<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Julie Yin on Medium]]></title>
        <description><![CDATA[Stories by Julie Yin on Medium]]></description>
        <link>https://medium.com/@julie.yin?source=rss-fb56e9a16f2c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*bhvacH9nGGwVHh1zO2pAuw.jpeg</url>
            <title>Stories by Julie Yin on Medium</title>
            <link>https://medium.com/@julie.yin?source=rss-fb56e9a16f2c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 18:06:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@julie.yin/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Twelve Essential Docker Commands for Beginners]]></title>
            <link>https://medium.com/@julie.yin/twelve-essential-docker-commands-for-beginners-2c28dd4aace6?source=rss-fb56e9a16f2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/2c28dd4aace6</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[docker]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[amazon-web-services]]></category>
            <dc:creator><![CDATA[Julie Yin]]></dc:creator>
            <pubDate>Wed, 24 Oct 2018 03:16:03 GMT</pubDate>
            <atom:updated>2018-10-26T22:01:38.073Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*op11vRA6Z3jCRmdWXwxPkw.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/photos/r3SGf7uvOA0?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Paola Aguilar</a> on <a href="https://unsplash.com/search/photos/whale?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>Docker has become more and more trendy in recent years. Compared to a traditional virtual machine, it is more lightweight and portable to use. However, the learning curve can be a little steep for people who are not familiar with this type of concept. Below are twelve basic commands for you to get started.</p><p>The “<strong>docker images</strong>” command lists all the docker images.</p><p>I deleted all the docker images stored on my local machine before I ran “docker images”, so my output looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UQStVg9tHznF1Uex49q5og.png" /></figure><p>Now let’s pull the <a href="https://hub.docker.com/_/ubuntu/">official Docker image of ubuntu</a> from <a href="https://hub.docker.com/">Docker Hub</a> by issuing “<strong>docker pull</strong>”:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dsxyi-FIf87_lHtQJD2uAQ.png" /></figure><p>Then run “docker images” again:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*giSGXtItRhTp06RhvwDpcA.png" /></figure><p>The docker image of ubuntu we just pulled is now listed with all the image name, ID, created date and size information.</p><p>We can also create a Docker container based on this image by issuing “<strong>docker run</strong>”:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GwmR0aJD7cA5e39KaL9Z8Q.png" /></figure><p>The -it flag allows us to run an interactive bash shell inside the container. To understand the difference between a Docker container and a Docker image, the following video could be helpful:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FYFl2mCHdv24%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DYFl2mCHdv24&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FYFl2mCHdv24%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/0ce208f319de4672bf9750db98ff2fcc/href">https://medium.com/media/0ce208f319de4672bf9750db98ff2fcc/href</a></iframe><p>To leave the container, you can type “exit” and then issue “<strong>docker ps -a</strong>”:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-tnfc_kBQD9C2DgTxwQ-KQ.png" /></figure><p>The command “docker ps -a” displays all the docker containers that have been created. The container we just created is listed here, and the status is shown as “exited”. In contrast, “<strong>docker ps</strong>” only lists the running containers.</p><p>Now let’s try the “docker run” command one more time:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LEo9Ok8WZGRpdpk1R6GqkA.png" /></figure><p>As seen above, the second “docker run” command created a new container, and its status is shown as “up 10 seconds”. The only difference is how we exited the container: first ctrl+p, then ctrl+q, which allows the container to continue running in the background.</p><p>To stop this running container, type “<strong>docker stop </strong>container-ID.”</p><p>If you’d like to restart one of the exited containers, simply issue “<strong>docker restart</strong> container-ID” and “<strong>docker exec</strong> -it container-ID bash.”</p><p>Or, if you want to save your docker container as a new docker image, you can use the command “<strong>docker commit </strong>container-ID new-image-name.” Then you can share the new image with other people or upload it to the Docker Hub.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b7rYCOgIWEI2302mnwawmw.png" /></figure><p>If you want to share data between a host directory and the docker container, volumes (-v) come in handy</p><p>(docker run -it -v host_directory:/data image_name):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Zt9a6jTQuGhMjkJl5t_h8Q.png" /></figure><p>In the end, if you’d like to free up some disk space, use “ <strong>docker rm</strong>” to remove obsolete docker images:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YAzplQo5vMEwLMuXUIyhMA.png" /></figure><p>Note that a container needs to be stopped before the deletion.</p><p>Similarly, You could delete the docker images using “<strong>docker rmi.”</strong></p><p>If you have any secret docker recipes that you would like to share, you are welcomed to provide them in the replies below. I am also working on a tutorial about how to simulate a small protein on AWS using Docker image, so definitely stay tuned for that!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2c28dd4aace6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding the data splitting functions in scikit-learn]]></title>
            <link>https://medium.com/@julie.yin/understanding-the-data-splitting-functions-in-scikit-learn-9ae4046fbd26?source=rss-fb56e9a16f2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ae4046fbd26</guid>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[scikit-learn]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Julie Yin]]></dc:creator>
