Alright. Now this question is pretty hard. I am going to give you an example.
Now the left numbers are my algorithm classification and the right numbers are the original class numbers
177 86
177 86
177 86
177 86
177 86
177 86
177 86
177 86
177 86
177 89
177 89
177 89
177 89
177 89
177 89
177 89
So here my algorithm merged 2 different classes into 1. As you can see it merged class 86 and 89 into one class. So what would be the error at the above example ?
Or here another example
203 7
203 7
203 7
203 7
16 7
203 7
17 7
16 7
203 7
At the above example left numbers are my algorithm classification and the right numbers are original class ids. As can be seen above it miss classified 3 products (i am classifying same commercial products). So at this example what would be the error rate? How would you calculate.
This question is pretty hard and complex. We have finished the classification but we are not able to find correct algorithm for calculating success rate :D
Here's a longish example, a real confuson matrix with 10 input classes "0" - "9"
(handwritten digits),
and 10 output clusters labelled A - J.
Confusion matrix for 5620 optdigits:
True 0 - 9 down, clusters A - J across
-----------------------------------------------------
A B C D E F G H I J
-----------------------------------------------------
0: 2 4 1 546 1
1: 71 249 11 1 6 228 5
2: 13 5 64 1 13 1 460
3: 29 2 507 20 5 9
4: 33 483 4 38 5 3 2
5: 1 1 2 58 3 480 13
6: 2 1 2 294 1 1 257
7: 1 5 1 546 6 7
8: 415 15 2 5 3 12 13 87 2
9: 46 72 2 357 35 1 47 2
----------------------------------------------------
580 383 496 1002 307 670 549 557 810 266 estimates in each cluster
y class sizes: [554 571 557 572 568 558 558 566 554 562]
kmeans cluster sizes: [ 580 383 496 1002 307 670 549 557 810 266]
For example, cluster A has 580 data points, 415 of which are "8"s;
cluster B has 383 data points, 249 of which are "1"s; and so on.
The problem is that the output classes are scrambled, permuted;
they correspond in this order, with counts:
A B C D E F G H I J
8 1 4 3 6 7 0 5 2 6
415 249 483 507 294 546 546 480 460 257
One could say that the "success rate" is
75 % = (415 + 249 + 483 + 507 + 294 + 546 + 546 + 480 + 460 + 257) / 5620
but this throws away useful information —
here, that E and J both say "6", and no cluster says "9".
So, add up the biggest numbers in each column of the confusion matrix
and divide by the total.
But, how to count overlapping / missing clusters,
like the 2 "6"s, no "9"s here ?
I don't know of a commonly agreed-upon way
(doubt that the Hungarian algorithm
is used in practice).
Bottom line: don't throw away information; look at the whole confusion matrix.
NB such a "success rate" will be optimistic for new data !
It's customary to split the data into say 2/3 "training set" and 1/3 "test set",
train e.g. k-means on the 2/3 alone,
then measure confusion / success rate on the test set — generally worse than on the training set alone.
Much more can be said; see e.g.
Cross-validation.
You have to define the error criteria if you want to evaluate the performance of an algorithm, so I'm not sure exactly what you're asking. In some clustering and machine learning algorithms you define the error metric and it minimizes it.
Take a look at this
https://en.wikipedia.org/wiki/Confusion_matrix
to get some ideas
You have to define a error metric to measure yourself. In your case, a simple method should be to find the properties mapping of your product as
p = properties(id)
where id is the product id, and p is likely be a vector with each entry of different properties. Then you can define the error function e (or distance) between two products as
e = d(p1, p2)
Sure, each properties must be evaluated to a number in this function. Then this error function can be used in the classification algorithm and learning.
In your second example, it seems that you treat the pair (203 7) as successful classification, so I think you have already a metric yourself. You may be more specific to get better answer.
Classification Error Rate(CER) is 1 - Purity (http://nlp.stanford.edu/IR-book/html/htmledition/evaluation-of-clustering-1.html)
ClusterPurity <- function(clusters, classes) {
sum(apply(table(classes, clusters), 2, max)) / length(clusters)
}
Code of #john-colby
Or
CER <- function(clusters, classes) {
1- sum(apply(table(classes, clusters), 2, max)) / length(clusters)
}
Related
I have a 2d grid with some of the tiles being obstacles (walls), I want to be able to find the shortest path that allows you to go around the grid being able to see all of the other grids in the map with a radius of view. Here is a pixel art example (Blacks are the obstacles, gray is an arbitrary path).
Set R to be the "radius of view"
Create orthogonal point grid with separation 2 * R
Remove grid points that collide with obstacles
Add connections from each point to the 9 or less closest points. Do not connect across obstacles
Calculate minimum spanning tree of remaining points.
Specify an "obstacle course" with a 20 by 20 gris, a view radius of 2 and three groups of abstacles
20 20 2
4 4
5 5
6 6
14 9
14 10
14 11
14 12
14 13
14 14
14 15
14 16
10 16
11 16
12 16
13 16
4 14
4 15
5 15
It looks like this
Adjacency list ( link id, node1 id, node2 id, distance )
l 42 47 5
l 42 142 5
l 47 52 5
l 47 152 7.07107
l 52 57 5
l 52 147 7.07107
l 52 152 5
l 52 157 7.07107
l 57 152 7.07107
l 57 157 5
l 142 242 5
l 142 247 7.07107
l 147 152 5
l 147 242 7.07107
l 147 247 5
l 147 252 7.07107
l 152 247 7.07107
l 152 252 5
l 157 257 5
l 242 247 5
l 242 342 5
l 247 252 5
l 247 347 5
l 257 357 5
l 342 347 5
l 347 352 5
l 352 357 5
Pass this graph into your favorite graph theory library to get the minimum spanning tree.
Here is the result when I use the PathFinder application
The spanning tree is useful if, for example, the robot needs to return to a base frequently to recharge. However, if your robot does not need to do this, then it involves a lot of unnecessary backtracking.
Calculation of a tour of the nodes that visits all of them with minimal backtracking can be done in various ways ( see travelling salesman problem ). I use depth first searching of the spanning tree plus using the Dijsktra algorithm to find the nearest unvisited node when the robot gets trapped at a leaf of the spanning tree. This quickly gives a reasonably effective result.
Application coded in C++, available at https://github.com/JamesBremner/obstacles
I downloaded Stanford NLP 3.5.2 and run sentiment analysis with default configuration (i.e. I did not change anything, just unzip and run).
java -cp "*" edu.stanford.nlp.sentiment.Evaluate -model edu/stanford/nlp/models/sentiment/sentiment.ser.gz -treebank test.txt
EVALUATION SUMMARY
Tested 82600 labels
66258 correct
16342 incorrect
0.802155 accuracy
Tested 2210 roots
976 correct
1234 incorrect
0.441629 accuracy
Label confusion matrix
Guess/Gold 0 1 2 3 4 Marg. (Guess)
0 323 161 27 3 3 517
1 1294 5498 2245 652 148 9837
2 292 2993 51972 2868 282 58407
3 99 602 2283 7247 2140 12371
4 0 1 21 228 1218 1468
Marg. (Gold) 2008 9255 56548 10998 3791
0 prec=0.62476, recall=0.16086, spec=0.99759, f1=0.25584
1 prec=0.55891, recall=0.59406, spec=0.94084, f1=0.57595
2 prec=0.88982, recall=0.91908, spec=0.75299, f1=0.90421
3 prec=0.58581, recall=0.65894, spec=0.92844, f1=0.62022
4 prec=0.8297, recall=0.32129, spec=0.99683, f1=0.46321
Root label confusion matrix
Guess/Gold 0 1 2 3 4 Marg. (Guess)
0 44 39 9 0 0 92
1 193 451 190 131 36 1001
2 23 62 82 30 8 205
3 19 81 101 299 255 755
4 0 0 7 50 100 157
Marg. (Gold) 279 633 389 510 399
0 prec=0.47826, recall=0.15771, spec=0.97514, f1=0.2372
1 prec=0.45055, recall=0.71248, spec=0.65124, f1=0.55202
2 prec=0.4, recall=0.2108, spec=0.93245, f1=0.27609
3 prec=0.39603, recall=0.58627, spec=0.73176, f1=0.47273
4 prec=0.63694, recall=0.25063, spec=0.96853, f1=0.35971
Approximate Negative label accuracy: 0.646009
Approximate Positive label accuracy: 0.732504
Combined approximate label accuracy: 0.695110
Approximate Negative root label accuracy: 0.797149
Approximate Positive root label accuracy: 0.774477
Combined approximate root label accuracy: 0.785832
The test.txt file is downloaded from http://nlp.stanford.edu/sentiment/trainDevTestTrees_PTB.zip (contains train.txt, dev.txt and test.txt). The download link is get from http://nlp.stanford.edu/sentiment/code.html
However, in the paper "Socher, R., Perelygin, A., Wu, J.Y., Chuang, J., Manning, C.D., Ng, A.Y. and Potts, C., 2013, October. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the conference on empirical methods in natural language processing (EMNLP) (Vol. 1631, p. 1642)." which sentiment analysis tool is based on, the authors reported that the accuracy when classify 5 classes is 0.807.
