1.Research Center of Information Technology,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China
2.National Engineering Research Center for Information Technology in Agriculture,Beijing 100097,China
3.Key Laboratory of Digital Village Technology,Ministry of Agriculture and Rural Affairs,Beijing 100097,China
Show less
文章历史+
Received
Published
2025-01-06
2025-07-01
Issue Date
2026-04-15
PDF (1556K)
摘要
针对我国动植物新品种资源信息服务存在的信息不对称、更新不及时、新品种资源难以落地等问题。提出一种结合深度神经网络和注意力机制的两步推荐算法。该算法首先基于动植物品种资源全文数据,构建动植物品种词汇库,搭建Siamese BERT(Bidirectional Encoder Representations from Transformers)网络,通过BERT获取文本上下文关联语义特征,运用双向最大匹配算法获取品种词典特征向量,融合语义及词典特征,通过相似度计算训练回归模型,实现动植物新品种与历史品种相似判断。接下来融合用户静态属性特征、用户行为特征、动植物品种资源特征等品种推荐影响因子,综合分析农业领域用户的行业差异、习惯差异,提出面向动植物品种推荐的用户特征及品种特征表征方式,基于CNN(Convolutional Neural Network)构建用户兴趣网络模型,并引入注意力机制来实现用户与品种的精准匹配推荐。在互联网实时监控新品种资源,实现全自动化的动植物新品种自动推荐。性能测试与效果验证结果表明,该算法在准确率和F1值评价指标上分别达到84.1%和0.832,相比基于协同过滤、矩阵分解等传统推荐算法,能更精准实现动植物新品种的推荐。
Abstract
In response to the problems of information asymmetry, untimely delayed updates and difficulty in implementing information services for new animal and plant species resources in China, a two-step recommendation algorithm that combines deep neural networks and attention mechanisms is proposed in this study. Firstly, based on the full-text data of animal and plant species resources, the algorithm constructs a vocabulary library of animal and plant species, and builds a Siamese BERT(Bidirectional Encoder Representations from Transformers) network. BERT was used to obtain text context-related semantic features. The bidirectional maximum matching algorithm is used to obtain variety dictionary feature vectors.Integrates semantic and dictionary features, a regression model was trained through similarity calculation to achieve the similarity judgment between new animal and plant varieties and historical varieties. Next, by integrating factors influencing variety recommendation, such as user static attribute characteristics, user behavior characteristics, and animal and plant variety resource characteristics, the industry differences and habits of users in the agricultural field are comprehensively analyzed. A user feature and variety feature representation method is proposed for animal and plant variety recommendation, and a user interest network model is constructed based on Convolutional Neural Network (CNN) and the attention mechanism is introduced to achieve accurate matching and recommendation between users and varieties. By monitoring new variety resources in real time on the internet, fully automatic recommendation of new animal and plant varieties is realized. The experimental performance testing and effectiveness verification results show that the algorithm achieves the accuracy and F1 score evaluation indicators of 84.1% and 0.832, respectively. Compared with the traditional recommendation algorithms based on collaborative filtering and matrix decomposition, the algorithm proposed in this study can more accurately recommend new varieties of animals and plants.
2)使用双向最大匹配算法进行实体识别并向量化。首先执行正向最大匹配,算法从文本起始位置开始扫描,初始化当前字符位置为0。在每一步处理中,算法截取从当前位置开始的最长10字符词组,查询预构建的动植物品种资源词汇库。若未匹配成功,则逐步缩短词组长度(每次减1字符)继续查询,直至匹配成功或缩减至单字符。匹配成功后,将该词组标记为实体,并将扫描位置更新至该实体末尾继续后续处理。接着执行逆向最大匹配,算法从文本末尾开始反向扫描,采用与正向匹配相同的处理逻辑,区别在于扫描方向和词组截取顺序改为从右向左。逆向匹配能有效捕捉正向匹配可能遗漏的特殊词组组合。最后,将正向匹配和逆向匹配结果进行对比、合并、去重,得到最终的实体识别结果。本研究仅保留与词汇库匹配的实体词,通过Word2Vec(Word embedding via neural networks)进行向量化,得到整体文本的词典特征向量。将该特征向量与BERT层获取的特征向量拼接,作为最终的品种资源特征向量。
用户特征提取首先是通过嵌入层获得维度为16的特征向量,例如性别、年龄、行业分别为 h {gender}、 h {age}、 h {industry},输入第1层全连接层,输出维度为128维的特征向量,再输入第2层全连接层,最终输出特征为256维的用户特征。通过多层全连接层,模型将低维用户特征进行组合和转换,学习到更加丰富的高阶特征表示。