            <pubDate>Mon, 03 Sep 2018 03:22:32 GMT</pubDate>
            <atom:updated>2018-09-03T22:57:14.624Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/629/1*_HoMKjrWahRiI-JmwYW6zg.png" /></figure><p>In this post, I am going to walk you through a simple exercise to understand two common ways of splitting the data into the training set and the test set in scikit-learn. The Jupyter Notebook is available <a href="https://github.com/janeyin600/Data_analysis/blob/master/train_test_split.ipynb">here</a>.</p><p>Let’s get started and create a Python list of numbers from 0 to 9 using range():</p><pre>X =  list(range(10))<br>print (X)</pre><pre>[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]</pre><p>Then, we create another list which contains the square values of numbers in X using list comprehension:</p><pre>y = [x*x <strong>for</strong> x <strong>in</strong> X]<br>print (y)</pre><pre>[0, 1, 4, 9, 16, 25, 36, 49, 64, 81]</pre><p>Next, we will import model_selection from scikit-learn, and use the function <a href="http://scikit-learn.org/stable/modules/generated/sklearn.model_selection.train_test_split.html">train_test_split( )</a> to split our data into two sets:</p><pre><strong>import</strong> <strong>sklearn.model_selection</strong> <strong>as</strong> <strong>model_selection</strong></pre><pre>X_train, X_test, y_train, y_test = <strong>model_selection.train_test_split</strong>(X, y, <strong>train_size</strong>=0.75,<strong>test_size</strong>=0.25, <strong>random_state</strong>=101)</pre><pre>print (&quot;X_train: &quot;, X_train)<br>print (&quot;y_train: &quot;, y_train)<br>print(“X_test: &quot;, X_test)<br>print (&quot;y_test: &quot;, y_test)</pre><pre>X_train:  [4, 9, 3, 5, 7, 6, 1]<br>y_train:  [16, 81, 9, 25, 49, 36, 1]<br>X_test:  [8, 2, 0]<br>y_test:  [64, 4, 0]</pre><p>By specifying the train_size as 0.75, we aim to put 75% of the data into our training set, and the rest of the data into the test set. Because we only have ten data points, the program automatically rounded the ratio to 7:3. It’s okay to omit the test_size parameter, if you already got the train_size specified, and you don’t mind the annoying warning message :)</p><p>Another observation is that the numbers in the lists after splitting do not follow the same ascending order as before. In another word, by default, the program ignores the original order of data. It randomly picks data to form the training and test set, which is usually a desirable feature in real-world applications to avoid possible artifacts existing in the data preparation process. To disable this feature, simply set the shuffle parameter as False (default = True).</p><p>You’ve probably seen people using cross_validation.train_test_split( ) for smiliar tasks. In fact, it’s just an old way of doing the same thing. Do the following</p><pre><strong>import</strong> <strong>sklearn.cross_validation</strong> <strong>as</strong> <strong>cross_validation</strong></pre><pre>X_train, X_test, y_train, y_test = <strong>cross_validation.train_test_split</strong>(X, y, <strong>train_size</strong>=0.75, <strong>random_state</strong>=101)</pre><p>will generate exactly the same outputs as above, given that we assigned the same number to <strong>Random_state</strong>. If you want your results to be stochastic each time, simply leave it as the default value “None”.</p><p>I assume that you are already familiar with the concept of <a href="https://en.wikipedia.org/wiki/Cross-validation_(statistics)">cross validation</a>. In scikit-learn, you can use the <a href="http://scikit-learn.org/stable/modules/generated/sklearn.model_selection.KFold.html">KFold( )</a> function to split your dataset into n consecutive folds.</p><pre><strong>from sklearn.model_selection import KFold<br>import numpy as np</strong></pre><pre>kf = <strong>KFold</strong>(<strong>n_splits</strong>=5)<br>X = np.array(X)<br>y = np.array(y)<br><strong>for</strong> train_index, test_index <strong>in</strong> kf.split(X):<br>    X_train, X_test = X[train_index], X[test_index]<br>    y_train, y_test = y[train_index], y[test_index]<br>    print(“X_test: &quot;, X_test)</pre><pre>X_test: [0 1]<br>X_test: [2 3]<br>X_test: [4 5]<br>X_test: [6 7]<br>X_test: [8 9]</pre><p>By specifying the n_splits parameter as 5, both of the X and y sets were divided into five folds (the y sets now shown here). You probably noticed that, this time, the program always picked two neighboring numbers from the original data sets, which means the data points were not shuffled (Why is the default setting of the shuffle parameter here different from that in <strong>train_test_split </strong>?) Nevertheless, use</p><p>kf = <strong>KFold</strong>(n_splits=5, <strong>shuffle=True</strong>)</p><p>will give you the same mixing effect for the original data sets as what we’ve seen before.</p><p>In addition, scikit-learn provides useful built-in functions to calculate the error metrics of multiple folds of test sets to evaluate machine learning models. For example,</p><p><a href="http://scikit-learn.org/stable/modules/generated/sklearn.model_selection.cross_val_score.html#sklearn.model_selection.cross_val_score"><strong>model_selection.cross_val_score</strong></a>(model, X, y, cv=kf, <a href="http://scikit-learn.org/stable/modules/model_evaluation.html"><strong>scoring=‘neg_mean_absolute_error’</strong></a>)</p><p>will report one score of mean absolute error.</p><p>Feel free to create your own jupyter Notebook and play around with those functions! Have fun!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ae4046fbd26" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Use Steepest Descent to Build a Better Water Model for Computer-Aided Drug Design]]></title>
            <link>https://medium.com/@julie.yin/use-steepest-descent-to-build-a-better-water-model-for-computer-aided-drug-design-df3717298604?source=rss-fb56e9a16f2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/df3717298604</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[gradient-descent]]></category>
            <category><![CDATA[drug-design]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[steepest-descent]]></category>
            <dc:creator><![CDATA[Julie Yin]]></dc:creator>
            <pubDate>Sat, 16 Jun 2018 20:34:59 GMT</pubDate>
            <atom:updated>2018-10-03T02:40:48.718Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*orLqSXytcWplrjZV4ln2pg.jpeg" /></figure><p>A great part of my research is devoted to improving the accuracy of binding affinity prediction of drug candidates. You may think that the medicines we take are all made of substances with overwhelmingly complex chemical formulas. On the contrary, orally active drugs are often relatively small molecules (<a href="https://en.wikipedia.org/wiki/Molecular_mass">molecular mass</a> &lt;500 <a href="https://en.wikipedia.org/wiki/Atomic_mass_unit">daltons</a>) with rather simple structures (10 or fewer rotatable bonds), according to the famous <a href="https://en.wikipedia.org/wiki/Lipinski%27s_rule_of_five"><strong>Lipinski’s rule</strong></a>.</p><p>Below shows one of the most important drugs ever developed in human history: <strong>penicillin</strong>. More accurately, it is penicillin G, one of the many derivatives of penicillin (derivatives are a group of compounds with small structural tweaks in organic chemistry). The penicillin drugs have been extensively used as antibiotics, and have saved numeruous lives over the years.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L5Kl75Q_7Q1Dum3Qinb1Bw.png" /><figcaption>The 3 dimensional ball-and-stick representation of <a href="https://en.wikipedia.org/wiki/Penicillin"><strong>Penicillin</strong></a>. Dark grey: Carbon; Light grey: Hydrogen; Red: Oxygen; Blue: Nitrogen; Yellow: Sulfur. Image generated by the author using MOE software.</figcaption></figure><p>Oftentimes, the very first step for such a small molecule to serve as a drug is to interact with certain types of proteins in the human body. Once the drug candidate interacts with proteins, it has a shot at triggering a cascade of physiological events aimed to cure diseases. If the molecule does not even interact with the proteins, most likely it won’t do much about the targeted disease. There is a term for this type of interactions between the protein and small drug-like compounds called “<strong>binding</strong>”, and we use another term called “<strong>binding affinity”</strong> to quantitatively measure how tight the protein and drug candidates interact with each other. An excellent introduction on protein-small molecule binding, drug design, and the measurement of binding affinities can be found <a href="http://bindingdb.org/bind/BindingDB-Intro.pdf">here</a>. Of course, to make a good drug, a strong binding is only the beginning of a long story: there are many other aspects a medicinal chemist needs to pay attention to, such as absorption, distribution, metabolism, and excretion (<a href="https://en.wikipedia.org/wiki/ADME"><strong>ADME</strong></a>). But like I mentioned, binding affinity measurement is a critical component in the early stage of drug discovery process.</p><p>One way of measuring the binding affinities of chemical compounds and identifying <strong>lead compounds</strong> (the starting point of a drug, promising but not good enough, plenty of room to improve) is through experiment. Millions of compounds can be screened withn a few days in big pharmaceutical companies, and this step is highly automated nowadays, mostly done by robotic arms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GyWFT9mD146JmjS552213A.jpeg" /><figcaption>Source: <a href="https://en.wikipedia.org/wiki/High-throughput_screening#/media/File:Chemical_Genomics_Robot.jpg">Wikipedia</a></figcaption></figure><p><strong>Alternatively, binding affinities can be calculated through computational modeling and simulations.</strong> If good and fast enough predictions can be generated <em>in silico</em>, it can save a lot of cost and time of synthesizing the compounds and testing them using biochemical assays. In addition, computational work can provide a lot of details about how the drug-like molecule and protein interaction with each other at atomic level, which is usually lacking in experiment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TnIggXS0UC3iUroAAAGWpQ.png" /><figcaption>Reproduced by the author according to a figure in Nate Silver’s book: “the signal and the noise”</figcaption></figure><p>Most computational predictions can be boiled down to two aspects: <strong>accuracy and precision</strong>. <strong>Accuracy</strong> refers to the closeness of a measured value to a standard or known value. <strong>Precision</strong> refers to the closeness of two or more measurements to each other (<a href="https://labwrite.ncsu.edu/Experimental%20Design/accuracyprecision.htm">definition</a>). Your measurements can be accurate but not precise, precise but not accurate, neither, or both. For example, the water density at 1 atm is 1.000 kg/liter at room tempature. If someone measures the water density under the same temperature and pressure conditions and reports 0.807, 0.806 and 0.808 kg/liter, I would consider those data points as precise but not accurate. For comparison, 1.208, 1.008 and 0.814 are realtively more accurate because the averaged value is very close to our target value, but still not precise enough. Smaller variance among measurements are expected given that this is not such a hard experiment to do. In contrast, data sets including 1.005, 1.008 and 0.997 are more in the realm of “accurate and precise”.</p><p>In the real-world modeling and simulations, you can generate very rough predictions using oversimplified model in no time, but the outcomes are likely to be neither accurate nor precise. Or, you could use sophisticated model and increase the sampling size, but you could still end up with not totally satisfying results, due to the limitations of your model and parameters. And that’s sort of where we are right now on computing binding affinities using complex physics-based methods — the awkward space between the third and the fourth targets: good on precision but still sturuggling a lot on accuracy. The scatter plot shows the correlation between the computed and experimental binding affinties of a series of host-guest complexes. In brief, host-guest molecules are model systems we used to study protein-ligand binding.