Creating new column with new "other variable" - dataframe

There are seven "Species" of fish in this data set, some have very few observations . To make analysis of Species a little easier, I need to create a new column in the data called Species.grouped that indicates if the fish is "Perch", "Bream", or "Other". So I have to group the 5 smallest groups of "Species" into a single group called "Other". The resultant column (Species.grouped) should have the value "Perch" if the fish is a Perch, "Bream" if it's a Bream, and "Other" if it's anything else.
Then I need to run a regression predicting the Weight of a fish using Species.grouped and Width as independent predictor variables (no interaction).


Looks like a problem that could be solved by mapping 'Species' values to desired values/categories like 'Perch', 'Bream' and 'Other' using a dictionary and then applying that onto pandas.
This Answer shows plenty of examples that could help you achieve your requirement.

Related

Calculate variable based on values of another variable

I'm learning SPSS for a research methods class but I'm a bit confused on how to enter and define values that represent a meaning after a certain point. For example, the problem I am working on states:
Suppose the following indexed scores represent performance on a new survey meant to understand an individual’s level of depression. Suppose a score of above 20 represents a depressed individual based on the survey design.
Scores: 13.5, 15.7, 14.3, 16.7, 21.2, 20.7, 22.3, 17.4, 16.8, and 12.4
What is the relative frequency of those individuals that represent depressed individuals?
How would one define or make the values over 20 marked as "depressed" to accurately calculate the relative frequency?
Please and thank you!
Picture of the variables and question

You need to calculate a new variable based on the existing one which will have a value of 1 if the original variable is over the threshold, and 0 if not. There are a few ways to do that, this is the simplest one:
compute depressed=(score>-20).
You can now add labels and analyse:
value labels depressed
1 "depressed: score over 20"
0 "score not over 20".
frequencies depressed.

How to populate all possible combination of values in columns, using Spark/normal SQL

