<?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 Dion Banno on Medium]]></title>
        <description><![CDATA[Stories by Dion Banno on Medium]]></description>
        <link>https://medium.com/@dionb?source=rss-7049f6ac3f26------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*RcPfEHMwU1-Xj33jbXbJ9g@2x.jpeg</url>
            <title>Stories by Dion Banno on Medium</title>
            <link>https://medium.com/@dionb?source=rss-7049f6ac3f26------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 18:25:08 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@dionb/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[Measuring the efficacy of puppy’s potty training]]></title>
            <link>https://medium.com/@dionb/measuring-the-efficacy-of-puppys-potty-training-ba5a42b08d58?source=rss-7049f6ac3f26------2</link>
            <guid isPermaLink="false">https://medium.com/p/ba5a42b08d58</guid>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[dogs]]></category>
            <dc:creator><![CDATA[Dion Banno]]></dc:creator>
            <pubDate>Sun, 13 Jun 2021 03:48:07 GMT</pubDate>
            <atom:updated>2021-11-02T04:05:12.411Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4cCUVndEdsPzBZe1QQVb4w.png" /></figure><p><em>Originally published on </em><a href="https://www.linkedin.com/pulse/measuring-efficacy-puppys-potty-training-dion-banno/"><em>https://www.linkedin.com/pulse/measuring-efficacy-puppys-potty-training-dion-banno/</em></a></p><p>On March 27th we welcomed a new member to the family: a 7-week old Golden Retriever, Harley. One of the major challenges for new puppy owners is that of potty training. As a first step, we opted to introduce the use of training pads in the first three weeks before attempting outdoor potty training. The question of interest is as follows: does the introduction of training pads seem to influence Harley’s potty behaviour?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/350/1*XMP6yTQsroh_BB8vooeGoA.png" /><figcaption>Harley at 8 weeks</figcaption></figure><p>Let’s turn to statistics to answer this question!</p><h3><strong>The Data</strong></h3><p>The data for this analysis was generated by recording Harley’s potty locations in her enclosed space across the first three weeks. A bird’s eye image was annotated across the three-week period for each observed incident.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aZAmbY-zcbMd5HFqWzIQxg.png" /><figcaption>Example of annotated image, and Harley’s enclosed space</figcaption></figure><p>Here’s the first five rows of the data set:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fZY97U8c_-p2vzKLmPq6Og.png" /><figcaption>First five rows of cleaned data set</figcaption></figure><h3>Exploratory Analysis</h3><p>The goal of the exploratory analysis is to identify the spatial structure of the incidents on the enclosed space and compare it to the scenario of no potty training.</p><p>However, since potty training was started from day one, we have to rely on simulation methods to construct this baseline process. Under no potty training, we may hypothesize that Harley’s potty preferences are indifferent across her enclosed space; hence, potty incidents would be randomly distributed. Complete spatial randomness was simulated in Harley’s enclosed space through a homogenous poisson process using an acceptance-rejection framework.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B6F5o_Z7JfMBA62KmIYNrQ.png" /><figcaption>Potty incidents on Harley’s enclosed space</figcaption></figure><p>At first glance, the data from week 1 seems to follow spatial randomness (the baseline) with an increased incident rate in the designated potty location. In contrast, weeks 2 and 3 are more clustered and show a significantly higher proportion of successful incidents.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/754/1*TNLF-lbPHpKdTBrdBhk9iA.png" /><figcaption>Proportion of incidents on the designated potty location</figcaption></figure><p>We can extract more insight by estimating the K and G functions of the spatial process across the three week period:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JOTBGs3CRKs-fTZWmdAwUg.png" /><figcaption>Quick analysis of K and G estimates</figcaption></figure><p>The upward shift in the K and G estimates in weeks 2 and 3 indicates that Harley’s potty behaviour has shifted to identify certain regions within her enclosed space as appropriate potty regions versus being indifferent between locations in the first week. This can be seen in the resulting clustered process post week 1. In potty training, this is half the battle — the next component is examining whether these hotspots seem to lie reasonably close to the designated potty location.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*77RB_Hdfykp03YgiXZVLPg.png" /><figcaption>Density of distance of potty incidents to designated location</figcaption></figure><p>The distribution of distance from observed incidents to the pee pad is relatively flat in week 1 and in the simulated baseline, which supports uniformity in locations. There is a clear shift to a more peaked (albeit, somewhat bimodal) distribution in weeks 2 and 3 indicating that Harley has identified the potty location; however, she still incorrectly potties on certain hotspots somewhat frequently. We can formally test this by designing a randomized permutation test since the data does not seem to follow a known probability distribution.</p><h3>Designing a randomized permutation test</h3><p>Statistically, we want to test if there is a difference in proximity of incidents to the designated potty location in comparison to the baseline process. Constructing a test statistic, T as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B8CR6CR3tjSrRP12W86i7w.png" /><figcaption>Designed test statistic</figcaption></figure><p>The test statistic is bounded from 0 to 1 where a value close to 1 indicates that all incidents are on the potty location (good); close to 0, indicates divergence of incidents from the potty location (bad).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Xc5TnjjKwuQryhBQrTIZZg.png" /></figure><p>Intuitively, weeks 2 and 3 seems to show an improvement to week 1 and the baseline. Are these improvements statistically significant, or can they be attributed to randomness?