# Are We in a TV Golden Age?

I recently found myself in a argument with my wife regarding whether TV was better now than previously. I believed that TV was better now than 20 years ago. My wife contended that there was simply more TV content being produced, and that this led to more good shows, but shows are not inherently any better.

This struck me as a great opportunity to do some quick data science. For this post, I scraped the names (from wikipedia) and ratings (from TMDb) of all American TV shows. I did the same for major American movies, so that I could have a comparison group (maybe all content is better or worse). The ratings are given by TMDb’s users and are scores between 1 and 10 (where 10 is a great show/movie and 1 is a lousy show/movie).

All the code for this post can be found on my github.

I decided to operationalize my “golden age of TV” hypothesis as the average TV show is better now than previously. This would be expressed as a positive slope (beta coefficient) when building a linear regression that outputs the rating of a show given the date on which the show first aired. My wife predicted a slope near zero or negative (shows are no better or worse than previously).

Below, I plot the ratings of TV shows and movies across time. Each show is a dot in the scatter plot. Show rating (average rating given my TMBb) is on the y-axis. The date of the show’s first airing is on the x-axis. When I encountered shows with the same name, I just tacked a number onto the end. For instance, show “x” would become show “x_1.” The size of each point in the scatter plot is the show’s “popularity”, which is a bit of a black box, but it’s given by TMBb’s API. TMDb does not give a full description of how they calculate popularity, but they do say its a function of how many times an item is viewed on TMDb, how many times an item is rated, and how many times the item has been added to watch or favorite list. I decided to depict it here just to give the figures a little more detail. The larger the dot, the more popular the show.

Here’s a plot of all TV shows across time.

To test the “golden age of TV” hypothesis, I coded up a linear regression in javascript (below). I put the regression’s output as a comment at the end of the code. Before stating whether the hypothesis was rejected or not, I should note that that I removed shows with less than 10 votes because these shows had erratic ratings.

As you can see, there is no evidence that TV is better now that previously. In fact, if anything, this dataset says that TV is worse (but more on this later).

I wanted to include movies as a comparison to TV. Here’s a plot of all movies across time.

It’s important to note that I removed all movies with less than 1000 votes. This is completely 100% unfair, BUT I am very proud of my figures here and things get a little laggy when including too many movies in the plot. Nonetheless, movies seem to be getting worse over time! More dramatically than TV shows!

Okay, so this was a fun little analysis, but I have to come out and say that I wasn’t too happy with my dataset and the conclusions we can draw from this analysis are only as good as the dataset.

The first limitation is that recent content is much more likely to receive a rating than older content, which could systematically bias the ratings of older content (e.g., only good shows from before 2000 receive ratings). It’s easy to imagine how this would lead us to believing that all older content is better than it actually was.

Also, TMDb seems to have IMDB type tastes by which I mean its dominated by young males. For instance, while I don’t like the show “Keeping up the Kardashians,” it’s definitely not the worst show ever. Also, “Girls” is an amazing show which gets no respect here. The quality of a show is in the eye of the beholder, which in this case seems to be boys.

I would have used Rotten Tomatoes’ API, but they don’t provide access to TV ratings.

Even with all these caveats in mind, it’s hard to defend my “golden age of TV” hypothesis. Instead, it seems like there is just more content being produced, which leads to more good shows (yay!), but the average show is no better or worse than previously.

# My First Kodi Addon - PBS NewsHour (a Tutorial)

I’ve been using Kodi/XBMC since 2010. It provides a flexible and (relatively) intuitive interface for interacting with content through your TV (much like an apple TV). One of the best parts of Kodi is the addons - these are apps that you can build or download. For instance, I use the NBA League Pass addon for watching Wolves games. I’ve been looking for a reason to build my own Kodi addon for years.

