## Monday, September 4, 2017

### How Pets Get Admitted and Later Leave Dallas Animal Shelters

Thanks to Dallas OpenData anyone has access to the city animal shelter records.  If you lost or found a pet it could be that he or she spent some time in a shelter - I personally took lost dogs there. It's unfortunate but every year tens of thousands of animals find their way to shelters with significant fraction never finding way out.

### What and How Many Animals are Admitted?

City of Dallas animal shelter dataset contains 5 different animal types with solid lead belonging to dogs and cats (hardly any surprise to anyone):

For consistency and plausibility of analysis we will focus on cats and dogs only

Each shelter record has animal's intake type (reason animal was admitted) and outcome (cause for animal disharge). Top 2 reasons why cats and dogs turn up at shelters are Stray (lost or abandoned) and Owner Surrender (willingly brought in by owner) while Confiscated (abused, no owner, etc.) is #3 for dogs but not cats

### How Animals Leave Shelter

Animals leave shelters (either alive or dead) for 4 main reasons (outcomes):  Adoption (good),  Euthanized (bad), Returned to Owner (good), and Transfer (neutral):

Unfortunately, for both cats and dogs the top reason to leave shelter is being euthanized. But that's where similarity between them ends:

• cats don't get returned to owner anywhere near as often as dogs;
• dogs' adoption and euthanized rates are almost the same while cats get adopted far less.

### From Admissions to Outcomes with Sankey

So what is the relationship between intake types and outcomes? Which and to what extent intake types drive outcomes? The good news there is some causality effect because each stay begins with intake type and ends with outcome.

We begin analyzing this relationship with higher level (in that case) but visually appealing visualization called sankey diagram (or just sankey). It is a specific type of flow diagram, in which the width of the arrows is shown proportionally to the flow quantity. In our case each dog shelter stay contributes to the pipe size flowing from left (an intake type) to right (an outcome). With this we basically visualize conditional probabilities of dog leaving shelter with certain outcome given its admission intake type (first image illustrates transitions for cats and second does the same for dogs):

While Owner Surrender intake type flows similarly for both, Stray animals don't: cat outcomes are dominated by Euthanized but dogs are dominated by Adoption with Transfer and Returned to Owner outcomes together matching Euthanized.

### Correlations Between Admissions and Outcomes

Next, we go beyond overall totals used in the sankey and compute correlations. To correlate intake types and outcomes we construct time series by computing monthly totals for each intake type and outcome obtaining monthly trends. Then we correlate between monthly trends (separate for cats and dogs) animals brought in and removed from Dallas animal shelters for each pair of top intake types (Confiscated, Owner Surrender, and Stray) and outcomes (Adoption, Euthanized, Returned to Owner, and Transfer) - 12 coefficients in total:

In this case strong correlation implies (at least to some extent) causation effect due to presence of temporal relationship, consistency, and plausibility criteria (see here and here). Few observations to note:
• The highest correlation for cats (0.91) and second highest for dogs (0.77) are observed between intake Surrendered by Owner and outcomeEuthanized which is almost as obvious as unfortunate.
• Correlation between Stray and Returned to Owner for dogs is the highest at 0.86. This is great news because it means the more dogs get lost the more of them are found. The higher this correlation the healthier the city for 2 reasons: a) lost animals return home and b) larger share of stray dogs are lost ones and not abandoned (given that the city keeps collecting them).
• Unfortunately trend in Stray cats correlates highly with Euthanized. So while Stray dog trend drives adoptions and returns, Stray cat trend affects euthanizations the most (we've seen that in sankey as well).
• No trends are affected by variations in Confiscated dogs, but this is likely due to smaller share of such admissions.
• Variation in Stray dogs admitted affect every outcome (but Euthanized). Indeed Stray intake type is the largest and is almost twice as big as the 2d largest dog type Owner Surrender
• Low correlation for dogs between Stray and Euthanized needs additional analysis because it's counter-intuitive.

### Monthly Trends

But can we do better than correlations of these trends which technically are sophisticated but still aggregates? Next visual places time series instead of correlation coefficients inside the same matrix grid allowing to see and compare actual monthly trends:

Note that each plot is a 3 x 4 matrix - the same dimension as correlations matrices before. But instead of correlation coefficient each cell contains a pair of monthly trends (in fact, each correlation was computed for these exact pairs of trends, hence, a reference to its aggregation origin). Each row corresponds to an intake type (the same blue line in each) and each column to an outcome (the same red line in each). Being able to see trends over time let's record a few observations (following the matrix order top down):
• Confiscated intake trends flat for both cats and dogs with only significant spike for dogs in January 2016. This spike is so unusual, relatively big, and contained within single month or two that it begs additional investigation into probable external event or procedural change that may have caused it.
• Number of Confiscated animals is relatively low to noticeably affect outcomes. Still, if we can reduce effect of other intake types some relationships are possible.
• Owner Surrender trend correlation with Euthanized outcome is so obvious that this type of visualization is sufficient to find it. Yes, it is unfortunate but people bring their old or unhealthy pets for a reason.
• The same applies to Stray and Owner Surrender for cats only.
• Owner Surrender has significant seasonal component spiking in summer possibly due to hot weather or holiday season or both. For cats only seasonal component is also strong in Stray trend.
• Euthanized trends together with Owner Surrender which causes it to a large degree.
• Stray dogs trend slowly upwards in Dallas and it's alarming.
• Adoption also trends upwards but not steep enough to compensate for inflow of dogs into shelters.  Targeted campaign to encourage more adoptions of pets in the city is due.
• Transfer outcome trending upward also compensates  for the growth in stray dogs. It's not clear if it's positive or negative though as there is no means to track what happens to dogs after transfer (or is it?).
• Stray trend for dogs dipped in January 2016 exactly when confiscated trend spiked - it could be a coincidence or related - for sure something to consider when investigating further.
• For dogs Euthanized trend correlates strongly with Stray intake until the summer of 2016 when they start to diverge in opposite directions - again some policy or procedural change apparently caused it. Indeed, if we observe other outcomes we notice that Returned to Owner trend began its uptick at around the same time (indeed, after I observed this I found out about this and this - significant changes in Dallas Animal Services leadership and policies around summer and fall of 2016).
I will be back with more analysis (survival analysis). R code for data processing, analysis, and visualizations from this post can be found here.

## Wednesday, July 5, 2017

### The Role of Small Data and Vacation Recap Example

Wikipedia defines small data 'small' enough for human comprehension but then it goes further by qualifying data in a volume and format that makes it accessible, informative and actionable. I am not certain the latter is always true: smaller footprint doesn't automatically qualify data as informative and actionable without more work. In my book small data usually scales to kilobytes and has just a handful of dimensions. But its main feature remains human comprehension which really means there is simple story behind it.

In the grand scheme of big data things small data story is the last mile of data science analysis. It still requires interpretation (or representation) in the form of visualization or application.

Case in point could be Google spreadsheet I kept while on vacation in Italy with daily recordings of miles and steps walked. Later I added main attractions for each day. The result was my personal small data covering about 2 weeks of touring Italy with bases in Rome and later in Sicily (this sentence was the story):

 Google sheet of activities while on vacation in Italy

As-is this spreadsheet is destined to Google archives contributing to ever growing collection of docs I created and happily forgot about. So I created this visualization that represents both most of data and the story:

 Small data visualization

Before explaining how this visualization was created with R I ought to acknowledge that Google spreadsheets offer adding a chart or graph to a document. But its functionality appears rather limited without resorting to JavaScript API.

Using R googlesheets package to source Google docs makes them integral part of data sources available from within R code:
For details on how code above authenticates with Google servers and processes documents see very detailed vignettes

Now we can get back to small data and its simple story. Which means single visualization may include most if not all of it. In case of small data the goal is designing such chart without sacrificing clarity.

Core attributes days (Date) and miles walked (Distance; I chose miles over Steps for simplicity) suggest a line chart with timeline along x-axis and distance for y-axis. But there are 2 more factors to incorporate: Place indicating where the base city was each day and Label for major attractions.

Base city receives color identification with deep red for Rome and olive green for Syracuse. Major attractions text was attached to each point with smart justifications to fit inside the chart:

Had I kept more detailed log I would have ended up with more dimensions to use. For example, miles driven by car or train, time spent at leisure versus touring, number of cities and places visited, historical marker attributes and so on. But that moves us further away from small data domain as footprint and dimensions grow and story becomes less comprehensible. One of indicators of this is that it becomes harder to collect data manually. Instead, there are apps that would do it for me, for example, Life Cycle or Apple Health.

Ultimately any big data problem is reduced to one or more small data ones by aggregating, regressions, clustering or some other data science method. The path to big data insights is a journey from big to small data in search of simple story. So learning how to deal with small data is where it all both ends and begins.