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*shTcyCA-DZdVEXHhjt1o9w.png" /><figcaption>Distribution of test statistics under randomization with week i and baseline</figcaption></figure><p>The observed test statistic (red dotted line) seems to be probable for week 1 (first panel) indicating that there is little evidence against an improvement when comparing week 1 with the simulated baseline. That is, that the observed visual “improvement” in week 1 can be attributed to randomness.</p><p>On the other hand, the observed test statistic (red dotted line) seems to be an improbable result when we mix the weekly data with the baseline for weeks 2 and 3 (seen in the second and third panel). Hence, the shift in Harley’s potty behaviour is <em>“statistically significant”</em> which means that there is very strong evidence to suggest that there is a weekly improvement in proximity to potty locations in weeks 2 and 3 compared to the baseline process.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X7pv0yZfn7IuXS_ll6bO8Q.png" /><figcaption>p values for testing difference between week i and baseline</figcaption></figure><p>We may also wish to compare week 2 and 3 due to their observed similarities:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YmR9sBZQHr-4AjLm5ki2iA.png" /><figcaption>Comparing week 2 and 3</figcaption></figure><p>We see that the observed test statistics (blue and red lines on the density) lie within the mode of the randomized test statistic’s distribution. This means that the observed data from weeks 2 and 3 are probable outcomes when randomly mixing the data for these two weeks.</p><p>This implies that there is no evidence against a difference in week 2 and 3 on the basis of proximity of incidents to the designated potty location. That is, there is no evidence of an improvement (or regression) in training progress from weeks 2 and 3 suggesting that Harley’s pee pad training seems to exhibit diminishing returns — easy to pick up on, but take longer to master. This seems to be a reasonable outcome for any learned skill — not only in puppy training.</p><h3>What to expect in the coming weeks?</h3><p>To make inferences on the upcoming weeks, we can model incident counts probabilistically using a simple Bayesian Poisson-Gamma framework. Mathematically, the number of potty incidents in each sub-region follows a Poisson distribution which is dependent on another random variable which encodes information about the rate of potty incidents:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*siw_s-gK76kJ-K7fqagQmw.png" /></figure><p>Under a Bayesian framework, we consider this rate as a random variable (as opposed to being a fixed constant under a frequentist framework) which then follows a Gamma distribution with hyper-parameters alpha and beta.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qBzNjLnkpOsxB23_MGgYXA.png" /></figure><p>We assume a distribution of the rate parameter before observing data primarily to get better count estimates for sub regions with little to no observed incidents. Combining information from the data distribution and this assumed prior knowledge gives us a new distribution from which we can infer future counts (this new distribution is called the posterior distribution).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iqvPLZWUB4kGsYaiAvYNOQ.png" /></figure><p>This Bayesian method of incident count estimation allows us to obtain not only averages for incident counts but construct a 95% credible interval for each sub regions. That is, we can infer that incident counts will be in these intervals with probability 0.95 (or 95%). Setting the hyper-parameters of the gamma distribution with the maximum likelihood estimates for post week 1 gets us the following Bayesian count estimates across Harley’s enclosed space:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qnE3_afp13IIUYOcud_Oxg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5I84BCWLZXsnuwxzqxooOA.png" /></figure><p>While there is definitely an improvement with pee pad training (as shown in the permutation tests), these Bayesian count estimates suggest that there will still be significant potty accidents in the coming weeks — so it looks like we’ll still need to be vigilant on the floor cleaning duties for now *sigh*!</p><p>Cool — now, we have some quantitative evidence and insights to support progress in Harley’s potty training. Data and statistics seem to indicate that Harley is one smart puppy for sure! :) ❤</p><p>Let me know what you think, and leave a clap if you enjoyed it!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ba5a42b08d58" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Creating cohesive Spotify playlists using Hierarchical Clustering]]></title>
            <link>https://medium.com/swlh/exploring-unsupervised-learning-with-my-spotify-playlist-cdcc3e2ef0b7?source=rss-7049f6ac3f26------2</link>
            <guid isPermaLink="false">https://medium.com/p/cdcc3e2ef0b7</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[unsupervised-learning]]></category>
            <dc:creator><![CDATA[Dion Banno]]></dc:creator>
            <pubDate>Sun, 30 Aug 2020 00:01:32 GMT</pubDate>