Enter PBS NewsHour. If you’re not watching PBS NewsHour, I’m not sure what you’re doing with your life because it’s the shit. It rocks. PBS NewsHour disseminates all their content on youtube and their website. For the past couple years, I’ve been watching their broadcasts every morning through the Youtube addon. This works fine, but it’s clunky. I decided to stream line watching the NewsHour by building a Kodi addon for it.

I used this tutorial to build a Kodi addon that accesses the PBS NewsHour content through the youtube addon. This addon can be found on my github. The addon works pretty well, but it includes links to all NewsHour’s content, and I only want the full episodes. I am guessing I could have modified this addon to get what I wanted, but I really wanted to build my own addon from scratch.

The addon I built is available on my github. To build my addon, I used this tutorial, and some code from this github repository. Below I describe how the addon works. I only describe the file default.py because this file does the majority of the work, and I found the linked tutorials did a good job explaining the other files.

I start by importing libraries that I will use. Most these libraries are used for scraping content off the web. I then create some basic variables to describe the addon’s name (addonID), its name in kodi (base_url), the number used to refer to it (addon_handle - I am not sure how this number is used), and current arguments sent to my addon (args).

The next function, getRequest, gathers html from a website (specified by the variable url). The dictionary httpHeaders tells the website a little about myself, and how I want the html. I use urllib2 to get a compressed version of the html, which is decompressed using zlib.

This next function actually fetches the videos (the hard part of building this addon). This function fetches the html of the website that has PBS’s video. It then searches the html for “coveplayerid,” which is PBS’s name for the video. I use this name to create a url that will play the video. I get the html associated with this new url, and search it for a json file that contains the video. I grab this json file, and viola I have the video’s url! In the final part of the code, I request a higher version of the video than PBS would give me by default.

If I fail to find “coveplayerid,” then I know this is a video with a youtube link, so I grab the youtube id. Some pages have a coveplayerid class, but no actual coveplayerid. I also detect these cases and find the youtube id when it occurs.

This next function identifies full episodes that have aired in the past week. It’s the meat of the addon. The function gets the html of PBS NewsHour’s page, and finds all links in a side-bar where PBS lists their past week’s episodes. I loop through the links and create a menu item for each one. These menu items are python objects that Kodi can display to users. The items include a label/title (the name of the episode), an image, and a url that Kodi can use to find the video url.

The most important part of this listing is the url I create. This url gives Kodi all the information I just described, associates the link with an addon, and tells Kodi that the link is playable. In the final part of the function, I pass the list of links to Kodi.

Okay, thats the hard part. The rest of the code implements the functions I just described. The function below is executed when a user chooses to play a video. It gets the url of the video, and gives this to the xbmc function that will play the video. The only hiccup here is I check whether the link is for the standard PBS video type or not. If it is, then I give the link directly to Kodi. If it’s not, then this is a youtube link and I launch the youtube plugin with my youtube video id.

This final function is launched whenever a user calls the addon or executes an action in the addon (thats why I call the function in the final line of code here). params is an empty dictionary if the addon is being opened. params being empty causes the addon to call list_videos, creating the list of episodes that PBS has aired in the past week. If the user selects one of the episodes, then router is called again, but this time the argument is the url of the selected item. This url is passed to the play_video function, which plays the video for the user!

That’s my addon! I hope this tutorial helps people create future Kodi addons. Definitely reach out if you have questions. Also, make sure to check out the NewsHour soon and often. It’s the bomb.

# Sifting the Overflow

In January 2017, I started a fellowship at Insight Data Science. Insight is a 7 week program for helping academics transition from academia to careers in data science. In the first 4 weeks, fellows build data science products, and fellows present these products to different companies in the last 3 weeks.

At Insight, I built Sifting the Overflow, a chrome extension which you can install from the google chrome store. Sifting the Overflow identifies the most helpful parts of answers to questions about the programming language Python on StackOverflow.com. To created Sifting the Overflow, I trained a recurrent neural net (RNN) to identify “helpful” answers, and when you use the browser extension on a stackoverflow page, this RNN rates the helpfulness of each sentence of each answer. The sentences that my model believes to be helpful are highlighted so that users can quickly find the most helpful parts of these pages.

