Description
Q1 [30 pts] Scalable singlemachine PageRank on 70M edge graph In this question, you will learn how to use your computer’s virtual memory to implement the PageRank algorithm that will scale to graph datasets with as many as billions of edges using a single computer (e.g., your laptop). As discussed in class, a standard way to work with larger datasets has been to use computer clusters (e.g., Spark, Hadoop) which may involve steep learning curves, may be costly (e.g., pay for hardware and personnel), and importantly may be “overkill” for smaller datasets (e.g., a few tens or hundreds of GBs). The virtualmemorybased approach offers an attractive, simple solution to allow practitioners and researchers to more easily work with such data (visit the NSFfunded MMap project’s homepage to learn more about the research). The main idea is to place the dataset in your computer’s (unlimited) virtual memory, as it is often too big to fit in the RAM. When running algorithms on the dataset (e.g., PageRank), the operating system will automatically decide when to load the necessary data (subset of whole dataset) into RAM. This technical approach to put data into your machine’s virtual memory space is called “memory mapping”, which allows the dataset to be treated as if it is an inmemory dataset. In your (PageRank) program, you do not need to know whether the data that you need is stored on the hard disk, or kept in RAM. Note that memorymapping a file does NOT cause the whole file to be read into memory. Instead, data is loaded and kept in memory only when needed (determined by strategies like least recently used paging and anticipatory paging). You will use the Python modules mmap and struct to map a large graph dataset into your computer’s virtual memory. The mmap() function does the “memory mapping”, establishing a mapping between a program’s (virtual) memory address space and a file stored on your hard drive we call this file a “memorymapped” file. Since memorymapped files are viewed as a sequence of bytes (i.e., a binary file), your program needs to know how to convert bytes to and from numbers (e.g., integers). struct supports such conversions via “packing” and “unpacking”, using format specifiers that represent the desired endianness and data type to convert to/from. Q1.1 Set up Pypy Install PyPy, which is a JustInTime compilation runtime for Python, which supports fast packing and unpacking. C++ and Java are generally faster than Python. However, several projects aim to boost Python speed. PyPy is one of them. Ubuntu sudo aptget install pypy MacOS Install Homebrew Run brew install pypy Windows Download the package and then install it. Run the following code in the Q1 directory to learn more about the helper utility that we have provided to you for this question. $ pypy q1_utils.py help Q1.2 Warm Up (10 pts) Get started with memory mapping concepts using the codebased tutorial in warmup.py. You should study the code and modify parts of it as instructed in the file. You can run the tutorial code asis (without any modifications) to test how it works (run “python warmup.py” on the terminal to do this). The warmup code is setup to pack the integers from 0 to 63 into a binary file, and unpack it back into a memory map object. You will need to modify this code to do the same thing for all odd integers in the range of 1 to 42. The lines that need to be updated are clearly marked. Note: You must not modify any other parts of the code. When you are done, you can run the following command to test whether it works as expected: $ python q1_utils.py test_warmup out_warmup.bin It prints True if the binary file created after running warmup.py contains the expected output. Q1.3 Implementing and running PageRank (20 pts) You will implement the PageRank algorithm, using the power iteration method, and run it on the LiveJournal dataset (an online community with millions of users to maintain journals and blogs). We recommend you revisit the MMap lecture to refresh your memory about the PageRank algorithm and the data structures and files that you may need to memorymap. (For more details, read the MMap paper.) You will perform three steps (subtasks) as described below. Step 1: Download the LiveJournal graph dataset (an edge list file) The LiveJournal graph contains almost 70 million edges. It is available on the SNAP website. We are hosting the graph, to avoid high traffic bombarding their site. Step 2: Convert the graph’s edge list to binary files (you only need to do this once) Since memory mapping works with