I have a scenario, where my original dataset looks like below
Data:
Country,Commodity,Year,Type,Amount
US,Vegetable,2010,Harvested,2.44
US,Vegetable,2010,Yield,15.8
US,Vegetable,2010,Production,6.48
US,Vegetable,2011,Harvested,6
US,Vegetable,2011,Yield,18
US,Vegetable,2011,Production,3
Argentina,Vegetable,2010,Harvested,15.2
Argentina,Vegetable,2010,Yield,40.5
Argentina,Vegetable,2010,Production,2.66
Argentina,Vegetable,2011,Harvested,15.2
Argentina,Vegetable,2011,Yield,40.5
Argentina,Vegetable,2011,Production,2.66
Bhutan,Vegetable,2010,Harvested,7
Bhutan,Vegetable,2010,Yield,35
Bhutan,Vegetable,2010,Production,5
Bhutan,Vegetable,2011,Harvested,2
Bhutan,Vegetable,2011,Yield,6
Bhutan,Vegetable,2011,Production,3
Image of the above csv:
Now there is a very small country lookup table which has all possible countries the source data can come with, listed. PFB:
I want to have the output data's number of columns always fixed (this is to ensure the reporting/visualization tool doesn't get dynamic number columns with every day's new source data ingestions depending on the varying distinct number of countries present).
So, I've to somehow join the source data with the country_lookup csv and populate all those columns with default value as F. Every country column would be binary with T or F being the possible values.
The original dataset from the above has to be converted into below:
Data (I've kept the Amount field unsolved for column Type having Derived Yield as is, rather than calculating them below for a better understanding and for you to match with the formulae):
Country,Commodity,Year,Type,Amount,US,Argentina,Bhutan,India,Nepal,Bangladesh
US,Vegetable,2010,Harvested,2.44,T,F,F,F,F,F
US,Vegetable,2010,Yield,15.8,T,F,F,F,F,F
US,Vegetable,2010,Production,6.48,T,F,F,F,F,F
US,Vegetable,2010,Derived Yield,(2.44+15.2)/(6.48+2.66),T,T,F,F,F,F
US,Vegetable,2010,Derived Yield,(2.44+7)/(6.48+5),T,F,T,F,F,F
US,Vegetable,2010,Derived Yield,(2.44+15.2+7)/(6.48+2.66+5),T,T,T,F,F,F
US,Vegetable,2011,Harvested,6,T,F,F,F,F,F
US,Vegetable,2011,Yield,18,T,F,F,F,F,F
US,Vegetable,2011,Production,3,T,F,F,F,F,F
US,Vegetable,2011,Derived Yield,(6+10)/(3+9),T,T,F,F,F,F
US,Vegetable,2011,Derived Yield,(6+2)/(3+3),T,F,T,F,F,F
US,Vegetable,2011,Derived Yield,(6+10+2)/(3+9+3),T,T,T,F,F,F
Argentina,Vegetable,2010,Harvested,15.2,F,T,F,F,F,F
Argentina,Vegetable,2010,Yield,40.5,F,T,F,F,F,F
Argentina,Vegetable,2010,Production,2.66,F,T,F,F,F,F
Argentina,Vegetable,2010,Derived Yield,(2.44+15.2)/(6.48+2.66),T,T,F,F,F,F
Argentina,Vegetable,2010,Derived Yield,(15.2+7)/(2.66+5),F,T,T,F,F,F
Argentina,Vegetable,2010,Derived Yield,(2.44+15.2+7)/(6.48+2.66+5),T,T,T,F,F,F
Argentina,Vegetable,2011,Harvested,10,F,T,F,F,F,F
Argentina,Vegetable,2011,Yield,90,F,T,F,F,F,F
Argentina,Vegetable,2011,Production,9,F,T,F,F,F,F
Argentina,Vegetable,2011,Derived Yield,(6+10)/(3+9),T,T,F,F,F,F
Argentina,Vegetable,2011,Derived Yield,(10+2)/(9+3),F,T,T,F,F,F
Argentina,Vegetable,2011,Derived Yield,(6+10+2)/(3+9+3),T,T,T,F,F,F
Bhutan,Vegetable,2010,Harvested,7,F,F,T,F,F,F
Bhutan,Vegetable,2010,Yield,35,F,F,T,F,F,F
Bhutan,Vegetable,2010,Production,5,F,F,T,F,F,F
Bhutan,Vegetable,2010,Derived Yield,(2.44+7)/(6.48+5),T,F,T,F,F,F
Bhutan,Vegetable,2010,Derived Yield,(15.2+7)/(2.66+5),F,T,T,F,F,F
Bhutan,Vegetable,2010,Derived Yield,(2.44+15.2+7)/(6.48+2.66+5),T,T,T,F,F,F
Bhutan,Vegetable,2011,Harvested,2,F,F,T,F,F,F
Bhutan,Vegetable,2011,Yield,6,F,F,T,F,F,F
Bhutan,Vegetable,2011,Production,3,F,F,T,F,F,F
Bhutan,Vegetable,2011,Derived Yield,(2.44+7)/(6.48+5),T,F,T,F,F,F
Bhutan,Vegetable,2011,Derived Yield,(10+2)/(9+3),F,T,T,F,F,F
Bhutan,Vegetable,2011,Derived Yield,(6+10+2)/(3+9+3),T,T,T,F,F,F
The image of the above expected output data for a structured look at it:
Part 1 -
Part 2 -
Formulae for populating Amount Field for Derived Type:
Derived Amount = Sum of Harvested of all countries with T (True) grouped by Year and Commodity columns divided by Sum of Production of all countries with T (True)grouped by Year and Commodity columns.
So, the target is to have a combination of all the countries from source and calculate the sum of respective Harvested and Production values which then has to be divided. The commodity can be more than one in the actual scenario for any given country, but that should not bother as the summation of amount happens on grouped commodity and year.
Note: The users in the frontend can select any combination of countries. The sole purpose of doing it in the backend rather than dynamically doing it in the frontend is because AWS QuickSight (our visualisation tool), even though can populate sum on selected column filters but doesn't yet support calculation on those derived summed fields. Hence, the entire calculation of all combination of countries has to be pre-populated (very naive approach) in order to make it available in report on dynamic users selection of countries.
Also if you've any better approach (than the above naive approach mentioned in note) to solve this problem, you are most welcome to guide me. I've also posted a question on the same problem without writing my expected approach for experts to show me the path on how we can solve this kind of a problem better than this naive approach. If you want to help solve it with some other technique, you're most welcome, here is the link to that question.
Any help shall be greatly acknowledged.

