file is a text file with each line representing the list of neighbours of a node:

<Node><TAB><Neighbours>

We provide a sample dataset in data/soc-LiveJournal1Adj.txt.

The output contains the 10 pairs with the highest number of mutual neighbours, where each line is in the format

<Node1><TAB><Node2><TAB><CountOfMutualNeighbours>

In order to run the algorithm on our provided dataset, run the following command:

bin/spark-submit code/mutual_neighbour.py data/soc-LiveJournal1Adj.txt

You can compare the result with an alternative, non-MapReduce algorithm we implemented by running the following command:

python code/mutual_neighbour_scipy.py data/soc-LiveJournal1Adj.txt

Given a collection of itemsets, we find frequent itemsets using the A-Priori algorithm.

Each line in the input file is a set of items separated by a spaces

<Item1><SPACE><Item2><SPACE><Item3> ...

We provide a sample dataset in data/browsing.txt.

The output contains the number of frequent items, the number of frequent pairs, and the 10 most frequent pairs:

<NumberOfFrequentItems>
<NumberOfFrequentPairs>
<Item><TAB><Item><TAB><Count>
...
<Item><TAB><Item><TAB><Count>

In order to run the algorithm on our provided dataset, run the following command:

python code/a_priori.py data/browsing.txt

Given a collection of documents, we find similar documents using the MinHash-based locality-sensitive hashing (LSH) algorithm.

Each line in the input file has the following format:

<DocumentID><SPACE><Document>

We provide a sample dataset in data/articles.txt.

The output contains document pairs with similarity greater than 0.9, where each line has the format

<DocumentID1><TAB><DocumentID2>

In order to run the algorithm on our provided dataset, run the following command:

python code/lsh.py data/articles.txt

We implement the k-Means algorithm on large datasets using Spark.

Each line in the input file has the following format:

<FEATURE 1><SPACE><FEATURE 2>...

We provide a sample dataset in data/kmeans.txt.

The output contains a single line with the average diameter of the created clusters:

<AverageDiameter>

In order to run the algorithm on our provided dataset, run the following command:

bin/spark-submit code/kmeans.py data/kmeans.txt k_value

We implement item-based and user-based collaborative filtering.

Each line in the input file has the following format:

<USER ID>,<MOVIE ID>,<RATING>,<TIMESTAMP>

We provide a sample dataset in data/ratings.txt.

The output contains 10 lines, where the first 5 lines and the last 5 lines contain the top 5 recommendations for the user with ID 600 using item-based and user-based collaborative filtering, respectively:

<MOVIE ID><TAB><PREDICTED RATING>

In order to run the algorithm on our provided dataset, run the following command:

python code/cf.py data/ratings.txt

We implement the PageRank algorithm using Spark.

Each line in the input file has the following format:

<Source Node><TAB><Destination Node>

We provide a sample dataset in data/graph.txt.

The output contains 10 lines, corresponding to the top 10 nodes with the highest PageRank score:

<Node><TAB><PageRank Score>

In order to run the algorithm on our provided dataset, run the following command:

bin/spark-submit code/pagerank.py graph.txt

We implement an efficient algorithm for counting triangles in a large graph, where brute-force algorithms are not feasible.

Each line in the input file has the following format:

<Node1><TAB><Node2><TAB><TimeStamp>

We provide a sample dataset in data/facebook.txt.

The output contains a single line with the number of triangles:

<NumberOfTriangles>

In order to run the algorithm on our provided dataset, run the following command:

python code/triangle.py data/facebook.txt

We implement a gradient descent SVM classifier.

There is one file containing the features and one file containing the labels. Each line in the features file has the following format:

<Feature 1>,<Feature 2>...

Each line in the labels file contains a single label that is either 1 or -1. We provide a sample dataset in data/svm-features.txt and data/svm-labels.txt.

The output contains three lines with the 10-fold cross validation accuracy of the model and the optimal C and eta parameters for learning the model.

<Accuracy>
<C>
<eta>

In order to run the algorithm on our provided dataset, run the following command:

python code/svm.py data/svm-features.txt data/svm-labels.txt

For hyperparameter tuning, you can use the functions svm.grid_search() and plot the results using svm.plot_grid_search():

We implement the Datar-Gionis-Indyk-Motwani algorithm for estimating the number of 1s in the last k elements of a binary stream.

Each line of the input contains either a 0 or a 1. We provide a sample dataset in data/stream.txt.

Each line of the output contains the estimated number of 1s in the last k elements, where k is a parameter of the algorithm to be specified by the user (see example usage)

In order to run the algorithm on our provided dataset, run the following command:

python code/dgim.py data/dgim.txt k1 k2 k3 ...

Each ki is a parameter of the algorithm.

Anand Rajaraman, Jure Leskovec, and Jeffrey D. Ullman. Mining of Massive Datasets