I wrote a quick post here about how I built Sifting the Overflow, so check it out if you’re interested. The code is also available on my github.

# Simulating the Monty Hall Problem

I’ve been hearing about the Monty Hall problem for years and its never quite made sense to me, so I decided to program up a quick simulation.

In the Monty Hall problem, there is a car behind one of three doors. There are goats behind the other two doors. The contestant picks one of the three doors. Monty Hall (the game show host) then reveals that one of the two unchosen doors has a goat behind it. The question is whether the constestant should change the door they picked or keep their choice.

My first intuition was that it doesn’t matter whether the contestant changes their choice because its equally probable that the car is behind either of the two unopened doors, but I’ve been told this is incorrect! Instead, the contestant is more likely to win the car if they change their choice.

How can this be? Well, I decided to create a simple simulation of the Monty Hall problem in order to prove to myself that there really is an advantage to changing the chosen door and (hopefully) gain an intuition into how this works.

Below I’ve written my little simulation. A jupyter notebook with this code is available on my github.

Here I plot the results

Probability of choosing correctly if change choice: 0.67
Probability of choosing correctly if do not change choice: 0.33
Probability of difference arising from chance: 0.00000


Clearly, the contestant should change their choice!

So now, just to make sure I am not crazy, I decided to simulate the Monty Hall problem with the contestant choosing what door to open after Monty Hall opens a door with a goat.

Probability of choosing correctly if change choice: 0.51
Probability of choosing correctly if do not change choice: 0.49
Probability of difference arising from chance: 0.57546


Now, there is clearly no difference between whether the contestant changes their choice or not.

So what is different about these two scenarios?

In the first scenario, the contestant makes a choice before Monty Hall reveals which of the two unchosen options is incorrect. Here’s the intution I’ve gained by doing this - because Monty Hall cannot reveal what is behind the chosen door, when Monty Hall reveals what is behind one of the unchosen doors, this has no impact on how likely the car is to appear behind the chosen door. Yet, the probability that the car is behind the revealed door drops to 0 (because Monty Hall shows there’s a goat behind it), and the total probability must be conserved so the second unchosen door receives any belief that the car was behind the revealed door! Thus, the unchosen and unrevealed door becomes 66% likely to contain the car! I am still not 100% convinced of this new intuition, but it seems correct given these simulations!

# SFN 2016 Presentation

I recently presented at the annual meeting of the society for neuroscience, so I wanted to do a quick post describing my findings.

The reinforcement learning literature postulates that we go in and out of exploratory states in order to learn about our environments and maximize the reward we gain in these environments. For example, you might try different foods in order to find the food you most prefer. But, not all novelty seeking behavior results from reward maximization. For example, I often read new books. Maybe reading a new book triggers a reward circuit response, but it certainly doesn’t lead to immediate rewards.

In this poster we used a free viewing task to examine whether an animal would exhibit a novelty preference when it was not associated with any possible rewards. We found the animal looked at (payed attention to) novel items more often than he looked at familiar items, but this preference for paying attention to novel items fluctuated over time. Sometimes the animal had a large preference for looking at the novel items and sometimes he had no preference for novels items.

Neurons that we recorded in the dlPFC and area 7a encoded whether the animal was currently in a state where he prefered looking at novel items or not and this encoding persisted across the entire trial period. Importantly, while neurons in these areas also encoded whether the animal was currently looking at a novel item or not, this encoding was distinct from the encoding of the current preference state. These results demonstrate that the animal had simultaneous neural codes representing whether he was acutely attending to novel items and his general preference for attending to novel items or not. Importantly, these neural codes existed even though there were no explicit reward associations.

