基於超連結圖譜表示法學習之跨領域音樂推薦演算法 - 政大學術集成

全文

(1)˙ºÖ#P \h:’x“KË⇠flÛ⇥®¶ó’ Cross-domain Music Recommendation Based on Superhighway Graph Embedding  v ⇢J⌥≥ ⌥ Yà⇢!ò. Student⇢Sheng-Fang Yang Advisor⇢Ming-Feng Tsai. ↵À?ª'x «⌦—x˚. 立. 治政 ©Î÷á 大. ‧ 國. 學 ‧. A Thesis submitted to Department of Computer Science. n. er. io. sit. y. Nat. National Chengchi University in partial fulfillment of the Requirements a l for the degree of i v. n C h Master U i engch in. Computer Science. -Ô⌘↵ ~ˆk t ] September 2019. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(2) Ù. ®WÍÒû!0 å⇣ «ÍÒÑ÷á ⌘ à'Ñ⇣1 ⇥fi ñ⇡itÜEÊÑx“⌥⌅.ìW ø∫ó⇡µBìà<ó Í/ ( m⇠fl⌦ Ü. öÑ˙. çX. _. ÍÒ*Ü1⇢Ô˝. ⇢Ü. ⌘Ñ⌥ 治+!ò + û'xB„ 政大 Ó“ +Ñ↵✏- ≤↵ãÀ ìN tÑNÊ◆Ù⌘`ö⌘Ñ↵ 立 ✏˝õ˙ _ì⌘(⌘v@Ñ≤m⌦ÙJ⇤ ⇠⇥ (v@Ñ dÑä·⇥ñH⌘ÅH. m⇠flv⌦bÑ˙p. ‧ 國. +B8⇤fà⌘(. ⇤ ì⌘ÔÂ2eKKBOX"x. \. k·b. 學. Ø-. _fà⌘_. Êõ⌥mL•Ã&¿fl. ÷⌦. y. õ∞ÑÛ’⇥çÜ. (œ!Ñ✏D. ⌘Å. ÊW§Ñ%4⌘. al. n. xπbÑ. ≥K. ‘flá”. _ì8˝ /∞. ’›(⌘ dÜ‹≥⌘(x. er. io. 6⌘2e⇡Ñ∞Ñ⇠fl _iB0fà⌘ õ˙p v@Ñ Ø-⌥⌘í¯%* q æ2 ª˝. sit. fà⌘. _⇤Êõ˙ªãã b. Nat. áì⌘_. ‧. mL⌦ º«ôÑU⌃π✏⌥>⇢ÊŸ⌦ÑìWI ⌘◊ o⇢ Ê _Å -bÑãÁ9 + 6⇠⌘⌘D⇣eå⇣ ⇤Ñ÷. Ch. ( ;⌦. Íí¯¿ı. e n g c hÑi. fi¥Ü⌘(v@B. i Un. v. B_í„⌘à⇢iÒ ;. ⌘Ñ. ;⇢Ü. õ⇥£ ◊ xw ÷=/˝(⌘ Åk©ÑB⇡fà⌘à'Ñ kŸ _–õÜ⌘à⇢ (Ñ«⌦⌥ÂX ‘xw dÜ6⇠ ⌘(KK H⌦ÑÂ\⌦ÃSK Vëxw. _fÜ⌘à⇢UãÑìW⌥˙p. °/(KK H⌦. õ∞Ñπ⌘ ˙p ã ˝h√ïe⇥ ⌘(*Ü∫ ÑÔ⌦. ÑÅ Ü÷⌘. /(. ⌘Ñ∂∫⌘ Ô⌦Ñj4. ˝ |Ó@w. ⇤÷á⌦. ˝˝f⌘. ì⌘ (‘√ ;⌦Ñc M Ü˛(Ñ⌘ _ 1. ⇣>⇤2eÑ. ˝P⇥ J⌥≥. ↵À?ª'x«⌦—x˚ July 2019. 1. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(3) ˙ºÖ#P. \h:’x“KË⇠flÛ⇥®¶ó’. -áXÅ —tÜ'x⁄Â _hx“ÄSÑÏ√|U ®¶˚q´„€… (º⌅.ÊŸ⌦ Û⇥2A˚q-ÑÛ⇥®¶_ä⇣ ⇧w ⌘0 'ÑÂ\ $v(⌅↵ ⇥4- §‘ÑF}“c_⇤ @ ⇥ ‡d ⌘⌘(ÜpÍ'≤Ôh:’x“ Heterogeneous Information Network Embedding ÔÂ⌥≤Ô^ãK¿fiïqºN≠¶⌘ œzì- &˙ºdzìÜå⇣åå¯‹KÛ⇥®¶Â\⇥»‡ º ∞ã|⇥4 (6⌥LÚF} ⌅Ií’Ñ«⌦u∫ ⌘⇤‡⌘x (6 qˇt‘®¶Ñæ⌘ ⇡ø1∫«ôÑ è' OL « ôÑ è'⇢8/ÊŸ⌦ ↵àw ⌘0'Ñ˚Ÿ v º®¶˚q t‘Ñ®¶Húqˇ/àË'Ñ⇥º/ ,÷á–˙Ü ↵˙ºpÍ '≤Ôh:’x“ÑÛ⇥®¶˚q ✏N†e≤Ô«⌦⇤∫PÃÑ⇥ 4\∫©Ük©92∞ã|⇥4K®¶Hú⇥. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. 2. i Un. v. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(4) Cross-domain Music Recommendation Based on Superhighway Graph Embedding. Abstract In recent years, big data and machine learning technology have been rapidly growing, and recommendation systems have been widely used in various real-world applications, such as music recommendation in music streaming services. However, for different domains, the recommneder systems will be different, because of the distinct user behavior data. Therefore, this thesis aims to use Heterogeneous Information Network Embedding to project the nodes in a network/domain into another network/domain on the basis of the low-dimension representations of the nodes. Therefore, this paper proposes a cross-domain music recommendation approach based on heterogeneous information network representation learning, the idea of which is to enrich the new domain/market data by using a well developed domain/market.. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. 3. i Un. v. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(5) Ó⌅ Ù. 1. -áXÅ. 2. Abstract , ‡ “÷ . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 M . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 vÓÑ . . . . . . . . . . . . . . . . . . . . . . . . ,å‡ ¯‹á{¢ . . . . . . . . . . . . . . . . . . . . . . . 2.1 ≤Ôh:’x“ . . . . . . . . . . . . . . . . . . . . 2.2 ®¶˚q . . . . . . . . . . . . . . . . . . . . . . . . 2.3 w˚✏x“ . . . . . . . . . . . . . . . . . . . . . . . , ‡ vπ’. . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 OLö© . . . . . . . . . . . . . . . . . . . . . . . . 3.2 pÍ'≤Ô˙ . . . . . . . . . . . . . . . . . . . . 3.3 ˙ÀÖ#P \ . . . . . . . . . . . . . . . . . . . . 3.4 ≤Ôh:’x“ . . . . . . . . . . . . . . . . . . . . 3.4.1 Deepwalk . . . . . . . . . . . . . . . . . . . . 3.4.2 Large-Scale Information Network Embedding . 3.4.3 Heterogeneous Preference Embedding . . . . . 3.5 ®¶˚q . . . . . . . . . . . . . . . . . . . . . . . . ,€‡ ÊWPú⌥ ÷ . . . . . . . . . . . . . . . . . . . . . 4.1 «ô∆ . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 ÊW-ö . . . . . . . . . . . . . . . . . . . . . . . . 4.3 U0⇡ñ . . . . . . . . . . . . . . . . . . . . . . . . 4.4 ÊWPú⌃ê⌥ ÷ . . . . . . . . . . . . . . . . . 4.4.1 ñ∫áh˛ . . . . . . . . . . . . . . . . . . 4.4.2 Ïfiáh˛ . . . . . . . . . . . . . . . . . . 4.4.3 sGñ∫áG<h˛ . . . . . . . . . . . . . 4.4.4 ∞N¶h˛ . . . . . . . . . . . . . . . . . . 4.5 Êã⌃ê . . . . . . . . . . . . . . . . . . . . . . . . 4.5.1 Êã⌃ê-Â®¶˚q∫ã . . . . . . . . . . 4.5.2 Êã⌃ê-Â≤Ôh:’x“∫ã . . . . . . ,î‡ P÷ . . . . . . . . . . . . . . . . . . . . . . . . . . . . √⇤á{ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 立. 政治大. n. al. er. io. sit. Nat. y. ‧. ‧ 國. 學. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. Ch. engchi. 4. i Un. v. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 3 1 1 2 4 4 5 6 8 8 8 10 12 12 12 14 14 16 16 17 19 21 21 21 22 23 24 24 24 29 30. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(6) Ó⌅ Û⇥2A ŸKpÍ'≤Ô . . Ó⇡⌥Üê⇥4ÑpÍ'≤Ôå⌃ Ö#P \K#•:✏ . . . . . LINEÑ¯<¶≤Ô . . . . . . .. 立. . . . . PÀ . . . . . . . .. . . . .. . . . .. . . . .. . . . .. . . . .. . . . .. . . . .. . . . .. . . . .. . . . .. . . . .. . . . .. . . . .. . . . .. 9 9 11 13. 政治大. 學 ‧. ‧ 國 io. sit. y. Nat. n. al. er. 3.1 3.2 3.3 3.4. Ch. engchi. 5. i Un. v. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(7) hÓ⌅ A B⇥4«ô∆x⁄q . . . . . . . . . ◆Ù∆-q LÚT◆Ù∆KLÚÕäá ñ∫áh˛ . . . . . . . . . . . . . . . . . Ïfiáh˛ . . . . . . . . . . . . . . . . . sGñ∫áG<h˛ . . . . . . . . . . . Top-N ∞N¶h˛ . . . . . . . . . . . . . ®¶˚qKÊã . . . . . . . . . . . . . . ≤Ôh:’x“Êã K . . . . . . . . ≤Ôh:’x“Êã Kå . . . . . . . .. 立. . . . . . . . . .. 政治大. . . . . . . . . .. . . . . . . . . .. . . . . . . . . .. . . . . . . . . .. . . . . . . . . .. . . . . . . . . .. . . . . . . . . .. . . . . . . . . .. . . . . . . . . .. . . . . . . . . .. . . . . . . . . .. . . . . . . . . .. . . . . . . . . .. 16 17 21 22 22 23 25 26 27. 學 ‧. ‧ 國 io. sit. y. Nat. n. al. er. h 4.1 h 4.2 h 4.3 h 4.4 h 4.5 h 4.6 h 4.7 h 4.8 h 4.9. Ch. engchi. 6. i Un. v. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(8) ,. ‡. “÷ 1.1. M. 政治大. 立. Ü. ÜÑ. (⇧Ô)(Û⇥2A ⌅_. ∑0". Recommondation system. ¯M. ∞^. al. (º⇣,(6. i¡Ñ. „€K(º⌅^F¡⌥. n. F¡⇢Û⇣Û. Ch. er. —tÜ‡∫≤ÔÑ|U. io. }. ^ã⇢T N˛ Recommondation. ®¶˚q. y. Nat. .⌦oN˛˚q. Û⇥. 'œÑÛ. ⇥ÇU)(⇡õ(⇧⌥Û⇥Ñ«⌦Ü9Ñ. ŸÑ¡Í1ä⇣ ⇧^8ÕÅÑpL »(v-nWÕÅÑ“r⇥ ®¶˚q/. ®K. Ÿ-. iv n ˙ºgπ®¶ U. Ñ(6§Ñ(L∫Ü®¶. O. œ/˚q. —çF¡I⇥®¶˚q;ÅÔÂ⌃⇣. Collaborative Filtering [24] [19]⌥˜ ✏®¶ Hybrid Recommondation. π’9⁄(6ÑwÚL∫ãÇ¿↵. U⌃. sit. ⇥«⌦⌥(⇧. ŸÜF}⌅✏⌅#ÑÛ⇥. ‧. ‧ 國. Û⇥_880 Ÿ_Ï√|Uw. 學. ®W⌅✏—ÄÂ∞ pÑ|U ∫⌘Ñ⌅.L∫_( ∑I€ ûÊ‘FóI⌘≤Ô⌦1Ô ⌃÷ó F} B Û⇥2A. engchi F}. fi . U˘IL∫. .. Content-based [4]⇥T N˛ ç)(¯<ú}. ˙ºgπ®¶;Å(i¡Ñ‚cyµãÇ^%. l'IÜ®¶^<'ÍÑ¯<i¡ ˜ T⇣ÙsÑ®¶Hú⇥(≤Ô ŸÎ. ✏®¶G/⌥⌦i.π’ç wÜÂ |UÑ˛ 'œÑi¡ (⇧ ⌅. ˝´®¶˚q@(⇥÷6⇡õ. PÃÑ⌦o. Ñ⇣,u∫ê'. ÇU(⇤œ. ⌅+. F(l‘ÑP6↵. óÑBì⌥zì⌥‹¶⇣,↵)(⇡õ«⌦. Kó /. ↵¯vÃ ⌘0'Ñ˚Ÿ⇥ h:’x“. Embedding. [2]/ÓMà±ÄÑÄS. (˛ ∫Âzg⇠fl. -1⇢ÑOL˝„€Ñ´(⇥h:’x“ÑZ’;Å/⌥ v0—«⌦”.⇣. ↵⌘œÑh:b✏. (Í6û. U⌃. ↵i¡Ñl' Natural Language. Processing ⇠fl· œ/ word2vec [17] É✏N^^ì≤Ôx“ÂP·œ↵^ &⌥KI ⇣⌘œÑh:✏ ÍøºZååÑ…( _w à'ÑH'⇥≤. 1. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(9) Network Embedding. Ôh:’x“. G/>§≤Ô⇠fl-˙ºh:’x“Ñ…. (⇥ãÇ⌘⌘ÔÂ⌥(6⌥(6KìÑ§í‹¬˙⇣ 5 ⌥(6¿fi v0 EÑ‹¬ ^‘⇣^<ºÂP-^⌥^Ñ‹¬&(h:’x“ÄS⌥(6¿fi 1⇢˙º word2vec Ñ≤. ïqÛ⇤N≠¶Ñ⌘œzì-ÂøZååÑ…(⇥ÓM Ôh:✏Ñπ’œ/ DeepWalk [20]. LINE [25]I˝´„€0K((«ô¢ÿ. >§≤Ô⌃êI⇠fl⌦⇥ ≤Ôh:’x“Ñπ’'⇢…(º Network. -. '⇢ÊŸ⌦Ñ«ô&. geneous Information Network Ñ œxS>§Â œ/ ↵∫ÔÂ «⌦Ôõ(. Û⇥®¶˚q_. ‡d. /pÍ'≤Ô. .⇧+i.Â⌦l'¿fi⌥#P. (6. x! \⇧. vÆMI⇢.¿fi «÷áÔÂ (⇢«. LÚ. LK. ©ºååÑ®¶⇥. 立. Ù/˛Ê«ô-. ↵à¥ÕÑOL. è'/⌥(6⌥i. Ù!ã(®¶B⇤æ⌘º®¶£õÚì. y. ®¶£õfl(6ú}¯<Fí´¿fl0Ñi¡. ⌅Ñi¡. ⇤. sit. ⇤. ^ã. 政治(!ã-ÔÂ±+Ù⇢ÙPÃÑ« 大. Nat. ¡ìí’N⌘. I. ‧. Spasity. û. ⌥K˙⇣pÍ'≤Ô. vÓÑ. è'. Hetero-. 學. 1.2. Í'≤Ô. ⇥pÍ'≤Ô/. ‧ 國. ⌦ _. /. Ô˝ \⇧ ÷á ⇠fl ↵x! '«÷á ⇢. ÷áI⌅.§í‹¬. Homogeneous Information. Í'≤Ô. ⇡B. ⌘⌘ÔÂ)(¯<^ãF. er. io. (6 ⌅PÃÑ⇠flÜk©&û7⇤1⇠flÑ®¶Hú ⇡↵K(Üw˚x“ Transfer Learning π’Ñ®¶˚q ⌘⌘1K∫Ë⇠fl®¶ Cross-domain. al. n. iv n C Recommendation ⇥Ê ⌘⌘_ÔÂ)(M h e n g c hÑi UË«⌦ Üû7®¶˚qÑHú⇥ (Û⇥2A. Ÿ-. (6⌥LÚxœu∫ê'. External Information. 9⁄q. ↵(6sG. F}LÚœ Nx~ÛxCñ ⇡h:Û⇥2A Ÿ-‹º(6F} ⌅Ñ«ô ⇤X(Wà'Ñ è'OL (6F} ⌅ÄÄ⇢⇤O⌘º∆-ºF}±ÄÑ LÚ. ≥q(Èc⌃„. N®¶ÑHú ∫¶. Matrix Factorization. v(πG9§Ó. ÑZ’⇤‡. Mean Square Error. è'OL. M. ÑU⌃⇡ñÍ(Nñ. F(Û⇥®¶⌦ ⌘⌘& (6´∑P(£õ÷Úì}NÑL / ˝”ULÚÑ„¶Ñ B_˝®¶fi(6@ú}ÑÛ⇥⇥‡d ÇU(–G. LÚ„¶». Ûº*|c˝®fi(⇧. £ÑÛ⇥. ø/Û⇥®¶˚q-. ↵ÕÅÑpL⇥ ,÷áË(˙À ↵Û⇥®¶˚q ✏N ↵ H'Ñ˙ Ve )(+ ⇤PÃK(6F} ⌅Ñ A ⇥4 (ÂPÃ B ⇥4K«⌦&⌥K˙⇣pÍ'≤Ô &)(≤Ôh:’x“. ⌥(6⌥LÚïqÛN≠¶Ñ⌘œzì2. ç✏N. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(10) ˙ºÂb®¶. Query-based Recommendation. ¶Pú⇥dπ’(º KKBOX. 立. Ñπ✏. ÊÑ«ô∆-. T⇣(6KºLÚÑ®. &˝®¶ÑHú. @–G⇥. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. 3. i Un. v. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(11) ,å‡ ¯‹á{¢ 2.1. ≤Ôh:’x“. 立. Word Embedding. ‧ 國. (˛ Í6û. U⌃I⇠fl-Ú⇣∫. 學. ^⌘œ. 政治大 .8(Ñy. ⌘œÑ≠¶⇤äóuÿà. ‡d^⌘œÑÓÑ. ø/✏N^^ì≤Ôx“. è. (. ó⌦⇤A⌃0„. &⌥^ïqÛN≠¶Ñ#å⌘œ. y. Nat. 'œÑá,«ô-. ‧. µ<&´„€Ñ(⇥QºÂM≥q!ãh:’⇢∫ One-hot h:’ One-hot Representation / TF-IDF Term Frequency-Inverse Document Frequency (. sit. er. ÷ó≤Ô¿fiÑ⌘œh:’\∫d¿fiÑyµ<&(ºååÙ⇢Ñ…(. al. n. PÀ. io. zì- ⌘œäó ∆ ©º¿ ó⌦Ñzì⌥Bì⇥ ≤Ôh:’x“ Network Embedding GûdÇıˆ8 ⌥≤Ô )(≤Ô¿fiì§í#PÑ -⇥. Ch. engchi. Í û word2vec – ˙ ^ ⌘ œ Ñ Ç ı å. i Un. v. 1⇢∫M∆å|09⁄dÇıª. Z | U ⇥ Perozzi I ∫ / , ↵ ⌥ ^ ⌘ œ Ñ Û ( º ≤ Ô - Ñ ÷ ⌘ √ ⇤ Ü word2vec Ñy' ⌥≤Ô¿fi^‘⇣^ ç✏N®_Jp Random Walk Ñπ✏". ¿fiè⌫Â!Ï^<ºÂPÑb✏. ç⇢N Skip-Gram !ã◆Ùó. ˙¿fiÑ⌘œh:✏ [20]⇥ Grover I∫Gö©ÜÊ. .˙º„¶*H⌧↵. Breadth-first Search ⌥Ò¶*H⌧↵ Depth-first Search Ñp’Ve ✏NÔ ß6√x p ⌥ q Üøtp’Ñπ✏ Ô!ãw Ù}ÑH' [11]⇥ Tang I∫G /–˙Ü. º word2vec ÑÓ⇡˝✏Ü*. <¶Ñ˝✏Ü›Y. Ñ@Ë⌥h@«⌦. ∫G/› (⇧ÑO} h:’Ñ⌘œ [5]⇥. ≤Ôh:’Ñ⌘œ. ✏N6ö. ¯. Â92≤Ôh:’Ñ⌘œ [25]⇥ Chen I. ⌥pÍ'≤Ô✏N. 4. ⌦ÕÑ®_JpVeÜó0≤Ô. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(12) 2.2. ®¶˚q. ®¶˚q. ÈvÊ/. .«⌦¢"Âw. 9⁄`8eÑÀB. ⌥iˆãÇ≤. áÎZ¯‹¶íèå®fi¯‹ÑPú⇥⇡ÄxO(⇡—åAtg⇣8⇣∫ Ä hÀÑxO&(xS ÊŸ⌦˝ 1⇢P©Ñ⇣ú⇥ 1994 t GroupLens v⇠ ä®˙, ↵Í’ ®¶˚q&–˙ÜT ⇡_/. È(T. Collaborative Filtering. N˛. ÄS. [24]⇥ 2005 tAdomavicius I∫⌥®¶˚q. N˛Ñ˚qK. qt&⌃∫ 3 ↵;Å^% s˙ºgπÑ®¶ T N˛Ñ®¶å˜ ®¶Ñπ ’[1]⇥2009t Koren I∫–˙Ü˙ºÈc⌃„ Matrix Factorization ÑT N˛˚q. [12, 13]⇥˙ºgπÑ®¶;Å. ⇡_/ÓMmLnM8(ÑÄSK. /(i¡Ñ‚cyµ. Metadata. /C«ô. ãÇ^%. l'IÜ®¶^<'. ÍÑ¯<i¡ T N˛;ÅÔÂ⌃⇣˙º(6 User-based å˙ºi¡ Item-based Ii.^ã⇥˙º(6ÑT N˛/9⁄(6 i¡Ñú} ®¶. 政治大 ⌥(6à¯<Ñ(6@ú}Ñi¡ ˙ºi¡ÑT N˛G/ÓMmL 立 (Ñπ’ v8√Çı1/˙Àº .G- s º ↵i¡ A B. „€ C. Ç. ˚q⇤®¶↵N⇡«1. à⇢ÊŸ⌦˝. /ÑHú. FÑ/⇤88⇤G0«ô*. OLÑ⌘0⇥. ÷6⇡⇧ÄS( è. /∑_’. sit. y. Nat. Cold Start. à. Ñ∫_↵NÑv÷^<Ñ∞^. ‧. ‧ 國. 1. 學. ú '⇢xúa A Ñ(6 B_úa B „hi¡ A ⌥ B à¯< ‡d ºú a A Ñ(61⇤®fii¡ B /i¡ C ãÇ(∞^®¶- (6↵N⇡«. er. io. —tÜl‘⌥≤ÔÑÎ 2U ãÀ ∫fW(Ù⇢v÷ ËÜê(⇧ i¡IÑM «⌦ Meta Information ÜP†7®¶ÑHú õv⇧_. n. a l Ñ®¶ k© ‡d v⌫f†e⇡õ«⌦Üû iv n C h e n g cº˚qÑU⌃Ü92(⇧ †®¶Ññ∫¶⇥ Li I∫⌫f((⇧ º¯M hi U |˛⇡õ«⌦. º9Ñ↵∫. U⌃ÑT N˛®¶˚q |˛_ @9Ñ [15]⇥ dÜ⇡õá,«ôÂ v_ãÀ⌫f( œ«⌦Üû2(⇧ÑO}⇥ McAuley I∫(Ü µÜû7T. N˛Ññ∫' [16]⇥Ê. ‹¬ÑPÀ«⌦. Zhang I∫G/t. i¡Ñá,yµÂ i¡Ñ. ì≤Ôx“(⇧⌥i¡Ñh:’. ˙ºÂb®¶/®¶˚q(Êõ®¶4o‡∫. º(6Ü™. ‹Ñi¡. i¡ÑxœÊ(*⇢. ‡d–.↵¶⌦. Ü(6⌥(⇧§í. œ«⌦2•(. (®¶OL⌦ó0. œy. wìN⇢dÑ^. /Ñ⇣H [26]⇥. 8(ÑÄSK. ü‡/. (6Í⇤‹Ë(Mb. ®¶OLÔñ∫«⌦¢"OL⇥‡d. Ñ¯. Learning to. Rank ÄSø(Êõ®¶Ñ…(⌦wÜà'Ñ\( ã.^'ÙÔ⌃∫ PointWise PairWise ListWise. Learning to Rank ùgv! .⇥ PointWise Ñ rank ‘. ⇤‹ËÑfi/Æ. î¡⌥(6. î¡Ñ‹¬. (◆ÙBÍ›. Matrix Factorization ). IÑ¯‹'2L*. œ/Èc⌃„. Factorization Machine. [22]1lº⇡.^ã⇥Èc⌃„⌥(6 5. Âbi¡. /⌃„✏_h ºi¡Ñ(. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(13) ¿↵O}ñ∫U⌃&Z⇣U⌃Èc. Stochastic gradient. &…1®_Ø¶↵M. descent ’⌥dU⌃Èc⌃„⇣i↵N≠¶Ñ(6yµ< U ⌥i¡yµ< V ⇡õN≠Ñyµ<ÔÂ– ↵¶⌦„h⇡↵(6 i¡ Ñy' ç(⇡õy µ⌘œZåå¯‹'ÑU⌃ _h⌥⇢.yµœ/U⌃ hhÀ / 6⇢8∫s' µ. Èc⌃„‘⇤nMÑH,G/⌃„✏_h .^. l'I⇤n2Ü. &⇤nyµKìvÊ. öÑ‹¬Ñ ãÇ⇢(≤ÔFó- ¸∑ ù¡ ¸∑K’^F¡Ñ⇢8∫7'⇥‡d Çú˝. ®¶i¡⇤à. ⌃„✏ /å. ›⌦¡Ñ( ~˙⇡õy. k©⇥⌃„✏_h…1x“⇡õyµìÑ§…‹¬Ü˙. À(6⌥i¡Ñyµ⌘œÜU0⇡õi¡Ñ¯‹'⇥ PairWise Ñ rank ‹ËÑfi ∫iˆ⇡xi¡Kì¯‹'Ñíè iî¡KìÑ¯‹'íè2L* Bayesian Personalized Ranking ⌃Èc. ù⁄⌥(6. (◆ÙB Í/9⁄i¡Ñ¯‹'Ñ ,8ãÑπ’G/ù✏↵∫ íè. !ã [23]⇥ù✏↵∫. íè. º(6⌥i¡ÑU. c⌘fiKÑi¡. U⌃ ⌥(6 †⌘fiK !U 治政 ⌃ Ñi¡Z⇣ ↵◆Ù Üx“KìÑ‹¬ Ù¿⌦Ñ✏©s∫⇢(⇧¯ 大º PointWise PairWise ‡ ºí U⌃Ñi¡Ü™Ùúa U⌃NÑi¡⇥¯ 立 è. /˝. ñ. ›. ‧ 國. i¡ìÑ¯‹'Zíè ¡Ñç∫í. i↵!¯‹Ñi¡ù6. 學. ∫. ‹¬. ↵O}Ñ⌃. @Â˝ÙæñÑó0(6. ºi¡K. ⌥œ!Ì„˙ÜÑzV92L†=. ⇥. sit. 1˝xÑØ¶2L*. OL. er. io. w˚✏x“ a l iv n Ch engchi U. n. 2.3. Nat. & v. Decision Tree Regression. y. xzV9. ‧. ìÑO}⇥ ListWise ⇤nÑ/ (6Ü™ t‘i¡®¶Ñíè& v2L* LambdaMART ∫v- .π’⇥ LambdaMART ⌧↵PúÑíèOLI⇣fi. ⌘⌘∫^(\∫. ↵x“⇧. ÑÂXÜx“∞ÂS ↵∞˚ŸB. (x“ÑB⇡⇢8⇤. /Ù•. ˙öÑ!✏ÜK(. ÑûˆãÀx“⇥_1/™. I˚ü. v⌘⌘G0. ⌘⌘⇤…(KMx“ìW-Ñ¯‹ÂXÜx“⇥Çú∞˚Ÿ. ⌘⌘. ÂMÑìWä¯‹ ⌘⌘1äπ◆å·É ⇡1/w˚✏x“ [18]⇥w˚✏x“ / ↵_hx“⌦ÑvpL ≥q⌦ ,Ñ_hx“˝ Ë( ↵yöÑ˚ Ÿ. ⇠fl⌦Zx“. Ûv÷Ñ˚ŸB. ‡d÷⌘⇤(⇡↵˚Ÿ-h˛ào} &í. ¶’. ÂX›Y&I˚Û∞˚Ÿ⌦ 1/⌥x“ ↵ÂXÑN↵ Ñ˚Ÿ⌦. ¯. Ñh˛. Target Task. ‡∫÷⌘!’⌥M. ˚Ÿx“0Ñ. s∞˚Ÿ⌥⌦˚Ÿà^<⇥ w˚✏x“ÑÇı !✏ èãI¯‹˝õ⌥ÄSˇÜ„z^< ¯. _1/Ó⇡/⇢N)(Üê˚Ÿ. 2Ó⇡˚Ÿ. Fv÷⌘⌥!ãK(. Source Task. -ÑË⌃ÂXÜ9. -Ñx“⇥ÇúÜê˚Ÿ⌥Ó⇡˚Ÿ¯‹'äÿ. £. w˚ÑHú1ä} äπ◆⇥ !ÆÑãPãÇ⇢⌘⌘Hx⇤ s✏ £⌘⌘ (x _ B1⇤‘⇤π◆ ‡∫⌘⌘ÚìHûNªx“ s✏ Ñ≈É-Hx ⇤ÜÇUsa. Õ1IÄS. GÇÜê˚Ÿ⌥Ó⇡˚Ÿ¯‹'äN⇢Û 6. ¯‹. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(14) G. Ô˝⇤. †w˚Ñ|. Üê˚Ÿ@w˚˙ÑÂXÕ. ;⇡ÜÓ⇡˚ŸÑx. “⇥w˚x“9⁄Üê⌥Ó⇡Ñ⇠flç˚Ÿ/&¯ ÔÂ⌃⇣x w˚x“ Inductive Transfer Learning ^„c✏w˚x“ Unsupervised Transfer Learning åI€w˚x“. Transductive Transfer Learning. .. x. w˚x“Í. º✏NûÜê˚Ÿ-w˚¯‹ÂXÂø(Ó⇡˚Ÿ-róÙÿÑHú. Ë. ^„c✏. w˚x“ Ëº„zÓ⇡⇠fl-Ñ^„cx“OL ÇZ^ M≠ ∆¶0 I I⇥w˚✏x“ Ù˝´„€K(º⌅✏˚Ÿ⌦ —tÜ1ºÒ¶x“ÄSÑ w. Ù∫Ï√|U⇥RainaI∫⌫f✏N'œ*⇡ËÑ«ôyµÜT0. Ñ⌃^Hú[21]⇥ Dai I∫G/. B⌥Üê«ô⌥Ó⇡«ô. Ó⇡«ô. wZ⌃§ÜqˇÓ. ⇡«ôÑ⌃§ [8]⇥ Ganin I∫G/✏N⌃^h(i↵⇠flìÑ¯<'Üyµ ˜÷(í Ó⇡«ô⇡⇠Ñ≈¡↵_˝x“0yµ [9]⇥ Bocsi I∫)(NÄA¡ LÑ_hK¬Ñ‹¿#’I«⌦K(ÛB¡LÑ_hK¬⌦. _. Wo}Ñh˛. [3]⇥ Cremonesi I∫⌥Ë⇠flÑ«⌦Ü9Ñ®¶˚qÑ¡Í [7]⇥ He I∫)(i M. Ë✏õ_6. 立. 治政 Attention ⌥«⌦ ⌃0w˚ÛÓ⇡«ôZx 大學 ‧. ‧ 國 io. sit. y. Nat. n. al. er. ¡¯‹Ñ«⌦ “ [10]⇥. Ch. engchi. 7. i Un. v. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(15) ,. ‡. vπ’ (d‡¿-. 政治大 ( 3.2✏¿⌘⌘⇤œ˙ Kπ’ 立. ⌘⌘⌥⇤s00À9⌘⌘@(Ñπ’⇥( 3.1✏¿⇠fl⇥4&˙À‹¬. ( 3.3✏¿- ⌘ ( 3.4✏¿- ⌘⌘⇤À9⌘⌘. 學. ‧ 國. ⌘⌘ÛÅ„zÑOL ⌘⌥À9⌘⌘ÇU#•. @(Ñ≤Ôh:’x“F∂⇥. ‧. 3.1. sit. Nat. N˛π’-. y. OLö©. ,(T. ⌘⌘⇤X. ⌘⌘88⇤⌥(6⌥i¡Ñ§í‹¬˙⇣. ↵0•È. io. er. c M = (mij ) 2 R v- U ∫(6∆ I ∫i¡∆ mij ∫(6 i ⌥ i¡ j Ñ‹¬ ⇡↵‹¬ÔÂ/ !↵N U⌃ fi !xI ^ã⇥ |U |⇥|I|. al. n. iv n C º⇡↵0•Èc M ⌘⌘_˝Â h↵å⌃ i U s G = (U, I, E) e n g cGhÜh: E = {(Ui , Ij)|i 2 U, j 2 I} ⇥(,÷á-. ⌘⌘ö©i↵. v-. ⇥4 GT ⌥ GS ⇥. GT ∫⌘⌘®¶˚qÅ„zÑÓ⇡⇥4 GS G∫⌘⌘(ÜZË⇠fl®¶ÑÜê ⇥4 ⌘⌘ÑÓ⇡1/Å⌥⇡i↵⇥4✏N ↵˝✏ F Ü˙ÀvKìÑ§í‹ ¬. 2. û7≤ÔÑ. ∆'. ç⌥v)(≤Ôh:’x“˙œ↵¿fiÑ⌘œh:. ✏ çÜZååÑ®¶⇥. 3.2. pÍ'≤Ô˙. pÍ'≤Ô Heterogeneous Information Network ∫ .w ⇢.¿fil'Ñ« ⌦≤Ô⇥ ↵pÍ'≤ÔÔÂh:⇣ G = {V, E, A} v- V ∫¿fi∆ E∫ äÑ∆. A ∫¿fi¯. ÚÔ˝_⇤6. 1⇢l'. …Kl'∆ œ/LK. |A|. L^. 2 ⇥ãÇ(Û⇥2A. Ú®. û. I«⌦⇥(˛Ê. Ñ«ô'⇢˝/Â⇡#PÑ‚c«ôÑb✏2X(«ô´8. Ÿ-. L. L·. ⇡õ‚c«ôà. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(16) „Ù•(≤Ôh:’ÑF∂2L◆Ù. ‡d⌘⌘≈. H⌥«ô2L⇤÷. ÷˙. ⌘⌘‹ËÑ«ô&⌥K˙⇣ 5≤Ô M˝|åZåbÑ≤Ôh:’x“◆Ù⇥ (d✏¿- ⌘⌘⇤s0X⌘⌘ÇU⌥üÀ«ô˙⇣pÍ'≤Ô &(Ö# P. \2LË⇠flP. ⇥. 政治大 3.1: 立Û⇥2A ŸKpÍ'≤Ô. ‧ 國. ⌘⌘ÅH∫ç⌘⌘Å(Í.^ãÑ«⌦⇥(,÷áºLÚÑí’‹¬. i↵Ë⇠fl⇥4ÑF}. s(6. ºLÚF}Ñ. ⌅Z–÷&⌥œ. ↵(6. ⌘⌘;Å/. ⌅Ü˙. ⇥‡d⌘⌘›. ‧. )((6. 學. ñH. ⌅. ºœ. ñLÚÑí’. Z/M†= ó0i↵0•Èc⇢Ó⇡⇥4 MT = (mij ) 2 R ⌥Üê⇥4 |US |⇥|IS | MS = (mxy ) 2 R Kå⌘⌘øÔ9⁄⇡i↵Èc⌥«ôI€⇣i↵pÍ'. al. GT ⌥ GS ⌃%. ÍÒÑ(6§∆. UT. GT. US Â LÚ. iv v- UT \ US =C? ET ⌥ ES ⌃%∫i↵⇥4ÑF} ⌅!x n U h i ⌘⌘⌥)(Ö#P e\n g ⌥i↵Ë⇠fl⇥4ÑpÍ'≤Ô Z# ch n. ∆ IT IS K∆ ⇥•W. ⇥. GT ∫Ó⇡⇥4K≤Ô. er. ∫Üê⇥4K≤Ô. io. GT = (UT , IT , ET ) ⌥ GS = (US , IS , ES ). sit. y. Nat. ≤Ô. |UT |⇥|IT |. P⇥. 3.2: Ó⇡⌥Üê⇥4ÑpÍ'≤Ôå⌃ 9. PÀ. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(17) 3.3. ˙ÀÖ#P. Ö#P. \. flìii Ö#PÑä. \. Superhighway Graph. [14]/. ↵w˚x“ÑVe. 9⁄i↵⇠. (6Kì@q´ÑLÚ ⌃0˙À(6ìÑä⇥ ✏N⇡↵ ÔÂó⌘⌘(w˚Üê⇠flÑÂX «⌦ÑN↵˝ ⇤}Ñ. Hú⇥Â↵⌘⌘ö©¿º/Ö#P. \Â Ö#P. \ÑZ’⇢. ºi↵⇥ ↵q´ÑLÚ∆ Iˆ. 4 GT = (UT , IT , ET ) ⌥ GS = (US , IS , ES ) ⌘⌘ Iˆ = IT \ IS 6= ? G⌘⌘ÔÂ✏N ↵LÚ i 2 Iˆ ó(6 ui 2 UT ⌥(6 uj 2 US @#P ⇡↵Ôë⌘⌘1∫K⌘ bridge ⇥⌥Üê⇥4«⌦P 2 Ó⇡⇥4. !Æ_ Ù∫Ñπ✏1/äi↵⇥4. G ç2L≤Ôh:’Ñ◆Ù⇥F/⇡↵⇤. GT ⌥ GS Ù•. ↵OL⇢(≤Ô. ⌦. u⇣. 5. GT ⇥4Ñ. (6Å⌥ GS ⇥4Ñ(6Z#P≈6óÅ✏N⇡ßK⌘⌦ÑLÚM˝T⇣ _1 /t5 ·i↵⇥4≈˝✏N Iˆ ∆ ÑLÚÜ2L#P ⇡#PÑ«⌦≥≠⌘⌘. 政治大 ∫óNºÑ1 –πb⌦Íû7ÜK⌘⌦¿fiÑ«⌦ v÷¿fiG‡‚KÅN` 立 Ùù6Í¡ Æ ⇥4Ñ«⌦⇥∫ÜÅPÃË⇠fl⇥4Ñ#P' ⌘⌘( Superhighway Graph. Ü˙À#P. ↵b⌘⌘À9Ö#P. \ÑZ. i↵Ë⇠flÑpÍ GT ⌥ GS ⌘⌘ÔÂ…1Ö#P \ó’˝✏ F ˙À ↵⇡x⇧∆ UˆT ⌥ UˆS v- UˆT 2 UT UˆS 2 US 6å⌘⌘ º⇡x ⇧⌫h-œ↵(6 ui 2 UˆT ⌥(6 uj 2 UˆS iiKì˙ÀÙ•Ñ#P‹¬ ⇡x⇧. ‧. ›. ‧ 國. \. 學. Ö#P ’⇥. sit. y. Nat. n. a(l. Uˆd =. er. io. ∆ ˙ÀÑπ✏/9⁄ 3.1✏. )v ˆ ni |N (u) \ I| C u | uh2e Ud , hi U ↵ , n gNc(u). Üx«(62e⇡x⇧∆ Ñ∆ d. ↵ /⌘⌘ÑÔøß√x. v- d 2 {S, T }. (3.1). N (u) /(6 u Ñ@. 0E. s⌘⌘x«–(6\∫⇡x⇧ÑÄª<⇥Ç ⌘⌘øÔû3.1✏ó0Ñi↵⇡x⇧∆ UˆT ⌥ UˆS -› iiKìÑ(6˙. À∞ÑÖ#Pä ∫⇢. xœ∫ |UˆT | ⇥ |UˆS | ù. w= v- ui 2 UˆT. )(Ö#P \ó0. uj 2 UˆS. ↵∞Ñ. ii(6. Ñ#PÑ⌦Õ w ⌘⌘ö©. ⇥ |N (ui ) \ N (uj )|,. (3.2). ∫⌘⌘øß⌦Õ‘áÑ√x⇥. å. ⌘⌘øÔ. G⇢. 10. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(18) G = F(GT , GS ) = {Ud \ I, EUd \I. (3.3). Esuper }. v- d 2 {S, T } ⇥⇡õì1 3.1✏#PÑ(6 — (6Ñ#Pä‡∫@q´Ñ. LÚ«⌦‘ãàÿ ✏©⌦s/÷⌘ÑF}O}⇤à•— ‡d⇡õ(6 — ( 6Ñä1±+Üà7»ÑO}‹¬ •W ⌘⌘1ÔÂ(≤Ôh:’◆Ù⇡5 ⇥. 學. (a). ‧. ‧ 國. 立. 政治大. n. er. io. sit. y. Nat. al. Ch. e n(b)g c h i. i Un. v. (c). 3.3: Ö#P. Ç. 3.3@:. \K#•:✏. ⌘⌘9⁄i↵⇠flìii 11. (6Kì@q´ÑLÚ. ùg⌘. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(19) ⌘@-öÑ√x. ⌃0˙À. ↵(6ìÑä⇥. ✏N⇡↵Ö#PÑä. ó⌘⌘(w˚Üê⇠flÑÂX «⌦ÑN↵˝ Ù†⌃) äèG_˝ HÑ2b†w˚ Negative transfer Ñ|. 3.4. -ö ⇥. ↵. ÔÂ ⌃Ñ˙. ≤Ôh:’x“. (,÷á-. ⌘⌘⌘xÜ. LÚÑ≤Ôh:’ ’⇥. 3.4.1. .≤Ôh:’x“Ñπ’Ü◆ÙpÍ. ⌃%∫ Deepwalk. LINE HPE. Üó0(6⌥. ↵b⌘⌘⇤À9⌅↵π. Deepwalk. 政治大 Deepwalk [20]⇤÷Ü word2vec 立 (U⌃Í6û -á,ÂPÆ^ÑZ’&⌥dÇ .b✏(º>§≤ÔI. K-⇥8e. ↵föÑ≤Ô. G-. Deep-. 學. ‧ 國. ıÂÊ. walk Â®_JpÑπ✏ ⌥ -Ñ¿fiJpÑÔëÂ^<ÂPb✏Ñ¿fiè⌫ª „ê &* Â↵Ó⇡˝✏⇢. n. Ch. i Un. v2V. (3.5). exp(Vv · Vi ). t ∫vñó'✏. window size. Vi ∫ i. Vj ∫v0E¿fiK⌘œh:’⇥dπ’ã_Äå≤Ôh:’ÑH. ≥. π’(. s. Ê\⌦_à≈. Û⌦Ô™/àu∞(⌃ÔL-ó0. /Ñ92. _‡∫ÙU'. _8´…(ºÊõ˚Ÿ-⇥. Large-Scale Information Network Embedding. ⌥ Deepwalk ^<. LINE [25] –õÜ. Ñ'œ«ô⇥÷⌘ö©Üi↵ ’. v. e nexp(V g c hj ·iVi). v- j /¿fi i 2 V Ñ context. 3.4.2. (3.4). y. tji+t,j6=i. Pr(vj |vi ) = P. Ñ⌘œh:’. log Pr(vj |vi ). sit. io. al. X. er. Nat Â vùˆ_á⇢. 1 X |V | i=1 i. ‧. V. L(V ) =. Â↵⌘⌘⌃%ÜZ™. svi↵¿fi. Ù•#PB. ↵ÔÙUÑπ’Âø(ºU⌃˛Ê. ∫ÑÓ⇡˝✏Ü›Y≤ÔPÀ&*. ⇥,. ↵1∫. é¯<¶. d≤Ôh:. First-order Proximity. é¯<¶1/⇡i↵fiÑäÑ⌦Õ 12. L-. Âifií. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(20) Ù•#PÑä. é¯<¶∫ 0⇥‡dvä⌦Ñ⌦Õäÿ. G. ⌘⌘ø™⇡i↵fi. ä¯<⇥,å↵1∫åé¯<¶ Second-order Proximity svi↵¿fiKì q´Ñ0E¿fiä⇢B ⌘⌘ø™⇡i↵fiä¯<⇥Â 3.4∫ãP⇢. 政治大. 3.4: LINEÑ¯<¶≤Ô v-äÑó0„häÑ⌦Õ. Âfi 1 ∫ãP. Â. flfi 1. é¯<¶Ü™. 學. ‧ 國. 立. ¯<Ñfi∫fi 6. ‡∫fi 1 ⌥. é¯<¶Ü™. y. º. Éh:Ñ/≤Ô. -. io. sit. 5⇥. @ËÑ⌦o. ‡d º≤Ô ÑPÀ É1!’›XÑàåt⇥ ‡∫ÉU⌃Ñ/¿fi⌥vhä0EÑ‹¬ ‡d‘⇤˝. n. al. ⌦(≤Ôh:’(åé¯<¶. ó-. äÑ¯#⌥&. er. 4. Nat. fi3. ‧. fi 6 Ù•¯#fi 1 ⌥fi 6 äÑ⌦Õ äÑó0 'ºv÷⌥fi 1 #•Ñfi Âåé¯<¶Ü™ flfi 1 ¯<Ñfi∫fi 2 ‡∫fi 1 ⌥fi 2 ¡ q Ñ0E. Ch. n U i engch. @Â(,÷á-. ºåé¯<¶Ü™. i v ›X⇤⇢≤ÔPÀÑ«. ⌘⌘(åé¯<¶Ü◆Ù≤Ôh:’⇥. œ↵¿fi˝⇤. i.bK. ↵/¿fiÍÒ. Ê. ↵G/. ⇤⇣∫v÷¿fiÑ context ‡d ºœ ↵fi ⌘⌘˝⇤ i.⌘œh:’⇢fi ,´Ñ⌘œh:’Â fiÑ context Ñ⌘œh:’⇥› ↵ ⌘ä (vi , vj ) ºû vi ÷. ⇣Ñ context ¿fi vj Ñ_á∫⇢. Pr(vj |vi ) = P. exp(Vvj · Vvi ) vk 2V exp(Vvk · Vvi ). (3.6). ⌘⌘ÑÓ⇡/Å* ↵⌫Ó⇡˝✏⇢. OLIN E2 =. X. (i,j)2E. 13. wi,j log Pr(vj |vi ). (3.7). ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(21) v’. E ∫Ω#0ÑäÑ∆. 3.4.3. Vvj ∫ context ¿fi vj K context Ñ⌘œh:. Vvi ∫ vi Ñ⌘œh:’. wi,j /äÑ⌦Õ⇥. Heterogeneous Preference Embedding. HPE [5] ⌥⌦i⇧Â\¯<F intention. ü, HPE ø/Â˙ºÂb✏. ·¯. Ñ®¶OLZ∫ñÅÑ˚ŸÓ⇡. ⇧Â\¯‘. @Â÷Ñ≤Ôh:’x“fl⌦i. õÙ’⇥ HPE ;Å/9⁄(⇧ÑO}Ü2Lx“. ZÜ. ⌥ Deepwalk ^<F p ⇣◆ÙÑ¿fi. @Â. Ñ/÷∫Üx“0(⇧ÑO} ✏N ⌦ÕÑ®_J ‡dΩ#↵ eÑ¿fiB ⌦ÕäÿÑä _⇤´x-⇣ œ Deepwalk _á/G˚⌃HÑ⇥Â. ∫◆ÙÑ¿fi fi 1 2L↵. query. 3.4∫ã. fi 6 ‘wfi 1 v÷Ñ¿fi. eÑΩ#B. fi3. 政治大. v⌘⌘x-. 4. 5. Ùπ. ◆´Ω-Z∫◆ÙÑ¿fi Ê ⌥ LINE ^< G-⌘⌘ÚìΩ0fi 6 v\ fi 1 Ñ context B ⌘⌘_⇤ BΩ#fi 6 Ñ¿fi sfi 1 7 8 v\ õO}å«⌦⇥ HPE ùg⇡↵. io. ^<º LINE. HPE ;Å*. ↵⌫Ó⇡˝✏⇢. n. al. if vj 2 Context(vi ) otherwise. 1, 0,. OHP E =. Ch X. (i,j)2E. v-. engchi. iv n UX. wi,j log Pr(vj |Vvi ) +. E h:⌘⌘Ω#0ÑäÑ∆. (3.8). sit. Nat. Pr(vj |Vvi ) =. (. _ál✏Ç↵⇢. y. pÍ'≤Ô¿fi2LΩ#&Ù∞¿fiÑ⌘œh:✏. ‧. ‧ 國. ì•x“Ü⇡↵¿fiÑ. 學. fi context _⇤n2Ü. B_⇤⌥Ω#ÑÑ¿fiD—Ñ¿. er. 立. fi 1 Ñ context ⇥⇡_„hW⌘⌘(ZΩ#Ñ. i. (3.9). kVvi k2. wi,j „häÑ⌦Õ. †e. ↵cè. Ñ√x/∫ÜMN¶Ï Ñ≈b| ⇥d HPE Ñ†eÜpeÑ®_Ø¶ ↵M’ Asynchronous Stochastic Gradient Descent 3.9✏2L s ⇥. 3.5. ®¶˚q. ì1⌦ 3.3✏¿⌥3.4✏¿ÑA↵å #P. \7. å. ⌘⌘Ôó0Ë⇠fl⇥4≤ÔpÍ. ç✏N≤Ôh:’x“◆Ù˙Ü. ì1Ö. œ↵(6Ñ⌘œh:’ ru ⌥. LÚÑ⌘œh:’ ri v-⌘œh:’Ñ≠¶Ü∫ d Ü`✏º(6 LÚÑx œ⇥(®¶˚q- ⌘⌘(◆Ùå⇣Ñh:’ 9⁄œ↵(⇧ u vF}K LÚ q ⇣,˙(⇧ ºœñLÚ i Ñ⌃x 14. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(22) Scoreu,q,i = p ⇥ (ru · riT ) + (1 q, i 2 |I|. u 2 |U |. íè⌃x ÿÑLÚ. ⇣↵∫. 立. p 2 [0, 1]. Kå⌘⌘ç9⁄d⌃x. ÷˙M k ñ. Ñ®¶⌫h⇥. 政治大. ‧. ‧ 國. 學. io. sit. y. Nat. n. al. er. v-. (3.10). p) ⇥ (rq · riT ). Ch. engchi. 15. i Un. v. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(23) ,€‡ ÊWPú⌥ (d‡¿. ÷. 政治大 ↵⌘⌘@(Ñ«ô∆Â v«ôy' 立. ⌘⌘⌥⇤. ⌘⌘⇤™. &U:v¯‹Ñq &–õ ÊWÑ!ã@(Ñ√xÂ ⌘⌘Ñ‘. ⇤π’⇥( 4.3✏¿. 學. ‧ 國. -⌘⌘⇤X x⁄⇥ ( 4.2✏¿. ÷⌥⌃ê⌘⌘Ñ∂Àåπ’@ó0Ñh˛Pú⇥( 4.1✏¿. ⌘⌘G⇤À9⌘⌘@(óU0⌥ ⇡⇥( 4.4✏¿-. ⇤‘⇤ ⌃ê⌘⌘π’⌥v÷π’Ñh˛⇥. å. ( 4.5-. y. sit. io. er. «ô∆. Nat. 4.1. ⌘⌘⇤✏NÊãÑU. ‧. :Ü¿fld˚qÑh˛⇥. ⌘⌘. al. n. i v ⌅\∫«ô∆ÜÊ n C hengchi U (6(Í∂Û⇥2As-@Z˙ÑF}L∫ œ. (d«÷á-⌘⌘( KKBOX ºÛ⇥2AsÑF} W⇥ KKBOX 6∆Ü@ F«ô˝s00⇠⌅Ü. ↵(6(¿ºBìfi}ÜÍ. ñL. ûÍ.›n{. e F}BìÂ 0@⇥,÷áx˙Ü∞ Ñ A ⇥4\∫Ó⇡«ô∆ U⌃ Ü 2018 t 5 Û 8 º A ⇥4Ñ@ F} ⌅ &⌥ 5 Û 7 Ñ«ôvZ◆Ù «ô∆ ∆A ê«ô∆ ∆B ⇥. Â↵!1◆Ù∆ A. 8. Ñ«ôvZ,f«ô∆. Ê πb ⌘⌘_x˙Ü⌘⌘(ÜP #_U⌃Ü 2018 t 5. h 4.1: A. Û7. Â↵!1,f. PÃ«⌦Ñ B ⇥4\∫Ü. Ñ«ôvZ◆Ù∆. Â↵!1◆Ù. B⇥4«ô∆x⁄q. 16. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(24) «ô∆. ◆Ù∆ A (2018/5-2018/7). ,f∆ A (2018/8). ◆Ù∆ B (2018/5-2018/7). 1580 22170 63938 40.47 288.40 0.182%. 1548 12276 31848 20.57 111,48 0.168%. 249640 558601 22258342 89.16 629.40 0.016%. |U | |I| |S| t(·) k(·) è¶. v- |S| ∫@ (6 - LÚ Õ⌥KF}D t(·) ∫(6sGF}ÑLÚ xœ k(·) ∫(6 ºLÚÑsGF}!x è¶ö©∫@ (6 — LÚ Õ ⌥KF}D T@ Ô˝D (d!«ô∆Ñ-ö⌥ 5. Ñ‘á. s∫. |S| ⇥ |U |⇥|I|. Ü\∫◆Ù∆Ü⇣, 8 Ñü‡/‡∫⌘ 治 *Ü\zV ⇡_/(ÊŸ 政 ⌘ )(NÄÑwÚ«ôÜ⇣,*ÜÑ®‚ 大 ⌦‘⇤ÊõÑZ’⇥ (⌘⌘–˙ÑÖ#P \˙ ’LÚ( ⇥4ÑÕä 立 ⇧àÕÅÑ‡. ∆A. ÷⌘}NÑLÚ∆. ÑÕäá/&ÖN. öÑÄª<⇥↵h@:∫◆Ù. B t‘LÚÑÕäá. 1dhÔÂ|˛0. LÚÑÕä'⌃%. WoWÑÓp⇥ º‘⇤✏Ñ«ô∆Ñ(⇧Ü™ º@#•Ñ(⇧ ÷⌘ÑO}ø⇤⌥ÍÒä¯—. Nat. n. 4.2. sit. LÚT◆Ù∆KLÚÕäá. «ô∆. Õäá. er. io. h 4.2: ◆Ù∆-q. al. ⇡õÕäÑLÚxä ‡d˝ ö↵¶Ñ. y. k©0✏«ô∆xóÙ}⇥. ºi↵«. ‧. ô∆ ÿ. «ô∆-Ñ. 學. (⇧. ‡∫⌘⌘˙ÀÖ#PÑ˙ñ(º⇡i↵. ‧ 國. á/. 07. i n C◆Ù∆A 78.09% U hengchi ◆Ù∆B 3.09%. v. ÊW-ö. d‡¿⇤™ ⌘⌘⌘xÑ~↵h:’x“π’ (Üv⇣⌘⌘®¶˚qÑ◆Ù˙ !ã⌥†⌦⌘⌘Ñ˙ π’å@x“Ñ!ã&s†X¯‹Ñ√x-ö⇥ • DeepWalk⇢ ⇢DeepWalk )(®_p’ ºÂPÑG-K◆Ù π✏Ü" -öh:’Ñ≠¶. Dimensions. ⇣. 2¿fiè⌫. &Â≤ÔPÀ^<. ≤Ôh:’⇥ (◆Ù Deepwalk B ∫ 100. ⌘⌘. œ↵¿fivZwfiÑ!x. Walk Times ∫ 20 ! œ!JpÑ›‚ Walk Step ∫ 40 e föÑ ñó'✏ Window Size ∫ 3 œ!†Ω#Ñxœ Negative Samples ∫ 5 !⇥ 17. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(25) • LINE⇢ ⇢ ( LINE. ⌘⌘(åé¯<¶\∫⌘⌘Ñ≤Ôh:’◆Ù⇡ñ. v✏©∫¿fiKì@¡ Ñ¯ 0Eä⇢ ⌘⌘1™÷⌘ä¯<⇥ (◆ Ù LINE ⌦ ⌘⌘-öh:’Ñ≠¶ Dimensions ∫ 100 ÷#!x Samples Times. -ö∫ 20 ⌅ !. Negative Samples. œ!†Ω#Ñxœ. ∫ 5 !⇥ • HPE⇢ ⇢ Heterogeneous Preference Embedding ÔÂ⌥(⇧ÑpÍ«⌦P Üx“. w. &9⁄x“0Ñ(⇧ÑO}2L¯‹Ñ®¶⇥ ⌘⌘-ö HPE h Dimensions. :’Ñ≠¶. ∫ 100. Samples Times. ÷#!x. -ö. ∫ 20 ⌅ ! œ!†Ω#Ñxœ Negative Samples ∫ 5 ! œ!JpÑ ›‚ Walk Step ∫ 3 e cè Ñ√x Regularization Term ∫ 0.01 ⇥ Â↵∫⌘⌘˙ π’⌥h:’x“Ñ√x-ö⇥. 政治大 \9⁄i↵ «ô∆-. • Superhighway⇢ ⇢ Ö#P. ↵(⇧ ºÊ 立 ↵«ô∆K(⇧Kìi¡ÑÕä¶Üzö/&Å˙À#P (˙À#P Ôh:’Ñx“⇥. (Ö#P. \˙. Ú‘ã ↵ <∫ 0.9. ˙ÀÖ#PÑ⌦Õ. _˝2. e*. v≤. -. ⌘⌘-öi↵(6Kìq. ÑL. ∫ 1.0. Â↵. Ü(d-ö⇥. ÑLÚ‘ã ↵ <∫ 0.9. y. ⌘⌘-öi↵(6Kì. ˙ÀÖ#PÑ⌦Õ. io. walk B. -. ⌘⌘-öh:’Ñ≠¶. Dimensions. ∫ 1.0 ⇥ (◆Ù Deep-. ∫ 100. er. q. \˙. sit. Nat. • Superhighway+DeepWalk⇢ ⇢ (Ö#P. .≤Ôh:’x“. ‧. ‧ 國. º(i↵hÀ≤ÔPÀKìJp⌥⇤Ù†π◆. 學. å. œ↵¿fiv. n. al iv ZwfiÑ!x Walk Times ∫ 20 ! œ!JpÑ›‚ Walk Step n C U ∫ 40 e föÑñó'✏ h e Window h i ∫ 3 œ!†Ω#Ñxœ n g cSize Negative Samples. ∫ 5 !⇥. • Superhighway+LINE⇢ ⇢ (Ö#P \˙ - ⌘⌘-öi↵(6Kìq ÑLÚ‘ã ↵ <∫ 0.9 ˙ÀÖ#PÑ⌦Õ ∫ 1.0 ⇥(◆Ù LINE ⌦ ⌘ ⌘-öh:’Ñ≠¶ -ö∫ 20 ⌅ !. Dimensions. ∫ 100. Samples Times. ÷#!x. Negative Samples. œ!†Ω#Ñxœ. ∫ 5 !⇥. • Superhighway+HPE⇢ ⇢(Ö#P \˙ - ⌘⌘-öi↵(6Kìq Ñ LÚ‘ã ↵ <∫ 0.9 ˙ÀÖ#PÑ⌦Õ ∫ 1.0 ⇥ ⌘⌘-ö HPE h:’ Ñ≠¶ !. Dimensions. ∫ 100. œ!†Ω#Ñxœ. Walk Step. ∫ 3 e cè. ÷#!x. Negative Samples Ñ√x. Samples Times ∫5!. -ö∫ 20 ⌅. œ!JpÑ›‚. Regularization Term. ∫ 0.01 ⇥. º⌦≤Ôh:’ ⌘⌘Ü( ProNet Wˆ [6]1 Z◆Ù⇥ 1. ProNet-core: https://github.com/cnclabs/proNet-core. 18. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(26) 4.3. U0⇡ñ. ∫Ü‘⇤aœ⌘⌘Ñ92π’⌥üπ’(Û⇥®¶Hú⌦Ñh˛. ⌘⌘°(Â↵. €.U0π✏ÜU0˙ºÂbÑ®¶˚qÑh˛⇢. Precision. ñ∫á. ñ∫áh:(fö,f«ô∆ÆÑ~⌃á vl✏Ç↵⇢. (⇧úaÑLÚ´⇣ü®¶Ñ∆. Precision =. P. u2U. T®¶⇧. |Tu | N. |U | 政治大∫®¶fiÑ⇧Æxœ v- T ∫(⇧úaÑLÚ(®¶⇧ÆÑ∆ N @ (⇧⇥ñ∫á ÿ∫立 1 ñ∫áäÿ h:⌘⌘Z˙Ñ®¶⇧Æ˝ u. (⇧úaÑLÚ´⇣ü®¶Ñ~⌃á⇥vl. sit. n. al. i u2U |I | n C Recall = U h e n g |U | i ch P. |Tu |. v. (4.2). u. Tu ∫(⇧úaÑLÚ(®¶⇧ÆÑ∆ Iu ∫(⇧úaÑ@ LÚ U ∫@ (⇧⇥Ïfiá ÿ∫ 1 ⌘⌘(Ïfiá/∫Ü¿fl«ô∆. -⌅↵(⇧´®¶˙ÜÑLÚ⇧Æ1Â⌦i.U0Ñl✏Ü↵ sa‹¬Ñ. Åv-. Gö º ↵(⇧Ü™ õ¿fl«ôå|˛ vÊ fi÷ fi@. er. io. vÑ∆. ö↵¶0U˛˙LÆÑ¡Í⇥. y. Nat. Ïfiáh:(fö,f«ô∆✏Ç↵⇢. ‡d˝. ®fi. ‧. Recall. Ïfiá. £ÑLÚ. ‧ 國. (⇧Ü™. U∫. 學. ä⇢. (4.1). £ÑLÚ⇥¯ ÷. £ÑLÚ⇥. ⇢⌘‘ã´˚q®¶˙Ü⇥. ⌘⌘ÔÂ|˛0. .⌃x–ÿ. ‚≈⇤Ê. LÆÑñ∫á∫ 1 ⇡ ↵(⇧÷. ñ∫á⌥Ïfiá/. .U0Ñ⌃x↵M⇥ãÇ⇢. ↵wÜ®¶Hú<Nà} FÊ £ÑLÚà⇢ Ùàπ◆1®. Ñ⇡↵(⇧ÑÏfiá1⇤àN œM(⇧. öÑ. ‡∫˚qí. £ÑLÚxœ_˝. #. ¶’® ‡d≈. Å B⇤œ⇡i.x<ÑÿNÜU0®¶˚qÑ}fi⇥. 19. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(27) Mean-average Precision. sGñ∫áG<. sGñ∫áG<h:(fö,f«ô∆∆. T®¶⇧ÆÑ~⌃á. U ∫@ (⇧. (⇧úaÑLÚ´⇣ü®¶Ñ. vl✏Ç↵⇢. MAP = v-. @. P. APu |U |. u2U. (4.3). APu ∫œ↵(⇧ÑsGñ∫á APu =. PN. Pi. j=1. i=1. s⇢. R(Itemj ) i. (4.4). N. v- Item ∫®¶fiÑLÚ⇧Æ R(·) ∫ ⌥:˝✏ v Itemj ((⇧úa ÑLÚ∆ BG∫ 1 &G∫ 0 ⇥(¢"˚q- sGñ∫á/ ⇧⌥⇡'ÑU 0⇡ñ. ⌘⌘(®¶fiÑ⇧Æ-. 治政 ÍÅ⇤ni¡Ñ¯‹' 大. ÙÕÅÑ/_Å⇤œ. 立,ÅU0⌘⌘⌧↵fiÑ⇧ÆÑ¡ÍB. i¡(⇧Æ-Ñ⌃è‹¬⇥. ì8(Ñ. FÊõ⌦&. ⇤. ˚qÑñ∫á∫ 0.4. ⇡ºåéÑh˛. ¯‹Ñá‡. 4«. G-˚q⌧fiÑá‡Ü. °íè(Í↵Mn˝. ‧. ¯‹. ‧ 國. ñ∫á∫ 1. 學. /ñ∫á s(⇧Æ- ¯‹Ñ⇧ÓT⇧ÆÑ‘á Fñ∫á& ⇤n⌃èÑ‹ ¬ ãÇ⌘⌘⌧↵fi 10 «á‡ ⌃ÛÑ≈¡/⇧Ægœ «á‡˝/¯‹Ñ ⇤qˇñ∫á. y. ºäåbÑMnfà⇤'ÑÚp. io. N. sit. á⇤œ⇧Ó(⇧Æ-ÑMn ‡d. º¯‹Ñá‡(⇧ÆÑMnä`M. ⇡↵⌃x1ä'. er. Nat. Ñ⌃x F⌘⌘⇤ }Ñ˚q º¯‹Ñá‡íè…rÅ( ¯‹Ñá‡K M ñ∫á!’U0⇡⇧⌥⇡ ‡d1 sGñ∫áÑU0π✏˙˛⇥sGñ∫. n. al ä¯‹Ñq≈ ÅäÈ˙˛⇡↵Çı⇥ (⇡· ⌘⌘⌥@ iv n C hengchi U ∫á÷G<ÜU0t↵˚q®¶Ñ¡ÍÓp⇥. ó0Ñ⌃xä ÷‘˛Ü (⇧ÑsGñ. Novelty. ∞N¶. ∞N¶h:(fö,f«ô∆-. @. (⇧. ºv®¶⇧ÆÑLÚO⌘±Ä⌥. &Ñ↵¶ vl✏Ç↵⇢. v-. N 1 1 XX ln bs Novelty = |U | N u2U i=1 ln max(bs , Ci ). Ci ∫LÚ(«ô∆-Ñ=F}!x. '∫ 1. (4.5). bs ∫ÔøtKæ<⇥∞N¶Ñx<. ✏G®—º 0 ⇥∞N¶(qÛ2AI. Ÿ-mLà8(Ñ. .U. 0®¶˚q¡ÍÑ⇡ñ⇥(®¶˚q- ