王嘯飛,鮑勝利*,陳炯環(huán)
基于潛在因子模型在子空間上的缺失值注意力聚類算法
王嘯飛1,2,鮑勝利1,2*,陳炯環(huán)1,2
(1.中國科學(xué)院 成都計(jì)算機(jī)應(yīng)用研究所,成都 610041; 2.中國科學(xué)院大學(xué),北京 100049)(?通信作者電子郵箱 baoshengli@casit.com.cn)
針對傳統(tǒng)聚類算法在對缺失樣本進(jìn)行數(shù)據(jù)填充過程中存在樣本相似度難度量且填充數(shù)據(jù)質(zhì)量差的問題,提出一種基于潛在因子模型(LFM)在子空間上的缺失值注意力聚類算法。首先,通過LFM將原始數(shù)據(jù)空間映射到低維子空間,降低樣本的稀疏程度;其次,通過分解原空間得到的特征矩陣構(gòu)建不同特征間的注意力權(quán)重圖,優(yōu)化子空間樣本間的相似度計(jì)算方式,使樣本相似度的計(jì)算更準(zhǔn)確、泛化性更好;最后,為了降低樣本相似度計(jì)算過程中過高的時(shí)間復(fù)雜度,設(shè)計(jì)一種多指針的注意力權(quán)重圖進(jìn)行優(yōu)化。在4個(gè)按比例隨機(jī)缺失的數(shù)據(jù)集上進(jìn)行實(shí)驗(yàn)。在Hand-digits數(shù)據(jù)集上,相較于面向高維特征缺失數(shù)據(jù)的K近鄰插補(bǔ)子空間聚類(KISC)算法,在數(shù)據(jù)缺失比例為10%的情況下,所提算法的聚類準(zhǔn)確度(ACC)提高了2.33個(gè)百分點(diǎn),歸一化互信息(NMI)提高了2.77個(gè)百分點(diǎn),在數(shù)據(jù)缺失比例為20%的情況下,所提算法的ACC提高了0.39個(gè)百分點(diǎn),NMI提高了1.33個(gè)百分點(diǎn),驗(yàn)證了所提算法的有效性。
潛在因子模型;缺失值;注意力機(jī)制;聚類算法;子空間
隨著互聯(lián)網(wǎng)的快速發(fā)展以及企業(yè)數(shù)據(jù)業(yè)務(wù)的大規(guī)模擴(kuò)張,數(shù)據(jù)資源呈現(xiàn)爆炸式增長的趨勢,對數(shù)據(jù)挖掘、信息過濾和搜索的研究得到了廣泛關(guān)注。聚類算法作為數(shù)據(jù)挖掘技術(shù)的一種重要形式,能夠根據(jù)相似度將一組數(shù)據(jù)(通常表現(xiàn)成一組數(shù)據(jù)張量或者多維數(shù)據(jù)空間當(dāng)中的點(diǎn)集)聚類成簇。同一個(gè)簇內(nèi)的數(shù)據(jù)密切相關(guān),不同簇內(nèi)的數(shù)據(jù)存在一定的差異[1]。
聚類算法作為一種無監(jiān)督的機(jī)器學(xué)習(xí)算法,可以在缺少先驗(yàn)信息的條件下,通過計(jì)算樣本間的相似度對數(shù)據(jù)進(jìn)行聚類。目前常見的聚類算法包括K均值(K-means)聚類算法[2]、高斯模型聚類[3]、DBSCAN(Density-Based Spatial Clustering of Applications with Noise)[4]、層次聚類[5]和子空間聚類[6]。目前聚類算法被廣泛應(yīng)用到多個(gè)領(lǐng)域,如生物醫(yī)學(xué)分析[7]、推薦系統(tǒng)[8]、工程優(yōu)化和社交網(wǎng)絡(luò)[9]等。
聚類算法在聚類過程中需要衡量數(shù)據(jù)樣本間的相似程度,盡可能地保證相似樣本分類在一個(gè)簇;但是在現(xiàn)實(shí)世界的數(shù)據(jù)收集過程中,底層數(shù)據(jù)通常不完整,存在部分缺失值樣本。缺失值產(chǎn)生的原因可能有多種:1)人為因素,例如在收集調(diào)查的數(shù)據(jù)集中,調(diào)查參與的人員可能拒絕回答或者忽略某些敏感問題;2)非人為因素,如設(shè)備的故障、不正確的測量、數(shù)據(jù)采集過程中存在限制或數(shù)據(jù)的損壞都會(huì)導(dǎo)致產(chǎn)生缺失數(shù)據(jù)[10]。文獻(xiàn)[11]中闡述了數(shù)據(jù)的缺失是聚類算法過程中常見的一類數(shù)據(jù)質(zhì)量問題,有效缺失值填充可以減少數(shù)據(jù)缺失造成的聚類錯(cuò)誤,提高聚類的準(zhǔn)確度。