New answers tagged

0

You can use Robust Squared Mahalanobis Distance to detect outliers in Multivariate. Then run your model one time using all data values and compute the Mean square error. Run it for the second time without these outliers and compute again MSE. See the difference. If you have so many outliers, you can use first principle component (PC) to reduce the ...


0

One way to approach it: Define a center tendency of the documents, a location in vector space. Then, define a distance metric (e.g., cosine, Minkowski, or Mahalanobis). Lastly, set a threshold in the distance metric that would define an outlier.


0

It sounds like you need a second data structure like an index. An index tracks important metadata information about the data. Graphs are typically not structured for fast querying. Indexes can be structured for fast querying.


3

That would be a part of feature selection. There are many methods to find out if there are relationships between the dependent variable and independent variables. To name a few: plots, measures of correlation, measures of mutual information.


0

One of the best non-deep learning systems for MNIST is k-nearest neighbors algorithm (k-NN) with non-linear deformation (P2DHMDM).


1

Precision and recall are "hard" metrics. They are measure if the model's prediction is exactly the same as the target label. Often times systems like yours can use a more flexible metric such as top-5 error rate, the model is considered to have generated the correct response if the target label is one of the model’s top 5 predictions.


1

The default quantiles with quantile are 0%, 25%, 50%, 75%, 100%. This means that the quartiles Q1 and Q3 are the second and fourth value in the two vector you obtain. Since indexes start at 1 it means that you can obtain Q1 with ar_quantiles[2] and Q3 with ar_quantiles[4]. Currently you're using indexes 1 and 2 and that's certainly what causes the problem. [...


0

Association rules are a relatively straightforward class of algorithms. Those earlier papers cover most of the interesting properties association rules. The field of recommender systems moved towards collaborative filtering and matrix factorization around that time. Those methods have increased empirical performance and are more interesting to study.


1

There is an overall inverse relationship, but not a strictly monotone one. See e.g. the precision-recall curves in the sklearn examples. A model that declares every record [to be positive class] has high recall but low precision. If the model declares every record positive, then $TP=P$ and $FP=N$ (and $FN=TN=0$). So recall is 1; and the precision is $P/(...


0

Before explaining I just want to point out that these points are only about the advantages of NB classification, there are also disadvantages (in particular NB is very prone to overfitting). They are robust to isolated noise points because such points are averaged out when estimating conditional probabilities from data. An "isolated noise point" ...


Top 50 recent answers are included