## Friday, June 23, 2017

### Logarithmic Scale Explained with U.S. Trade Balance

Skewed data prevail in real life. Unless you observe trivial or near constant processes data is skewed one way or another due to outliers, long tails, errors or something else. Such effects create problems in visualizations when a few data elements are much larger than the rest.
Consider U.S. 2016 merchandise trade partner balances data set where each point is a country with 2 features: U.S. imports and exports against it:
Suppose we decided to visualize top 30 U.S trading partners using bubble chart, which simply is a 2D scatter plot with the third dimension expressed through point size. Then U.S. trade partners become disks with imports and exports for xy coordinates and trade balance (abs(export - import)) for size:

China, Canada, and Mexico run far larger balances compared to the other 27 countries which causes most data points to collapse into crowded lower left corner. One way to "solve" this problem is to eliminate 3 mentioned outliers from the picture:

While this plot does look better it no longer serves its original purpose of displaying all top trading partners. And undesirable effect of outliers though reduced still presents itself with new ones: Japan, Germany, and U.K. So let us bring all countries back into the mix by trying logarithmic scale.
Quick refresher from algebra. Log function (in this example log base 10 but the same applies to natural log or log base 2) is commonly used to transform positive real numbers. All because of its property of mapping multiplicative relationships into additive ones. Indeed, given numbers A, B, and C such that

A*B=C and A,B,C > 0

applying log  results in additive relationship:

log(A) + log(B) = log(C)

For example, let A=100, B=1000, and C=100000 then

100 * 1000 = 100000

so that after transformation it becomes

log(100) + log(1000) = log(100000)  or   2 + 3 = 5

Observe this on 1D plane:

Logarithmic scale is simply a log transformation applied to all feature's values before plotting them. In our example we used it on both trading partners' features - imports and exports which gives bubble chart new look:

The same data displayed on logarithmic scale appear almost uniform but not to forget the farther away points from 0 the more orders of magnitude they are apart on actual scale (observe this by scrolling back to the original plot). The main advantage of using log scale in this plot is ability of observing relationships between all top 30 countries without loosing the whole picture and avoiding collapsing smaller points together.

For more detailed discussion of logarithmic scale refer to When Should I Use Logarithmic Scales in My Charts and Graphs? Oh, and how about that trade deficit with China?

This is a re-post from the original blog on LinkedIn.

## Friday, May 26, 2017

### MapReduce in Two Modern Paintings

Two years ago we had a rare family outing to the Dallas Museum of Art (my son is teenager and he's into sport after all). It had an excellent exhibition of modern art and DMA allowed taking pictures. Two hours and dozen of pictures later my weekend was over but thanks to Google Photos I just stumbled upon those pictures again. Suddenly, I realized that two paintings I captured make up an illustration of one of the most important concepts in big data.

There are multiple papers, tutorials and web pages about MapReduce and to truly understand and use it one should study at least a few thoroughly. And there are many illustrations of MapReduce structure and architecture out there.

But the power of art can express more with less with just two paintings. First, we have work by Erró Foodscape, 1964:

 Foodscape, 1964 Oil on canvas
It illustrates variety, richness, potential of insight (if consumed properly), and of course, scale. The painting is boundless with no ends to the table surface in all 4 directions. Also observe many types of food and drinks, packaging, presentations, varying in colors, texture and origin (better quality image here). All these represent big data so much better than any kind of flowchart diagram.

The 2d and final painting is by Wayne Thiebaud Salads, Sandwiches, and Desserts, 1962:

 Salads, Sandwiches, and Desserts, 1962 Oil on canva
Should we think of how MapReduce works this seemingly infinite table (also fittingly resembling conveyor line) looks like result of split-apply-combine executed on Foodscape items. Indeed, each vertical group is combination of the same type of finished and plated food combined into variably sized groups and ready to serve (better quality image here).

As with any art there is much about MapReduce that was left out of the picture. That's why we still have papers, books, and Wikipedia.  And again, I'd like to remind of importance of taking your kids to a museum.

## Tuesday, December 20, 2016

### Correlation Primer with Aster and R

Calculating correlations is often starting point before more advanced analytical steps take place. Big data (long data) always presents computational challenges of both scale and distributed nature. In turn they may get aggravated by the presence of large number of features (wide data). But challenges do not stop here as complex relationships induce analysis of correlations across subsets and groups.

Such mix of long and wide becomes more common in the age of internet-of-things, sensor and machine data with non-human data sources dominating analytical use cases.
Thus, when computing correlations on big data the following capabilities matter:
• scale on large distributed data sets (long data)
• scale on wide distributed data sets (wide data / large number of features)
• flexibility on wide data sets (ability to permutate features such as Cartesian combinations, one-to-many, etc.)
• correlations on subsets and groups.
Correlations in R comes standard with stats function cor but it doesn't meet most of the capabilities above. As always Teradata Aster big data analytical platform offers both scalability and functionality far exceeding capabilities above. And thanks to Aster R (TeradataAsterR) package it is available without leaving R environment.

With Aster and R integration there are multiple ways of correlating on datasets. Before sending you to the link for detailed discussion I summarized approaches discussed there by the capabilities:

Method / Solution features Variable (columns) Permutations Calculating for Groups SQL-MR In-database R
Aster R ta.cor
N
N
Y
N
Aster R in-database ta.tapply
N
Y
N
Y
toaster computeCorrelations
Y
Y
Y
N

Please visit my latest RPubs post for detailed discussion and comparison of these methods.

## Tuesday, May 31, 2016

### Running similar but independent jobs in parallel on Aster with R

No surprise that Teradata Aster runs each SQL, SQL-MR, and SQL-GR command in parallel on many clusters with distributed data. But when faced with the task of running many similar but independent jobs one has to do extra work to parallelize them in Aster. When running a SQL script the next command has to wait for the previous to finish. This makes sense when commands contribute to the pipeline with results of each job passed down to next one. But what if the jobs are independent and produce their own results each. For example, cross-validation of linear regression or other models is divided into independent jobs each working with its respective partition (of total K in case of K-fold cross-validation). These jobs could run in parallel in Aster with little help from R. This post will illustrate how to run K linear regression models in parallel in Aster as part of the K-fold cross-validation procedure.

### The Problem

Cross-validation is important technique in machine learning that receives its own chapters in the textbooks (e.g. see Chapter 7 here). In our examples we implement a K-fold cross-validation method to demonstrate how to run parallel jobs in Aster with R. The implementation of K-fold cross-validation that will be given is neither exhaustive nor exemplary as it introduces certain bias (based on month of the year) into the models. But this approach could definitely lead to a general solution for cross-validation and other problems involving execution of many similar but independent tasks on Aster platform.

Further more, the examples will be concerned only with the step in K-fold cross-validation that creates K models on overlapping but different partitions of the training dataset. We will show how to construct K independent linear regression models in parallel on Aster, each for one of the K partitions of the table (not the same as table partitioning in Aster).

### Data and R Packages

We will use Dallas Open Data data set available from here (including Aster load scripts).
To simplify examples we will also use R package toaster for Aster and several other packages - all available from CRAN:

### Data set, Model and K Folds

Dallas Open Data has information on building permits across city of Dallas for the period between January 2012 through May 2014 stored in the table dallasbuildingpermits. We can quickly analyze this table from R with toaster and see its numerical columns:

which results in:
[1] "area" "value" "lon" "lat"
These 4 fields will make up our simple linear model to determine the value of construction using its area and location. And now the same in R terms:

This problem is not beyond R memory limits but our goal is to execute linear regression in Aster. We enlist toaster's computeLm function that returns R lm object:

Lastly, we need to define the folds (partitions) on the table to build linear regression model on each of them. Usually, this step performs equal and random division into partitions. Doing this with R and Aster is actually not extremely difficult but will take us beyond the scope of the main topic. For this reason alone we propose quick and dirty method of dividing building permits into 12 partitions (K=12) using issue date's month value (in SQL):

Again, do not replicate this method in real cross-validation task but use it as a template or a prototype only.
To make each fold's compliment (used to train 12 models later) we simply exclude each month's data, e.g. selecting the compliment to the fold 6 in its entirety (in SQL):

### Computing Cross-Validation Models in Aster with R

Before we get to parallel execution with R we show how to script in R Aster cross-validation of linear regression. To begin we use standard R for loop and computeLm with the argument where that limits data to the required fold just like in SQL example above:

This results in the list fit.folds that contains 12 linear regression models for each fold respectively.
Next, we replace the for loop with the specialized foreach function designed for parallel execution in R. There is no parallel execution yet but all necessary structure for transition to parallel processing:

foreach performs the same iterations from 1 to 12 as for loop and combines results into list by default.

### Parallel Computing in Aster with R

Finally, we are ready to enable parallel execution in R. For more details on using package doParallel see here, but the following suffices to enable a parallel backend in R on Windows:

After that foreach with operator %dopar% automatically recognizes parallel backend cl and runs its iterations in parallel:

Comparing with foreach %do% earlier notice extra handling for ODBC connection inside foreach %dopar%. This is necessary due to inability of sharing the same database connection between parallel processes (or threads, depending on the backend implementation). Effectively, with each loop we reconnect to Aster with a brand new connection by reusing original connection's configuration in function odbcReConnect.

### Elaspsed Time

Lastly, let's see if the whole thing was worth the effort. Chart below contains elapsed times (in seconds) for all 3 types of loops: for loop in R, foreach %do% (sequential), and foreach %dopar% (parallel):

## Sunday, April 24, 2016

### Map of the Windows Fonts Registered with R

If you already found package extrafont then you probably found how to load and use Windows fonts in R visualizations. But just in case, everything to get started with extrafont is found here and summarized for using fonts in Windows for on-screen or bitmap output below:

One thing to add is a summary of all Windows fonts registered in R. This will come handy when designing new visualizations and deciding on which font or combination of fonts and their faces to use. The code below produces a table where rows are fonts and columns are faces with font name printed using both the font and the face (if available) in each table cell:

The resulting table is this handy visual: