# Tag Info

0

The ROC-AUC curves are used to find the best threshold that optimizes True Positive Rate vs False Positive Rate. Using it in a K-Fold cross-validation is a good practice to determine the best threshold to use. Then, your final test is here to validate that you did not overfit on some hyperparameters, including this threshold. So ROC-AUC must not be used ...

0

Transformers. Most NMT systems since 2018 use the Transformer architecture. All relevant NMT frameworks support Transformers, including: Marian Fairseq OpenNMT

0

Although there are several nearest neighbor tool as one mention by @oneextrafact. The problem with that tool is you need to index the external database for a logic operation like you mentioned that you want to build some logic using over date. I will recommend you is to extract a document vector. Although there are several approaches to do that either by ...

1

The binary cross-entropy loss function is based on the assumption that there is only one output node and it can have a value between 0-1. If you have more than two outputs and using the softmax activation function, you should use a categorical cross-entropy loss function in order to handle the multiclass situation. But in your scenario, there is a binary ...

0

It sounds likes you have non-stationary data which can be a challenge to model. One option is to completely discard all older data and only train/test on newer data. This way the model will only capture the newer relationships. This approach assumes there is enough newer data to train a model. Another option is heavily regularize the model. The goal of ...

0

This is an active area of research called Human Activity Recognition. There are several public datasets available to cross-validate your methods, and you might want to start here: UCI HAR Dataset. There's a paper that accompanies the dataset that describes their preprocessing methods, so you'll want to have a look at that and see if anything helps in terms ...

0

The way you are imputing your feature can't be replicated in the test set, because it needs knowledge of the target classes! You need to select a different imputation strategy, that doesn't rely on your target feature. Assuming that you are using another feature, the same way you were using your target, you need to store the value(s) you are imputing each ...

0

Sqlova reach 99.8% in training and 90.0% in testing. This work improve Sqlova further in testing.

0

If the labeling people is sufficient, the best solution is to re-label the noisy data.

0

The scoring function is used as an objective measure of performance. The choice of scoring function itself is subjective and should reflect what you or the problem deems to be important in terms of the balance between whatever metrics you are tracking (e.g., precision & recall, or sensitivity & specificity, or BLEU & ROUGE). Arithmetic mean, ...

0

I think you can speak of imbalanced targets if (in case of a binary classification problem) the classes are not represented in a 50:50 manner. This is almost always the case. With about 25/75 in your case, I would see this as „imbalanced“. There are some strategies to deal with this problem, such as (re)sampling so that you achieve a 50:50 balanced sample (...

0

Will 64 filters be created for each channel (red, green and blue) of the image? No, Rather filters also will have 3 channels. It would be (333). Its work something like this- Borrowed from- http://datahacker.rs/convolution-rgb-image/ More here- https://cs231n.github.io/convolutional-networks/.

0

It is not the number of features that is the problem with Gaussian Naive Bayes (GaussianNB). It is the decision boundary that GaussianNB is learning. Naive Bayes is constrained to the learn the marginal distribution of the data because "naive" assumption. Often times the conditional distribution is useful to make predictions. Given the performance ...

0

Oh, well, let's say it depends on your task and model. For instance, in autoencoders, there are two main choices for upsampling. You can employ transposed convolution or rescaling. The better choice is the latter case due to the fact that the former can lead to checker-board artifacts. About zero padding, it is usually done in almost all CNNs if you use the ...

0

It depends on application. For upsampling, Keras documentation says: Repeats the rows and columns of the data by size[0] and size[1] respectively. For zero padding, Keras documentation says: This layer can add rows and columns of zeros at the top, bottom, left and right side of an image tensor. So, imagine You have picture of sharp edge shapes, and You ...

1

The task you describe corresponds exactly to Named Entity Recognition (NER). This is a standard task for which there are many available libraries. NER is usually done with sequence labelling models such as Conditional Random Fields. Sequence labelling implies that the data is provided as a sequence which looks like this: During <features ...> O the ...

0

I think these are often used colloquially as synonyms, but let's try to find the differences. Each of them begins with "Time Series" (TS). So the difference lies in the three following terms. here with my interpretation: Analysis - wanting to describe and understand characteristics the observed data coming from the generating function$^1$. ...

0

Have a look at the Common Voice project from Mozilla. They have an open source dataset of multiple languages, Spanish currently has 324 hours of validated speech data (out of 579 total). You can download the dataset for Spanish and other languages at https://commonvoice.mozilla.org/en/datasets.

0

To answer the question so it is not presented as unanswered: How does this relate to our neural network? A cost function $J(W,b)$ has also a number of local maxima and minima, as you can see in this picture, for example. The fact that $J$ has multiple minima can also be interpreted in a nice way. In each layer, you use multiple nodes which are assigned ...

2

As far as I know, the dataset has not an official or a common name. In their text book "Data Mining" Witten et al. refer to it as "The Weather Problem". The example is based in on a dataset described in "Induction of Decision Trees" by Quinlan in 1986. Accordingly, it has been used in many books as an example to illustrate ...

0

This question is similar to this one and this one, but seems to be ill-posed. Either because it implies an unknown (undefined) way how the neurons process the inputs, or because the provided solution (5, 3) is wrong. Concretely, usual neurons only sum their inputs and the bias and pass the sum through a step function. These are the typical neurons used in ...

1

This specific example is trying to predict a variable which can have two values, either yes or no. In data science terms this would be a classification problem. Part a asks to use a naive bayes classifier for this data, which can easily be done using the scikit-learn library.

1

What you are trying to do here is named-entity recognition. Namely, the task consists of classifying substrings into a set of named entities (i.e. person, location, etc.). From a more formal perspective, this is a sequence labeling task that classifies parts of a sequence. This task can be approached in different ways: gazetteers/string matching regular ...

1

To begin with, I don't quite get it why you want to flip them. In the binary case, you flip Negatives and Positive, so True Negative becomes True Positive and so do FP/FN. Hence you flip specificity/true negative and sensitivity/recall values, so overall accuracy and F1 stay the same.

1

I think that the issue depends on what you'd expect the model to learn: If the model is supposed to "know" the users it has seen during training, i.e. exploit the user id in order to infer particular choices for a specific user, then I don't see the point in adding this kind of frequency feature: the model already "knows" what choices ...

1

I primarily agree with @NikosM., but take issue with the question's assertion (my emphasis) without losing any performance Having very high correlation means they are very nearly in a linear relationship, but it is possible that the deviation from linearity is not random, and in fact is predictive. Here's a simple example: import numpy as np import ...

2

This post seems similar to yours and may help. It seems that what you are looking for is a derivative-free optimization method. The Wikipedia page for the concept lists such methods. Intuitively, these techniques will sample the function (the network in your case) with various inputs (pressure, temperature, speed) and will figure out which inputs optimize it....

1

In a pure random search, 60 points is often given as a rule of thumb, because provably with probability 95% such a search finds a hyperparameter combination in the top 5%. However, that 5% is as a percent of the volume of the search space, so giving much-too-broad a search space, the best 5% might not be a fantastic score for the model. So it does seem to ...

0

The other answers make sense but I would be more categorically negative about the idea: Is this approach a correct approach, or logical with respect to machine learning principles ? No, it's not. The parameters of a ML model (whether supervised or unsupervised) are estimated using a particular set of features designed as the input for the problem. Changing ...

3

Is this approach a correct approach, or logical with respect to machine learning principles? It will affect the performance of the model in the sense that your algorithm learned to separate the clusters based upon distance according to all the features. I have read discussions about how to calculate feature importance on unsupervised problems like yours, so ...

1

Interesting question. The answer is: It depends. The best way to find out how it would affect your model is with the shap package. You can use it to uncover the importance of features and reveal interaction effects in the model. There could be a very different effect depending on how „important“ the excluded features are. Let‘s assume a very simple decision ...

0

General options: Group your features into groups by generating new labels that represent multiple features Use dimensionality reduction, e.g. PCA, Autoencoder, etc. A lot of these are implemented in Sklearn and the downside is that your features become confusing to analyze once it converts them to pure mathematically representations that can have a ...

