Sage Journals: Discover world-class research

Abstract

Noise data in text are one of the main factors affecting the quality of text categorization. A parallel noise data elimination algorithm based on principal component analysis method and term frequency-inverse document frequency method for the noise data issue of massive text categorization is proposed. Five types of noise data which may occur during text categorization process are analyzed and summarized in this paper. Before text categorization, a redundant noise elimination algorithm based on key feature selection is presented for redundant noise features. During the process of text categorization, the error noise detection algorithm is given for inaccurate noise features. The proposed method is compared with other four typical noise processing methods in different noise ratios on two common corpora. The results show that the proposed method is feasible and can maintain more stable and excellent classification performance and lower running time.

Keywords

Massive data text categorization noise feature reduction error feature key feature selection parallelization

Introduction

Text representation for text categorization is not clean and contains noise data (such as feature noise, category noise, sample noise, etc.).¹ The existing noise data increase the computational burden and reduce the accuracy of the text categorization performance. With the arrival of Big data and “Internet+” era, a wide variety of text information has been appeared competing, and massive text data have been formed. Therefore, how to effectively reduce the noise data and improve the classification performance of massive texts is a worthy subject.

The noise reduction methods that reduce the impact of noise data in text are divided into two kinds: noise toleration method and noise elimination method.² The former method is provided to increase the noise sensitivity of text representation and used to ignore the noise data in text processing.^3–5 For another perspective, the noise elimination method is proposed to identify the noise data and remove those identified noise data. Therefore, it can improve the rate of the key features selection and reduce the impact of noise features.^6–8 In general, different methods have different angle and focus and have different applications in different fields. However, to solve the text classification problem, it has not been able to achieve the ideal classification accuracy, especially in the realization of high accuracy for massive text classification. In this paper, from the point of the noise elimination, we proposed a parallel noise eliminate (PNE) algorithm based on principal component analysis (PCA) method and term frequency-inverse document frequency (TF-IDF) method for the massive text. We divided the algorithm into two stages, pre-classification and during classification process, to detect and eliminate noise data. The PNE algorithm was tested by using common corpora and compared with other similar algorithms.

The rest of this paper is organized as follows: the next section presents some related works about the solution of the noise data elimination algorithm. Section “PNE algorithm” proposes the PNE algorithm, and in section “Experiments and analysis,” experiment and analysis are done to verify the performance of PNE algorithm compared with the other similar methods. The final section concludes this paper.

Related work

In this section, we present a brief description of PCA method and TF-IDF method and give a summary of the noise data types in the text classification.

PCA method

PCA method is a multivariate statistical method for converting a set of related variables into another set of irrelevant variables by linear transformation. The new variables are arranged in descending order of variance to establish a principal component sequence.^9,10 The PCA method can calculate and rank the contribution of all the variables according to the evaluation method of the variable contribution and select several variables with the highest contribution. In recent years, the PCA method has been applied to feature selection and feature extraction in text classification to reduce redundant features and feature dimensionality. Uguz employed genetic algorithm and PCA method to extract the key features, which are used to reduce the influence of noise feature and to improve the performance of k-Nearest Neighbor (KNN) algorithm.¹¹ Bharti and Singh applied the PCA method to reduce dimensions in the feature space to select the most important information of the text.¹² Xu et al. compared the PCA method with the multi-label dimensionality reduction method via dependency maximization and conducted experimental verification and analysis.¹³ Javed et al. compared and analyzed the PCA-based feature selection algorithm with other two-stage algorithms in dimensionality reduction.¹⁴ We use the principal component selection to remove the redundant features in the original feature set and preserve the most important features for text categorization. In PCA method, each text is set as a n-dimensional vector Y . Y = (y₁, y₂, …, y_n)^T, where n is the number of original features in a text. According to the word frequency method, a feature y_i, arbitrary one element in Y , denotes the word frequency of y_i in the text. This is the generation method of vector element. Calculate the covariance matrix ∑ of Y , i.e. ∑= YY ^T. Then, we can obtain the eigenvalue λ_i (λ₁≥λ₂≥…≥λ_n≥0) and the corresponding eigenvector p_i (i = 1, 2, …, n). The i-th principle component of Y is X_i=p_i^T Y . The contribution rate of X_i is defined as $λ_{i} / \sum_{k = 1}^{n} λ_{k}$ . The cumulative contribution rate of the principal component is defined as $θ = \sum_{i = 1}^{m} λ_{i} / \sum_{j = 1}^{n} λ_{j}$ . The number of principal components is determined by θ. In general, θ > 0.95.