此外,为验证模型对冷启动问题的处理能力,设计2组对比试验:1)新品种冷启动测试。从试验数据中提取2023年新审定的320个零行为数据品种作为测试集,其首推准确率达到68.5%(传统协同过滤方法为42.3%),且通过Siamese BERT的特征映射,AUC(Area Under ROC Curve)仅比常规品种下降5.7%(传统方法下降15.2%)。2)新用户模拟试验。随机屏蔽500名用户的历史行为,仅基于静态属性推荐,前3次交互后推荐准确率提升至72.1%,显著高于基线方法的53.4%。这表明本研究提出的两步推荐机制能有效利用品种语义关联和用户属性特征缓解冷启动问题,尤其在新品种推荐场景表现突出,但新用户冷启动的初始准确率(61.2%)仍存在优化空间,可能与农业用户的地域特异性相关。
LiH J, MengC X, ZhangH W, SongW B, LiuX W. Comparison and implication of crop germplasm resource management in China and the United States based on multi-source data[J]. Journal of China Agricultural University, 2024, 29(10): 161-171 (in Chinese)
[3]
WuF, LyuF, RenJ, YangP, QianK, GaoS J, ZhangY X. Characterizing internet card user portraits for efficient churn prediction model design[J]. IEEE Transactions on Mobile Computing, 2024, 23(2): 1735-1752
[4]
WuH R, LiuC, ZhaoC J. Personalized agricultural knowledge services: A framework for privacy-protected user portraits and efficient recommendation [J]. The Journal of Supercomputing, 2024, 80(5): 6336-6355
[5]
LengH Y, ShaoJ X, ZhangZ, QianY R, MaM N, LiZ C. Research on dual-channel user portrait construction method based on DPCNN-BiGRU and attention mechanism [J]. Journal of Intelligent & Fuzzy Systems, 2023, 45(2): 2579-2591
HuangY M, ZhaoC C. Study on three-level short video user portrait based on improved topic model method [J]. Computer Science, 2024, 51(S1): 698-704 (in Chinese)
WangS Q, LiuZ F, WangJ M. A review of scholar profiling research[J]. Library and Information Service, 2022, 66(20): 73-81 (in Chinese)
[10]
岳怡然. 面向知识服务的用户画像构建与应用研究[D]. 北京: 中国农业科学院, 2021
[11]
YueY R. Research on the construction and application of user profiles oriented to knowledge service[D]. Beijing: Chinese Academy of Agricultural Sciences, 2021 (in Chinese)
ZhaoY L, ChengL J, PangS, XuJ N, LuY J. A review on spatiotemporal data of user profile driven by social sensing[J]. Information Science, 2022, 40(9): 185-192 (in Chinese)
ZhangY J, DongZ, MengX W. Research on personalized advertising recommendation systems and their applications[J]. Chinese Journal of Computers, 2021, 44(3): 531-563 (in Chinese)
ChenZ Y, HuangB. Research on user portrait of improved word vector model[J]. Computer Engineering and Applications, 2020, 56(1): 180-184 (in Chinese)
[20]
SongC X, DongH Y. Application of intelligent recommendation for agricultural information: A systematic literature review[J]. IEEE Access, 2021, 9: 153616-153632
GengJ, LiuC L, WeiX M, ChengM Y, YuanK, LiY, LiuY Z. Product recommendation method based on user repurchase behavior[J]. Journal of Computer Research and Development, 2023, 60(8): 1795-1807 (in Chinese)
ZhaoJ C, SunS F, GuoJ X, ZhongY, WangH B, WangM, QinY. Research on the construction of knowledge graph and recommendation technology of agricultural online learning resources[J]. Computer Applications and Software, 2022, 39(8): 69-75 (in Chinese)
ChenL H, PanX H. Recommending books based on knowledge graph and reader profiling[J]. Data Analysis and Knowledge Discovery, 2023, 7(12): 164-171 (in Chinese)
HouC, NiuP Y. Review of research status and prospects of agricultural knowledge graphs[J]. Transactions of the Chinese Society for Agricultural Machinery, 2024, 55(6): 1-17 (in Chinese)
LiuC, WangC S, MiaoY S, ZhuH J, GuoW, WuH R. An active recommendation method for vegetable production knowledge services[J]. Journal of Hebei Agricultural University, 2024, 47(6): 127-135 (in Chinese)
[35]
GopiP S S, KarthikeyanM. Multimodal machine learning based crop recommendation and yield prediction model[J]. Intelligent Automation & Soft Computing, 2023, 36(1): 313-326
[36]
MusanaseC, VodacekA, HanyurwimfuraD, UwitonzeA, KabandanaI. Data-driven analysis and machine learning-based crop and fertilizer recommendation system for revolutionizing farming practices [J]. Agriculture, 2023, 13(11): 2141