Exploring Mangalyaan tweets with R

Mangalyaan is the spacecraft of Indian Space Research Orgnization’s Mars Orbiter Mission that entered the orbit of Mars last week. There were several tweets in Twitter with hashtag #Mangalyaan about it last week. I wanted to use R to explore those tweets. Tiger Analytics had done an interesting post on this topic last year when Mangalyaan launched. I found their analysis to infer topics particularly interesting. I do hope they repeat their analysis with the latest tweets. My goals and methods of analysis here are much more basic. I wanted to do the following:

  • Extract tweets containing the hashtag #Mangalyaan and create a word cloud
  • Attempt to find some topics/groupings from tweets
    • Try out R topicmodels package to infer toics
    • Find groupings using hierarchical clustering
    • Find groupings using graph community detection algorithms

The full code and explanation is in the following location. I was able to extract about 1000 tweets spanning 4 days from Sep 23, 2014 to Sep 26, 2014 and used it for the analysis below. All the analysis below should be viewed in the context that it is based on a small sample size. The word cloud of the frequent terms in the tweets is:

Next, I used R package topic Models with number of topics set to 5 (no particular reason) and got the following result for top 10 words in each topictopics
I had done only basic preprocessing and ran the model with default parameters. Better preprocessing and model parameter tuning might give better results.

Applying hierarchical clustering on frequent terms gives the following grouping:hierClust

I found that igraph package has some easy to use functions for community detection and plotting. Here the co-occurence of words across tweets is used to construct a graph and the community detection algorithm is applied to that graph. These are plotted both as a dendrogram and a graph plot.

In summary, this was a fun exercise. I got to learn a bit of the following R packages: twitteR, topicModels, igraph.

Posted in R | Tagged , , , , | Leave a comment

Exploring Hotel Review Data from Trip Advisor with R

I wanted to use R to explore hotel review data. I chose to explore reviews for 3 hotels from Trip Advisor. First, I had to scrape the review data. I have described how I scraped the data here. I used the extracted review data and did the following exploratory analysis:
* Check ratings over time
* Check the frequent words in the top quotes for each review grouped by star rating
* Check if I can find any themes in reviews with simple k-means clustering

I have described the exploratory analysis of data here.

I think my analysis was probably a bit simplistic. Right now I didn’t find anything non-obvious from this exploratory analysis. But it was still a fun exercise. In the future, I will explore how topic model packges work with this data.



Posted in R | 1 Comment

Attempt at R version of Peter Norvig’s Spelling Corrector

I recently came across a short tutorial by Peter Norvig on natural language processing using some interesting examples. I found it very interesting and really liked Peter Norvig’s explanations and the fact that it was all in ipython notebook for somebody to work alongside reading the tutorial. I wanted to see how a R version would work and so this is my attempt at a R version for the first part of his tutorial on spelling corrector. The knitted document is in the following location.

Posted in R | Tagged , , , , | 3 Comments

Exploring census and demographic data with R

I wanted to use R to explore how to access and visualize census data, home value data and school rating data using Indianapolis metro area as an example. My learning objectives here are to learn how to use R to:

  • Make choropleth maps at zip code level
  • Process data in XML or JSON format that is returned when using API provided by websites (here I chose Zillow and education.com as an example)

The data that I wanted to overlay over the Indianapolis map was median age, median income, median home listing price/value per sq feet, and school rating at the zip code level. I got the census data using US census data API, zip level home demographic data using Zillow demographic API, school rating data using education.com API. (Disclaimer: This is only a one off use of Zillow API and education.com API. I believe I am in compliance with the terms of use of these API for Zillow and education.com. However, if somebody notices any violation of terms of use of these API, please bring it to my attention and I will remove this material). I should also note here that there are several nice interactive tools and widgets on the web that get the same information. My purpose here is just to play with R to extract and visualize this data.

The full R code along with explanation can be found here.

The maps of median age, median household income, median home listing price/value per sq feet and school rating overlaid over Indianapolis map is shown below




Posted in R | 1 Comment

Using R in Used Car Search

Recently I was looking to buy a used car. I was using the Edmunds website to estimate the true market value of a used car. It has a nice click through interface where it guides you to input make, model, year etc. and will give an estimated value. While the interface is nice, it would be nice to get a broader view across years, models etc. Luckily, Edmunds also has a developer API which is nicely documented. Here I am listing my basic attempts at using the developer API to get some data into R and get exploratory plots of car value by year. For using this API, you need to get a API key (specifically the vehicle API key is used here).