Is my results I obtained normal?
I get the same results when I run it out of the box. It would not surprise me if the version of their system they made for Stanford CoreNLP differs slightly from the version in the paper.
According to Data Structures Using C by Tenenbaum, one of the improvements of bubble sort is to have successive passes go in opposite direction so that the small elements move quickly to the front which will reduce the required number of passes [pg 336].
I worked out two examples, one which supports this statement and other which is against this one.
Supports: 25 48 37 12 57 86 33 92
iterations using usual Bubble sort :
25 48 37 12 57 86 33 92
25 37 12 48 57 33 86 92
25 12 37 48 33 57 86 92
12 25 37 33 48 57 86 92
12 25 33 37 48 57 86 92
iterations using improvement:
25 48 37 12 57 86 33 92
25 37 12 48 57 33 86 92
12 25 37 33 48 57 86 92
12 25 33 37 48 57 86 92
against: 3 4 1 2 5
iterations using usual Bubble sort:
3 4 1 2 5
3 1 2 4 5
1 2 3 4 5
iterations using improvement:
3 4 1 2 5
3 1 2 4 5
1 3 2 4 5
1 2 3 4 5
So is the statement incorrect that this improvement will always help? Or I am doing something wrong here ?
The example you gave above shows that this algorithm isn't a strict improvement over a standard bubble sort.
The advantage of this approach (sometimes called "cocktail sort," by the way) is that in cases where there are a lot of small elements at the end of the array, it rapidly pulls them to the front compared against normal bubble sort. For example, consider this array:
2 3 4 5 6 7 8 9 10 11 12 ... 10,000,000 1
With a normal bubble sort, it would take 9,999,999 passes over this array to sort it because the element 1, which is way out of place, only gets swapped one step forward on each iteration. On the other hand, with a cocktail sort, this would take just two passes - one initial pass and then a reverse pass.
While the above example is definitely contrived, in a randomly-shuffled array, there are likely going to be some smaller elements toward the end of the array and the number of passes of bubblesort is going to have to be large to move them back. Going in both directions helps speed this up.
That said, bubblesort is a pretty poor choice of a sorting algorithm, so hopefully this is just a theoretical discussion. :-)
Suppose we have a 2D (5x5) matrix:
test =
39 13 90 5 71
60 78 38 4 11
87 92 46 45 35
40 96 61 17 1
90 50 46 89 63
And a second 2D (5x2) matrix:
tidx =
1 3
2 4
2 3
2 4
4 5
And now we want to use tidx as an idex into test, so that we get the following output:
out =
39 90
78 4
92 46
96 17
89 63
One way to do this is with a for loop...
for i=1:size(test,1)
out(i,:) = test(i,tidx(i,:));
end
Question:
Is there a way to vectorize this so the same output is generated without a for loop?
Here is one way:
test(repmat([1:rows(test)]',1,columns(tidx)) + (tidx-1)*rows(test))
What you describe is an index problem. When you place a matrix all in one dimension, you get
test(:) =
39
60
87
40
90
13
78
92
96
50
90
38
46
61
46
5
4
45
17
89
71
11
35
1
63
This can be indexed using a single number. Here is how you figure out how to transform tidx into the correct format.
First, I use the above reference to figure out the index numbers which are:
outinx =
1 11
7 17
8 13
9 19
20 25
Then I start trying to figure out the pattern. This calculation gives a clue:
(tidx-1)*rows(test) =
0 10
5 15
5 10
5 15
15 20
This will move the index count to the correct column of test. Now I just need the correct row.
outinx-(tidx-1)*rows(test) =
1 1
2 2
3 3
4 4
5 5
This pattern is created by the for loop. I created that matrix with:
[1:rows(test)]' * ones(1,columns(tidx))
*EDIT: This does the same thing with a built in function.
repmat([1:rows(test)]',1,columns(tidx))
I then add the 2 together and use them as the index for test.