            <atom:updated>2020-09-03T03:41:57.955Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ez51cAT1aRMhj3QX-146xg.png" /></figure><p>One day, I was listening to my Spotify playlist of my favourite obscure Mariah Carey tracks — appropriately named <em>Obscure Mariah Faves</em>— like any normal person would, and noticed an interesting problem. My compilation seemed to lack a quintessential element of good playlists: cohesiveness.</p><p>Because the playlist only had 46 songs, it was pretty easy for me to manually create cohesive sub-playlists especially given that I am pretty well-versed on her discography.</p><p>But what happens when there are more than 100 songs? What about 100,000 — or even “infinitely” many songs — from different artists? Surely, there has to be a more automated — perhaps data-driven — way of solving this problem. This is where statistical methods of unsupervised learning come into play; and in particular, methods of clustering.</p><h3>The Potpourri Playlist</h3><p>With this problem in my mind, I compiled a new playlist of random songs that I’ve liked over the course of my 5 year (and counting) Spotify tenure — I called it the <a href="https://open.spotify.com/playlist/3L2HToam3LDyxwd3vkWID0"><em>Potpourri</em></a> playlist.</p><p>Spotify provides the following attributes about each song with the following descriptions:</p><ul><li><strong>danceability:</strong> how suitable a track is for dancing</li><li><strong>energy:</strong> intensity and activity</li><li><strong>loudness: </strong>relative loudness of a track</li><li><strong>speechiness:</strong> presence of spoken words</li><li><strong>acousticness: </strong>how acoustic a track sounds</li><li><strong>instrumentalness: </strong>presence of vocals (including rapped words)</li><li><strong>liveness: </strong>presence of a live audience</li><li><strong>valence:</strong> how positive a track sounds (feelings of happiness, euphoria, or cheerfulness)</li><li><strong>tempo:</strong> beats per minute</li></ul><p>Examining the marginal distribution of each standardized attribute for the songs in the <em>Potpourri</em> playlist gives us the following estimated probability densities:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MCi13LTG4YVhHT2oAP6XcA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uTLGca3yqeNjuJBSFC1WWA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BJzwb2rbQqJz2GourytBdQ.png" /></figure><p>If the playlist was cohesive, we would expect a fairly peaked distribution around 0, which implies that values are concentrated within a small neighbourhood of the mean. However, the marginal distributions are often quite skewed, heavy-tailed, and indicate a wide range of attribute values. This suggests evidence of dissimilarity within the songs in the <em>Potpourri</em> playlist; hence, clustering may be beneficial.</p><p>Our data set of interest is the <em>Potpourri</em> playlist which describes the aforementioned attributes of 182 songs. We wish to gain insight on how to segment these songs into smaller playlists that create a more cohesive sound. This problem is unsupervised because the songs in the <em>Potpourri </em>playlist don’t belong to an explicitly defined segment— rather, the goal is to find such a suitable segmentation.</p><h3>Preliminary Analysis</h3><p>For the first step of the analysis, I looked at the correlation matrix of the attributes in the <em>Potpourri </em>playlist:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YMnDFUGyBM5SNPzbqQCUaA.png" /><figcaption>Potpourri Playlist: Correlation Matrix</figcaption></figure><p>For the most part, the pair-wise correlation between the attributes seem to be weak (i.e, close to 0 or no correlation). However, there are some attributes that are highly correlated— namely, energy and loudness (0.68), and energy and acousticness (-0.66).</p><p>Statistically, highly correlated attributes can be troublesome as it leads to the issue of multicollinearity. In the context of clustering, this may have negative consequences as having strong correlation between attributes means that the dissimilarity metric used may be biased towards highly correlated attributes — giving them more weight, so to speak.</p><p>A naive solution may be to drop attributes which are highly correlated. However, this would imply potentially dropping an attribute like energy — which I’m reluctant to do as my knowledge of the <em>Potpourri </em>playlist indicates that this is an important attribute for segmentation.</p><p>Another alternative is to do feature extraction by projecting the original data into an orthogonal space. In doing so, we get a new representation of the original data while retaining the same information. This linear transformation gives us the principal components of the <em>Potpourri</em> data set. Mathematically, this transformation is simply the eigenvalue decomposition of the variance-covariance matrix of the original, standardized data set— where the principal components are, themselves, eigenvectors.</p><p>Each principal component contributes a level of “information” about the original data. This contribution can be seen through its corresponding eigenvalue.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OF2vvBatWElRBXoGNCVwyA.png" /><figcaption>Eigenvalue associated with the i-th eigenvector (or principal component) of the Potpourri playlist</figcaption></figure><p>We can interpret the eigenvalues as a metric for the amount of “information” or variation that a principal component holds about the songs in the <em>Potpourri </em>playlist.</p><p>Alternatively, we can look at a cumulative view of variation explained:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h9BSbShc1KTcZzLuuF3FMw.png" /><figcaption>Explained variability as a function of principal components</figcaption></figure><p>The figure above shows the explained variation associated with the principal components of the <em>Potpourri</em> playlist. For instance, the first principal component accounts for about 30% of total variability. Moreover, about 90% of variability can be explained by using only six principal components.</p><p>This implies that we can reduce the dimensions of the original data set without losing too much information. This may be beneficial in clustering as dissimilarity metrics are often calculated based on euclidean distance — which loses meaning as the number of dimensions increase.</p><p>This also implies that we can extract a lower-dimensional visualization of the <em>Potpourri</em> playlist using the first two principal components.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4n29LFSv0EXjVfILdWIFQA.png" /><figcaption>Two-Dimensional Representation of the Potpourri playlist</figcaption></figure><p>The blue points indicate songs within a neighbourhood with radius of 1.96 around (0,0) —I chose 1.96 to be based on a 95% Wald-type “interval” (or neighbourhood) around (0,0).