The code used for creating the plots below can be found here.

The plots below show the market value across styles for a given model/make of a car by year. The y axis in these plots is the Edmunds estimated True Market Value for a typically equipped vehicle (not sure how they define it). The bars in the graph span the range of market value.

TMV for Honda Accord


TMV for Toyota Corolla


Edmunds also makes several adjusments for car condition, car mileage and additional car options. The plot below show the TMV adjustment for each option for a 2008 Toyota Corolla (LE 4dr Sedan (1.8L 4cyl 4A)

adjOptionThe plot below shows the adjustment made for car condition for a 2008 Toyota Corolla (LE 4dr Sedan (1.8L 4cyl 4A)


The plot below shows the mileage adjustment again for  a 2008 Toyota Corolla (LE 4dr Sedan (1.8L 4cyl 4A)


Posted in R | 3 Comments

Learning R: Parameter Fitting for Models Involving Differential Equations

It looks like MATLAB, Octave and Python seem to be the preferred tools for scientific and engineering analysis (especially those involving physical models with differential equations). However as part of my learning R experience, I wanted to check out some of R tools for parameter fitting of models involving ordinary differential equations. R has packages deSolve for solving differential equations and FME for parameter fitting. The specific example here is taken from the computational appendix (A.6) of the book Chemical Reactor Analysis and Design Fundamentals by Rawlings and Ekerdt. In fact, all examples in this book are available in Octave and MATLAB.

This example involves two reactions in series

A \xrightarrow{k_1} B \xrightarrow{k_2} C

where we need to estimate the rate constants of the the two reactions k_1 and k_2  from data

Here the required libraries are loaded

# set working directory

# load libraries
library(ggplot2) #library for plotting
library(reshape2) # library for reshaping data (tall-narrow <-> short-wide)
library(deSolve) # library for solving differential equations
library(minpack.lm) # library for least squares fit using levenberg-marquart algorithm

The data available is concentration of A, B, and C over time and is loaded below and plotted

#load concentration data

# plot data


The system being solved is




with inital concentrations being c_A=1, c_B=0, c_C=0.

The rate equations are captured in a function that is an input parameter to the ODE solver

# rate function

 # rate constant passed through a list called parms

 # c is the concentration of species

 # derivatives dc/dt are computed below
 r[1]=-k1*c["A"] #dcA/dt
 r[2]=k1*c["A"]-k2*c["B"] #dcB/dt
 r[3]=k2*c["B"] #dcC/dt

 # the computed derivatives are returned as a list
 # order of derivatives needs to be the same as the order of species in c


Computing the predicted concentration for a given set of rate constants involves just solving the ODEs with intial conditions. This is illustrated for the parameter values k1=2, k2=1 (which is the actual parameters based on which data was generated).

# predicted concentration for a given parameter set

      time     A          B         C
[1,] 0.000 1.00000000 0.0000000 0.00000000
[2,] 0.263 0.59096447 0.3555550 0.05348051
[3,] 0.526 0.34923934 0.4834497 0.16731099
[4,] 0.789 0.20638798 0.4958219 0.29779013
[5,] 1.053 0.12172441 0.4543303 0.42394528
[6,] 1.316 0.07193491 0.3925423 0.53552284

This is wrapped into a function whose input is the parameters to be estimated and the output is the residuals. Here there are three concentrations that are fitted. In general, we may want to put different weights to these but  in this example, they have the same weight. Also, the package FME has several nice features for fitting and uncertainty estimation. Here I have not used FME but directly tried to do the fit more to learn how this works.


 # inital concentration
 # time points for which conc is reported
 # include the points where data is available
 # parameters from the parameter estimation routine
 # solve ODE for a given set of parameters

 # Filter data that contains time points where data is available
 outdf=outdf[outdf$time %in% df$time,]
 # Evaluate predicted vs experimental residual

 # return predicted vs experimental residual


The parameter fitting is done using levenberg-marquardt routine in package minpack.lm.

# parameter fitting using levenberg marquart algorithm
# initial guess for parameters
# fitting

The fitval object has information on estimated parameter and variance covariance matrix

# Summary of fit

Estimate  Std. Error  t value  Pr(>|t|)
k1 2.01906  0.04867     41.49      <2e-16 ***
k2 0.99297  0.01779     55.82      <2e-16 ***
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.0212 on 58 degrees of freedom
Number of iterations to termination: 7
Reason for termination: Relative error in the sum of squares is at most `ftol'.