Note: I am still looking for a fast solution. Two of the solutions below are wrong and the third one is terribly slow.
I have N toys from 1....N. Each toy has an associated cost with it. You have to go on a shopping spree such that on a particular day, if you buy toy i, then the next toy you can buy on the same day should be i+1 or greater. Moreover, the absolute cost difference between any two consecutively bought toys should be greater than or equal to k. What is the minimum number of days can I buy all the toys.
I tried a greedy approach by starting with toy 1 first and then seeing how many toys can I buy on day 1. Then, I find the smallest i that I have not bought and start again from there.
Example:
Toys : 1 2 3 4
Cost : 5 4 10 15
let k be 5
On day 1, buy 1,3, and 4
on day 2, buy toy 2
Thus, I can buy all toys in 2 days
Note greedy not work for below example: N = 151 and k = 42
the costs of the toys 1...N in that order are :
383 453 942 43 27 308 252 721 926 116 607 200 195 898 568 426 185 604 739 476 354 533 515 244 484 38 734 706 608 136 99 991 589 392 33 615 700 636 687 625 104 293 176 298 542 743 75 726 698 813 201 403 345 715 646 180 105 732 237 712 867 335 54 455 727 439 421 778 426 107 402 529 751 929 178 292 24 253 369 721 65 570 124 762 636 121 941 92 852 178 156 719 864 209 525 942 999 298 719 425 756 472 953 507 401 131 150 424 383 519 496 799 440 971 560 427 92 853 519 295 382 674 365 245 234 890 187 233 539 257 9 294 729 313 152 481 443 302 256 177 820 751 328 611 722 887 37 165 739 555 811
You can find the optimal solution by solving the asymmetric Travelling Salesman.
Consider each toy as a node, and build the complete directed graph (that is, add an edge between each pair of nodes). The edge has cost 1 (has to continue on next day) if the index is smaller or the cost of the target node is less than 5 plus the cost of the source node, and 0 otherwise. Now find the shortest path covering this graph without visiting a node twice - i.e., solve the Travelling Salesman.
This idea is not very fast (it is in NP), but should quickly give you a reference implementation.
This is not as difficult as ATSP. All you need to do is look for increasing subsequences.
Being a mathematician, the way I would solve the problem is to apply RSK to get a pair of Young tableaux, then the answer for how many days is the height of the tableau and the rows of the second tableau tell you what to purchase on which day.
The idea is to do Schensted insertion on the cost sequence c. For the example you gave, c = (5, 4, 10, 15), the insertion goes like this:
Step 1: Insert c[1] = 5
P = 5
Step 2: Insert c[2] = 4
5
P = 4
Step 3: Insert c[3] = 10
5
P = 4 10
Step 4: Insert c[4] = 15
5
P = 4 10 15
The idea is that you insert the entries of c into P one at a time. When inserting c[i] into row j:
if c[i] is bigger than the largest element in the row, add it to the end of the row;
otherwise, find the leftmost entry in row j that is larger than c[i], call it k, and replace k with c[i] then insert k into row j+1.
P is an array where the lengths of the rows are weakly decreasing and The entries in each of row P (these are the costs) weakly increase. The number of rows is the number of days it will take.
For a more elaborate example (made by generating 9 random numbers)
1 2 3 4 5 6 7 8 9
c = [ 5 4 16 7 11 4 13 6 5]
16
7
5 6 11
P = 4 4 5 13
So the best possible solution takes 4 days, buying 4 items on day 1, 3 on day 2, 1 on day 3, and 1 on day 4.
To handle the additional constraint that consecutive costs must increase by at least k involves redefining the (partial) order on costs. Say that c[i] <k< c[j] if and only if c[j]-c[i] >= k in the usual ordering on numbers. The above algorithm works for partial orders as well as total orders.
I somewhat feel that a greedy approach would give a fairly good result.
I think your approach is not optimal just because you always pick toy 1 to start while you should really pick the least expensive toy. Doing so would give you the most room to move to the next toy.
Each move being the least expensive one, it is just DFS problem where you always follow the least expensive path constrained by k.