</p><p>The songs that fall outside this neighbourhood are as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AFJ39HeCCjVPwO2ULUDwlA.png" /><figcaption>Songs outside the 95% neighbourhood around (0,0)</figcaption></figure><p>It’s interesting to see how a majority of these songs have a pretty mellow vibe with the exception of <em>Juice</em> and <em>Good as Hell </em>by Lizzo, and <em>Make It Happen</em> by MC. It’s also interesting how <em>Think About Me</em> by dvsn seems to be a clear numerical outlier when, personally, it doesn’t seem to sound radically different to other songs in the <em>Potpourri</em> playlist.</p><p>Looking at the plot, there doesn’t seem to be a natural clustering present— however, this two-dimensional representation only accounts for 46% of variability. In other words, over half of the information is still unaccounted for. A natural question is as follows: how many principal components should be used for clustering?</p><p>There isn’t exactly a clear-cut answer as principal component analysis is also a method within unsupervised learning. Instead, we can choose k heuristically by looking at the variation explained. For this problem, I believe that 90% of variation is enough for a good playlist segmentation. Hence we may choose k=7 (i.e, cluster based on 7 principal components). By no means can we argue optimality by choosing this value of k, but it’s a start.</p><p>Before proceeding with clustering, I want to take a slight detour from unsupervised learning and look at some of my favourite songs in the <em>Potpourri </em>playlist. In particular, I want to look at the Nearest Neighbours of my favourite songs — in doing so, we can see similar songs, where similarity is defined through euclidean distance.</p><p>One of my absolute favourites is an urban-inspired track by Mariah Carey called <em>The Roof. </em>It’s definitely a hidden hip-hop infused gem that showcases the pinnacle of Mariah’s lyricism. To me, this is the epitome of song-writing excellence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/652/1*HfkWv-z4EvuzPwaHe3kH6A.png" /><figcaption>5 Nearest Neighbours of The Roof in the Potpourri Playlist</figcaption></figure><p>Another obscure gem that I love is <em>Breakdown (Mo’ Thugs Remix)</em> which is an R&amp;B/hip-hop collaboration between Mariah and Bone Thugs-N-Harmony. It’s such a pleasure to listen to these two distinct artists merging stylistically — and doing so in a way that doesn’t feel contrived. And of course, it’s also extremely well-written and well produced.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/924/1*8x3vLP2rHta5IbOYhZvZuQ.png" /><figcaption>5 Nearest Neighbours of Breakdown (Mo’ Thugs Remix) in the Potpourri Playlist</figcaption></figure><p>It’s pretty cool to see how the nearest neighbours of <em>Breakdown (Mo’ Thugs Remix) </em>contains other favourites such as <em>Do It to Me, Babydoll</em>, <em>Love Like That, </em>and <em>We Belong Together — </em>additionally, it’s quite insightful to see the similarity metric between these songs<em>.</em></p><p>Another favourite of mine is <em>Bad Girl</em> by Usher — the <em>Confessions </em>album is definitely one of my favourites ever.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hL-20MgGvAZs-Wj7wsNwWQ.png" /><figcaption>5 Nearest Neighbours of Bad Girl in the Potpourri Playlist</figcaption></figure><p>It’s pretty neat to see that Lil’ Kim records are close in euclidean distance to <em>Bad Girl</em>. I never would have put <em>Crush on You </em>and <em>Bad Girl </em>together on my own — but now, I can’t un-hear the similarity.</p><h3>Clustering</h3><p>The clustering method that I chose for this particular problem is agglomerative hierarchical clustering — which is really just a fancy way of referring to a bottom-up approach to clustering.</p><p>Agglomerative hierarchical clustering initializes each song as its own cluster and iteratively merges with other clusters based on a similarity metric. The algorithm terminates once all the tracks are clustered as one.</p><p>We can define similarity between clusters through a number of ways. However, since the goal of the problem is to create cohesive sub playlists, we may want to choose a measure of variability as the merging mechanism. One such measure is Ward’s method:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bRCVAId0JApwveZJzekT1w.png" /><figcaption>Ward’s Distance Method (source: shorturl.at/ivDLO)</figcaption></figure><p>The formula gives the distance between two clusters A and B. It is defined as the sum of squares of the merged cluster minus the sum of squares of each disjoint clusters. This quantity can be thought of as a measure of cluster variability and corresponds to the cost of merging two clusters. Hierarchical clustering will choose to merge clusters that minimize this quantity per iteration.</p><p>For the purpose of illustration, here is an example using four songs:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ke1OLisuZB7tKr44WfU5xw.png" /><figcaption>Agglomerative Hierarchical Clustering for 4 tracks</figcaption></figure><p>The merging order is as follows: <em>Ex-Factor </em>and <em>Get You</em>; <em>Triggered (freestyle) </em>and <em>Back to Back; </em>(<em>Ex-Factor, Get You</em>) and (<em>Triggered (freestyle), Back to Back</em>).</p><p>Clearly, this hierarchical clustering has a natural binary tree-like data structure where the tree itself is a cluster with nodes that are also clusters. The root of the tree is an individual track — which can be viewed as a cluster with no preceding cluster (i.e, a node with no child). Or in Python:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QDVEtLEe8DNan1ChbVn-lQ.png" /><figcaption>Cluster Object as a binary tree (see GitHub below for more deets)</figcaption></figure><p>With this definition, we can build functions and instance methods to implement the hierarchical clustering on the <em>Potpourri </em>playlist without using libraries like scikit-learn.</p><p>Here’s the final call on the function that I created that facilitates the hierarchical clustering:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IkVqbHbUg6eaJqEzrRi3yw.png" /><figcaption>Implementation of hierarchical clustering (see GitHub below for deets)</figcaption></figure><p>The result is a Cluster object which can be converted into an adjacency matrix through more Python, and is visualized as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TB2EfHVubGriwnIgb51Xhw.png" /><figcaption>Potpourri network from hierarchical clustering (visualized through Gephi)</figcaption></figure><p>The figure above shows the resulting network from applying hierarchical clustering. The merging cost between two clusters are represented through the edges’ width: the merging cost is the highest when all songs are in one cluster; the lowest, when all songs are distinct clusters. As previously alluded to, the algorithm terminates once all the songs are in one cluster — this final cluster is indicated by the green node in the figure above.</p><p>A natural question that follows is that of how many clusters should be used? Well, the answer lies between 1 to 182 (the number of songs in the playlist). To gain insight on how many should be used, we can examine the merging cost as a function of the number of clusters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t7uyEN3KpMndTttbLqnSEQ.png" /><figcaption>Merging cost as a function of number of clusters</figcaption></figure><p>As expected, the figure above shows that merging cost is a decreasing function of the number of clusters. This makes sense since the cluster variability is at the minimum when there is only one song in each playlist, and increases as more songs are clustered together. Another observation worth mentioning is that the “gains” (or reduction in variability) seems to exhibit marginal decrease. This is more clearly seen by taking a first difference:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y7jJ5u5Do2M-mesj2yiaqQ.png" /><figcaption>Reduction in merging cost as a function of number of clusters</figcaption></figure><p>The figure above shows more explicitly that the reduction in variability stabilizes at 8 clusters. After, the reduction in variability tails off to 0 as the number of clusters approaches the number of songs in the playlist (182). For this reason, we may heuristically choose 8 clusters. In doing so, we are attempting to balance the trade off between minimizing variability in each cluster while grouping as many similar tracks as possible.</p><h3>Clustering Results</h3><p>We can compare the marginal distributions for each cluster across the original song attributes to examine the results. Note that Cluster 1 has been removed from the following figures since it only contains one song:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e9bAOktw1xaaxSob3QvJ8A.png" /><figcaption>Cluster Attribute Comparison (1/3)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M1uToVbgMYxHFj3RYUBcHw.png" /><figcaption>Cluster Attribute Comparison (2/3)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YKVB4rIf0a82nwcFsY_MQQ.png" /><figcaption>Cluster Attribute Comparison (3/3)</figcaption></figure><p>There seems to be noticeable variation in distribution across the clusters on attributes like danceability, energy, loudness, speechiness, acousticness, and liveness. In particular, their central tendency and shape seems to vary — this is a good sign of effective clustering.</p><p>Another observation worth noting is that there does seem to be high variability within some clusters for all the attributes. This is somewhat troublesome as this suggests evidence of dissimilarity.</p><p>Visualizing the 8 clusters in the <em>Potpourri </em>playlist:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zwLMu1JneTEIdOtnC2hKYA.png" /><figcaption>Two-Dimensional Representation of Potpourri Clusters</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-ojvb0v8dpwCWUH8QHRl4Q.png" /><figcaption>Three-Dimensional Representation of Potpourri Clusters</figcaption></figure><p>It’s also quite insightful to look at the tracks of some of the resulting sub-playlists. For the sake of brevity, I’m only going to look at two playlists.</p><p>The distribution plots above indicate that Playlist 2 seems to have high variability across the song attributes. The resulting playlist is as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ET8qFtuNc3iDBQUPiMR5wQ.png" /><figcaption>Playlist 2: Bouncy Laid Back Beats</figcaption></figure><p>Looking at Playlist 2, the tracks seem to have a common laid-back beat and generally has a mellow vibe while maintaining that musical “bounce”. The tracks that best exemplifies this are <em>Sober, Family Business, EARFQUAKE, God’s Plan</em>, <em>The Distance</em>, <em>Get Me Home</em>, and <em>I Know What You Want</em>. In addition, the vocals in the songs favour a more subtle sound. It’s interesting as I wouldn’t have put these songs together on my own.</p><p>However, there are glaring outliers: in particular, <em>Juice</em>, <em>Good as Hell</em>, <em>Make It Happen-Live at Madison Square Garden, </em>and <em>Treasure </em>seem to be on the more upbeat and energetic side; while <em>Ex-Factor</em> is on the slower side. These tracks may contribute to Playlist 2’s high variability.</p><p>On the other hand, Playlist 7 seems to produce a more cohesive cluster:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DcEdVvWFK8XnUtEmO917zQ.png" /><figcaption>Playlist 7: Ballad with a Beat</figcaption></figure><p>Playlist 7 has a more consistent beat accompanied by more present vocals. This playlist embodies that mid-tempo R&amp;B sound that I really enjoy — which can be categorized as a ballad-with-a-beat sound. Almost like a hip-hop and R&amp;B fusion. This is exemplified by songs like <em>Follow Me,</em> <em>Top of the World</em>, <em>We Belong Together</em>, and <em>Are You That Somebody</em>.</p><p>Cool. Now I have some idea on a good segmentation for the <em>Potpourri</em> playlist. However, the analysis indicates some inconsistencies using this method of clustering. This is somewhat common in unsupervised learning as the results are somewhat subjective, and tends to have reliance on domain knowledge. In order to get a better segmentation, we may want to experiment with different algorithms and dissimilarity metrics.</p><h3>Inconsistencies and recommendations for further analysis</h3><p>Although the clusters seem to show improved cohesion compared to the original <em>Potpourri </em>playlist, there are still clusters that have a non-trivial level of dissimilarity.