binary files, you will convert the graph’s edge list into its binary format by running the following command at the terminal/command prompt: $ python q1_utils.py convert <pathtoedgelist.txt> Example: Consider the following toygraph.txt, which contains 7 edges: 0 1 1 0 1 2 2 1 3 4 4 5 5 2 To convert the graph to its binary format, you will type: $ python q1_utils.py convert toygraph/toygraph.txt This generates 3 files: toygraph/ toygraph.bin: binary file containing edges (source, target) in littleendian “int” C type toygraph.idx: binary file containing (node, degree) in littleendian “long long” C type toygraph.json: metadata about the conversion process (required to run pagerank) In toygraph.bin we have, 0000 0000 0100 0000 # 0 1 (in littleendian “int” C type) 0100 0000 0000 0000 # 1 0 0100 0000 0200 0000 # 1 2 0200 0000 0100 0000 # 2 1 0300 0000 0400 0000 # 3 4 0400 0000 0500 0000 # 4 5 0500 0000 0200 0000 # 5 2 ffff ffff ffff ffff … ffff ffff ffff ffff ffff ffff ffff ffff In toygraph.idx we have, 0000 0000 0000 0000 0100 0000 0000 0000 # 0 1 (in littleendian “long long” C type ) 0100 0000 0000 0000 0200 0000 0000 0000 # 1 2 … ffff ffff ffff ffff ffff ffff ffff ffff Note: there are extra values of 1 (ffff ffff or ffff ffff ffff ffff) added at the end of the binary file as padding to ensure that the code will not break in case you try to read a value greater than the file size. You can ignore these values as they will not affect your code. Step 3: Implement and run the PageRank algorithm on LiveJournal graph’s binary files Follow the instructions in pagerank.py to implement the PageRank algorithm. You will only need to write/modify a few lines of code. Next, run the following command to execute your PageRank implementation: $ pypy q1_utils.py pagerank This will output the 10 nodes with the highest PageRank scores. For example: $ pypy q1_utils.py pagerank toygraph/toygraph.json node_id score 1 0.4106875 2 0.2542078125 0 0.1995421875 5 0.0643125 4 0.04625 3 0.025 (Note that only 6 nodes are printed here since the toy graph only has 6 nodes.) Step 4: Experiment with different number of iterations. Find the output for the top 10 nodes for the LiveJournal graph for n=10, 25, 50 iterations (try the iterations n argument in the command above; the default number of iterations is 10). A file in the format pagerank_nodes_n.txt for “n” number of iterations will be created. For example: $ pypy q1_utils.py pagerank toygraph/toygraph.json iterations 25 You may notice that while the top nodes’ ordering starts to stabilize as you run more iterations, the nodes’ PageRank scores may still change. The speed at which the PageRank scores converge depends on the PageRank vector’s initial values. The closer the initial values are to the actual pagerank scores, the faster they converge. Deliverables 1. warmup.py [6pt]: your modified implementation. 2. out_warmup.bin [3pt]: the binary file, automatically generated by your modified warmup.py. 3. out_warmup_bytes.txt [1pt]: the text file with the number of bytes, automatically generated by your modified warmup.py. 4. pagerank.py [14pt]: your modified implementation. 5. pagerank_nodes_n.txt [6pt]: the 3 files (as given below) containing the top 10 node IDs and their pageranks for n iterations, automatically generated by q1_utils.py. ○ pagerank_nodes_10.txt [2pt] for n=10 ○ pagerank_nodes_25.txt [2pt] for n=25 ○ pagerank_nodes_50.txt [2pt] for n=50 Q2 [50 pts] Random Forest Classifier Note: You must use Python 3.x for this question. You will implement a random forest classifier in Python. The performance of the classifier will be evaluated via the outofbag (OOB) error estimate, using the provided dataset. Note: You must not use existing machine learning or random forest libraries like scikitlearn. The dataset you will use is extracted from the UCI Bank Marketing dataset where each record is data related with direct marketing campaigns of a Portuguese banking institution. The dataset has been cleaned to remove missing attributes. The data is stored in a commaseparated file (csv) in your Q2 folder as hw4data.csv. Each line describes an instance using 20 columns: the first 19 columns represent the attributes of the application, and the last column is the ground truth label for the term deposit subscription (0 means “not subscribed”, 1 means “subscribed”). Note: The last column should not be treated as an attribute. You will perform binary classification on the dataset to determine if a client will subscribe to a term deposit ot not. Essential Reading Decision Trees To complete