# Estimated parameter
[1] 2.019065

[1] 0.992973

# degrees of freedom: # data points - # parameters
[1] 58
# mean error
[1] 0.02119577

# variance Covariance Matrix

      k1              k2
k1 0.0023685244 -0.0003605831
k2 -0.0003605831 0.0003164724

The estimated parameters are (k1=2.019,k2=0.993) which is close to the real parameters (k1=2,k2=1) used to generate the data. The predicted profiles is overlaid with experimental data

# plot of predicted vs experimental data

# simulated predicted profile at estimated parameter values

# Overlay predicted profile with experimental data


Estimation of Parameter Uncertainty

Usually though the model is nonlinear, the \alpha confidence region is approximated using the following ellipsoid

(\theta-\hat{\theta})^TS^{-1}(\theta-\hat{\theta}) \leq pF(p,n-p,\alpha)

where \hat{\theta} is the estimated value of parameter, S is the variance covariance matrix of estimated parameters and p=2 is the number of parameters and n=60 is the total number of data points.

# Get the 95% confidence region

# Inverse of covariance matrix

# get points for a circle with radius r
# transform points of circle into points of ellipse using
# svd of inverse covariance matrix
Sinv_svd=svd(Sinv) # inverse of covariance matrix
xt=t(Sinv_svd$v)%*%diag(1/sqrt(Sinv_svd$d))%*%t(z) # transform from circle to ellispse
# translate the ellipse so that center is the estimated parameter value



Another way of estimating the uncertainty is using bootstrapping procedure. Here several simulated datasets are generated using the current estimated model and adding random normal noise to each data point (with mean=0 and variance = mean square error from model). Then the parameter is estimated for each simulated dataset. The set of parameters thus generated indicate the uncertainty in parameter estimates.

# Simulation based estimation of uncertainty

# store original experimental data in a separate dataframe

# conc profile based on estimated k1 and k2


# number of simulations


for (i in 1:nsim){

 # Simulate data set by adding normal random variable with mean 0 and stdev from fit


 # get parameter estimate for the simulated dataset
 # store estimated parameters in the ith row


# plot the parameter estimates from the 1000 simulations
# overlay the 95% ellipse computed previously


Next, the percentage of parameter estimates that fall within the ellipse is  computed and found to be 93% (expected is 95%). In this case, the ellipsoidal approximation of parameter uncertainty seems adequate. But this might not be the case in general.

# percentage of parameters from simulation within the 95% ellipse
for(i in 1:nsim){
sum(tmp <= qf(0.95,2,58)*2)/nsim

[1] 0.933

As I mentioned previously, the package FME has several functions for parameter fitting, MCMC simulation, sensitivity analysis but was not used in this post. Also, when running multiple simulations using ODE models, it might be better to use the C compiled version of reaction rate function. deSolve package provides that ability and describes it in this vignette. The data and R code is available here.

# session Info

R version 3.0.1 (2013-05-16)
Platform: i386-w64-mingw32/i386 (32-bit)

[1] LC_COLLATE=English_United States.1252 LC_CTYPE=English_United States.1252
[3] LC_MONETARY=English_United States.1252 LC_NUMERIC=C
[5] LC_TIME=English_United States.1252

attached base packages:
[1] stats graphics grDevices utils datasets methods base

other attached packages:
[1] minpack.lm_1.1-7 deSolve_1.10-6 reshape2_1.2.2 ggplot2_0.9.3.1

loaded via a namespace (and not attached):
 [1] colorspace_1.2-2 dichromat_2.0-0 digest_0.6.3 grid_3.0.1 gtable_0.1.2
 [6] labeling_0.2 MASS_7.3-26 munsell_0.4 plyr_1.8 proto_0.3-10
[11] RColorBrewer_1.0-5 scales_0.2.3 stringr_0.6.2 tools_3.0.1

Posted in R | 4 Comments

Learning R Using a Chemical Reaction Engineering Book: Part 4

The links to previous parts are listed here. (Part 1, Part 2, Part 3). In this part, I tried to recreate the examples in sections A.3.1 of the computational appendix in the reaction engineering book (by Rawlings and Ekerdt).