# PCA Tutorial

Principal Component Analysis (PCA) is an important method for dimensionality reduction and data cleaning. I have used PCA in the past on this blog for estimating the latent variables that underlie player statistics. For example, I might have two features: average number of offensive rebounds and average number of defensive rebounds. The two features are highly correlated because a latent variable, the player’s rebounding ability, explains common variance in the two features. PCA is a method for extracting these latent variables that explain common variance across features.

In this tutorial I generate fake data in order to help gain insight into the mechanics underlying PCA.

Below I create my first feature by sampling from a normal distribution. I create a second feature by adding a noisy normal distribution to the first feature multiplied by two. Because I generated the data here, I know it’s composed to two latent variables, and PCA should be able to identify these latent variables.

I generate the data and plot it below.

The first step before doing PCA is to normalize the data. This centers each feature (each feature will have a mean of 0) and divides data by its standard deviation (changing the standard deviation to 1). Normalizing the data puts all features on the same scale. Having features on the same scale is important because features might be more or less variable because of measurement rather than the latent variables producing the feature. For example, in basketball, points are often accumulated in sets of 2s and 3s, while rebounds are accumulated one at a time. The nature of basketball puts points and rebounds on a different scales, but this doesn’t mean that the latent variables scoring ability and rebounding ability are more or less variable.

Below I normalize and plot the data.

After standardizing the data, I need to find the eigenvectors and eigenvalues. The eigenvectors point in the direction of a component and eigenvalues represent the amount of variance explained by the component. Below, I plot the standardized data with the eigenvectors ploted with their eigenvalues as the vectors distance from the origin.

As you can see, the blue eigenvector is longer and points in the direction with the most variability. The purple eigenvector is shorter and points in the direction with less variability.

As expected, one component explains far more variability than the other component (becaus both my features share variance from a single latent gaussian distribution).

Next I order the eigenvectors according to the magnitude of their eigenvalues. This orders the components so that the components that explain more variability occur first. I then transform the data so that they’re axis aligned. This means the first component explain variability on the x-axis and the second component explains variance on the y-axis.

Finally, just to make sure the PCA was done correctly, I will call PCA from the sklearn library, run it, and make sure it produces the same results as my analysis.

(-1.0, 0.0)
(-1.0, 0.0)


# Attention in a Convolutional Neural Net

This summer I had the pleasure of attending the Brains, Minds, and Machines summer course at the Marine Biology Laboratory. While there, I saw cool research, met awesome scientists, and completed an independent project. In this blog post, I describe my project.

In 2012, Krizhevsky et al. released a convolutional neural network that completely blew away the field at the imagenet challenge. This model is called “Alexnet,” and 2012 marks the beginning of neural networks’ resurgence in the machine learning community.

Alexnet’s domination was not only exciting for the machine learning community. It was also exciting for the visual neuroscience community whose descriptions of the visual system closely matched alexnet (e.g., HMAX). Jim DiCarlo gave an awesome talk at the summer course describing his research comparing the output of neurons in the visual system and the output of “neurons” in alexnet (you can find the article here).

I find the similarities between the visual system and convolutional neural networks exciting, but check out the depictions of alexnet and the visual system above. Alexnet is depicted in the upper image. The visual system is depicted in the lower image. Comparing the two images is not fair, but the visual system is obviously vastly more complex than alexnet.

In my project, I applied a known complexity of the biological visual system to a convolutional neural network. Specifically, I incoporated visual attention into the network. Visual attention refers to our ability to focus cognitive processing onto a subset of the environment. Check out this video for an incredibly 90s demonstration of visual attention.

In this post, I demonstrate that implementing a basic version of visual attention in a convolutional neural net improves performance of the CNN, but only when classifying noisy images, and not when classifying relatively noiseless images.

Code for everything described in this post can be found on my github page. In creating this model, I cribbed code from both Jacob Gildenblat and this implementation of alexnet.

I implemented my model using the Keras library with a Theano backend, and I tested my model on the MNIST database. The MNIST database is composed of images of handwritten numbers. The task is to design a model that can accurately guess what number is written in the image. This is a relatively easy task, and the best models are over 99% accurate.

I chose MNIST because its an easy problem, which allows me to use a small network. A small network is both easy to train and easy to understand, which is good for an exploratory project like this one.

Above, I depict my model. This model has two convolutional layers. Following the convolutional layers is a feature averaging layer which borrows methods from a recent paper out of the Torralba lab and computes the average activity of units covering each location. The output of this feature averaging layer is then passed along to a fully connected layer. The fully connected layer “guesses” what the most likely digit is. My goal when I first created this network was to use this “guess” to guide where the model focused processing (i.e., attention), but I found guided models are irratic during training.

Instead, my current model directs attention to all locations that are predictive of all digits. I haven’t toyed too much with inbetween models - models that direct attention to locations that are predictive of the N most likely digits.

So what does it mean to “direct attention” in this model. Here, directing attention means that neurons covering “attended” locations are more active than neurons covering the unattended locations. I apply attention to the input of the second convolutional layer. The attentionally weighted signal passes through the second convolutional layer and passes onto the feature averaging layer. The feature averaging layer feeds to the fully connected layer, which then produces a final guess about what digit is present.

I first tested this model on the plain MNIST set. For testing, I wanted to compare my model to a model without attention. My comparison model is the same as the model with attention except that the attention directing signal is a matrix of ones - meaning that it doesn’t have any effect on the model’s activity. I use this comparison model because it has the same architecture as the model with attention.

I depict the results of my attentional and comparison models below. On the X-axis is the test phase (10k trials) following each training epoch (60k trials). On the Y-axis is percent accuracy during the test phase. I did 3 training runs with both sets of models. All models gave fairly similar results, which led to small error bars (these depict standard error). The results are … dissapointing. As you can see both the model with attention and the comparison model perform similarly. There might be an initial impact of attention, but this impact is slight.

This result was a little dissapointing (since I’m an attention researcher and consider attention an important part of cognition), but it might not be so surprising given the task. If I gave you the task of naming digits, this task would be virtually effortless; probably so effortless that you would not have to pay very much attention to the task. You could probably talk on the phone or text while doing this task. Basically, I might have failed to find an effect of attention because this task is so easy that it does not require attention.

I decided to try my network when the task was a little more difficult. To make the task more difficult, I added random noise to each image (thank you to Nancy Kanwisher for the suggestion). This trick of adding noise to images is one that’s frequently done in psychophysical attention expeirments, so it would be fitting if it worked here.

The figure above depicts model performance on noisy images. The models are the as before, but this time the model with attention is far superior to the comparison model. Good news for attention researchers! This work suggests that visual attentional mechanisms similar to those in the brain may be beneficial in convolutional neural networks, and this effect is particularly strong with the images are noisy.

This work bears superficial similarity to recent language translation and question answering models. Models like the cited one report using a biologically inspired version of attention, and I agree they do, but they do not use attention in the same way that I am here. I believe this difference demonstrates a problem with what we call “attention.” Attention is not a single cognitive process. Instead, its a family of cognitive processes that we’ve simply given the same name. Thats not to say these forms of attention are completely distinct, but they likely involve different information transformations and probably even different brain regions.

# Revisting NBA Career Predictions From Rookie Performance...again

Now that the NBA season is done, we have complete data from this year’s NBA rookies. In the past I have tried to predict NBA rookies’ future performance using regression models. In this post I am again trying to predict rookies’ future performance, but now using using a classification approach. When using a classification approach, I predict whether player X will be a “great,” “average,” or “poor” player rather than predicting exactly how productive player X will be.

Much of this post re-uses code from the previous posts, so I skim over some of the repeated code.

As usual, I will post all code as a jupyter notebook on my github.

Load the data. Reminder - this data is available on my github.

