Intelligent Grouping Method of Science and Technology Projects Based on Data Augmentation and SMOTE

Figures & data

Figure 1. The process of text classification.

Figure 2. The overall model architecture.

Figure 3. Paraphrasing.

Figure 4. Mixup.

Figure 5. Generating data by GPT-3.

Table 1. Partial experimental data set.

category	keywords
畜牧-饲料 (livestock - feed)	旱生牧草; 选育; 退化草地; 改良; 应用示范 (xerophyte forage; selective breeding; degraded grassland; improvement; application demonstration)
食品包装与储藏 (food packaging and storage)	复合材料; 蛋白质; 聚吡咯; 晶体结构; 光电性能 (composite material; protein; polypyrrole; crystal structure; optoelectronic properties)
计算机软件与计算机应用 (computer software and computer applications)	老龄人口主动健康; 数据可信管理; 健康状态评估; 智能服务导航; 云端协同 (active health of the aging population; data trusted management; health status assessment; intelligent service navigation; cloud collaboration)
微生物学 (microbiology)	食品有机废物; 乳酸; 微生物菌肥; 全组分资源化利用 (food organic waste; lactic acid; microbial fertilizer; full-component resource utilization)

Figure 6. The reduced categories and corresponding quantity. (IS: Information Science, CS: Chemical Science, MS1: Medical Science, ES: Earth Science, EMAS: Engineering and Materials Science, MS2: Mathematical Science, LS: Life Science, MS: Management Science).

Table 2. Partial experimental data set after merging categories.

category	keywords
生命科学(Life Sciences)	旱生牧草; 选育; 退化草地; 改良; 应用示范(Xerophyte forage; selective breeding; degraded grassland; improvement; application demonstration)
工程与材料科学(Engineering and Materials Science)	复合材料; 蛋白质; 聚吡咯; 晶体结构; 光电性能(composite material; protein; polypyrrole; crystal structure; optoelectronic properties)
信息科学 (Information Science)	老龄人口主动健康; 数据可信管理; 健康状态评估; 智能服务导航; 云端协同(Active health of the aging population; data trusted management; health status assessment; intelligent service navigation; cloud collaboration)
生命科学 (Life Sciences)	食品有机废物; 乳酸; 微生物菌肥; 全组分资源化利用(Food organic waste; lactic acid; microbial fertilizer; full-component resource utilization)

Table 3. Confusion matrix of the classification result.

	predict value is positive	predict value is negative
the actual value is positive	TP	FN
the actual value is negative	FP	TN

Table 4. Performance of the different models.

Model	Original	Paraphrasing	Mixup	Generating data by GPT-3
SVM+TF-IDF	47.52%	79.64%	62.96%	69.96%
XGBoost+TF-IDF	38.02%	63.38%	51.33%	61.12%
SVM+Word2Vec	61.56%	76.67%	56.34%	65.04%
XGBoost+Word2Vec	60.01%	75.27%	58.95%	69.52%
SVM+TF-IDF weighted Word2Vec	56.92%	71.92%	57.68%	67.91%
XGBoost+TF-IDF weighted Word2Vec	53.55%	71.04%	58.19%	69.40%

Figure 7. F1-score corresponding to each category. (IS: Information Science, CS: Chemical Science, MS1: Medical Science, ES: Earth Science, EMAS: Engineering and Materials Science, MS2: Mathematical Science, LS: Life Science, MS: Management Science).

Table 5. Performance of the different models.

Model	Original	Paraphrasing	Mixup	Generating data by GPT-3
SVM+TF-IDF	77.82%	96.78%	91.83%	94.65%
XGBoost+TF-IDF	77.18%	90.82%	86.01%	89.85%
SVM+Word2Vec	91.76%	95.08%	84.43%	91.84%
XGBoost+Word2Vec	89.46%	93.91%	86.17%	91.45%
SVM+TF-IDF weighted Word2Vec	82.42%	94.07%	85.20%	90.95%
XGBoost+TF-IDF weighted Word2Vec	81.62%	92.61%	86.33%	93.28%

Figure 8. F1-score corresponding to each category after introducing SMOTE. (IS: Information Science, CS: Chemical Science, MS1: Medical Science, ES: Earth Science, EMAS: Engineering and Materials Science, MS2: Mathematical Science, LS: Life Science, MS: Management Science).