this question, you need to develop a good understanding of how decision trees work. We recommend you review the lecture on decision tree. Specifically, you need to know how to construct decision trees using Entropy and Information Gain to select the splitting attribute and split point for the selected attribute. These slides from CMU (also mentioned in lecture) provide an excellent example of how to construct a decision tree using Entropy and Information Gain. Random Forests To refresh your memory about random forests, see Chapter 15 in the “Elements of Statistical Learning” book and the lecture on random forests. Here is a blog post that introduces random forests in a fun way, in layman’s terms. OutofBag Error Estimate In random forests, it is not necessary to perform explicit crossvalidation or use a separate test set for performance evaluation. Outofbag (OOB) error estimate has shown to be reasonably accurate and unbiased. Below, we summarize the key points about OOB described in the original article by Breiman and Cutler. Each tree in the forest is constructed using a different bootstrap sample from the original data. Each bootstrap sample is constructed by randomly sampling from the original dataset with replacement (usually, a bootstrap sample has the same size as the original dataset). Statistically, about onethird of the cases are left out of the bootstrap sample and not used in the construction of the kth tree. For each record left out in the construction of the kth tree, it can be assigned a class by the kth tree. As a result, each record will have a “test set” classification by the subset of trees that treat the record as an outofbag sample. The majority vote for that record will be its predicted class. The proportion of times that a predicted class is not equal to the true class of a record averaged over all records is the OOB error estimate. Starter Code We have prepared starter code written in Python for you to use. This would help you load the data and evaluate your model. The following files are provided for you: ● util.py: utility functions that will help you build a decision tree ● decision_tree.py: a decision tree class that you will use to build your random forest ● random_forest.py: a random forest class and a main method to test your random forest What you will implement Below, we have summarized what you will implement to solve this question. Note that you MUST use information gain to perform the splitting in the decision tree. The starter code has detailed comments on how to implement each function. 1. util.py: implement the functions to compute entropy, information gain, and perform splitting. 2. decision_tree.py: implement the learn() method to build your decision tree using the utility functions above. 3. decision_tree.py: implement the classify() method to predict the label of a test record using your decision tree. 4. random_forest.py: implement the functions _bootstrapping(), fitting(), voting() Note: You must achieve a minimum accuracy of 80% for random_forest. As you solve this question, you will need to think about multiple parameters in your design, some may be more straightforward to determine, some may be not (hint: study lecture slides and essential reading above). For example, ● Which attributes to use when building a tree? ● How to determine the split point for an attribute? ● When do you stop splitting leaf nodes? ● How many trees should the forest contain? Note that, as mentioned in lecture, there are other approaches to implement random forests. For example, instead of information gain, other popular choices include Gini index, random attribute selection (e.g., PERT Perfect Random Tree Ensembles). We decided to ask everyone to use an information gain based approach in this question (instead of leaving it openended), to help standardize students solutions to help accelerate our grading efforts. Deliverables 1. hw4data.csv: The dataset used to develop your program. Do not modify this file. 2. util.py [10 pts]: The source code of your utility functions. 3. decision_tree.py [30 pts]: The source code of your decision tree implementation. 4. Random_forest.py [10 pts]: The source code of your random forest implementation with appropriate comments. Q3 [30 points] Using ScikitLearn Note: You must use Python 3.x for this question. Scikitlearn is a popular Python library for machine learning. You will use it to train some classifiers on the Epileptic Seizure Recognition[1] dataset in the folder, called seizure_dataset.csv. Q3.1 Classifier Setup [7 pts] Train each of the following classifiers on the dataset, using the classes provided in the links below. You will do hyperparameter tuning in Q3.2 to get the best accuracy for each classifier on the dataset. 