Load more data, and normalize the data for the PCA transformation.

In the past I used k-means to group players according to their performance (see my post on grouping players for more info). Here, I use a gaussian mixture model (GMM) to group the players. I use the GMM model because it assigns each player a “soft” label rather than a “hard” label. By soft label I mean that a player simultaneously belongs to several groups. For instance, Russell Westbrook belongs to both my “point guard” group and my “scorers” group. K-means uses hard labels where each player can only belong to one group. I think the GMM model provides a more accurate representation of players, so I’ve decided to use it in this post. Maybe in a future post I will spend more time describing it.

For anyone wondering, the GMM groupings looked pretty similar to the k-means groupings.

In this past I have attempted to predict win shares per 48 minutes. I am using win shares as a dependent variable again, but I want to categorize players.

Below I create a histogram of players’ win shares per 48.

I split players into 4 groups which I will refer to as “bad,” “below average,” “above average,” and “great”: Poor players are the bottom 10% in win shares per 48, Below average are the 10-50th percentiles, Above average and 50-90th percentiles, Great are the top 10%. This assignment scheme is relatively arbitrary; the model performs similarly with different assignment schemes.

[0.096314496314496317,
0.40196560196560199,
0.39950859950859952,
0.10221130221130222]


My goal is to use rookie year performance to classify players into these 4 categories. I have a big matrix with lots of data about rookie year performance, but the reason that I grouped player using the GMM is because I suspect that players in the different groups have different “paths” to success. I am including the groupings in my classification model and computing interaction terms. The interaction terms will allow rookie performance to produce different predictions for the different groups.

By including interaction terms, I include quite a few predictor features. I’ve printed the number of predictor features and the number of predicted players below.

(1703, 1432)
(1703,)


Now that I have all the features, it’s time to try and predict which players will be poor, below average, above average, and great. To create these predictions, I will use a logistic regression model.

Because I have so many predictors, correlation between predicting features and over-fitting the data are major concerns. I use regularization and cross-validation to combat these issues.

Specifically, I am using l2 regularization and k-fold 5 cross-validation. Within the cross-validation, I am trying to estimate how much regularization is appropriate.

Some important notes - I am using “balanced” weights which tells the model that worse to incorrectly predict the poor and great players than the below average and above average players. I do this because I don’t want the model to completely ignore the less frequent classifications. Second, I use the multi_class multinomial because it limits the number of models I have to fit.

0.738109219025


Okay, the model did pretty well, but lets look at where the errors are coming from. To visualize the models accuracy, I am using a confusion matrix. In a confusion matrix, every item on the diagnonal is a correctly classified item. Every item off the diagonal is incorrectly classified. The color bar’s axis is the percent correct. So the dark blue squares represent cells with more items.

It seems the model is best at predicting poor players and great players. It makes more errors when trying to predict the more average players.

Lets look at what the model predicts for this year’s rookies. Below I modified two functions that I wrote for a previous post. The first function finds a particular year’s draft picks. The second function produces predictions for each draft pick.

Below I create a plot depicting the model’s predictions. On the y-axis are the four classifications. On the x-axis are the players from the 2015 draft. Each cell in the plot is the probability of a player belonging to one of the classifications. Again, dark blue means a cell or more likely. Good news for us T-Wolves fans! The model loves KAT.

# Creating Videos of NBA Action With Sportsvu Data

All basketball teams have a camera system called SportVU installed in their arenas. These camera systems track players and the ball throughout a basketball game.

The data produced by sportsvu camera systems used to be freely available on NBA.com, but was recently removed (I have no idea why). Luckily, the data for about 600 games are available on neilmj’s github. In this post, I show how to create a video recreation of a given basketball play using the sportsvu data.

This code is also available as a jupyter notebook on my github.

The data is provided as a json. Here’s how to import the python json library and load the data. I’m a T-Wolves fan, so the game I chose is a wolves game.