如果采用不恰當(dāng)?shù)奶畛浞绞?,則數(shù)據(jù)填充只是為了在算法上運(yùn)行缺失數(shù)據(jù),對提高算法準(zhǔn)確度并沒有幫助;如果算法準(zhǔn)確度較低,則數(shù)據(jù)填充對后續(xù)工作也沒有幫助,仍然存在數(shù)據(jù)質(zhì)量問題。對存在缺失值的樣本進(jìn)行聚類的結(jié)果通常不準(zhǔn)確,因?yàn)闃颖鹃g的相似度度量受缺失值的影響導(dǎo)致樣本分類錯(cuò)誤。為了提高聚類算法準(zhǔn)確度,解決數(shù)據(jù)質(zhì)量問題,使用聚類算法對帶有缺失值的數(shù)據(jù)進(jìn)行聚類,準(zhǔn)確衡量缺失數(shù)據(jù)中樣本的相似度就顯得尤為重要。
對于缺失數(shù)據(jù)的處理,比較簡單且常用的方式有直接刪除或者不處理缺失值;但是在聚類過程中必然會(huì)使樣本相似度的計(jì)算準(zhǔn)確率降低,導(dǎo)致聚類的精度降低。目前應(yīng)用較為廣泛的數(shù)據(jù)插補(bǔ)法主要有特殊值替換(均值替換、0替換和最大最小值替換等)、多重插補(bǔ)法、K最近鄰(K-Nearest Neighbors, KNN)插補(bǔ)法、期望最大化(Expectation Maximization, EM)插補(bǔ)法和其他機(jī)器學(xué)習(xí)算法等[12]。
針對數(shù)據(jù)不完整的情況,對缺失數(shù)據(jù)進(jìn)行聚類分析引起了國內(nèi)外學(xué)者的廣泛討論。Albayrak等[13]通過極大似然估計(jì)填充缺失數(shù)據(jù),但需要一部分的完整數(shù)據(jù)進(jìn)行聚類,再對缺失數(shù)據(jù)進(jìn)行處理;如果數(shù)據(jù)的缺失位分布均勻,將很難篩選完整的數(shù)據(jù)進(jìn)行建模。Pattanodom等[14]通過隨機(jī)填充缺失數(shù)據(jù)構(gòu)建不同的聚類模型,再衡量模型的效果以篩選最佳的填充方式;然而該方法構(gòu)建的模型可解釋性很難被評估,且對計(jì)算量的要求過高,如果數(shù)據(jù)規(guī)模較大,將產(chǎn)生大量的子模型。Huayan等[15]通過EM算法結(jié)合K-means算法處理缺失數(shù)據(jù),但是EM算法存在收斂慢和對初始化依賴較嚴(yán)重的缺點(diǎn),所以存在較大的不穩(wěn)定性。喬永堅(jiān)等[16]提出了一種面向高維特征缺失數(shù)據(jù)的KNN插補(bǔ)子空間聚類(KNN Interpolation Subspace Clustering, KISC)算法,利用數(shù)據(jù)子空間的近鄰關(guān)系填充數(shù)據(jù);然而在樣本相似度計(jì)算上僅計(jì)算了子空間中系數(shù)矩陣的歐氏距離進(jìn)行填充,并未考慮基數(shù)矩陣的關(guān)聯(lián)關(guān)系。王一棠等[17]采用了模糊聚類和非線性回歸插補(bǔ)的算法填充缺失數(shù)據(jù),針對盾構(gòu)機(jī)數(shù)據(jù)的效果良好;但是應(yīng)用具備一定的領(lǐng)域限制,且實(shí)現(xiàn)較復(fù)雜。
針對以上算法存在的問題,本文提出了一種基于潛在因子模型(Latent Factor Model, LFM)在子空間上的缺失值注意力聚類算法。LFM結(jié)合矩陣分解可以很好地處理數(shù)據(jù)稀疏問題[18]。該算法主要分為以下幾個(gè)部分:1)針對缺失數(shù)據(jù)樣本間相似度難以度量的問題,通過LFM將原始數(shù)據(jù)空間映射到子空間,緩解樣本的稀疏性,且子空間包含樣本的結(jié)構(gòu)特征信息,使得樣本相似度得以計(jì)算且具備一定可解釋性。2)在樣本相似度計(jì)算方式上引入特征矩陣進(jìn)行權(quán)重注意力分配,使得樣本相似度的度量方式更精確,插補(bǔ)的泛化性更好。3)為了緩解在特征點(diǎn)對點(diǎn)相似度計(jì)算方式上高額的時(shí)間復(fù)雜度問題,改進(jìn)了注意力權(quán)重矩陣的結(jié)構(gòu),使時(shí)間復(fù)雜度達(dá)到了指數(shù)級別的優(yōu)化。
表1符號和定義
Tab.1 Symbols and definitions