Solving a system of ordinary differential equations

This example involves reaction (Benzene pyrolysis) in a plug flow reactor. The actual reactions happening are:

\begin{aligned}  (Rxn1) \; &2B = D + H \\  (Rxn2) \; &B + D = T+H    \end{aligned}

The rate of each reaction is given by:

r_2 = k_2(c_Bc_D-\frac{c_Tc_H}{K_2})

The feed to the reactor consists of 60kmol/hr of Benzene (B). The temperature of the reactor is T=1033K and the pressure is P=1atm.The rate constants and equilibrium constants for this example are:
k_1=7\times 10^5\; L/mol.hr,\; k_2=4\times 10^5 \; L/mol.hr,\; K_1=0.31,\; K_2=0.48

# Appendix A.3.1: Solution of Differential Equations

# Benzene pyrolysis example

# Parameters
# NBf - feed benzene rate - mol/h
# R - Universal gas constant
# T - Reactor temperature K
# P - Reactor pressure atm
# k1 - rxn1 forward rate constant L/mol.h
# k2 - rxn2 forward rate constant L/mol.h
# Keq1 - rxn1 equilibrium constant
# Keq2 - rxn2 equilibrium constant

The governing equations for conversion versus volume in a plug flow reactor is based on extent of each of the reactions:
\frac{d\epsilon_1}{dV}=r_1,\; \frac{d\epsilon_2}{dV}=r_2
The initial conditions (corresponding to feed conditions N_B(0)=60kmol/h,\;N_D(0)=N_H(0)=N_T(0)=0) are that the extent of reaction is zero.
\epsilon_1(0)=0, \; \epsilon_2(0)=0

The flow rates of each component along the reactor volume can be calculated from reaction extent
N_B=N_B(0)-2\epsilon_1-\epsilon_2, \; N_D=\epsilon_1-\epsilon_2, \; N_H=\epsilon_1+\epsilon_2, \; N_T=\epsilon_2

These are setup in a function that can be passed to an ODE solver. In this case the ODE solver we use is lsode from the R package deSolve. The inputs to the function are:

  • Variable over which the integration is done (Volume in this case)
  • The state variables of the system (Extent of the two reactions)
  • Parameters that are needed for description of the system (Rate constants, Temperature, Pressure, etc.)

The output from this function is the rate of change as described by the equations previously.

# load library deSolve for solving ODEs

# function that will be passed to odesolver
# vol is the variable over which the system is integrated (equivalent of time in batch reactions)
# ext is the extent of reactions 1 and 2
# params are the parameters passed to the system
rxnrate=function(vol,ext,params) {

Since the reaction start only after the feed enters the reactor, the extent of reaction is zero for both reactions at the beginning of the reactor (V=0L). The set of volumes where the concentration and reaction extent is computed is chosen in this case to be from 0L to 1600L at every 50L. The ODE solver lsode from deSolve package is used to solve the system of equations.

# initial extent of reaction (zero in this case for both reactions)
# Volumes where the concentration is reported (in this case 0 to 1600L at every 50L)
# Solution of the set of differential equations using lsode solver in deSolve package

extout contains the extent of reaction vs volume data. That is used to compute mole fraction and conversion at different volumes along the reactor.

# Calculation of mole fraction and conversion from extent of reaction at different volumes

Next conversion and mole fraction is plotted as a function of reaction volume

# load library ggplot2 for plotting
# load library reshape2 for data reshaping
# plot of conversion vs volume
scale_x_continuous(breaks=seq(0,1600,by=200))+xlab('Volume (L)')+ylab('Conversion')+theme_bw(20)

# plot of mole fraction vs volume
scale_x_continuous(breaks=seq(0,1600,by=200))+xlab('Volume (L)')+ylab('moleFraction')+theme_bw(20)



R version 2.15.1 (2012-06-22)
Platform: i386-apple-darwin9.8.0/i386 (32-bit)

[1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base

other attached packages:
[1] reshape2_1.2.1  ggplot2_0.9.2.1 deSolve_1.10-4

loaded via a namespace (and not attached):
[1] colorspace_1.1-1   dichromat_1.2-4    digest_0.5.2       grid_2.15.1
[5] gtable_0.1.1       labeling_0.1       MASS_7.3-18        memoise_0.1
[9] munsell_0.3        plyr_1.7.1         proto_0.3-9.2      RColorBrewer_1.0-5
[13] scales_0.2.2       stringr_0.6.1

Posted in R | 2 Comments