TF-IDF method

TF-IDF method is a word frequency weighting method commonly used in information retrieval. Yang and Pedersen conducted a comprehensive performance analysis and comparison of the five feature selection methods and suggested that if the computation is large, DF thresholding method will be the best choice.¹⁵ Zhang et al. gave a comprehensive analysis and comparison of the TF-IDF method and the Latent Semantic Indexing (LSI) algorithm.¹⁶ Yazdani et al. proposed a feature selection method combined TF-IDF method and Gaussian radial basis function to enhance the classification accuracy.¹⁷ Razzaghnoori et al. used the TF-IDF method as the weighting method in the feature vector of the question conversion.¹⁸ We employ TF-IDF method to calculate the weighted value of each feature, that is, to evaluate the classification distinction ability of each feature. Then, we can retain the features with larger weighted value and delete the features that have the smallest weighted value. The following equation is used for calculating the weighted value of each feature in the text set.¹⁹

T F I D F_{i} = t f (i, j) \times i d f (i) = \sum_{j = 1}^{N} (1 + \log t f_{i, j}) \times \log (N / n_{i})

(1)

In equation (1), tf_i_, _j is the word frequency of the feature i in the text j. n_i represents the number of the text that included the feature i. N is the total text number in text set.

Noise data

In response to the noise data processing, researchers have given a lot of research in the past 10 years. Yang presented the text classification multi-class noise elimination method using the least squares fitting method to reduce the redundant features of the training text set by deleting the non-information word and singular value cutting.²⁰ Wang et al. proposed a noise data cutting method for the classification performance of the text, but its accuracy is not high enough.⁶ Gong et al. introduced an automatic noise reducer based on supervised learning method, it is used to delete the noise of the text set and improve the speed and quality of the digest algorithm.⁷ Altinel and Ganiz proposed a semantic semi-supervised learning algorithm to effectively eliminate noise data on a relatively large and complex training set, and it can effectively improve the classification accuracy under the condition that the marked samples are quite limited.⁸ In general, the study of noise cancellation is to better improve the quality of text classification.

For a specific text, the number of categories it belonged to and classifiers it needed is limited. When receiving the output information from all the classifiers, the unrelated output classifier is considered the redundant classifier. To increase the accuracy of the massive text, it is necessary to remove the noise data in the process of text classification. According to the noise data types before and after text classification process, the followings are the summary. (1) The redundant noise features that appear during the text segmentation process. (2) The erroneous noise features that emerge during critical feature selection process. (3) The redundant noise that exists in the feature matching process. (4) Inconsistent data appearing during categories matching. (5) Expired classifiers that arise from text categorization process.

Regarding the first and second noise data type, if the feature is filtered out by key feature selection, it will not have any effect on the text categorization. If the feature becomes a key feature, it can be removed by inconsistency and error detection. The third noise data type is filtered according to the category matching of the key feature set. To reduce the burden of the categorization process, the filtering work needs to be done before the category matching. We propose a parallel redundant feature noise elimination (RNE) algorithm to improve the performance of the key feature selection method. The fourth and fifth noise data belong to the error noise data type. For dealing with this kind of noise data, it is necessary to detect the inconsistency between features and categories, and the expired classifier, etc. To handle those noise data type, we propose a parallel error detection (PED) algorithm by employing historical error.

PNE algorithm

The PNE algorithm is presented for solving the problem of the noise elimination in massive text sets, and it is consisted of two parts: one part is the RNE algorithm, which is used to eliminate the redundant noise feature in the original feature set of the text before the text classification, thus useless features can be removed. Another part is to detect and distinguish the error noise data existed in text classification process. According to the historical error features, the exception detection is performed for the existing key features. Based on the number of classifier output that is incorrect to determine whether the classifier is expired or not. And, it is described in the PED algorithm in detail. The massive texts can be interpreted as a single text with more information, or the text set contains numerous texts. The PNE algorithm is proposed to reduce the dimension of original feature set, eliminate incorrect features, and detect and match the inconsistent key features and categories. So that the proposed algorithm can reduce the time consuming and storage space of the classification process. It would be improved the classification performance of massive text set greatly. For the convenience to identify and deal with the noise data, we formalize the entire massive classification system as follows.

Assume D is a text set, D= {D₁, D₂, …, D_i, …, D_T}, where 1 ≤ i ≤ T, and D_i denotes a text in the set D. Arbitrarily text D_i has a corresponding original feature set W_i after text segmentation and preprocessing. W_i= { f_i₁, f_i₂, …, f_in}, where f_ij (1 ≤ i ≤ T, 1 ≤ j ≤ n) denotes an original feature in W_i, and n denotes the total number of features in W_i. W_i is the input of the RNE algorithm (specific steps are shown in Algorithm 1). Carry out the first simplification of the original feature set based on the TF-IDF method and PCA method. Calculate the TF-IDF value of each feature that is served as the principle component and retain those features with greater text classification capacity. This is the second simplification for the feature set. After two-stage feature simplifications, the key feature set F_i is obtained, where F_i= { f_i₁, f_i₂, …, f_im} and n ≫ m. The main purpose of RNE algorithm is to eliminate useless redundant noise data before the text classification process.

Algorithm 1: RNE

Input: the original feature set W_i of an arbitrarily text D_i

Output: the key feature set F_i of D_i

Step 1. Assume the original feature set W_i is a column vector A = (f_i₁, f_i₂, …, f_in)^T, calculate the covariance matrix ∑= AA ^T and get the corresponding eigenvalue λ_j and eigenvector p_j.

Step 2. Map <Key: λ_j+ p_j +f_ij, Value: principalvalue >.

Step 3. Reduce <Key: λ_j+ p_j +f_ij, Value: principalvalue >.

Step 4. Set the cumulative contribution rate θ of the top k features, θ > 0.95, put the top k features to establish the reduced feature set W_i’.

Step 5. Calculate the weighted value of each feature in W_i’ according to equation (1).

Step 6. Map <Key: tf_ij+ idf_j +f_ij, Value: tfidfvalue>.

Step 7. Reduce <Key: tf_ij+ idf_j +f_ij, Value: tfidfvalue>.

Step 8. Select the feature that have the greatest weighted value and put it in key feature set F_i.

In this algorithm, the higher the word frequency of the feature is, the more important it is. The larger the weighted value is, the greater the classification performance is. The cumulative contribution rate is required larger than 0.95. That is used to preserve more important features. The size of the weighted threshold is set to 20% of the highest weighted value in the current features.

The key feature set F_i= { f_i₁, f_i₂, …, f_im} obtained by Algorithm 1 is a collection of features that can representing the document key content. The next step is to perform the classification matching for the features in F_i and the detailed steps are described in Algorithm 2. It is mainly to get the target category c_i for the text D_i, to detect and delete all the error noise features in the classification process, to record the expired classifier and set the classifier as a phasing out classifier.

Algorithm 2: PED

Input: the key feature set F_i, the error noise feature set S

Output: the target category c_i, the updated error noise feature set S

Step 1. Initialize the category set CS=∅.

Step 2. Calculate F_i’=F_i∩S, and remove F_i’ from F_i.

Step 3. If the category of feature f_ik is inconsistent with the labeled category, add the feature f_ik to S and delete f_ik from F_i.

Step 4. Map <Key: tf_ik+ idf_k +f_ik, Value: list contfvalue>.

Step 5. Reduce <Key: tf_ik+ idf_k +f_ik, Value: contfzvalue>.

Step 6. Put all the categories of contfzvalue into CS.

Step 7. Select the category with the highest contribution among CS as the target category c_i.

Step 8. Count the number of error classifiers. If the number of errors is out of the threshold, the classifier is labeled as the eliminating classifier.

Among Algorithm 2, the error feature set S is an empty set during the first running, i.e. S=∅, when F_i=F₁. The element number of S will increase as algorithm execution frequencies increases. CS denotes an alternative category set corresponding to the classification of the key feature set F_i. Due to the inconsistence between the semantic or emotional meaning of the features and the actual category, features in S are all abnormal in the process of classification matching. The elimination threshold is set to one-third of the number of key features.

Assume Z is the category set in the text classification system. Z= {z₁, z₂, …, z_L}, where z_i (1 ≤ i ≤ L) denotes a category in set Z. L denotes the total number of categories in set Z. The contribution of the feature f_ik to the category z_j is expressed as follows

\begin{matrix} {Cont}_{i j} = \frac{\log t f (f_{i k}) \times \log (\frac{V}{i d f (f_{i k})} + 0.01)}{\sqrt{\sum_{i - 1}^{m} {[\log t f (f_{i k}) \times \log (\frac{V}{i d f (f_{i k})} + 0.01)]}^{2}}}, \\ 1 \leq i \leq V, 1 \leq j \leq L, 1 \leq k \leq m \end{matrix}

(2)

where f(f_ik) indicates the word frequency of the feature f_ik in text D_i. idf(f_ik) indicates the reverse text frequency of the feature f_ik. m is the total number of features in key feature set. V is the total number of text in the text set. According to feature set of the labeled text, the inconsistent between features and the category is recorded. All actual output of the classifier is recorded to determine whether the classifier is expired.

Experiments and analysis

Experimental setup

We utilize MATLAB R2014b to construct a text classification system for this paper. The operating system is ubuntu12.04. Java version is jdk1.6-u21-x64. Hadoop version is hadoop-0.20.2. Eclipse version is eclipse-3.7. Central Processing Unit (CPU) is the Intel Core i5. The main frequency is 3.20 GHz and memory is 8G. We employ MATLAB compiler to translate the Map and Reduce operator into Hadoop executable program.

We choose three state-of-the-art noise data elimination methods, which are ECN algorithm,⁶ NTDSEC algorithm,²¹ and Double Centroids Noise Eliminate (DBCNE) algorithm,²² respectively, as baselines to compared and analyzed with the proposed PNE algorithm.

Evaluation metrics

We choose two common metrics, the noise feature elimination rate FP and the text classification precision P, to evaluate the PNE algorithm and baseline algorithms. According to the equation of noise cancellation rate p1 in Zhang et al.²¹ and the formula of noise elimination precision (NEP) in Wang et al.,⁶ we redefined the noise feature elimination rate FP in this paper as follows. For an arbitrary text D_i in the text set, the redundant noise feature set can be expressed as D_i–F_i after employing the PNE algorithm. S is the error noise feature set. Assume Noise denotes the full noise feature set in the text D_i. Then, FP can be described as follows.

F P = \frac{| S \cap Noise | + | (D_{i} - F_{i}) \cap Noise |}{| Noise |}

(3)

FP denotes the percentage of the number of detected redundant noise features and detected error noise features in the total number of noise features. The higher the value of FP is, the less the noise feature remaining in the key feature set has. The precision of the text classification P is a very common metric.^11,23 P is described as follows

P = T P / (T P + F P)

(4)

P denotes the percentage of texts with the right category in the whole text set. The larger the value of P is, the higher the text classification performance of the PNE algorithm is.

Text corpora

The common corpora, Reuters-21578 and 20Newsgroup, were selected as experimental datasets in this paper. The details are shown in Tables 1 and 2, where CN is the category no., C is the category name, and TN denotes the number of texts belonging to this category.

Table 1.

Documents of each category on Reuters-21578.

CN	C	TN	CN	C	TN
C1	Earn	3975	C6	Trade	518
C2	Acq	2408	C7	Interest	495
C3	Money-fx	759	C8	Ship	295
C4	Crude	606	C9	Wheat	294
C5	Grain	605	C10	Corn	245

CN: category No.; C: category name; TN: number of texts.

Table 2.

Documents of each category on 20Newsgroup.

CN	C	TN	CN	C	TN
C11	Comp.sys.ibm.pc.hardware	1000	C16	Rec.sport.hockey	1000
C12	Comp.sys.mac.hardware	1000	C17	Sci.space	1000
C13	Rec.autos	1000	C18	Soc.religion.christian	997
C14	Rec.motorcycles	1000	C19	Talk.politics.guns	1000
C15	Rec.sport.baseball	1000	C20	Talk.politics.mideast	1000

CN: category No.; C: category name; TN: number of texts.

Experimental results and discussion

In this section, we designed three experiments to assess the PNE algorithm. Firstly, we designed an experiment to verify the effectiveness of noise elimination of the PNE algorithm. Secondly, we evaluate the parallelization availability of the PNE algorithm. At last, we set another experimental environment to compare the noise elimination ability with other similar noise elimination methods under different noise ratios.

Verify the validity of the noise elimination of the PNE algorithm

We verified and analyzed the elimination rate of the PNE algorithm for the redundant noise features before text classification and the error noise features during the text classification process in this experiment. We record the running results and compare PNE algorithm with other noise elimination methods. 70% data of Reuters-21578 and 20Newsgroup are set as a training set, and 30% of the data are set as a test set. According to the principle of Pareto, we adjust the noise ratio as 20% in the test set. PNE algorithm, Noise-Tolerance Data Stream Ensemble Classifiers (NTDSEC) algorithm, Eliminating Class Noise (ECN) algorithm and DBCNE algorithm are running on two corpora by 10-fold cross-validation method. According to equations (3) and (4), FP and P of four algorithms are calculated and described in Tables 3 and 4, where Avg represents the average value of FP or P of each algorithm in each category set. In this paper, all the results of PNE algorithm are passed the significance testing. And, the evaluation metrics of PNE algorithm is more advantageous in 0.95 confidence interval than the other baseline algorithms on the two corpora.

Table 3.

The value of FP and P of four algorithms on corpus Reuters-21578.

	FP				P
	NTDSEC	ECN	DBCNE	PNE	NTDSEC	ECN	DBCNE	PNE
Earn	0.6752	0.9365	0.9132	0.9615	0.7679	0.9646	0.9827	0.9964
Acq	0.6593	0.9324	0.9168	0.9658	0.7985	0.9783	0.9882	0.9951
Money-fx	0.6697	0.9471	0.9251	0.9709	0.8487	0.9846	0.9978	0.9984
Crude	0.6854	0.9453	0.9578	0.9649	0.8558	0.9809	0.9965	0.9979
Grain	0.6891	0.9482	0.9546	0.9766	0.8571	0.9761	0.9973	0.9986
Trade	0.6989	0.9449	0.9687	0.9624	0.8649	0.9851	0.9954	0.9969
Interest	0.6573	0.9211	0.9321	0.9562	0.8206	0.984	0.9967	0.9978
Ship	0.7254	0.9497	0.9740	0.9725	0.8718	0.9754	0.9908	0.9997
Wheat	0.7165	0.9523	0.9798	0.9763	0.9211	0.982	0.9979	0.9989
Corn	0.7283	0.9586	0.9837	0.9682	0.9052	0.9885	0.9968	0.9993
Avg	0.6905	0.9436	0.9506	0.9675	0.8512	0.9800	0.9940	0.9979

PNE: parallel noise eliminate.

Table 4.

The value of FP and P of four algorithms on corpus 20Newsgroup.

	FP				P
	NTDSEC	ECN	DBCNE	PNE	NTDSEC	ECN	DBCNE	PNE
Comp.sys.ibm.pc.hardware	0.7021	0.9431	0.9647	0.9675	0.7621	0.9781	0.9961	0.9968
Comp.sys.mac.hardware	0.7059	0.9504	0.9485	0.9663	0.7887	0.9624	0.9972	0.9979
Rec.autos	0.7109	0.9496	0.9564	0.9717	0.8387	0.9714	0.9973	0.9981
Rec.motorcycles	0.7153	0.9528	0.9638	0.9705	0.8586	0.9756	0.9958	0.9984
Rec.sport.baseball	0.7176	0.9532	0.9596	0.9728	0.8485	0.9642	0.9979	0.9989
Rec.sport.hockey	0.7208	0.9547	0.9641	0.9698	0.8402	0.9753	0.9965	0.999
Sci.space	0.7197	0.9536	0.9678	0.9593	0.8164	0.9776	0.9957	0.9976
Soc.religion.christian	0.7168	0.9617	0.9536	0.9684	0.8513	0.9864	0.9964	0.9974
Talk.politics.guns	0.7223	0.9624	0.971	0.9755	0.9236	0.9835	0.998	0.9984
Talk.politics.mideast	0.7214	0.9622	0.9689	0.9802	0.9016	0.9819	0.9979	0.9986
Avg	0.7153	0.9544	0.9618	0.9702	0.8430	0.9756	0.9969	0.9981

It can be seen from Tables 3 and 4 that FP of the PNE algorithm is significantly higher than that of the NTDSEC algorithm and the ECN algorithm when the noise ratio is 20% on the corpora Reuters-21578 and 20Newsgroup, and the average value of FP of PNE algorithm is higher than the other three baseline algorithms. Those demonstrate that the PNE algorithm can effectively remove the redundant noise features in original feature set and the error noise features in key feature set to preserve the key features. Using the historical error feature set S, PNE algorithm can accurately detect the inconsistency in the key feature set, which is very helpful to improve classification precision. It can be observed from the two tables that the FP value of PNE algorithm is greater than the other baseline algorithms in category Earn, Acq, Money-fx, Crude, Grain, Interest and nine categories in 20Newsgroup except Sci.space. However, the FP value of PNE algorithm is slightly smaller than DBCNE algorithm in category Trade, Ship, Wheat, Corn, and Sci.space. This situation is due to the category ambiguity of features in these categories and one feature belonging to multiple categories, which can easily lead to misjudgment of the PNE algorithm. The detailed growth ratio of FP and P of PNE algorithm compared with DBCNE algorithm is shown in Table 5. However, for the P value, PNE algorithm is significantly higher than the DBCNE algorithm, ECN algorithm, and NTDSEC algorithm and has an obvious advantage. So, compared with the three baselines, the PNE algorithm can effectively improve the text classification precision by reducing the noise feature elimination rate.

Table 5.

The growth ratio of PNE compared with DBCNE.

Category	FP (%)	P (%)	Category	FP (%)	P (%)
Earn	5.29	1.39	Comp.sys.ibm.pc.hardware	0.29	0.07
Acq	5.34	0.7	Comp.sys.mac.hardware	1.88	0.07
Money-fx	4.95	0.06	Rec.autos	1.6	0.08
Crude	0.74	0.14	Rec.motorcycles	0.7	0.26
Grain	2.3	0.13	Rec.sport.baseball	1.38	0.1
Trade	−1.39	0.15	Rec.sport.hockey	0.59	0.25
Interest	2.59	0.11	Sci.space	−0.88	0.19
Ship	−0.15	0.89	Soc.religion.christian	1.55	0.1
Wheat	−0.36	0.1	Talk.politics.guns	0.46	0.04
Corn	−1.58	0.25	Talk.politics.mideast	1.17	0.07

PNE: parallel noise eliminate.

Verify the parallelization of the PNE algorithm

The running time of four algorithms in the above experiment is compared to verify the advantages of parallelization of PNE algorithm, which is shown in Figure 1. From Figure 1(a) and (b), the curve of the running time of PNE algorithm has a subtle gap compared with the DBCNE algorithm, and both are smaller than NTDSEC algorithm and ECN algorithm. This is because the two sub-algorithms of the PNE algorithm, RNE and PED, are the parallel processing algorithms, which use the Map and Reduce operator to perform the separate processing based on the divide and conquer method. It is shown that, compared with the other state-of-the-art algorithms, the PNE algorithm has the similar running time in the premise of the higher precision because of its parallelization characteristic. We can be observed that the PNE algorithm has an overall advantage over other baseline algorithms.

Figure 1.

The running time of four algorithms on two corpora. (a) Reuters-21578 corpus. (b) 20Newsgroup corpus.

Influence of changing noise ratio on PNE algorithm performance

The reduced noise ratio of the original feature set in each category is designed to verify and evaluate the influence of noise ratio on the performance of PNE algorithm. The FP value and P value of four algorithms are observed and analyzed under the decreasing noise ratio, which has four experimental analysis environments 20, 15, 10, and 5% noise ratio. We used the 10-fold cross-validation method for the implementation of the four algorithms on the two corpora. Averaging the 10 times running results from the four algorithms and obtaining their FP and P as shown in Figure 2. Under four noise ratios, the FP value and P value of the PNE algorithm are significantly better than the other baseline algorithms in 0.95 confidence interval.

Figure 2.

FP and P of each algorithm under different noise ratios. (a) The value of FP. (b) The value of P.

It can be seen from Figure 2(a) that FP of each algorithm is decreasing with the decreasing noise ratio. Compared with the other three algorithms, PNE algorithm can maintain a stable noise elimination rate, which indicates that PNE can effectively remove the noise features of the text set, especially when the noise ratio is small. Figure 2(b) shows that P of the four algorithms is increasing slightly with the decreasing noise ratio. Particularly, the classification precision of the PNE algorithm is the most excellent. It is shown that the PNE algorithm can maintain a good and stable classification performance with the decreasing noise ratio, which is more practical than the other baseline algorithms.

Conclusions

A parallel noise elimination algorithm PNE is proposed for noise data processing in massive text classification. Firstly, we summarized five types of noise data for recognizing the redundant noise features and the error noise features. Secondly, we presented a RNE algorithm to eliminate the redundant noise feature in the original feature set, using the PCA method and the TF-IDF method. Thirdly, we proposed the PED algorithm to detect and remove the error noise features in the key feature set based on the historical error features set. Finally, we design experiments to verify and evaluate the PNE algorithm on two common corpora. The experimental results show that: (1) compared with the baseline algorithms, PNE algorithm has higher noise elimination rate and text classification precision; (2) the parallel processing of the PNE algorithm makes it possible to maintain a smaller running time and a faster calculation speed in the case of massive texts; (3) with the gradually decreasing noise ratio, the PNE algorithm can maintain a good and stable noise elimination rate and text classification precision, which indicates the PNE algorithm has an obvious advantage over the other baseline algorithms.

In the next work, we will explore how to perform the effective selection of key features in the massive data and compare the implementation efficiency of the text classification when the data information is extremely larger.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Natural Science Foundation of Jilin Province (grant number 20150101054JC), Postdoctoral Research Fund of Jilin Province (grant number 40301919), the Key Program for Science and Technology Development of Jilin Province (grant number 20150204036GX), Provincial Industrial Innovation Special Fund Project of Jilin Province (grant number 2017C051) and Major Science and Technology Bidding Project of Jilin Province (grant number 20170203004GX).

References

Esuli

Sebastiani

Improving text classification accuracy by training label cleaning. ACM Trans Inf Syst 2013; 31: 1–31.

Katz

Ofek

Shapira

ConSent: context-based sentiment analysis. Knowl Base Syst 2015; 84: 162–178.

Aggarwal

Zhao

On the use of side information for mining text data. IEEE Trans Knowl Data Eng 2014; 26: 1415–1429.

Barigou

Improving k-nearest neighbor efficiency for text categorization. NNW 2016; 26: 45–66.

Song

Liang

et al . Taking advantage of improved resource allocating network and latent semantic feature selection approach for automated text categorization. Appl Soft Comput 2014; 21: 210–220.

Wang

Guan

Wang

XL.

A method for eliminating class noise in text classification based on feature class attribute. Acta Automatica Sinica 2007; 33: 809–816. (in Chinese)

Gong

Tian

SF.

Supervised learning of an automatic noisy semantic unit filter for multi-document summarization. J Comput Res Dev 2013; 50: 873–882. (in Chinese).

Altinel

Ganiz

MC.

A new hybrid semi-supervised algorithm for text classification with class-based semantics. Knowl Base Syst 2016; 108: 50–64.

Pan

JK.

Human motion recognition based on triaxial accelerometer. J Comput Res Dev 2016; 53: 621–631. (in Chinese).

10.

Feng

Wang

Zuo

Novel feature selection method based on random walk and artificial bee colony. IFS 2017; 32: 115–126.

11.

Uguz

A two-stage feature selection method for text categorization by using information gain, principal component analysis and genetic algorithm. Knowl Base Syst 2011; 24: 1024–1032.

12.

Bharti

Singh

PK.

Hybrid dimension reduction by integrating feature selection with feature extraction method for text clustering. Expert Sys Appl 2015; 42: 3105–3114.

13.

Liu

Yin

et al . A multi-label feature extraction algorithm via maximizing feature variance and feature-label dependence simultaneously. Knowl Base Syst 2016; 98: 172–184.

14.

Javed

Maruf

Babri

HA.

A two-stage Markov blanket based feature selection algorithm for text classification. Neurocomputing 2015; 157: 91–104.

15.

Yang

Pedersen

JO.

A Comparative Study on Feature Selection in Text Categorization. In: ICML ‘97 Proceedings of the Fourteenth International Conference on Machine Learning, 1997, pp. 412–420.

16.

Zhang

Yoshida

Tang

XJ.

A comparative study of TF* IDF, LSI and multi-words for text classification. Expert Syst Appl 2011; 38: 2758–2765.

17.

Yazdani

Murad

MAA

Sharef

et al . Sentiment classification of financial news using statistical features. Int J Patt Recogn Artif Intell 2017; 31: 1750006.

18.

Razzaghnoori

Sajedi

Jazani

IK.

Question classification in Persian using word vectors and frequencies. Cognit Syst Res 2018; 47: 16–27.

19.

Plansangket

Gan

JQ.

A query suggestion method combining TF-IDF and Jaccard coefficient for interactive web search. Air 2015; 4: 119–125.

20.

Yang

YM.

Noise Reduction in a Statistical Approach to Text Categorization. In: Proceedings of the 18th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 1995, pp.256–263.

21.

Zhang

Yang

Effectiveness analysis and application in data streams of cross validation noise-tolerance classification algorithm. Acta Electronica Sinica 2011; 39: 378–382. (in Chinese).

22.

Ping

Zhou

Yang

YX.

A novel scheme for accelerating support vector clustering. Comput Informat 2012; 31: 613–638.

23.

Jiang

Liou

Lee

SJ.

A fuzzy self-constructing feature clustering algorithm for text classification. IEEE Trans Knowl Data Eng 2011; 23: 335–349.

Parallel noise eliminate: A parallel noise elimination algorithm for massive text categorization

Abstract

Keywords

Introduction

Related work

PCA method

TF-IDF method

Noise data

PNE algorithm

Experiments and analysis

Experimental setup

Evaluation metrics

Text corpora

Experimental results and discussion

Verify the validity of the noise elimination of the PNE algorithm

Verify the parallelization of the PNE algorithm

Influence of changing noise ratio on PNE algorithm performance

Conclusions

Footnotes

Declaration of conflicting interests

Funding

References