• No se han encontrado resultados

Efecto del acondicionamiento térmico en la carotenogénesis

3. Carotenoides en frutos cítricos

3.2. Cambios en carotenoides durante la conservación postcosecha El aspecto visual de los frutos cítricos es un parámetro determinante en la

3.2.2. Efecto del acondicionamiento térmico en la carotenogénesis

The most popular machine learning techniques for sentiment analysis are Naive Bayes (NB) [17, 57, 63, 111, 129, 160], Multinomial Naive Bayes (MNB) [63, 170], Support Vector Machines (SVM) [17, 44, 63, 111] and Maximum Entropy (ME) [63, 111]. Complement Naive Bayes (CNB) [170], Sequential Minimal Optimization (SMO) [166] and Random Forest (RF) [170] are machine learning techniques that have been used in data mining research before, however, they have not been explored much in the field of sentiment analysis [166, 170]. Consequently, they may be promising in sentiment analysis for education. These techniques are reviewed in the following.

Naive Bayes (NB)

Naive Bayes uses a probabilistic model that originates from Bayes theorem. It assumes

Table 4.3 Summary of studies with features

Research Domain Feature(s)

Munezero et al. [118] Education Lexicon-based

Guha et al. [71] Twitter Unigrams

Lexicon-based Question marks Exclamation marks Emoticons

Number of hashtags

Pang and Lee [131] Movie reviews Unigrams

Barbosa et al. [12] Twitter Hashtags

All Punctuation marks (i.e. full stops, commas, question and exclama- tion marks)

Exclamation marks

Kouloumpis et al. [90] Twitter Unigrams

Lexicon-based Emoticons

e.g., [131, 136], as it only needs a small amount of training data to estimate parameters. Ad- ditionally, it can deal with discrete and continuous attributes and is fast and incremental [14]. Datasets sometimes consist of uneven classes, meaning one class has more instances than the other. Naive Bayes does not work well with uneven class sets and Complement Naive Bayes addresses this problem [65, 138].

Multinomial Naive Bayes (MNB)

Multinomial Naive Bayesis a specific case of a Naive Bayes classifier which uses a multino- mial distribution for each of the features. Multinomial Naive Bayes estimates the conditional probability of a particular feature (word) given a class as the relative frequency of term t in documents belonging to class c. Similar to Naive Bayes, it is fast, easy to implement and relatively effective [138].

Complement Naive Bayes (CNB)

Complement Naive Bayesestimates feature probabilities for each class based on the com-

plement of that class (i.e. based on the instances from all the other classes rather than the target class). The Complement Naive Bayes classifier improves the weaknesses in the Naive Bayes classifier (i.e. dealing with uneven classes). Research by Gokulakrishnan et al. [65]

about general Twitter sentiment analysis shows that Complement Naive Bayes gives a higher performance than Naive Bayes when the classes are uneven.

Support Vector Machines (SVM)

A Support Vector Machine (SVM) is a non-probabilistic classifier. Models that include just positive and negative classes are classified using a binary classifier, while models with the neutral class included are classified using a multi-class classifier. SVM finds hyperplanes that separate the classes [91]. SVM is highly effective at traditional text categorisation and generally outperforms Naive Bayes [14].

The effectiveness of the SVM partially depends on the kernel [28]. There are different types of kernels, of which the most popular are linear, polynomial and radial basis functions.

The linear kernel usually works best with larger amounts of data. The linear kernel is graphed as a straight line. In contrast, non-linear kernels are graphed as curved lines and paths, and generally give better results [28]. The polynomial and radial basis functions are two types of non-linear kernels. The polynomial kernel is popular with natural language processing [28, 66] and the radial basis kernel is popular for sentiment analysis applications [28]. For our experiments, the LibSVM model [28] in Weka [183] was used.

Maximum Entropy

The Maximum Entropy (ME) classifier is similar to the Naive Bayes classifier, except that instead of the features acting independently, the model finds weights for the features that maximize the likelihood of the training data using search-based optimization [131].

One advantage of ME is that it does not make assumptions about the relationships between features [131]. Consequently, in contrast to NB and SVM, it could potentially perform better when conditional independence assumptions are not met [14]. One drawback is that it is not very realistic in many practical problems, as real datasets contain random errors or noise, which create a less clean dataset [44]. Consequently, we expect that higher levels of preprocessing would lead to better results in the ME classifier.

Sequential minimal optimization (SMO)

Sequential minimal optimizationis an algorithm derived from SVM. It addresses the quadratic programming problem (i.e. the problem of optimizing a quadratic function of several variables subject to linear constraints on these variables) that arises during the training of support vector machines.

SMO utilizes the smallest possible quadratic programming optimization problem, which are addressed quickly and analytically, improving its scaling and computation time [133]. SMO does not require extra matrix storage and it is less susceptible to numerical precision problems [133]. Additionally, SMO performs well for linear SVMs because its computation time is controlled by SVM evaluation which can be expressed as a single dot product, rather than a sum of linear kernels [133]. Moreover, SMO performs well for SVMs with sparse inputs and for non-linear SVMs [133]. This is due to the reduced kernel computation time speeding up SMO.

Random Forest (RF)

Random Forestis based on Random Tree classifier, and uses a number of decision trees to

improve the classification rate. Random forests leave one third of the sample training set out of the trees set. When each tree is built, the data is run down the tree, and proximities are computed for each pair of cases. If two cases occupy the same terminal node, their proximity is increased by one. Eventually, the proximities are normalized by dividing by the number of trees. Proximities are used in locating outliers, producing low-dimensional views of the data, and replacing missing data. The advantages of RF are that it does not over-fit due to its capability of handling a large amount of data, it is also fast and memory efficient [18]. In the education domain, Naive Bayes is found to be the best technique by Troussas et al. [174]. Contrarily, Song et al. [156] identified Support Vector Machines as the best per- forming classifier. Thus, machine learning techniques can perform differently within the same domain. This prompts the need to investigate different techniques.