0

By using .values in your definition of X, you've converted to a numpy array and lost the column names. Just removing that, you'll provide a frame to SRS, and mlxtend will use the column names in the k_feature_names_ attribute; so that's probably the best approach. There are two other approaches: one is to add the custom_feature_names parameter of mlxtend....

0

Based on this scikit-learn documentation, you can get a boolean mask (in the same order) of the input features, via the get_support method:

1

Machine Learning is not something that can be mastered or learnt in a short time, and you need at least 3-4 months familiarize yourself with the basics and even after that you need at least 6-7 months to get to a good place with your ML knowledge. To get started you can first go through this course (Python for Everybody - Full University Python Course) to ...

1

1

News sentences will have more unique tokens than normal conversations. Conversations have more stop words than news articles. I think you can use bert or normal wordvect classification to train a baseline model here. I would play aroud the pipeline of fake news classifier and news-conversation classifier. like passing the text to news classifier first and ...

0

You can do the above task by making your own pattern for entity extraction- https://spacy.io/usage/rule-based-matching

1

I agree that it looks like an optimization problem, but the parameters are not completely clear. Here are some vague ideas I have about the design of such a system: Let's assume a unit of time for which you have data and by which you attempt to predict things, for instance by the hour. I think the first part needs to be about predicting energy needs by unit ...

0

Intuitively speaking, ensembles benefit most from diversity. Imagine being in a room of people making a decision together. If everyone more or less agrees, you don't benefit from having more people at the table. But if people tend to have different opinions, when they DO agree, it is a stronger message that the decision must be correct. The same applies to ...

0

You can calculate a feature importance for any classifier. Take a look at lime or shap Shap unifies seven different methods (cited from the shap GitHub page): 1 LIME: Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. "Why should i trust you?: Explaining the predictions of any classifier." Proceedings of the 22nd ACM SIGKDD International ...

0

You are misusing the idea of "test" dataset in machine learning. A test dataset should only be used once. You are re-using the test dataset and changing the modeling choices based on that re-use. This is an example of data leakage which will lower the generalization performance. Additionally, the number of samples for the P group (3 and 21 ...

0

You write ... the model needs to be evaluated against each product in real-time., which gets me thinking that you use a binary classification (sigmoid in the final layer) architecture with negative sampling for the user/item interactions when training your model. Have you considered using multi-class classification instead? Thus, for the user only predict ...

0

Sadly, this isn't possible. Simply put, if you could know for sure that your model made a mistake, then, your model will never make mistakes. This is the painful reality of deploying models in practice. You will need manual validations to confirm whether your model performed correctly. In some systems, you may be able to outsource this to the user itself (i....

0

I think there is a 2 step solution for what you are asking: Identifying the best performing models: paperswithcode is a website with the best performing models for different tasks and different datasets. For any specific task (e.g. sentiment analysis), you can go to the appropriate section in the site, and try to identify a benchmark dataset that is most ...

0

For matrix factorization I usually see it being initialized by a uniform distribution from [0, 1) like in this (pytorch) or a truncated normal with mean=0.0 and std=1.0 as in this (tensorflow).

0

One option is incremental training. Update the model weights as more recent data is available. This automatically assigns more importance to recent records. Incremental training works well within the Bayesian framework where historical data is the prior and more recent data updates the prior.

1

If you want to apply the result of fit_transform, you must assign to your columns. columns = ['S_LENGTH', 'S_WIDTH', 'P_LENGTH', 'P_WIDTH'] min_max = preprocessing.MinMaxScaler() df[columns] = min_max.fit_transform(df[columns]) df.head() Output ID S_LENGTH S_WIDTH P_LENGTH P_WIDTH SPECIES 0 1 0.0 0.0 1.0 0.0 VIRGINICA 1 ...

0

You can use df.select_dtypes(exclude='object') to exclude categorical columns. Also while importing dataset, set the index_col='ID' to use ID as index instead of column.

1

Actually I think that you are mostly correct, except that in my understanding "with/without replacement" applies only to selecting one subset, not across subsets. This means that if we have a training set of instances $T=\{t_1,..,t_N\}$: With bagging, a particular sample can contain duplicate instances, in other words it is not a subset of $T$ but ...

Top 50 recent answers are included