</p><p>We often use <strong>Linear regression</strong> model to fit the predicted binding affinities to the experimental values. In the scatter plot below, the solid black line is the “one-to-one” ratio line, which indicates a perfect binding prediction. The red dashed line is the trendline generated by the linear regression model for the values predicted using water model 1, and the green dashed line is the trendline for water model 2 predicted values. As you can see here, in both cases, there is a strong linear relationship between the predicted and measured values. However, the green spheres (affinities predicted using water model 2) are collectively closer to the one-to-one line than the red squares (affinities predicted using water model 1), which suggests a more accurate prediction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/962/1*WjeXR2d69feTtORDg6THbQ.png" /><figcaption>Scatter plot generated from a small subset of the research data, for demonstration purpose only. For a complete description of the research, please see <a href="https://chemrxiv.org/articles/Bind3P_A_Water_Model_to_Improve_the_Accuracy_of_Binding_Calculations/5848113">here</a></figcaption></figure><p>So how did we go from water model 1 to water model 2? Basically we used gradients to guide our parameter search. We wanted to simulate an aqueous environments because drugs will be dissolved in our body fluid, and an aqueous environment can be approximated as thousands of or more water molecules which look like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/772/1*qyiJRt2CXwNG3P_w2r9wKA.png" /><figcaption>Schematic drawing of a water model, generated using MOE software. Red: Oxygen; Gray: Hydrogen.</figcaption></figure><p>Whether we like it or not, we end up with spending most of the time simulating the water, other than the stuff we are actually interested in, because the water molecules take up most of the space in our simulation box. But, water is important for molecular dynamics simulations, and it is really, really hard to get it right.</p><p>The simplest three-site (one oxgyen atom and two hydrogen atoms, no virtual sites) rigid water model has several parameters, including van der Waals parameters of the oxygen atom, the partial charges, the bond length and the angle parameters. To optimize the water parameters, we computed the first-order gradients of the binding affinities with respect to the parameters we wanted to refine, and the gradients can tell us how much and to what direction the computed binding affinity will change, once we make a small perturbation of the water parameters. We then used those information to reduce the RMSE value (the root-mean-square deviation). Another round of simulations were carried out, using the perturbed parameters, and gradients were computed once again from new simulation data. Those steps can be done iteratively until the error could not be reduced further by changing parameters. And the backbone of this optimization procedure is <a href="https://towardsdatascience.com/gradient-descent-in-a-nutshell-eaf8c18212f0"><strong>gradient descent</strong></a>, i.e. the <strong>steepest descent</strong>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=df3717298604" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Machine Learning: the Myth of Titanic]]></title>
            <link>https://medium.com/@julie.yin/machine-learning-the-myth-of-titanic-ac90c52ff774?source=rss-fb56e9a16f2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/ac90c52ff774</guid>
            <category><![CDATA[logistic-regression]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-analysis]]></category>
            <dc:creator><![CDATA[Julie Yin]]></dc:creator>
            <pubDate>Thu, 07 Jun 2018 20:45:51 GMT</pubDate>
            <atom:updated>2018-06-09T17:30:27.748Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/470/1*nvfFfffX0BHQvOTYnDvJTw@2x.jpeg" /></figure><p>There is currently <a href="https://www.kaggle.com/c/titanic">a prediction competition about the survival chance of Titanic passengers</a> going on at Kaggle. It immediately grabs my attention becuase Titanic is my favorite movie of all time (my second favorite is the original Jurassic Park).</p><p><strong><em>Molly Brown</em></strong><em>: Hey, uh, who thought of the name Titanic? Was it you, Bruce?<br></em><strong><em>J. Bruce Ismay</em></strong><em>: Yes, actually. I want to convey sheer size, and size means stability, luxury, and above all, strength.<br></em><strong><em>Rose DeWitt Bukater</em></strong><em>: Do you know of Dr. Freud, Mr Ismay? His ideas about the male preoccupation with size might be of particular interest to you.<br></em><strong><em>J. Bruce Ismay</em></strong><em>: Freud? Who is he? Is he a passenger?</em></p><p>How classic is that?</p><p>An excellent tutorial that walks you through a Kaggle competition can be found <a href="https://towardsdatascience.com/machine-learning-kaggle-competition-part-one-getting-started-32fb9ff47426">here</a>. The competition data were already divided into two sets: the training set and the test set. The training set provides passenger information that could be relevant to their survival chance, including age, gender, name, ticket class, ticket number, fare, cabin number, ports of embarkation and a few others. The facts that they were survived or not were also inlcuded. The test set includes the same information except for the survival list, and that is what the participants were asked to predict.</p><p>In this case, we will just assume that the train set and the test set data were randomly divided. In another word, we assume the training set is a good representation of the entire data set so that we can train our model on those data confidently. We also need to keep our fingers crossed that those data were accurately recorded, since according to the movie’s plot, Rose was registered as not survived because she discarded her identity and Jack boarded in the last minute with a ticket he won through gambling, so the information associated with his ticket was clearly collected from someone else!</p><p>My “Titanic” Jupyter Notebook is available <a href="https://github.com/janeyin600/Machine_learning_projects/blob/master/Titanic_survival_prediction/disaster_survival_prediction.ipynb">here</a>. Given that this is a binary classification problem, we will pick <a href="http://ml-cheatsheet.readthedocs.io/en/latest/logistic_regression.html">logistic regression </a>method for modeling. Another advantage of using this approach is that the coefficients generated form logistic regresion models can be used to rank features according to their importance in predicticing the target value:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/862/1*6YmUPTsbXA-Ub4e71AfPSA.png" /><figcaption>Logistic regression coefficients ranked in an ascending order, reflecting the importance of features in predicting the survival possbilities of Titanic passengers</figcaption></figure><p>As seen in this figure, the top three ranked features are gender, ticket class and age, which is quite in line with our impression of the Titanic disaster. It is well known that children and women had a much larger survival chance on Titanic, and the passengers in the first class also had a much higher survival rate, but that’s not interesting enough. Let’s take a closer look at those data by generating some <a href="https://jakevdp.github.io/PythonDataScienceHandbook/05.13-kernel-density-estimation.html">KDE (kernal density estimation) </a>plots:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*INQejhuthpm7bwqlHNxNUg.png" /></figure><p>The KDE plot can be considered as smoothed histogram. The blue line represents the age distribution of all female passengers, and the orange line indicates the survived females only. The first thing we noticed from the plot is that both curves are close to a Gaussian distribution, with majority of the population falling into the 20–40 age group, which is well expected. On the other hand, it might surprise you with the high similarlity and the almost identical heights of those two lines, but keep in mind that those are based on normalized data. So all it can tell us is that there is no significant difference of the age distribution between all female and the survived female passenger on Titanic. The KDE plot doesn’t really reflect the difference of two series in terms of absolute values. In another word, the survived chance of women has little to do with their age in general, although we did see that the teenage girls somehow had a higher mortality rate than other female age groups, unfortunately. In our training set the number of saved and died female passengers was counted as 233 and 81, respectively.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jIIl9LSjJCjy28c2TLkuZg.png" /></figure><p>The KDE plot of the male passengers tells a totally different story. We saw a great loss of gentlemen in their late 20s and early 30s. On the other hand, the survival chance of young boys were much higher. Totally 109 male passengers were recorded as saved and 418 as died in our training set. Let’s now compare the age distributions of all male and female passengers at different ticket classes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pi18ovPmskNPGiaHqH2Rog.png" /></figure><p>Note that the figures above are the age distributions of <strong>all</strong> passengers. The blue line represents male and the orange line represents female this time. Compared to the second and the third classes, the first class had more senior passengers. And men in their 20s and 30s were dominant in the third class. Nevertheless, those are still not too far away from the normal distribution pattern we are comfortable with.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AGb9z-lZqmiobqmKTq_7tQ.png" /></figure><p>For the saved ones, the age distributions of both male and female passengers closely resemble what they were before in the first class. There seems to be a decline in the population of senior male passengers. The most dramatic change happened in the second class and mainly for the males. In contrast, the distribution patterns of the third class seem not much affected, which is puzzling, too, in spite of the reduced peak height of males, since we all know the survival chance in the third class was dim.</p><p>To explain my observations, I found a <a href="https://pdfs.semanticscholar.org/c0df/071bca727c7e6ae663e9f1e0aa6d5c74c5eb.pdf">document</a> that provides a detailed breakdown of the mortality rate on Titanic:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*J5HfF0NDdBbI9l1m0hWe-g.png" /><figcaption>Data resource: Wayne Hall, <em>Social Science and Medicine </em>1986 22 (6): 687–690.</figcaption></figure><p>Based on those data, the men in the second class had a strikingly high mortality rate of 91.7%! And for the third class both men and women underwent high death rates of 83.8% and 57.8%. Calculating using our training set data, the mortality rates of Titanic passengers are: 63.1%(m) and 3.2% (w) for the first class; 84.3%(m) and 7.9%(w) for the second class; and 86.5%(m) and 50.0%(w) for the third class, which confirmed the representativeness of our training set data for the entire data set.</p><p>Clapping for this article if you like the story and want to see more!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ac90c52ff774" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Use One-Hot-Encoding To Analyze Adult Income Data]]></title>
            <link>https://medium.com/@julie.yin/use-one-hot-encoding-to-analyze-adult-income-data-and-some-bad-news-for-the-single-people-in-the-cef71f9d47b4?source=rss-fb56e9a16f2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/cef71f9d47b4</guid>
            <category><![CDATA[linear-regression]]></category>
            <category><![CDATA[encoding]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[logistic-regression]]></category>
            <dc:creator><![CDATA[Julie Yin]]></dc:creator>
            <pubDate>Mon, 04 Jun 2018 22:17:49 GMT</pubDate>
            <atom:updated>2018-06-06T00:57:18.431Z</atom:updated>
            <content:encoded><![CDATA[<p>In this post, I am going to illustrate how to use <strong>logistic regression</strong>, combined with the “<strong>one-hot-encoding</strong>” techniques to reveal certain interesting facts from the UCI adult income dataset. This dataset can be downloaded from <a href="https://www.kaggle.com/flyingwombat/logistic-regression-with-uci-adult-income">here</a>, along with the data description and some basic analysis.</p><p>My <a href="https://github.com/janeyin600/Machine_learning_projects/blob/master/adult_income_predictions/Adult_income_predictions.ipynb">Jupyter Notebook</a> is available here, and the analysis is built on <a href="https://www.valentinmihov.com/2015/04/17/adult-income-data-set/">a post</a> written by <a href="https://www.valentinmihov.com/">Valentin Mihov</a>. I then implemented the one-hot-encoding technique for further data interpretation.</p><p>The first step is to load the csv data into Pandas DataFrame:</p><pre>income_df = pd.read_csv(&quot;data/income_data.csv&quot;)<br><br><strong>print</strong> list(income_df)<br><strong>print</strong> income_df.shape</pre><pre>Output:[&#39;age&#39;, &#39;workclass&#39;, &#39;fnlwgt&#39;, &#39;education&#39;, &#39;educational-num&#39;, &#39;marital-status&#39;, &#39;occupation&#39;, &#39;relationship&#39;, &#39;race&#39;, &#39;gender&#39;, &#39;capital-gain&#39;, &#39;capital-loss&#39;, &#39;hours-per-week&#39;, &#39;native-country&#39;, &#39;income&#39;]<br>(48842, 15)</pre><p>The feature names are pretty self-explanatory (Phew!) Life is too short to read the data description, isn’t it?</p><p>Another thing we can immediately see is that this dataframe contains 15 features and 48842 rows (data points).</p><p>One little trick is to use “income_df.sample()” to get a quick glance of the raw data. Compared to “head()” or “tail()”, the sample() function selects and prints out serveral random row in the csv file:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MraNiKGqKB6WJe0zDn3UMQ.png" /></figure><p>One quick finding is that this dataset contains both numerical (e.g. “age”) and categorical data (e.g. “education”). I wrote a small funciton to get a better overview of the data:</p><pre><em># get some basic ideas about the data</em><br><strong>def</strong> summerize_data(df):<br>    <strong>for</strong> column <strong>in</strong> df.columns:<br>        <strong>print</strong> column<br>        <strong>if</strong> df.dtypes[column] == np.object: <em># Categorical data</em><br>            <strong>print</strong> df[column].value_counts()<br>        <strong>else</strong>:<br>            <strong>print</strong> df[column].describe() <br>            <br>        <strong>print</strong> &#39;<strong>\n</strong>&#39;<br>    <br>summerize_data(income_df)</pre><p>For the numerical features, this funciton prints out the mean values and standard deviations of data in the entire column. For the categorical features, it simply counts the numbers of samples in each subcategory. For example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/728/1*HEirWjq3tlbMfoelyTKy_w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/484/1*Giyx7uPhilqaAIBzq_7ozQ.png" /></figure><p>Let’s skip the tedious data cleaning, splitting and scaling process (see details in my <a href="https://github.com/janeyin600/Machine_learning_projects/blob/master/adult_income_predictions/Adult_income_predictions.ipynb">Jupyter Notebook</a>), and take a look at the output of our <a href="http://ml-cheatsheet.readthedocs.io/en/latest/logistic_regression.html">logistic regression</a> model. <strong>Compared to linear regression, logistic regression is a binary classifier algorithm.</strong> For example, in the current case, it can predict, based on the age, occupation, marital status and other factors, whether one’s income is above 50K per year, but it cannot predict a continuous number, like $75,636.</p><p>There is some i<a href="https://blog.myyellowroad.com/using-categorical-data-in-machine-learning-with-python-from-dummy-variables-to-deep-category-66041f734512">n-depth discussion</a> about how to convert categorical variables to something that the machine could better understand, i.e. <strong>encoding</strong>. The easiest way would be to assign a different integer to each subcategory . For instance, in our “workclass” category, we can assign “1” to “Private”, “2” to “Sel-emp-not-inc”, “3” to “Local-gov” and so on. Sciki-learn provides us a tool called preprocessing.<strong>LabelEncoder </strong>to do that. We can then use the encoded data as part of our training and test data.</p><pre>cls = linear_model.LogisticRegression()<br><br>cls.fit(X_train, y_train)<br>y_pred = cls.predict(X_test)<br><strong>print</strong> accuracy_score(y_test, y_pred)<br>Output:0.7992793883</pre><p>The accuracy score is around 0.80, you might get a slightly different number each time because it is a stochastic process. There are multiple things you can do to improve your prediction, if that’s your goal, but our interest here is to use the <a href="http://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html"><strong>coeff_</strong></a><strong> </strong>attribute, the coefficient of the features in the logistic regression function, to rank the importance of predicators.</p><p>After plotting the coeff_ attribute, we get the figure and the table below:</p><pre>coefs = pd.Series(cls.coef_[0], index=X_train.columns)<br><strong>print</strong> X_train.columns<br>coefs = coefs.sort_values()<br>plt.subplot(1,1,1)<br>coefs.plot(kind=&quot;bar&quot;)<br>plt.show()<br><strong>print</strong> coefs.sort_values(ascending = False)</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EcWvb_J_yRlb2WqAcDdYpg.png" /><figcaption>Features in the UCI adult income dataset, ranked by logistic regression coefficient</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/620/1*fAgeUnLO9X-DjxMVoFDC5Q.png" /></figure><p>Note that at this point the “educational-num” column has been removed from the dataframe, since it contains essentially the same information as “education”. This ranking is somewhat useful. It tells us features such as capital-gain, age, work hours, martical-status are more important than others like occupation, race and nationality when it comes to predicting one’ income, if we believe that the sampling is unbiased and our model is correct. However, It does not tell us how exactly those features play a role in prediction. And the relatively low ranking of occupation seems counterintuitive. Moreover, what does it mean that the relationship and marirtal-status “negatively” impact the income level?</p><p>Another problem is that the approach of encoding categorical features to ordinal variables is intrinsically flawed. It works fine when the feature has a binary nature, e.g. gender (in simplied scenarios). However, when the feature has more than two subcategories, the machine will make some arbitrary connections between unrelated values. For example, if we assign “1” to “Private”, “2” to “Sel-emp-not-inc”, and “3” to “Local-gov”, the machine will think “Private” are more similar to “Sel-emp-not-inc” than “Local-gov” because “1” is closer to “2” compared to “3”. But this is not our intention.</p><p>We can use one-hot-encoding to solve this problem — converting feature A to vector [0,0,1], feature B to [0,1,0], and feature C to [1,0,0]. It is recommended to process the numerical and categorical features separately, and join them later on.</p><pre>numeric_subset = income_df.select_dtypes(&#39;number&#39;)<br>categorical_subset = income_df.select_dtypes(&#39;object&#39;)<br><br><em># One hot encode</em><br>categorical_subset = pd.get_dummies(categorical_subset[categorical_subset.columns.drop(&quot;income&quot;)])</pre><p>After conversion, the new dataframe is expanded, with each subcategory being a new column:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qEgPqwiO4qzm2cGirQ_wmw.png" /><figcaption>Using One-Hot-Endoding, the “workclass” feature is now expanded to a matrix with lots of “0”s and some “1”s.</figcaption></figure><p>Using the new encoded data, we can regenerate our training and test set, and re-fit our logistic model. The accuracy score is increased to some extent. More importantly, the coefficients from the new encoded data tell us a more complete story about the data:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4di6aJDvFrjGOzGopdrDxg.png" /><figcaption>… …</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UagyU01X6bsJyfIZiYZugA.png" /></figure><p>Bad news for the singles: you are at the very bottom of the income ladder! And if you want to get rich, gain some capital, get married, and never ever get a Ph.D. :D</p><p>That seems fun but we might want to further ask that how would the age and work hours affect the income? To answer this question, I wrote a function to divide the sample population into different age groups:</p><pre><em># Further group numerical predictors such as &quot;age&quot; or &quot;hours-per-week&quot; to gain more insight!</em><br><em># Group the &quot;age&quot; column</em><br>age_group = []<br><strong>for</strong> age <strong>in</strong> income_df[&quot;age&quot;]:<br>    <strong>if</strong> age &lt; 25:<br>        age_group.append(&quot;&lt;25&quot;)<br>    <strong>elif</strong> 25 &lt;= age &lt;= 34:<br>        age_group.append(&quot;25-34&quot;)<br>    <strong>elif</strong> 34 &lt; age &lt;= 44:<br>        age_group.append(&quot;35-44&quot;)<br>    <strong>elif</strong> 44 &lt; age &lt;= 54:<br>        age_group.append(&quot;45-54&quot;)<br>    <strong>elif</strong> 54 &lt; age &lt;= 65:<br>        age_group.append(&quot;55-64&quot;)<br>    <strong>else</strong>:<br>        age_group.append(&quot;65 and over&quot;)</pre><pre>new_income_df = income_df.copy()<br>new_income_df[&quot;age_group&quot;] = age_group<br><strong>del</strong> new_income_df[&quot;age&quot;]</pre><p>You could do the same thing for work hours. Using the new dataframe, we get a more detailed output:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KAOoNHz7BI50emdkJmGQSg.png" /><figcaption>… …</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0fCFP0NpwXahtgMO1VdwQg.png" /></figure><p>Hmm…Now it looks like you have to work for 48–60 hours per week for maybe one or two decades to get resonably rich (Sorry!). For the single people, you still got plenty of room to improve on your income ;)</p><p>Edit: I shared this post with some of my colleagues, and one of them pointed out that the “importance scores” of the encoded features look suspicious because they are dependent of each other. The ranking does make a lot of sense to me, although I should probably remove either “relationship” or “marital-status”, because they overlap with each other to a large extent. It may be a good idea to combine some of the subcategories as well, depending on how detailed you want to get for your analysis. Keep in mind that you are at the risk of overfitting when you divide the data points into too many classes.</p><p>I later on summed up the coefficients of all the subcategories in the “marital-status” and “workclass” features, and it turns out the sum is about equal to the coefficient of their common feature before using One Hot Encoding. Does the ranking of feature importance including subcategorical predictors make sense to you? Thoughts?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cef71f9d47b4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Interpret Your Data Analysis Outputs]]></title>
            <link>https://medium.com/@julie.yin/how-to-tell-a-nice-story-using-your-data-ff3d12971566?source=rss-fb56e9a16f2c------2</link>
            <guid isPermaLink="false">https://medium.com/p/ff3d12971566</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[apc]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[support-vector-machine]]></category>
            <dc:creator><![CDATA[Julie Yin]]></dc:creator>
            <pubDate>Wed, 30 May 2018 18:36:47 GMT</pubDate>
            <atom:updated>2018-05-31T00:47:53.185Z</atom:updated>
            <content:encoded><![CDATA[<p>Your left your data science classroom, real or virtual… with a fancy Jupyter notebook saved to your laptop. You felt like the time and the dollars you spent may be well worth it … But deep inside you were insecure: What is this analysis for? What did it really tell me?</p><p>To illustrate how to make your data analysis sexier and more meaningful, I will give an example using the famous Iris dataset, analyzed with a classic supervised machine learning algorithms: support vector machine (SVM), as well as a frequently used dimensionality-reduction technique: principle component analysis (PCA).</p><p>The Iris dataset is a set of sepal length, sepal width, petal length and petal width data points for three flower species: <a href="https://en.wikipedia.org/wiki/Iris_setosa"><em>Iris setosa</em></a>, <a href="https://en.wikipedia.org/wiki/Iris_virginica"><em>Iris virginica</em></a> and <a href="https://en.wikipedia.org/wiki/Iris_versicolor"><em>Iris versicolor</em></a>. It must make perfect sense to you now why it is called the Iris dataset! Those data were published by the British biologist <a href="https://en.wikipedia.org/wiki/Ronald_Fisher">Ronald Fisher</a> in 1936. He took 50 samples for each species, so the data can be transformed to a 4 x 150 two-dimensional array.</p><p>Surprisingly, with this relativly small dataset you can apply so many different machine learning methods and draw interesting conclusions out of your analysis. There are multiple ways to download the Iris dataset. If you already installed Anaconda on your local workstation, you can load the data directly from sciki-learn datasets:</p><pre><strong>from</strong> sklearn <strong>import</strong> datasets<br>iris = datasets.load_iris()</pre><p>Alternatively, you can load that from the seaborn dataset:</p><pre><strong>import</strong> seaborn <strong>as</strong> sns<br>iris = sns.load_dataset(&quot;iris&quot;)</pre><p>You can also download the data from the UCI machine learning repository:<a href="https://archive.ics.uci.edu/ml/datasets/iris">https://archive.ics.uci.edu/ml/datasets/iris</a>.</p><p>One important concept in machine learning and data science is <strong>dimensionality reduction</strong>, which is useful when you got too many features in your data frame, and not every single feature is critical in answering the question you are interested in. To illustrate how would selecting different features affect the predicting accuracy of your model, let’s first train the model only using the data of two sepal features: <strong>the sepal length and width</strong>. <strong>The goal of the analysis is to predict whether a species is <em>Iris-versicolor</em> or not.</strong></p><p>This post is not intended as a tutorial on how to fit or test a SVM model, as you can quickly grab those techniques from the scikit-learn documentation page (<a href="http://scikit-learn.org/stable/auto_examples/svm/plot_iris.html">http://scikit-learn.org/stable/auto_examples/svm/plot_iris.html</a>) and many other online resources. I will just go ahead and show you the predictions for the entire dataset generated by both linear and non-linear SVM models trained with 75% of the data points.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i0PnCRI_hGLgam6M6eHmwA.png" /><figcaption>The dividing line or curve in each panel is a “hyperplane” generated by different SVM models to best separate <em>Iris-versicolor</em> (red dots) from non-<em>Iris-versicolor flowers (blue dots)</em></figcaption></figure><p>Apparently both of the linear or non-linear SVM models performed poorly on separating the <em>Iris-versicolor </em>from<em> </em>other two species. The inaccurate predictions may be related to our selections of predictors. Let’s try again with the other two features of the petal: <strong>the</strong> <strong>petal length and width</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GTi2d31Jwu-UVOM7H40qmw.png" /><figcaption>The dividing line or curve in each panel is a “hyperplane” generated by different SVM models to best separate <em>Iris-versicolor</em> (red dots) from non-<em>Iris-versicolor flowers (blue dots)</em></figcaption></figure><p>Our prediction got greatly improved! Hooray!</p><p>But how, you may ask, do I know which features to include in my analysis? It might be easier to make the choice in the current case, since we only have four features, but it will become incredibly difficult to decide which features to switch on and off when there are dozens of potential predictors.</p><p>There is a nice post on stackoverflow showing you how to conduct PCA analysis to extract relevant features: <a href="https://stackoverflow.com/questions/22984335/recovering-features-names-of-explained-variance-ratio-in-pca-with-sklearn">https://stackoverflow.com/questions/22984335/recovering-features-names-of-explained-variance-ratio-in-pca-with-sklearn</a></p><p>In brief, PCA analysis generates PCA components which are a set of values of <a href="https://en.wikipedia.org/wiki/Correlation_and_dependence">linearly uncorrelated</a> variables. The <strong>explained variance ratio</strong> shows the percentage of variance explained by each component (pca.explained_variance_ratio_). In the current case, PC1 alone can explain a majority of the variance in the data points (&gt; 70%). By generating four PCA components (pca = PCA(n_components=4)), I can get a table showing the correlation coefficents between each PCA component and each feature in the Iris dataset:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aJlxSJP9xAr_FSlk4YZ5qg.png" /></figure><p>As we can see in this table, the sepal width and PC1 are less correlated than other three features since the correlation coefficient is only -0.26. And that is consistent with what we saw from the SVM outcome — using the sepal length and sepal width collectively generated less accurate predictions than using the petal length and width together. Note that The PCA sign does not affect the variance explained by each component. Finally, we can happily conclude that it should be safe to drop the sepal width data and it shouldn’t affect the accuracy of our prediction to a significant extent.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff3d12971566" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>