圖1 本文算法的框架
1.3.1求解方法
1.3.2優(yōu)化方法
1.4.1特征相似度權(quán)重圖
通過計(jì)算不同特征的余弦相似度可以得到特征間的注意力權(quán)重值,不同特征的注意力權(quán)重可以構(gòu)建一個(gè)樣本注意力權(quán)重矩陣,它的結(jié)構(gòu)如圖2所示。
圖2 樣本注意力權(quán)重矩陣的結(jié)構(gòu)
1.4.2樣本注意力相似度計(jì)算方法
圖3 樣本相似度計(jì)算方式
如果相似集對應(yīng)位置存在缺失,則計(jì)算公式如式(13)所示:
算法1 相似樣本集算法。
圖4 多指針相似度權(quán)重圖
實(shí)驗(yàn)通過4個(gè)數(shù)據(jù)集驗(yàn)證本文算法的有效性,包括兩個(gè)圖像分類數(shù)據(jù)集和兩個(gè)小樣本標(biāo)準(zhǔn)分類數(shù)據(jù)集,數(shù)據(jù)集信息如表2所示。
表2 數(shù)據(jù)集信息
其中Hand-digits和COIL20為兩個(gè)被廣泛使用的圖像分類數(shù)據(jù)集,分別為手寫數(shù)字的圖像數(shù)據(jù)集和多個(gè)物體不同角度拍攝的圖像數(shù)據(jù)集[16]。Breast-cancer和Wine為Sklearn開源機(jī)器學(xué)習(xí)工具庫中的標(biāo)準(zhǔn)分類數(shù)據(jù)集。
實(shí)驗(yàn)在Ubuntu18.04,CPU 2.90 GHz 16 GB內(nèi)存 GPU RTX3060 12 GB顯存上進(jìn)行,其中深度學(xué)習(xí)框架采用Torch,對數(shù)據(jù)的預(yù)處理和分析采用numpy和pandas進(jìn)行。
同時(shí)為了驗(yàn)證多指針相似度權(quán)重圖對算法時(shí)間復(fù)雜度的提升,通過計(jì)算相似度算法每一個(gè)epoch數(shù)據(jù)插補(bǔ)的運(yùn)行時(shí)間進(jìn)行衡量,最終結(jié)果取10次運(yùn)行時(shí)間的平均值。
為了充分驗(yàn)證算法的有效性,實(shí)驗(yàn)選用聚類準(zhǔn)確度(Accuracy,ACC)和歸一化互信息(Normalized Mutual Information, NMI)作為評測指標(biāo),其中:ACC可以直接衡量聚類的準(zhǔn)確度,是常用的聚類度量指標(biāo);相較于ACC,NMI通過比較兩個(gè)聚類結(jié)果的相似程度,它的值域是[0,1],且NMI值不受簇類標(biāo)簽排列的影響[22]。NMI的計(jì)算公式如式(16)所示:
表3展示了不同數(shù)據(jù)集分別在ACC和NMI指標(biāo)上的實(shí)驗(yàn)結(jié)果。
表3不同算法聚類結(jié)果的ACC和NMI比較 單位:%
Tab.3 Comparison of ACC and NMI in clustering results between different algorithms unit:%
由表3可以發(fā)現(xiàn):1)在Hand-digits數(shù)據(jù)集上,當(dāng)缺失比例為10%~20%時(shí),本文算法在子空間上的ACC和NMI均高于KISC算法和特殊值填充的K-means算法;相較于次優(yōu)的KISC算法,在數(shù)據(jù)缺失比例為10%的情況下,本文算法的ACC提高了2.33個(gè)百分點(diǎn),NMI提高了2.77個(gè)百分點(diǎn),在數(shù)據(jù)缺失比例為20%的情況下,本文算法的ACC提高了0.39個(gè)百分點(diǎn),NMI提高了1.33個(gè)百分點(diǎn);當(dāng)缺失比例為30%~40%時(shí),本文算法在原始空間的聚類效果優(yōu)于本文算法在子空間的聚類效果,說明當(dāng)缺失比例過大時(shí),子空間的特征結(jié)構(gòu)被損壞,聚類精度降低;同時(shí),在該缺失比例下本文算法在原始空間上的聚類效果優(yōu)于KISC算法和特殊值填充的K-means算法在原始空間的聚類效果,說明本文算法采用的子空間特征注意力的缺失值填充算法是有效的。2)在COIL2數(shù)據(jù)集上,本文算法在子空間上的ACC值均高于KISC算法和特殊值填充的K-means算法,說明了本文算法的先進(jìn)性;但是NMI指標(biāo)在缺失比例為30%時(shí),本文算法相較于KISC算法NMI指標(biāo)精度有所下降,在缺失比例為40%時(shí),本文算法在原始空間上的NMI指標(biāo)最優(yōu),可以發(fā)現(xiàn)當(dāng)缺失比例為30%時(shí)本文算法在原始空間的NMI指標(biāo)和KISC算法在子空間的NMI指標(biāo)差距較小,進(jìn)一步說明了子空間特征注意力的缺失值填充算法的有效性。3)在Breast-cancer和Wine數(shù)據(jù)集上,本文算法的ACC值均高于KISC算法和特殊值填充的K-means算法,只有當(dāng)Wine數(shù)據(jù)集缺失比例為40%時(shí),本文算法的NMI值低于0填充的K-means算法。進(jìn)一步發(fā)現(xiàn),當(dāng)缺失比例擴(kuò)大時(shí),本文算法和其他算法的差距逐漸縮小,如在Wine數(shù)據(jù)集上,當(dāng)缺失比例位于10%~30%,本文算法的ACC和NMI指標(biāo)均為最優(yōu);但當(dāng)缺失比例達(dá)到40%,本文算法的NMI指標(biāo)相較于K-means算法低了1.39個(gè)百分點(diǎn),說明在缺失比例過大的情況下,子空間對特征的整體概括能力有所下降。所以本文算法在缺失比例位于10%~30%區(qū)間具備明顯優(yōu)勢。
本文針對傳統(tǒng)聚類算法在對缺失樣本進(jìn)行數(shù)據(jù)填充過程中存在樣本相似度難度量且填充數(shù)據(jù)質(zhì)量差的問題,提出了一種基于LFM在子空間上的缺失值注意力聚類算法。該算法通過LFM解決了因?yàn)闃颖鞠∈栊詫?dǎo)致相似距離無法度量、精度較低的問題;同時(shí)在樣本相似度的計(jì)算上考慮到子空間特征矩陣之間的相關(guān)性,并將它引入相似度計(jì)算方式,提高了樣本相似度計(jì)算方式的準(zhǔn)確度和泛化性。最后在多個(gè)數(shù)據(jù)集上通過ACC和NMI指標(biāo)驗(yàn)證了算法的有效性和可靠性。同時(shí)實(shí)驗(yàn)中發(fā)現(xiàn):當(dāng)數(shù)據(jù)缺失比例增大時(shí),子空間對特征信息的概括能力有所降低;當(dāng)數(shù)據(jù)缺失比例過大時(shí),當(dāng)前算法的聚類精度會(huì)有所下降。如何在數(shù)據(jù)缺失比例過大的情況下有效提高子空間下數(shù)據(jù)聚類算法的精度,需要進(jìn)一步研究。
[1] AHALYA G,PANDEY H M. Data clustering approaches survey and analysis [C]// Proceedings of the 2015 International Conference on Futuristic Trends on Computational Analysis and Knowledge Management. Piscataway: IEEE,2015: 532-537.
[2] XU H, YAO S, LI Q, et al. An improved K-means clustering algorithm[C]// Proceedings of the 2020 IEEE 5th International Symposium on Smart and Wireless Systems within the Conferences on Intelligent Data Acquisition and Advanced Computing Systems. Piscataway: IEEE, 2020: 1-5.
[3] 王一妹,劉輝,宋鵬,等.基于高斯混合模型聚類的風(fēng)電場短期功率預(yù)測方法[J].電力系統(tǒng)自動(dòng)化,2021,45(7):37-43.(WANG Y M, LIU H, SONG P, et al. Short-term power forecasting method of wind farm based on Gaussian mixture model clustering [J]. Power System Automation, 2021,45(7): 37-43.)
[4] JEBARI S, SMITI A, LOUATI A.AF-DBSCAN: an unsupervised automatic fuzzy clustering method based on DBSCAN approach[C]// Proceedings of the 2019 IEEE International Work Conference on Bioinspired Intelligence. Piscataway: IEEE,2019:1-6.
[5] RONG Y, LIU Y.Staged text clustering algorithm based on K-means and hierarchical agglomeration clustering[C]// Proceedings of the 2020 IEEE International Conference on Artificial Intelligence and Computer Applications. Piscataway: IEEE,2020:124-127.
[6] 任麗娜,秦永彬,黃瑞章,等.基于多層子空間語義融合的深度文本聚類[J].計(jì)算機(jī)應(yīng)用研究, 2023,40(1):70-74,79.(REN L N, QIN Y B, HUANG R Z,et al. Deep document clustering model via multi-layer subspace semantic fusion [J]. Application Research of Computers, 2023,40(1):70-74,79.)
[7] TAN L, GU S, WU C, et al. K-means clustering method based on node similarity in traditional Chinese medicine efficacy[C]// Proceedings of the 2020 39th Chinese Control Conference. Piscataway: IEEE,2020:742-747.
[8] YI Y. Design of intelligent recommendation APP for eco-tourism routes based on popular data clustering of points of interest[C]// Proceedings of the 2021 2nd International Conference on Smart Electronics and Communication. Piscataway: IEEE,2021: 1270-1273.
[9] SHRUTHI S, JOSE A M, ANUROOP P R. Improvisation of cluster efficiency using min-cut algorithm in social networks[C]// Proceedings of the 2017 International Conference on Communication and Signal Processing. Piscataway: IEEE,2017:1641-1644.
[10] LIU Q,HAUSWIRTH M. A provenance meta learning framework for missing data handling methods selection[C]// Proceedings of the 2020 11th IEEE Annual Ubiquitous Computing, Electronics & Mobile Communication Conference. Piscataway: IEEE,2020: 349-358.
[11] 徐宇明,陳誠,熊赟,等.APT-KNN:一種面向分類問題的高效缺失值填充算法[J].計(jì)算機(jī)應(yīng)用與軟件,2011,28(4):135-139.(XU Y M, CHEN C, XIONG Y,et al. APT-KNN: an efficient missing value imputation method oriented toward classification issue [J]. Computer Applications and Software, 2011, 28 (4): 135-139.)
[12] 徐鴻艷,孫云山,秦琦琳,等.缺失數(shù)據(jù)插補(bǔ)方法性能比較分析[J].軟件工程,2021,24(11):11-14.(XU H Y, SUN Y S, QIN Q L, et al. Comparative analysis of the performance of interpolation methods for of missing data[J]. Software Engineering, 2021,24 (11): 11-14.)
[13] ALBAYRAK M, TURHAN K, KURT B. A missing data imputation approach using clustering and maximum likelihood estimation[C]// Proceedings of the 2017 Medical Technologies National Congress. Piscataway:IEEE, 2017:1-4.
[14] PATTANODOM, IAM-ON N, BOONGOEN T. Clustering data with the presence of missing values by ensemble approach[C]// Proceedings of the 2016 Second Asian Conference on Defence Technology. Piscataway: IEEE, 2016: 151-156.
[15] HUAYAN S, YELI L, RUNFEI Z, et al. Accelerating EM missing data filling algorithm based on the K-means[C]// Proceedings of the 2018 4th Annual International Conference on Network and Information Systems for Computers. Piscataway: IEEE, 2018:401-406.
[16] 喬永堅(jiān),劉曉琳,白亮.面向高維特征缺失數(shù)據(jù)的K最近鄰插補(bǔ)子空間聚類算法[J].計(jì)算機(jī)應(yīng)用,2022,42(11):3322-3329.(QIAO Y J, LIU X L, BAI L. K-nearest neighbor interpolation subspace clustering algorithm for high-dimensional data with feature missing [J]. Journal of Computer Application, 2022,42(11): 3322-3329.)
[17] 王一棠,龐勇,張立勇,等.面向盾構(gòu)機(jī)不完整數(shù)據(jù)的模糊聚類與非線性回歸填補(bǔ)[J].機(jī)械工程學(xué)報(bào), 2023, 59(12): 28-37.(WANG Y T, PANG Y, ZHANG L Y ,et al. Fuzzy clustering and nonlinear regression filling for incomplete data of shield machine [J]. Journal of Mechanical Engineering, 2023, 59(12): 28-37.)
[18] 陳曄,劉志強(qiáng).基于LFM矩陣分解的推薦算法優(yōu)化研究[J].計(jì)算機(jī)工程與應(yīng)用,2019,55(2):116-120.(CHEN Y, LIU Z Q. Research on improved recommendation algorithm based on LFM matrix factorization [J]. Computer Engineering and Applications, 2019,55(2): 116-120.)
[19] XIONG Y, LI H. Collaborative filtering algorithm in pictures recommendation based on SVD[C]// Proceedings of the 2018 International Conference on Robots & Intelligent System. Piscataway: IEEE, 2018:262-265.
[20] 楊鏌. 基于偏置LFM和LSH的混合電影推薦算法研究[D].武漢:華中師范大學(xué),2021:13-15.(YANG M. Research on hybrid movie recommendation algorithm based on bias LFM and LSH[D]. Wuhan: Central China Normal University, 2021: 13-15.)
[21] GUO S, LI C. Hybrid recommendation algorithm based on user behavior[C]// Proceedings of the 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference. Piscataway: IEEE, 2020:2242-2246.
[22] 龍建武,王強(qiáng).反向近鄰構(gòu)造連通圖的聚類算法[J/OL].計(jì)算機(jī)科學(xué)與探索:1-15 [2023-02-07]. http://kns.cnki.net/kcms/detail/11.5602.TP.20220930.1744.006.html.(LONG J W, WANG Q. Clustering algorithm for constructing connected graphs by reverse nearest neighbors [J/OL]. Journal of Frontiers of Computer Science and Technology: 1-15 [2023-02-07]. http://kns.cnki.net/kcms/detail/11.5602.TP.20220930.1744.006.html.)
Missing value attention clustering algorithm based on latent factor model in subspace
WANG Xiaofei1,2, BAO Shengli1,2*, CHEN Jionghuan1,2
(1,,610041,;2,100049,)
To solve the problems that traditional clustering algorithms are difficult to measure the sample similarity and have poor quality of filled data in the process of filling missing samples, a missing value attention clustering algorithm based on Latent Factor Model (LFM) in subspace was proposed. First, LFM was used to map the original data space to a low dimensional subspace to reduce the sparsity of samples. Then, the attention weight graph between different features was constructed by decomposing the feature matrix obtained from the original space, and the similarity calculation method between subspace samples was optimized to make the calculation of sample similarity more accurate and more generalized. Finally, to reduce the high time complexity in the process of sample similarity calculation, a multi-pointer attention weight graph was designed for optimization. The algorithm was tested on four proportional random missing datasets. On the Hand-digits dataset, compared with the KISC (K-nearest neighbors Interpolation Subspace Clustering) algorithm for high-dimensional feature missing data, when the missing data was 10%, the Accuracy (ACC) of the proposed algorithm was improved by 2.33 percentage points and the Normalized Mutual Information (NMI) was improved by 2.77 percentage points; when the missing data was 20%, the ACC of the proposed algorithm was improved by 0.39 percentage points, and the NMI was improved by 1.33 percentage points, which verified the effectiveness of the proposed algorithm.
Latent Factor Model (LFM); missing value; attention mechanism; clustering algorithm; subspace
This work is partially supported by Western Young Scholars Project of Chinese Academy of Sciences (RRJZ2021003).
WANG Xiaofei, born in 1997, M. S. candidate. His research interests include machine learning, recommendation algorithm.
BAO Shengli, born in 1973, Ph. D., research fellow, His research interests include software engineering, big data.
CHEN Jionghuan, born in 1998, M. S. candidate. His research interests include machine learning, big data.
TP391.1
A
1001-9081(2023)12-3772-07
10.11772/j.issn.1001-9081.2022121838
2022?12?12;
2023?02?13;
2023?02?16。
中國科學(xué)院西部青年學(xué)者項(xiàng)目(RRJZ2021003)。
王嘯飛(1997—),男,湖南慈利人,碩士研究生,主要研究方向:機(jī)器學(xué)習(xí)、推薦算法;鮑勝利(1973—),男,安徽黃山人,研究員,博士,主要研究方向:軟件工程、大數(shù)據(jù);陳炯環(huán)(1998—),男,山東濰坊人,碩士研究生,主要研究方向:機(jī)器學(xué)習(xí)、大數(shù)據(jù)。