Merge two CSV and collate data

I have two CSV files, the first like so:
Book1:
ID,TITLE,SUBJECT
0001,BLAH,OIL
0002,BLAH,HAMSTER
0003,BLAH,HAMSTER
0004,BLAH,PLANETS
0005,BLAH,JELLO
0006,BLAH,OIL
0007,BLAH,HAMSTER
0008,BLAH,JELLO
0009,BLAH,JELLO
0010,BLAH,HAMSTER
0011,BLAH,OIL
0012,BLAH,OIL
0013,BLAH,OIL
0014,BLAH,JELLO
0015,BLAH,JELLO
0016,BLAH,HAMSTER
0017,BLAH,PLANETS
0018,BLAH,PLANETS
0019,BLAH,HAMSTER
0020,BLAH,HAMSTER
And then a second CSV with items associated with the first list, with ID being the common attribute between the two.
Book2:
ID,ITEM
0001,PURSE
0001,STEAM
0001,SEASHELL
0002,TRUMPET
0002,TRAMPOLINE
0003,PURSE
0003,DOLPHIN
0003,ENVELOPE
0004,SEASHELL
0004,SERPENT
0004,TRUMPET
0005,CAR
0005,NOODLE
0006,CANNONBALL
0006,NOODLE
0006,ORANGE
0006,SEASHELL
0007,CREAM
0007,CANNONBALL
0007,GUM
0008,SERPENT
0008,NOODLE
0008,CAR
0009,CANNONBALL
0009,SERPENT
0009,GRAPE
0010,SERPENT
0010,CAR
0010,TAPE
0011,CANNONBALL
0011,GRAPE
0012,ORANGE
0012,GUM
0012,SEASHELL
0013,NOODLE
0013,CAR
0014,STICK
0014,ORANGE
0015,GUN
0015,GRAPE
0015,STICK
0016,BASEBALL
0016,SEASHELL
0017,CANNONBALL
0017,ORANGE
0017,TRUMPET
0018,GUM
0018,STICK
0018,GRAPE
0018,CAR
0019,CANNONBALL
0019,TRUMPET
0019,ORANGE
0020,TRUMPET
0020,CHERRY
0020,ORANGE
0020,GUM
The real datasets are millions of records, so I'm sorry in advance for my simple example.
The problem I need to solve is getting the data merged and collated in a way where I can see which item groupings most commonly appear together on the same ID. (e.g. GRAPE,GUM,SEASHELL appear together 340 times, ORANGE and STICK 89 times, etc...)
Then I need to see if there is any change/deviation to the general results in common appearance when grouped by SUBJECT.
Tools I'm familiar with are Excel and SQL, but I also have PowerBI and Alteryx at my disposal.
Full disclosure: Not homework, or work, but a volunteer project, thus my unfamiliarity with this kind of data manipulation.
Thanks in advance.

An Alteryx solution:
Drag the two .csv files onto your canvas (seen as book1.csv and book2.csv in my picture; Alteryx will create "Input" tools for you.
Drag a "Join" tool on and connect the two .csv files to its inputs; select "ID" as the join field; unselect the "Right_ID" as output since it's merely a duplicate of "ID"
Drag a "Summary" tool on and connect the Join tool's output to the Summary tool's input; select all three of the outputs and add as a "group by"... then add the ID column with a "count"
Drag a browse tool on and connect the summary's output to the browse tool's input.
run the workflow
After all that, click on the browse tool and you should see what is seen in my screenshot: (which is showing just the first ten rows of output):

+1 for taking on a volunteer project - I think anyone who knows data can have a big impact in support of their favourite group or cause.
I would just pull the 2 files into Power BI as 2 separate tables (Get Data / From File). Create a relationship between the 2 tables based on ID (it might get auto-generated). It should be one to many.
Then I would add a Calculated Column to the Book1 table to Concatenate the related ITEM values, eg.
Items =
CALCULATE (
CONCATENATEX (
DISTINCT ( 'Book2'[ITEM] ),
'Book2'[ITEM],
", ",
'Book2'[ITEM], ASC
)
)
Now you can use that Items field in visuals (e.g. a Table), along with Count of ID to get the frequency.
Adding Subject to a copy of the table (e.g. to the Columns well of a Matrix) will produce your grouped scenario, or you could add a Subject Slicer.
As you will be comparing subsets of varying size, I would change Count of ID to Show value as - % of grand total.

Little different solution using Alteryx.
With this dataset, there are very few repeating 3 or 4 item groups. You can do the two item affinity analysis and get a probability of 3 or 4 item groups, or you can count the 3 and 4 item groups individually. I believe what you want is the latter as your probability of getting grapes with oranges may be altered by whether you have bananas in the cart or not.
Anyway, I did not join in the subject until after finding all of my combinations. I found all the combinations by taking the Cartesian join of two, then three, then four of the original set. I then removed all duplicates by ensuring items were always in alphabetical order in each row. I then counted occurrences of each combination. More joins can be added in the same pattern to count groups of 5,6,7...
Once you have the counts of occurrences, then I would join back with the subjects and perform this analysis on each group and compare to the overall results.
I'm supposed to disclose that I work for Alteryx.

first of all if you are using windows
just navigate to the directory which contains the CSV and write the following command:
copy pattern newfileName.csv
#example
copy *.csv merged.csv
now you created one csv file, the file is too large now you can't process it once, depending on your programming language you can use appropriate way, for python you can use generators to process line by line, or pandas you can read chunk by chunk it will be easy.
I hope this help you.

Build a Kibana Histogram with buckets dynamically created by ElasticSearch terms aggregation

I want to be able to combine the functionality of the Kibana Terms Graph (be able to create buckets based on uniqueness of values from a particular attribute) and Histogram Graph (separate data into buckets based on queries and then illustrate the date based on time).
Overall, I want to create a Histogram, but I only want to create the Histogram based on the results of one query, not multiple queries like it's being done in the Kibana demo app. Instead, I want each bucket to be dynamically created per unique value of my particular field. For example, consider the following data returned by my query:
{"myValueType": "New York"}
{"myValueType": "New York"}
{"myValueType": "New York"}
{"myValueType": "San Francisco"}
{"myValueType": "San Francisco"}
Also assume that each record has a timestamp field for separating histogram data by date. For that particular date, I want the data to be communicated as a count of 3 into the New York bucket and a count of 2 into the San Francisco bucket. However, I am only able to show a count of 5 for my one linked query. When I configure the Histogram, I am able to specify a field to use for my timestamp, but not to create buckets from. I could've sent a field to compute a total/min/max/mean, but this field would've had to be numeric, so that is not the solution either.
If I were to use a Term Graph to create a pie or bar graph, I am indeed able to separate my data into buckets based on the unique values of my specified field (in this case, "myValueType"), but this would total up the data for all-time, not split up the data by timestamp. Although this is good information to know, it is not ideal because I wouldn't be able to detect trends in my data.
I am looking for a solution that will do one of the following:
Let me dynamically create queries in my Kibana dash board to create "buckets" in a Histogram
Allow me to run an ElasticSearch Terms Aggregation to supposidly split up my data into buckets based on "myValueType" and integrate these results into my Histogram
Customize the JSON of my dashboard, but this doesn't look possible to me
Create my own custom panel, but this is not desirable
Link a Kibana "TopN" query in Kibana. Actually, this has proven to be a work-around for my problem because the TopN query dynamically created one query per unique value/term from the specified fieldName. However, the problem is that I can only link one colour to this TopN query and each unique term will be placed in a bucket that uses a different shade of the colour. Ideally, every bucket in my Histogram will have a completely different colour associated to it. Imagine how difficult it will be to distinguish unique terms as the number of buckets grows.
If all else fails, I make one query per unique value from my search field. This will allow me to have one unique colour per bucket, but as the number of unique terms in the "myValueType" field changes, I need to keep adding/removing queries from Kibana, which can get quite messy.
I'm sure there is someting that I am missing here. Please help me out. Many thanks.
A highly related SOF question: Is it Possible to Use Histogram Facet or Its Curl Response in Kibana

This would be a great feature. It looks like it will be supported in Kibana4, but there doesn't seem to be much more info out there than that.
For reference: https://github.com/elasticsearch/kibana/issues/1249

Maybe a little late but it is actually possible in the newest BETA release.
kibana 4 beta 3 installation download

How to do car search like autoscout24.de with / without SQL?

I am interested in the implementation of the search engine in autoscout24.de. It is a platform where you can sell/buy cars. Every car advert has properties: make, price, kilometers, color, etc. (in sum over 50 different properties) that can be searched for.
I am specifically interested in the detail search that works like this: every possible property is displayed on the page. In brackets behind each property there is the number of cars that will match the new search if the property is selected.
Example: I'll start with empty search criterias.
Property make:
BMW (100.000)
Volkswagen (200.000)
Ford (150.000)
...
Property color:
black (210.000)
silver (50.000)
white (100.000)
...
and so on for the other properties.
I'd like to know:
How would you implement this kind of search with SQL?
How would you implement it with an in-memory data structure?
Range queries should be supported, too (all cars with price from X to Y)
Update:
The numbers in brackets show the number of results after the addition of the search criteria. So it changes each time a property is added / removed...
So a naive algorithm would work like this:
find all cars with current search criteria (e.g. make Ford)
for each property do: find all cars that matches previous search criteria ("Ford") AND the search criteria for the chosen property. Write the count in brackets behind the property.
This algorithm is naive because it would execute 1 + N queries (N=#properties). Nobody wants to do that ;-)

I believe that this is referred to as "faceted search". The Apache Solr project might be worth looking at.

It's a basic code
Create a result object with one counter for each property that the cars have
Check all cars one by one, if the car match the filter then add one to each of the numbers
...But it's blasting fast !
I think they do it on several computers, shreading data across them. Each computer compute 5% of the data and send the result to the front computer wich sum all counts.
There are tools for that : look for "map reduce", "elastic search", "strom"...

Have a properties table:
+Properties
id
title
value
count
The count field allows you to "earn" an extra query , so instead of checking how much cars have a certain property , you can just update this field when adding new cars.
Example of rows in this table:
1 'color' 'white' 1000
2 'color' 'black' 122
3 'km' '5000' 1233
4 'km' '30000' 54
And for the cars table , for each property add a field.
+Cars
id
color
km
and the color and km fields will hold the ID's of the property's row in the Properies table.
EDIT: if you're planning not to use mysql db , you might consider using XML files to contain the properties data. But once again, you should update its count value anytime you add / remove or update a car.
<Properties>
<Property>
<Type>Color</Type>
<Value>White</Value>
<Count>1000</Count>
</Property>
</Properties>