I have confusion about the numbers at the end of the branches of a J48 tree. For example, using the weather.nominal data the tree looks the same, whether the Test options are set to Use training set or Cross-validation or Percentage split.
This is the output:
J48 pruned tree
------------------
outlook = sunny
| humidity = high: no (3.0)
| humidity = normal: yes (2.0)
outlook = overcast: yes (4.0)
outlook = rainy
| windy = TRUE: no (2.0)
| windy = FALSE: yes (3.0)
According to the textbook by the authors of this software, in an example using this exact data they say, "In the tree structure, a colon introduces the class label that has been assigned to a particular leaf, followed by the number of instances that reach that leaf, expressed as a decimal number because of the way the algorithm uses fractional instances to handle missing values. If there were incorrectly classified instances (there aren’t in this example) their number would appear, too: thus 2.0/1.0 means that two instances reached that leaf, of which one is classified incorrectly"
So this means that no instances were incorrectly classified in the above tree with the weather.nominal dataset.
On the other hand, when the test options are set to either 'Use training set' or 'Percentage split' (with the default random seed), there are incorrectly classified instances. For example, with a 60 percentage split, it shows the following
=== Evaluation on test split ===
=== Summary ===
Correctly Classified Instances 2 40 %
Incorrectly Classified Instances 3 60 %
There seems to be a contradiction here but I must be missing something. Is the tree shown initially not the tree that is built with the 60 percentage split?
That is not stated anywhere as far as I have seen but I can't think of any other explanation.
Just for completeness, the data is here:
outlook,temperature,humidity,windy,play
sunny,hot,high,FALSE,no
sunny,hot,high,TRUE,no
overcast,hot,high,FALSE,yes
rainy,mild,high,FALSE,yes
rainy,cool,normal,FALSE,yes
rainy,cool,normal,TRUE,no
overcast,cool,normal,TRUE,yes
sunny,mild,high,FALSE,no
sunny,cool,normal,FALSE,yes
rainy,mild,normal,FALSE,yes
sunny,mild,normal,TRUE,yes
overcast,mild,high,TRUE,yes
overcast,hot,normal,FALSE,yes
rainy,mild,high,TRUE,no
If you take a closer look at the output, you will see the following:
=== Classifier model (full training set) ===
The model that is being depicted there is the model that was trained on the full dataset, not your split.
The next section has the following heading:
=== Evaluation on test split ===
The statistics that you are referring to are based on a model trained and evaluated on your dataset split.
Related
I am working with a educational dataset called IPEDS from the National Center for Educational Statistics. They track students in college based upon major, degree completion, etc. The problem in Stata is that I am trying to determine the total count for degrees obtained by a specific major.
They have a variable cipcode which contains values that serve as "majors". cipcode might be 14.2501 "petroleum engineering, 16.0102 "Linguistics" and so forth.
When I write a particular code like
tab cipcode if cipcode==14.2501
it reports no observations. What code will give me the totals?
/*Convert Float Variable to String Variable and use Force Replace*/
tostring cipcode, gen(cipcode_str) format(%6.4f) force
replace cipcode_str = reverse(substr(reverse(cipcode_str), indexnot(reverse(cipcode_str), "0"), .))
replace cipcode_str = reverse(substr(reverse(cipcode_str), indexnot(reverse(cipcode_str), "."), .))
/* Created a total variable called total_t1 for total count of all stem majors listed in table 1*/
gen total_t1 = cipcode_str== "14.2501" + "14.3901" + "15.0999" + "40.0601"
This minimal example confirms your problem. (See, by the way, https://stackoverflow.com/help/mcve for advice on good examples.)
* code
clear
input code
14.2501
14.2501
14.2501
end
tab code if code == 14.2501
tab code if code == float(14.2501)
* results
. tab code if code == 14.2501
no observations
. tab code if code == float(14.2501)
code | Freq. Percent Cum.
------------+-----------------------------------
14.2501 | 3 100.00 100.00
------------+-----------------------------------
Total | 3 100.00
The keyword is one you use, precision. In Stata, search precision for resources, starting with blog posts by William Gould. A decimal like 14.2501 is hard (impossible) to hold exactly in binary and the details of holding a variable as type float can bite.
It's hard to see what you're doing with your last block of code, which you don't explain. The last statement looks puzzling, as you're adding strings. Consider what happens with
. gen whatever = "14.2501" + "14.3901" + "15.0999" + "40.0601"
. di whatever[1]
14.250114.390115.099940.0601
The result is a long string that cannot be a valid cipcode. I suspect that you are reaching towards
... if inlist(cipcode_str, "14.2501", "14.3901", "15.0999", "40.0601")
which is quite different.
But using float() is the minimal trick for this problem.
Following my earlier question, I have tried to work on a code to return a string if a search term in a certain list is in a string to be returned as follows.
import re
from nltk import tokenize
from nltk.tokenize import sent_tokenize
def foo():
List1 = ['risk','cancer','ocp','hormone','OCP',]
txt = "Risk factors for breast cancer have been well characterized. Breast cancer is 100 times more frequent in women than in men.\
Factors associated with an increased exposure to estrogen have also been elucidated including early menarche, late menopause, later age\
at first pregnancy, or nulliparity. The use of hormone replacement therapy has been confirmed as a risk factor, although mostly limited to \
the combined use of estrogen and progesterone, as demonstrated in the WHI (2). Analysis showed that the risk of breast cancer among women using \
estrogen and progesterone was increased by 24% compared to placebo. A separate arm of the WHI randomized women with a prior hysterectomy to \
conjugated equine estrogen (CEE) versus placebo, and in that study, the use of CEE was not associated with an increased risk of breast cancer (3).\
Unlike hormone replacement therapy, there is no evidence that oral contraceptive (OCP) use increases risk. A large population-based case-control study \
examining the risk of breast cancer among women who previously used or were currently using OCPs included over 9,000 women aged 35 to 64 \
(half of whom had breast cancer) (4). The reported relative risk was 1.0 (95% CI, 0.8 to 1.3) among women currently using OCPs and 0.9 \
(95% CI, 0.8 to 1.0) among prior users. In addition, neither race nor family history was associated with a greater risk of breast cancer among OCP users."
words = txt
corpus = " ".join(words).lower()
sentences1 = sent_tokenize(corpus)
a = [" ".join([sentences1[i-1],j]) for i,j in enumerate(sentences1) if [item in List1] in word_tokenize(j)]
for i in a:
print i,'\n','\n'
foo()
The problem is that the python IDLE does not print anything. What could I have done wrong. What it does is run the code and I get this
>
>
Your question isn't very clear to me so please correct me if i'm getting this wrongly. Are you trying to match the list of keywords (in list1) against the text (in txt)? That is,
For each keyword in list1
Do a match against every sentences in txt.
Print the sentence if they matches?
Instead of writing a complicated regular expression to solve your problem I have broken it down into 2 parts.
First I break the whole lot of text into a list of sentences. Then write simple regular expression to go through every sentences. Trouble with this approach is that it is not very efficient but hey it solves your problem.
Hope this small chunk of code can help guide you to the real solution.
def foo():
List1 = ['risk','cancer','ocp','hormone','OCP',]
txt = "blah blah blah - truncated"
words = txt
matches = []
sentences = re.split(r'\.', txt)
keyword = List1[0]
pattern = keyword
re.compile(pattern)
for sentence in sentences:
if re.search(pattern, sentence):
matches.append(sentence)
print("Sentence matching the word (" + keyword + "):")
for match in matches:
print (match)
--------- Generate random number -----
from random import randint
List1 = ['risk','cancer','ocp','hormone','OCP',]
print(randint(0, len(List1) - 1)) # gives u random index - use index to access List1
I am trying to calculate the perplexity for the data I have. The code I am using is:
import sys
sys.path.append("/usr/local/anaconda/lib/python2.7/site-packages/nltk")
from nltk.corpus import brown
from nltk.model import NgramModel
from nltk.probability import LidstoneProbDist, WittenBellProbDist
estimator = lambda fdist, bins: LidstoneProbDist(fdist, 0.2)
lm = NgramModel(3, brown.words(categories='news'), True, False, estimator)
print lm
But I am receiving the error,
File "/usr/local/anaconda/lib/python2.7/site-packages/nltk/model/ngram.py", line 107, in __init__
cfd[context][token] += 1
TypeError: 'int' object has no attribute '__getitem__'
I have already performed Latent Dirichlet Allocation for the data I have and I have generated the unigrams and their respective probabilities (they are normalized as the sum of total probabilities of the data is 1).