1. Linear Regression 2. MultiLayer Perceptron 3. Random Forest 4. Support Vector Machine (The link points to SVC, which is a particular implementation of SVM by scikit.) Scikit has additional documentation on each of these classes, explaining them in more detail, such as how they work and how to use them. Use the skeleton file called hw4q3.py to write your code. In report.txt, under section Q3.1, follow the skeleton and put your training and testing accuracies for each classifier. Report your accuracies as percentages and round them to the nearest whole number, e.g., 85%. As a reminder, the general flow of your machine learning code will look like: 1. Load dataset 2. Preprocess (you will do this in Q3.2) 3. Split the data into x_train, y_train, x_test, y_test 4. Train the classifier on x_train and y_train 5. Predict on x_test 6. Evaluate testing accuracy by comparing the predictions from step 5 with y_test. Here is an example. Scikit has many other examples as well that you can learn from. Q3.2 Hyperparameter Tuning [17 pts] Tune your Random Forest and SVM to obtain their best accuracies on the dataset. For Random Forest, tune the model on the unmodified test and train datasets. For SVM, either standardize or normalize the dataset before using it to tune the model. Note: If you are using StandardScaler: Pass x_train into the fit method. Then transform both x_train and x_test to obtain the standardized versions of both. The reason we fit only on x_train and not the entire dataset is because we do not want to train on data that was affected by the testing set. Tune the hyperparameters specified below, using the GridSearchCV function that Scikit provides: For random forest, tune the parameters “n_estimators” and “max_depth”. For SVM, tune “C” and “kernel” (try only ‘linear’ and ‘rbf’). Use 10 folds by setting the cv parameter to 10. You should test at least 3 values for each of the numerical parameters. For C, the values should be different by factors of at least 10, for example, 0.001, 0.01, and 0.1, or 0.0001, 0.1 and 100. In section Q3.2 of report.txt, state the values you tested for each hyperparameter. Also follow the skeleton in report.txt to report the best combination of hyperparameter values for each classifier tuned, its testing accuracy from Q3.1, and its best testing accuracy from tuning. For each classifier the best testing accuracy from tuning should be at least as high as the testing accuracy from Q3.1. Note: If GridSearchCV is taking a long time to run for SVM, make sure you are standardizing or normalizing your data beforehand. Q3.3 CrossValidation Results [2 pts] Let’s practice getting the results of crossvalidation. For your SVM (only), report the mean training score, mean testing score and mean fit time for the best combination of hyperparameter values that you obtained in Q3.2. The GridSearchCV class holds a ‘cv_results_’ dictionary that should help you report these metrics easily. Report the metrics in report.txt under the Q3.3 section. Report your accuracies as percentages and round them to the nearest whole number, for example 85%. Q3.4 Best Classifier [4 pts] Out of all 4 classifiers (for Random Forest and SVM take the best one from GridSearchCV for each), assess which one performed the best. Use testing accuracies, fit time or a combination of both in your reasoning. Put your explanation in report.txt under section Q3.4, using at most 50 words. Deliverables report.txt A text file containing your results and explanations for all parts. hw4q3.py Skeleton file filled with your code from Q3.1Q3.3. seizure_dataset.csv the original dataset. Submission Guidelines Submit the deliverables as a single zip file named HW4{GT account username}.zip. Write down the name(s) of any students you have collaborated with on this assignment, using the text box on the Canvas submission page. The zip file’s directory structure must exactly be (when unzipped): HW4{GT account username}/ Q1/ warmup.py out_warmup.bin out_warmup_bytes.txt pagerank.py pagerank_nodes_10.txt pagerank_nodes_25.txt pagerank_nodes_50.txt Q2/ hw4data.csv util.py decision_tree.py random_forest.py Q3/ report.txt hw4q3.py seizure_dataset.csv You must follow the naming convention specified above. Version 2 [1] Derived from https://archive.ics.uci.edu/ml/datasets/Epileptic+Seizure+Recognition 通过Google云端硬盘发布 – 举报滥用行为 – 每5分钟自动更新一次
