Text GCN-SW-KNN: a novel collaborative training multi-label classification method for WMS application themes by considering geographic semantics

Figures & data

Figure 1. Architecture of the proposed collaborative training model Text GCN-SW-KNN by cooperating ML-KNN and text GCN-SW, where the first two tiers show the working mechanism of Text GCN-SW

Figure 2. Exemplary shortest path between feature words and theme within SWEET and WordNet. (a) Feature word A is included in SWEET; (b) Feature word $A$ is not included in SWEET but in WordNet; (c) Feature word A and alternative word B are two subclasses away; (d) Feature word A and Theme T are two subclasses away

Table 1. Co-training algorithm based on Text GCN-SW and ML-KNN

Input: labeled training sample list (train_list) and unlabeled test sample list (test_list) Output: newly labeled test sample list test_list
1:	Initialize iteration counter r = 0 for the co-training model
2:	Copy a new list incon_list = test_list for test samples with inconsistent labels by base classifiers
3:	Initialize changing detection flag changed = true for incon_list
4:	while (incon_list.length() > 0 and changed ! = false and r < max_Iteration) do
5:	Reset changed = false
6:	Train ML-KNN using train_list as base classifier C1
7:	Train Text GCN-SW using train_list as base classifier C2
8:	for each i in range [0, incon_list.length()-1] do
9:	Assign incon_list[i].label = C1.classify(incon_list[i].document)
10:	Create a new label GCN_label = C2.classify(incon_list[i].document)
11:	if incon_list[i].label = = GCN_label then
12:	train_list.add(incon_list[i])
13:	incon_list.remove(incon_list[i])
14:	changed = true
15:	r++
16:	for each j in range [0, test_list.length()-1] do
17:	if test_list[j] corresponded test sample in train_list then
18:	Assign test_list[j].label = train_list.getItemById(test_list[j].id).label
19:	else if test_list[j] corresponded test sample in inconsistent_list then
20:	Assign test_list[j].label = incon_list.getItemById(test_list[j].id).label

Figure 3. An exemplary WMS metadata document in XML format, including URL, title, abstract and keyword, etc

Figure 4. Frequency histogram of the number of labels for each WMS and each layer on the selected 501 samples

Figure 5. Performance comparison of text GCN-SW-KNN and eight baselines for both service-level and layer-level classification, respectively. (a) is the result for WMS and (b) is for WMS layer

Figure 6. Stability of F1-score in repeating experiments for Text GCN and Text GCN-SW. (a) is the result for WMS and (b) is for WMS layer

Figure 7. Accuracy and F1 for Text GCN, Text GCN-SW, ML-KNN and LSTM with different proportions of the training data: (a) is for WMS and (b) is for WMS layer

Table A1. Performance metrics of text GCN-SW-KNN and eight baselines for WMS service-level classification

Method	Hamming loss	Accuracy	Jaccard	Precision	Recall	F1
Text GCN-SW-KNN	0.07009	0.64200	0.71968	0.77842	0.75975	0.74959
Text GCN-SW	0.07118	0.63800	0.71483	0.75475	0.77300	0.74460
Text GCN	0.07345	0.61900	0.69053	0.71975	0.74383	0.71744
ML-KNN	0.06818	0.60000	0.71050	0.76183	0.77200	0.74750
BRKNN	0.06909	0.60000	0.68850	0.74617	0.73450	0.72024
LSTM	0.08800	0.54200	0.62822	0.66520	0.68705	0.65940
Bi-LSTM	0.08436	0.56200	0.64490	0.68223	0.69900	0.67459
CNN	0.09872	0.54300	0.58083	0.58650	0.61300	0.59346
CLSTM	0.08709	0.56200	0.62833	0.64783	0.67475	0.65027

Table A2. Performance metrics of text GCN-SW-KNN and eight baselines for WMS layer-level classification

Method	Hamming loss	Accuracy	Jaccard	Precision	Recall	F1
Text GCN-SW-KNN	0.04256	0.74760	0.80543	0.84909	0.81787	0.82515
Text GCN-SW	0.04433	0.72164	0.79705	0.81880	0.85357	0.82263
Text GCN	0.04542	0.71411	0.78679	0.80596	0.83969	0.81050
ML-KNN	0.04545	0.73288	0.79663	0.83436	0.82095	0.81826
BRKNN	0.04483	0.74658	0.79561	0.82420	0.81935	0.81393
LSTM	0.05490	0.69863	0.75269	0.77853	0.78140	0.77082
Bi-LSTM	0.05327	0.70171	0.75628	0.77518	0.79255	0.77448
CNN	0.06235	0.67995	0.70570	0.70772	0.73241	0.71444
CLSTM	0.05355	0.69552	0.74349	0.75654	0.77834	0.75937

Data availability statement

The code and data that support the findings of this study are openly available in GitHub at https://github.com/ZPGuiGroupWhu/Text-based-WMS-Application-Theme-Classification/tree/master.