My unigrams and their probability looks like:
Negroponte 1.22948976891e-05
Andreas 7.11290670484e-07
Rheinberg 7.08255885794e-07
Joji 4.48481435106e-07
Helguson 1.89936727391e-07
CAPTION_spot 2.37395965468e-06
Mortimer 1.48540253778e-07
yellow 1.26582575863e-05
Sugar 1.49563800878e-06
four 0.000207196011781
This is just a fragment of the unigrams file I have. The same format is followed for about 1000s of lines. The total probabilities (second column) summed gives 1.
I am a budding programmer. This ngram.py belongs to the nltk package and I am confused as to how to rectify this. The sample code I have here is from the nltk documentation and I don't know what to do now. Please help on what I can do. Thanks in advance!
Perplexity is the inverse probability of the test set, normalized by the number of words. In the case of unigrams:
Now you say you have already constructed the unigram model, meaning, for each word you have the relevant probability. Then you only need to apply the formula. I assume you have a big dictionary unigram[word] that would provide the probability of each word in the corpus. You also need to have a test set. If your unigram model is not in the form of a dictionary, tell me what data structure you have used, so I could adapt it to my solution accordingly.
perplexity = 1
N = 0
for word in testset:
if word in unigram:
N += 1
perplexity = perplexity * (1/unigram[word])
perplexity = pow(perplexity, 1/float(N))
UPDATE:
As you asked for a complete working example, here's a very simple one.
Suppose this is our corpus:
corpus ="""
Monty Python (sometimes known as The Pythons) were a British surreal comedy group who created the sketch comedy show Monty Python's Flying Circus,
that first aired on the BBC on October 5, 1969. Forty-five episodes were made over four series. The Python phenomenon developed from the television series
into something larger in scope and impact, spawning touring stage shows, films, numerous albums, several books, and a stage musical.
The group's influence on comedy has been compared to The Beatles' influence on music."""
Here's how we construct the unigram model first:
import collections, nltk
# we first tokenize the text corpus
tokens = nltk.word_tokenize(corpus)
#here you construct the unigram language model
def unigram(tokens):
model = collections.defaultdict(lambda: 0.01)
for f in tokens:
try:
model[f] += 1
except KeyError:
model [f] = 1
continue
N = float(sum(model.values()))
for word in model:
model[word] = model[word]/N
return model
Our model here is smoothed. For words outside the scope of its knowledge, it assigns a low probability of 0.01. I already told you how to compute perplexity:
#computes perplexity of the unigram model on a testset
def perplexity(testset, model):
testset = testset.split()
perplexity = 1
N = 0
for word in testset:
N += 1
perplexity = perplexity * (1/model[word])
perplexity = pow(perplexity, 1/float(N))
return perplexity
Now we can test this on two different test sets:
testset1 = "Monty"
testset2 = "abracadabra gobbledygook rubbish"
model = unigram(tokens)
print perplexity(testset1, model)
print perplexity(testset2, model)
for which you get the following result:
>>>
49.09452736318415
99.99999999999997
Note that when dealing with perplexity, we try to reduce it. A language model that has less perplexity with regards to a certain test set is more desirable than one with a bigger perplexity. In the first test set, the word Monty was included in the unigram model, so the respective number for perplexity was also smaller.
Thanks for the code snippet! Shouldn't:
for word in model:
model[word] = model[word]/float(sum(model.values()))
be rather:
v = float(sum(model.values()))
for word in model:
model[word] = model[word]/v
Oh ... I see was already answered ...
I am using Weka for my internship but I have a little knowledge about data mining. So, maybe someone knows how can I apply the following results on my data-sets to get all data by cluster ? The method that I use now is to compute distances between my attributes and the mean value of each cluster then I classify them by the nearest value. But this method is too rough for me .
