June 11, 2014

Forecasting Retail Sales - Linear Regression with R and Hadoop

image from Responsemagazine.com
A retail store tracks the volume of sale for each stock-keeping-unit (SKU) that the store deals with. Given the sales for days 1 through 5, is it possible to predict the sales on days 6 through 10 ? Common sense dictates that sales will remain constant and the average sales per day for the first 5 days will be the same as the average sales per day for the next 5 days. However this may not always be the case, if there is a rising or falling trend. If there is a strongly rising trend, caused by a some strong promotional activity, then the assumption of constant sale will lead to a stock-out and loss of potential business. Similarly, if there is a strongly falling trend, then a similar assumption will lead to accumulation of dead stock and hence a loss related to excessive inventory. Instead of days, the same analysis can be done on the basis of weeks, fortnights or even month. Net-net given the sales over 5 periods of time, it is useful to be able to predict the same for the next 5 periods. How can we do this ? Without resorting to the simplistic "average daily sale" strategy ?

A simple solution is to use linear regression, a well known and widely used statistical tool. If you have the sales data for a particular SKU for the past 5 days, you can "fit" a regression line, determine the slope and the intercept of this line and use the resulting linear regression "model" to predict the expected sales for the next 5 days. Based on these predictions,  you can place orders for these SKUs so that the gap between expected and the actual is minimum. A software tool like R can be used to solve this problem very easily.

All this is well known. But when the number of SKUs is of the order of 50,000 - 70,000 then the time required  -- to build so many regression models, even with R, and then using each to estimate the sales quantity for different SKUs -- becomes enormous ! In fact, if one has to do this on a rolling-basis every day to predict the sales over the next 5 days, then it becomes impossible. Even before we have a solution to today's prediction, the next set of data is waiting and getting stale !

This is where Hadoop steps in. By splitting the regression problem for 50,000 - 70,000 SKUs across multiple computers, it is quite possible to solve the entire problem in a reasonable amount of time. This means that the person responsible for placing orders for the replenishment of inventory would know which of the SKUs would need to be ordered in a higher quantity and which to be ordered in a lower quantity. This is the Linear Regression problem that we will solve with R and Hadoop.

R is not necessary for regression. Any programming language like Java can be used but using R ( or for that matter, a similar tool like Python ) allows the ready-made function -- lm()  for linear regression -- to be used without re-inventing the wheel. In fact, R is a free and open-source statistical tool that is very widely used across the data analytics community. There are two ways to use R with Hadoop. First, we can use the streaming feature of Hadoop with R scripts or we can use the RHadoop set of packages from Revolution Analytics ( which include rhdfs, rmr2 and rhbase). The RHadoop path initially looks easier because it allows one to operate from within the familiar  R environment, but configuring RHadoop is difficult ( or at least, the author was unsuccessful despite a lot of effort). Moreover RHadoop is in reality using the same streaming feature of Hadoop to get the job done. So there is no loss if one ignores RHadoop and uses the native streaming feature of Hadoop directly.

So now we will see how to solve the Sales Forecasting Problem.

We have miniaturized the problem by assuming the Retail Store stocks and sells only 3 products, namely salt, soap and soda. On a particular day, arbitrarily designated as Day 08, the sales of these three SKUs was as follows and this was stored in a file called DailySales08.txt
8 soap 90
8 salt 90
8 soda 120
where the first column represents the day, the second the SKU name (or code) and the third column is the sales on Day 08. There are 4 other files, namely DailySales09.txt, DailySales10.txt, DailySales11.txt, DailySales12.txt.  In reality, each of these files will have very large number of records,  with one record for each SKU

Based on the data for 5 days, from day 08 to day 12, we need to estimate the data for 6th to 10th day, or for day 13 to day 17. Once we run the R program in Hadoop, the following output is generated
salt dates [ 8 - 12 ] : 400  next [ 6 - 10 ] 175 : 17  -- 575  
soap dates [ 8 - 12 ] : 600  next [ 6 - 10 ] 1025 : 239  -- 1625  
soda dates [ 8 - 12 ] : 620  next [ 6 - 10 ] 845 : 187  -- 1465 

where each row represents the picture for each SKU, where we can see in row 1
  • SKU is "salt"
  • cumulative actual sales on days 8 - 12 ( the first 5 days of the analysis ) is 400
  • cumulative expected sales from 6th to 10th day is 175
  • the estimated sale on the last, 10th day, that is day 17 is 17 ( just a coincidence !)
  • the total estimated sales over the 10 day period is 575 ( 400 actual, 175 estimated)
Why is the actual sales in the first 5 days 400 but the predicted sales in the next 5 days only 175 ? See what the regression data reveals :

The black dots represent the actual sales on the first 5 days [ day 8 - day 12 ] Based on this the model has created the regression line : sales = 170 -9*days and with this the estimated values for all the days can be calculated and shown as red dots on the graph. Because of the falling trend, the expected sale in the next 5 days is significantly lower than in the first 5 days. Or so says the regression data !

To run this program, a development environment was created an Ubuntu 14.04 laptop running R 3.0.2 and Hadoop 2.2.0 installed in a single cluster mode as described in my earlier post Demystifying Hadoop and MR with this DIY tutorial.

Section 4 of that tutorial showed how the Hadoop streaming utility was used to run a WordCount program in Python. The same strategy is used in this case, where we have replaced the python programs with two R scripts, LinReg-map.R and LinReg-red.R and a shell script runRetail.sh was used to execute the map-reduce job. The source code of all three scripts along with the 5 datafiles are available at the Git Repository prithwis/Retail.

Once the Mapper (LinReg-map.R) runs, the output looks like this, though in reality, this output will not be stored but instead "streamed" to the Reducer

salt 10$120 
salt 11$50 
salt 12$60 
salt 8$90 
salt 9$80 
soap 10$100 
soap 11$140 
soap 12$160 
soap 8$90 
soap 9$110 
soda 10$150 
soda 11$130 
soda 12$140 
soda 8$120 
soda 9$80 

here the Key is the SKU name, and the Value is a string formed by the concatenation of the date and the quantity sold, separated by the $ char.

In this case there were only 15 records ( 3 SKUs x 5 days ) but even if the number of SKUs is very high, the task of creating this sorted list of <key, value> can be distributed across multiple servers in the Hadoop cluster. This sorted list of records can now be distributed again to multiple servers for the second, reducer, program LinReg-red.R to execute. Hadoop ensures that all records pertaining to any one key ( or SKU) is sent to only one machine where the Linear Regression function is executed.

The reducer program reads through all the <Key, Value> pairs for each Key ( or SKU), splits the Val at the $ char isolate the date and the sale value for that date and create two lists one of dates and the other of the corresponding sale values. These two lists are passed, along with the key (SKU) to the user defined EstValue() function. The fourth parameter N, in our case 9, represents the number of days between the last day of the period and the first day for which data is available. In this case, first day was 8, N is 9, so the last day is the 10th day or day 17.

The EstValue() function is where the Linear Regression module lm() is finally called with the two lists for days, sales as input. For a quick recap of how Linear Regression is done in R, read this tutorial. A little bit of data manipulation is done in which, the days (8,9,10,11,12) are replaced by the more generic (1,2,3,4,5) and so the estimates are done for days (6,7,8,9,10) instead of (13,14,15,16,17). This transformation does not have any implication on the result.

There are 3 ways of testing / running this set of programs of which the first two can be done on a laptop

  • To test the R scripts without calling Hadoop, one can simply pipe the commands as follows : cat DailySales*.txt | ./LinReg-map.R | sort | ./LinReg-red.R > output.txt . This simulates the entire streaming process by sending the data from the 5 data files into the "stdin" of the mapper script that in turn streams the data to the Unix sort utility which in turn streams the sorted key-value pairs to the reducer script which in turn sends the "stdout" output into a file called output.txt This the output that you can see in the post above
  • To run the same scripts on the Hadoop Single Machine Cluster installed on a laptop, we use the following shell script runRetail.sh

#hdfs dfs -ls 
#hdfs dfs -mkdir /user/hduser/Retail-in
#hdfs dfs -copyFromLocal /home/hduser/RetailSales/DailySales*.txt /user/hduser/Retail-in
hdfs dfs -ls /user/hduser/Retail-in
hdfs dfs -rm -r /user/hduser/Retail-out
hadoop jar /usr/local/hadoop220/share/hadoop/tools/lib/hadoop-streaming-2.2.0.jar -D mapred.job.name='RetailR' -mapper /home/hduser/RetailSales/LinReg-map.R -reducer /home/hduser/RetailSales/LinReg-red.R -input /user/hduser/Retail-in/* -output /user/hduser/Retail-out 
hdfs dfs -ls /user/hduser/Retail-out

this scripts creates the Retail-in directory in HDFS and loads the DailySales files from local directory to the HDFS filesystem. It deletes the output directory, if it exists, and then calls the Hadoop streaming program with the 4 mandatory parameters : mapper script, reducer script, input directory and output directory ( all with fully qualified names, to avoid any ambiguity ). The only additional parameter is the job name (RetailR) that helps track the job on http://localhost:8088 and http://localhost:50070

In both these cases, the output is the same. [Update] - To see how to do this directly in RHadoop, see this post.

Now that we know that the program works fine, how do we scale up ? When we have thousands of SKUs and we want to use data from, say 15 or 20 days to build the regression model, the number of records will go up dramatically. One can of course procure multiple servers and configure all of them with Ubuntu, R and Hadoop but this is a very big, complicated and error-prone task. The simple solution is to use the
  • Amazon Web Services Elastic Map Reduce ( AWS/EMR) services, where the Mapper and Reducer programs can be run without any change on the same ( or if necessary on much, much larger) data to get identical results obtained in the first two methods.
To try out AWS/EMR, you need to visit the AWS website with a credit card and sign up for a loginid. Then follow the steps given in this tutorial by Raffael Vogler to run the LinReg map and reduce scripts. Follow the steps but instead of Vogler's programs, use the ones described post. You should also ignore the Bootstrapping step as well two lines of -jobconf stream.num.map.output.key.fields=2 and
-jobconf map.output.key.field.separator=\t that were meant to be placed in the Arguments box since these are not required for the Linear Regression programs. Running the programs with the test data given in this post will take around 10 mins, 7 to provision and configure the machine and 3 mins to run the job. This will result in a charge of around US$ 2 or US$ 3 that will be billed to the credit card used to create the loginid. Should you use AWS/EMR do remember to terminate the cluster at the end of the exercise as otherwise the billing will continue.

AWS/EMR really removes the hassles of configuring Hadoop and makes running Map Reduce jobs as easy as, well almost, send a Gmail message ! Everything is GUI oriented. You choose the number of type of machines and input the location of the data files and the map and reduce scripts. So after building and testing your R scripts on a laptop, you can scale up to hundreds of servers in  a few minutes  and that too for only a few minutes ! Who could ask for anything more ?

In this post, we have defined a simple sales prediction problem that could be faced in any retail store and we have shown how it can be solved with Hadoop and R. The approach taken has been adopted from a YouTube video created and uploaded by Fady El-Rukby and even though he solves a completely different problem and uses native Java, not streaming R, we have used the same data and compared results to make sure that the Linear Regression function of R is working correctly. To learn more about R and Data Science in general, please read this post on Data Science - A DIY approach and to get business perspective join the Business Analytics Program at Praxis Business School, Calcutta

1 comment:

Charanpreet said...

Prithwis - you have explained a typical retail industry problem and the analytics approach to solving it beautifully. This should help demystify analytics in general and big data in particular. Great job!