</p><p>This may arise from the fact that hierarchical clustering is a method of hard-clustering: songs are either assigned to a playlist or not. For this reason, we may want to look at soft-clustering methods such as mixture models which clusters data probabilistically.</p><p>Alternatively, another approach is to look at each song longitudinally and use correlation between attributes as the dissimilarity metric. This would capture song attributes in greater granularity — as opposed to aggregating the attribute measures throughout the entire song.</p><h3>Final Words</h3><p>If you’re interested, check out the code that I wrote for this project on my <a href="https://github.com/dionbanno/Spotify/upload">GitHub</a>. It includes cool things like:</p><ul><li>integrating Spotify Web API on Python</li><li>a documented interactive analysis “tool” if you want to analyze your own Spotify playlists with Python</li><li>as well as my implementations of PCA, KNN, and Hierarchical Clustering</li></ul><p>I find that implementing these methods from scratch is a great way to learn and gain a deeper understanding of these methods; it also removes the “black box nature” of using other packages.</p><p>Feel free to leave your comments and areas of improvement. Also, connect with me on <a href="https://www.linkedin.com/in/dion-banno-78b955192/">LinkedIn</a>!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cdcc3e2ef0b7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swlh/exploring-unsupervised-learning-with-my-spotify-playlist-cdcc3e2ef0b7">Creating cohesive Spotify playlists using Hierarchical Clustering</a> was originally published in <a href="https://medium.com/swlh">The Startup</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Analyzing my Tinder Data]]></title>
            <link>https://medium.com/swlh/analyzing-my-tinder-data-3b4f05a4a34f?source=rss-7049f6ac3f26------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b4f05a4a34f</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[dating]]></category>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[tinder]]></category>
            <dc:creator><![CDATA[Dion Banno]]></dc:creator>
            <pubDate>Fri, 28 Aug 2020 22:01:01 GMT</pubDate>
            <atom:updated>2020-11-03T18:08:24.424Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YAlCOoUIdjJAlJJQNW4J0w.png" /></figure><p><em>Originally published on </em><a href="https://www.linkedin.com/pulse/analyzing-my-tinder-data-dion-banno/"><em>https://www.linkedin.com</em></a><em> on March 27, 2020 (data up to date as of March 20, 2020).</em></p><p>Day 3 of social distancing.</p><p>As I sit on my couch scrolling through my Instagram feed to see yet another drawing of an orange — apparently the latest Instagram challenge to pass the time — I was starting to get… <em>bored</em>. Who would’ve thought that an INTP like myself would succumb to boredom in day 3 of social distancing?</p><p>With no cool restaurants to explore, no plans with friends to hang out, no gyms to go to anytime soon — why not start a project to pass the time?</p><p>But, what project? I know I wanted to do something that would allow me to gain insight on some aspect of my life, and what’s more relevant to a 20 year old’s life than dating? In today’s ultra-digital world, dating has become synonymous with Tinder. I mean, how else are we supposed to meet and connect with people nowadays? Through physical and social communities like friends, mutuals, school, and work as has literally been the case for hundreds of generations prior? Nope, that’s <em>crazy</em>.</p><p>Tinder allows us to connect with people within our communities that we would never have met otherwise — for better or for worse. And as with many social media apps, Tinder allows you to request your own personal data.</p><p>And so I did.</p><h3><strong>Tinder Data</strong></h3><p>The requested Tinder data was in JSON format, and follows this simplified structure:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/702/1*9pktMRdfTv0sPXBf68MYqw.png" /><figcaption>Tinder Data Structure</figcaption></figure><p>Reading in the data into Python with the following script:</p><pre>import json with open(&#39;data.json&#39;) as json_file: <br>   data ​= json.load(json_file)</pre><p>Now, the first problem was taking this data structure and converting it to one that I could easily traverse through to analyze. Because Usage is simply a count aggregated daily, it was natural to convert this into a standard tabular data structure with rows as dates and columns as the aforementioned features.</p><pre>import pandas as pd <br>df = pd.DataFrame(list(data[&#39;Usage&#39;][&#39;app_opens&#39;].keys())) </pre><pre>for x in list(data[&#39;Usage&#39;].keys()): <br>    df[x]= list(data[&#39;Usage&#39;][x].values())</pre><p>Here’s the first five rows of the data frame — aka my first 5 days on Tinder:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ym8XX3E-sGCzOf5JTrirQA.png" /><figcaption>My first 5 days on Tinder</figcaption></figure><p>With the Messages, however, I wanted to explore other alternatives. Since an individual message can be viewed as an object with attributes text and sent date, I defined a Message Class/Object, and stored these in a dictionary where the key indicated the unique match ID.</p><pre>class Message:<br>    &#39;&#39;&#39;Fields: Text (Str)<br>               Date (Datetime)&#39;&#39;&#39;<br>    <br>    def __init__(self, text, date):<br>        self.text = text<br>        self.date = date<br>    <br>    def __repr__(self):<br>        message_rep = &quot;{}: {}&quot;<br>        <br>        return message_rep.format(self.date, self.text)<br><br><br><br>message_dict={}<br><br>for x in data[&#39;Messages&#39;]:<br>    match_id=x[&#39;match_id&#39;].split()[-1]<br>    sent = []<br>    for messages in x[&#39;messages&#39;]:<br>        sent_date = &quot; &quot;.join(messages[&#39;sent_date&#39;].split()[0:-1])[:-3]<br>        sent.append(Message(messages[&#39;message&#39;].lower(),sent_date))<br>    message_dict[match_id]=sent</pre><p>Now, we need more Python to parse through the messages to derive basic insights. Here’s an excerpt containing the basic idea:</p><pre>day_count, time_count, emoji_count = {}, {}, {}<br>day_time, date_count, word_count = {}, {}, {}<br><br>for matches in message_dict:<br>   messages = message_dict[matches]<br>   for msg in messages:<br>       date=msg.date.split(&quot; &quot;)<br>       day=date[0][:-1]<br>       time = date[-1][:2]+&#39;:00&#39;<br>       dt=&quot;-&quot;.join(date[1:4])<br>       words=msg.text.split(&quot; &quot;)<br>       check_lst = [[day, day_count], [time, time_count],<br>                    [day_time, day_time],[dt, date_count], <br>                    [words, word_count]]<br><br>       i=0<br>       while i &lt; 4:<br>          x=check_lst[i]<br>          key=x[0]<br>          dictionary = x[1]<br>          if key not in dictionary.keys():<br>               dictionary[key]=1<br>               i=i+1<br>          else:<br>               dictionary[key]=dictionary[key]+1<br>               i=i+1<br><br>       for x in words:<br>          t = str.maketrans(dict.fromkeys(string.punctuation))<br>          x = x.translate(t)<br>          stripped = list(x)<br>          for char in stripped:<br>              if char in emojis:<br>                 if char not in emoji_count.keys():<br>                     emoji_count[char]=1<br>                 else:<br>                     emoji_count[char]=emoji_count[char]+1</pre><h3><strong>Analysis &amp; Insights</strong></h3><p>Quick stats as of March 20th 2020:</p><ul><li>10,083 total app opens</li><li>Swiped right on 3,331 profiles, with a daily max of 92 on January 4 2019</li><li>Swiped left on 38,132 profiles, with daily max of 2,145 profiles on January 4 2019</li><li>349 matches, with daily max of 12 matches on March 18 2020</li><li>1,164 total messages sent</li><li>1,289 total messages received</li><li>125 unique conversations</li><li>32 social media/number exchanges</li><li>16 meet ups</li><li>countless dollars spent on bubble tea</li></ul><p>Traversing through my sent messages, we get the following word cluster:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DhiFyC39PsQlCyJ-ecOOVg.png" /><figcaption>My word cloud of sent messages</figcaption></figure><p>Looking at my top words:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BDolGuIaZyAW9sNmDlBl9g.png" /></figure><blockquote><em>“damn you’re cute wanna grab bubbletea ? haha”</em></blockquote><p>Interesting. Seems fairly normal in the context of Tinder. Now, I’m curious as to why statistics is one of my top words…</p><p>Now, among the messages sent, about 4% of these were emojis. Evidently, emojis are well integrated into digital messaging. Here are my top 5 sent emojis:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w8UVp4vIVahd6lX4jU84lA.png" /></figure><p>Moreover, data indicates that 15% of my sent messages had only 6 words — with 38% of my sent messages falling between the 5–7 word count range.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*der9TBR9WV6Ml6vtZw6bjg.png" /><figcaption>Length of messages sent</figcaption></figure><p>Looking at the distribution of conversation length measured in days, we see a left-skewed distribution — with 67% of conversations having tenure of less than one day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*O1ojRvDb41MNPp9Zc_JFoQ.png" /><figcaption>Conversation tenure in days</figcaption></figure><p>Among these single day conversations, a majority of them are dead-end: in other words, no messages were sent after my initial recorded message.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iB18llHeQt2tmBeNUd75pA.png" /><figcaption>Response rate of single day conversations</figcaption></figure><p>Now, before hammering down on my one-liners, there is a slight caveat: because I only have data on my sent messages, I used my first and last message within a match as a proxy for conversation length. As such, it is unclear which participant actually ended the conversation. So these ‘no responses’ could have been messages that I didn’t follow up on.</p><p>In fact, looking at the count of messages sent versus received indicates that my messages are generally answered — at least when aggregating on the monthly level. So maybe my one liners are <em>somewhat</em> effective — <em>sureeee</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sahqZiij78DTANNWnDpq4w.png" /><figcaption>Monthly messages</figcaption></figure><p>When are these messages actually sent out?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PaEcF4PjuLV2hONQEZ1KHw.png" /><figcaption>Message activity</figcaption></figure><p>Data indicates that peak messaging time occurs at 9 pm.</p><p>Cool — but these insights are only applicable once a match has actually occurred. We all know that 90% of Tinder consists of swiping.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*z1RNNc0g0iMYGmIjJXrwZA.png" /><figcaption>Monthly swiping activity</figcaption></figure><p>It’s interesting to see that 18% of total swipes were done in my first month of Tinder.</p><p>Defining match rate as the proportion of matches to swipe rights, we see that my match rate generally hovers at around 12.5% — with the highest match rate of 45% in March 2019 despite its low matches.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-xNHmgMPG1uSxWfUQ1zWzw.png" /><figcaption>Monthly matches</figcaption></figure><p>Assuming independence in swipes and holding the probability of a match fixed, we can think of each swipe right as a Bernoulli trial — where a successful outcome is a match.</p><p>Mathematically, we have a random variable, Y, that follows a binomial distribution:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AmOm1xQHVHi3jbg7FuhNzQ.png" /></figure><p>Or in our context:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Wm6L35XN1Ybiwh9KedMKgg.png" /></figure><p>Given my Tinder data and assuming a fixed probability of success (p), the maximum likelihood estimate of the parameter p is simply the estimated match rate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VpJSSu1ZzRUQssFYxIWUzw.png" /></figure><p>Holding the number of my received swipe rights constant, we can construct the following cumulative binomial probability distributions:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1dV54GlrF-rHTCZg8ZvtwA.png" /><figcaption>Cumulative probability of at least one match as a function of swipe rights</figcaption></figure><p>The figure above shows the probability of at least one match given a fixed probability of success, p. We can see that the probability of at least one match increases with the number of swipe rights. In other words, a match is inevitable as you swipe right — this is, of course, holding the number of received swipe rights constant. This resulting convergence is a consequence of the Law of Large Numbers.</p><p>Given my current swiping behaviour (p=0.10), it would take at least 30 swipes to get at least one match — emphasis on <strong>at least</strong>: meaning the number of matches could range from 1 to the number of swipe rights inclusive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8aEKycPIA7sL_1LoRLNGkg.png" /><figcaption>Swiping behaviour</figcaption></figure><p>Holding the number of my received swipe rights constant, a quick way to increase the probability of at least one match is to increase the number of swipe rights given. However, more doesn’t necessarily mean better: the trade-off between quality and quantity is more nuanced, so I’ll leave it at that.</p><p>A natural question that follows is how many of these matches actually lead to coffee or bubble tea? Data indicates a 12.8% conversion rate among my engaged matches. A 95% confidence interval estimate indicates a lower bound of 7% and an upper bound of 19% — the 6% margin of error could be telling of external factors, such as proximity, that could affect one’s interest to meet up.</p><p>Now, assuming independence among engaged matches and that each person is equally open to meet up, we can think of this as yet again another Bernoulli trial — where a successful outcome is a meet up.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OCZv3faX_-TlOPP-E7jt5A.png" /></figure><p>Given my Tinder data and assuming a fixed probability of success (p), the maximum likelihood estimate of the parameter p is simply the estimated conversion rate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9j_iq5d4JvDhJZh8s38vHw.png" /></figure><p>With these assumptions, we can make inferences on future outcomes such as calculating the probability of getting x number of meet ups — in other words, Prob(meet up = x | p = 0.128).</p><p>Pretty cool.</p><p>This is especially useful when it comes to allocating budgets for dates. Personally, first dates for me are around the $10 — $20 ball park — though, the variance on that is somewhat high. Assuming that I allocate $35 per month on dates and each date is $15, we can run simulations with 100 engaged matches over 6 months:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B_TyddZYCYnxacHaMS9USQ.png" /><figcaption>Simulation expenses over 6 months</figcaption></figure><p>Since the number of engaged matches is large (n=100), the binomial distribution can be approximated by a Gaussian probability density. This resulting convergence in distribution is a consequence of the Central Limit Theorem.</p><p>The probability of a deficit can be calculated as the area under the curve to the left of the red dotted line. Hence we can calculate this probability easily using the Gaussian approximation:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uOuMF_g3-q4TjK43n3oFqA.png" /></figure><p>Since the budget remaining is a linear function of meet ups — which we estimate through a Gaussian random variable — then, the budget remaining also follows a Gaussian distribution:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TNXZyNs_-NJDJiva85-0zA.png" /></figure><p>With these assumptions, the probability of a deficit is 0.3594. Yikes — this is somewhat concerning given that I’m on a student budget.</p><p>So, it’s probably not financially viable to message 100 matches over 6 months given my current conversion rate. To stay on budget, I either: decrease the number of messaged matches or decrease my conversion rate. Hmm, tough call — I’d have to go with the former.</p><p>Tweaking the parameters in the binomial simulation we get the following results:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7D797NT_DdCc5qxkAV5uRQ.png" /><figcaption>Tweaking parameters in budget simulation</figcaption></figure><p>Now, the probability of going over budget when I reduce my engagement to 75 matches is 0.06 (green density) — much better. Having said that, I also don’t want to end up with a big surplus since that would imply little to no dates (yellow). Hence, I should engage with 75 to 85 matches over the course of 6 months to fully utilize my budget.</p><p>Cool. Now I have some new insights about my current Tinder behaviour — however, by no means is this analysis exhaustive. If you happen to have Python installed and have your own personal Tinder data — or if you just want to look at the back-end logic of the Python functions used in this analysis — feel free to check out the code that I wrote for this project:</p><p><a href="https://github.com/dionbanno/dion_creates">https://github.com/dionbanno/dion_creates</a></p><h3><strong>Recommendations for further projects:</strong></h3><ul><li>Can we perform A/B testing on certain key words and phrases to see if they increase the probability of a response/meet up?</li><li>It would be cool to have a repository of individual Tinder data classified per user attribute such as location, gender, age, etc. and doing a regression analysis to see if certain user attributes affect success</li></ul><h3><strong>Final words</strong></h3><p>Data from this analysis indicates that 64% of matches go un-messaged. So shoot your shot. Go ignite those matches — who knows? It might be worth while.</p><p>Feel free to leave your comments, and connect with me on <a href="https://www.linkedin.com/in/dion-banno-78b955192/">LinkedIn</a>. I’d also be curious to know — what metrics would you have chosen to analyze, and how?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b4f05a4a34f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swlh/analyzing-my-tinder-data-3b4f05a4a34f">Analyzing my Tinder Data</a> was originally published in <a href="https://medium.com/swlh">The Startup</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>