Let’s take a quick look at the data. It’s a dictionary with three keys: gamedate, gameid, and events. Gamedate and gameid are the date of this game and its specific id number, respectively. Events is the structure with data we’re interested in.

[u'gamedate', u'gameid', u'events']


Lets take a look at the first event. The first event has an associated eventid number. We will use these later. There’s also data for each player on the visiting and home team. We will use these later too. Finally, and most importantly, there’s the “moments.” There are 25 moments for each second of the “event” (the data is sampled at 25hz).

[u'eventId', u'visitor', u'moments', u'home']


Here’s the first moment of the first event. The first number is the quarter. The second number is the time of the event in milliseconds. The third number is the number of seconds left in the quarter (the 1st quarter hasn’t started yet, so 12 * 60 = 720). The fourth number is the number of seconds left on the shot clock. I am not sure what fourth number (None) represents.

The final matrix is 11x5 matrix. The first row describes the ball. The first two columns are the teamID and the playerID of the ball (-1 for both because the ball does not belong to a team and is not a player). The 3rd and 4th columns are xy coordinates of the ball. The final column is the height of the ball (z coordinate).

The next 10 rows describe the 10 players on the court. The first 5 players belong to the home team and the last 5 players belong to the visiting team. Each player has his teamID, playerID, xy&z coordinates (although I don’t think players’ z coordinates ever change).

[1,
1452903036782,
720.0,
24.0,
None,
[[-1, -1, 44.16456, 26.34142, 5.74423],
[1610612760, 201142, 45.46259, 32.01456, 0.0],
[1610612760, 201566, 10.39347, 24.77219, 0.0],
[1610612760, 201586, 25.86087, 25.55881, 0.0],
[1610612760, 203460, 47.28525, 17.76225, 0.0],
[1610612760, 203500, 43.68634, 26.63098, 0.0],
[1610612750, 708, 55.6401, 25.55583, 0.0],
[1610612750, 2419, 47.95942, 31.66328, 0.0],
[1610612750, 201937, 67.28725, 25.10267, 0.0],
[1610612750, 203952, 47.28525, 17.76225, 0.0],
[1610612750, 1626157, 49.46814, 24.24193, 0.0]]]


Alright, so we have the sportsvu data, but its not clear what each event is. Luckily, the NBA also provides play by play (pbp) data. I write a function for acquiring play by play game data. This function collects (and trims) the play by play data for a given sportsvu data set.

Below I show what the play by play data looks like. There’s a column for event number (eventnum). These event numbers match up with the event numbers from the sportsvu data, so we will use this later for seeking out specific plays in the sportsvu data. There’s a column for the event type (eventmsgtype). This column has a number describing what occured in the play. I list these number codes in the comments below.

There’s also short text descriptions of the plays in the home description and visitor description columns. Finally, I use the team column to represent the primary team involved in a play.

I stole the idea of using play by play data from Raji Shah.

EVENTNUM EVENTMSGTYPE HOMEDESCRIPTION VISITORDESCRIPTION TEAM
0 0 12 None None None
1 1 10 Jump Ball Adams vs. Towns: Tip to Ibaka None OKC
2 2 5 Westbrook Out of Bounds Lost Ball Turnover (P1... None OKC
3 3 2 None MISS Wiggins 16' Jump Shot MIN
4 4 4 Westbrook REBOUND (Off:0 Def:1) None OKC

When viewing the videos, its nice to know what players are on the court. I like to depict this by labeling each player with their number. Here I create a dictionary that contains each player’s id number (these are assigned by nba.com) as the key and their jersey number as the associated value.

Alright, almost there! Below I write some functions for creating the actual video! First, there’s a short function for placing an image of the basketball court beneath our depiction of players moving around. This image is from gmf05’s github, but I will provide it on mine too.

Much of this code is either straight from gmf05’s github or slightly modified.

The event that I want to depict is event 41. In this event, Karl Anthony Towns misses a shot, grabs his own rebounds, and puts it back in.

EVENTNUM EVENTMSGTYPE HOMEDESCRIPTION VISITORDESCRIPTION TEAM
37 41 1 None Towns 1' Layup (2 PTS) MIN

We need to find where event 41 is in the sportsvu data structure, so I created a function for finding the location of a particular event. I then create a matrix with position data for the ball and a matrix with position data for each player for event 41.

Okay. We’re actually there! Now we get to create the video. We have to create figure and axes objects for the animation to draw on. Then I place a picture of the basketball court on this plot. Finally, I create the circle and text objects that will move around throughout the video (depicting the ball and players). The location of these objects are then updated in the animation loop.

I’ve been told this video does not work for all users. I’ve also posted it on youtube.

# NBA Shot Charts: Updated

For some reason I recently got it in my head that I wanted to go back and create more NBA shot charts. My previous shotcharts used colored circles to depict the frequency and effectiveness of shots at different locations. This is an extremely efficient method of representing shooting profiles, but I thought it would be fun to create shot charts that represent a player’s shooting profile continously across the court rather than in discrete hexagons.

By depicting the shooting data continously, I lose the ability to represent one dimenion - I can no longer use the size of circles to depict shot frequency at a location. Nonetheless, I thought it would be fun to create these charts.

I explain how to create them below. I’ve also included the ability to compare a player’s shooting performance to the league average.

In my previous shot charts, I query nba.com’s API when creating a players shot chart, but querying nba.com’s API for every shot taken in 2015-16 takes a little while (for computing league average), so I’ve uploaded this data to my github and call the league data as a file rather than querying nba.com API.

This code is also available as a jupyter notebook on my github.

Here, I create a function for querying shooting data from NBA.com’s API. This is the same function I used in my previous post regarding shot charts.

You can find a player’s ID number by going to the players nba.com page and looking at the page address. There is a python library that you can use for querying player IDs (and other data from the nba.com API), but I’ve found this library to be a little shaky.

Create a function for drawing the nba court. This function was taken directly from Savvas Tjortjoglou’s post on shot charts.

Write a function for acquiring each player’s picture. This isn’t essential, but it makes things look nicer. This function takes a playerID number and the amount to zoom in on an image as the inputs. It by default places the image at the location 500,500.

Here is where things get a little complicated. Below I write a function that divides the shooting data into a 25x25 matrix. Each shot taken within the xy coordinates encompassed by a given bin counts towards the shot count in that bin. In this way, the method I am using here is very similar to my previous hexbins (circles). So the difference just comes down to I present the data rather than how I preprocess it.

This function takes a dataframe with a vector of shot locations in the X plane, a vector with shot locations in the Y plane, a vector with shot type (2 pointer or 3 pointer), and a vector with ones for made shots and zeros for missed shots. The function by default bins the data into a 25x25 matrix, but the number of bins is editable. The 25x25 bins are then expanded to encompass a 500x500 space.

The output is a dictionary containing matrices for shots made, attempted, and points scored in each bin location. The dictionary also has the player’s ID number.

Below I load the league average data. I also have the code that I used to originally download the data and to preprocess it.

I really like playing with the different color maps, so here is a new color map I created for these shot charts.

Below, I write a function for creating the nba shot charts. The function takes a dictionary with martrices for shots attempted, made, and points scored. The matrices should be 500x500. By default, the shot chart depicts the number of shots taken across locations, but it can also depict the number of shots made, field goal percentage, and point scored across locations.

The function uses a gaussian kernel with standard deviation of 5 to smooth the data (make it look pretty). Again, this is editable. By default the function plots a players raw data, but it will plot how a player compares to league average if the input includes a matrix of league average data.

Alright, thats that. Now lets create some plots. I am a t-wolves fan, so I will plot data from Karl Anthony Towns.

First, here is the default plot - attempts.