=== Run information ===
Scheme:weka.clusterers.EM -I 100 -N -1 -M 1.0E-6 -S 100
Relation: wcet_cluster6 - Copie-weka.filters.unsupervised.attribute.Remove-R1-3,5-weka.filters.unsupervised.attribute.Remove-R5-12
Instances: 467
Attributes: 4
max
alt
stmt
bb
Test mode:evaluate on training data
=== Model and evaluation on training set ===
EM
Number of clusters selected by cross validation: 6
Cluster
Attribute 0 1 2 3 4 5
(0.28) (0.11) (0.25) (0.16) (0.04) (0.17)
==================================================================
max
mean 9.0148 10.9112 11.2826 10.4329 11.2039 10.0546
std. dev. 1.8418 2.7775 3.0263 2.5743 2.2014 2.4614
alt
mean 0.0003 19.6467 0.4867 2.4565 44.191 8.0635
std. dev. 0.0175 5.7685 0.5034 1.3647 10.4761 3.3021
stmt
mean 0.7295 77.0348 3.2439 12.3971 140.9367 33.9686
std. dev. 1.0174 21.5897 2.3642 5.1584 34.8366 11.5868
bb
mean 0.4362 53.9947 1.4895 7.2547 114.7113 22.2687
std. dev. 0.5153 13.1614 0.9276 3.5122 28.0919 7.6968
Time taken to build model (full training data) : 4.24 seconds
=== Model and evaluation on training set ===
Clustered Instances
0 163 ( 35%)
1 50 ( 11%)
2 85 ( 18%)
3 73 ( 16%)
4 18 ( 4%)
5 78 ( 17%)
Log likelihood: -9.09081
Thanks for your help!!
I think no-one can really answer this. Some tips off the top of my head.
You have used the EM clustering algorithm, see animated gif on wikipedia page. From Weka's Documentation Synopsis:
"EM assigns a probability distribution to each instance which
indicates the probability of it belonging to each of the clusters. "
Is this complex output really what you want?
It also selects a number of clusters for you (unless you constrain that number).
In weka 3.7 you can use the unsupervised attribute filter "ClusterMembership" in the Preprocess dialog to replace your dataset with a result of the cluster assignments. You need to select one reference attribute, though. By default it selects the last one. This creates hard-to -interpret output.
I need to create a data base with Amazon, commission junction & link share API's & data feeds and then match the same products to create comparisons on product information.
My problem is related to the matching process.
I start by matching products via SKU/UPC/ASIN but this not perform well because many of the products doesn't contain this information.
I maked some research and the most popular techniques I found are :
-Measuring cosine similarity via TF-IDF
-Measuring edit distance/ levenshtein / Jaro-Winkler
In this technique i used cosine similarity and Jaro-Winkler
How I do the matching :
Step 1 : Preprocessing
Preprocessing to transform strings into a normal form :
Lowercase
Filter stop words (new, by, the …)
Strip whitespace
replace all whitespace occurrences with a single space character
Step 2, Indexing :
Index Amazon products in a Solr core [core A] and CJ/Linkshare [core B] in an other core. The goal of indexing is to limit the number of string comparisons (via TF-IDF and Jaro-Winkler)
Step 3, matching :
I start by retrieving a product title from core B, make a solr search in core A with this title and take the top 30 results.
I measure similarity via TF-IDF between the product i want to match (the query) and the 30 results retrieved by solr search. I keep the products with similarity > 80%
sort the tokens from each product alphabetically.I then compare the transformed strings with Jaro Winkler distance and keep the products with similarity > 80% (==> This perform a Jaro Winkler similarity between phrases)
Here, I tokenize both strings (query and product to match) , and perform a comparison between tokens.
But this techniques also don't perform well. Example :
Product 1 : Orange by Hugo Boss, 3 Ounce Eau de toilette Spray
Product 2 : In Motion Orange By Hugo Boss Eau De Toilette Spray 3 Ounces
Product 1 and 2 are similar via this techniques but actually they are different.
How can I improve this algorithm? Is that the right way to match products?
How if i train a classifier with token's weight (using Jaro Winkler) (learning data from matched products via UPC) and use this classifier to match products in a final step?
PS : I have products from different categories (health, beauty, electronics, books, movies...) and data is very unstructured or incomplete.
Any advice will be helpfull
Thanks
Smail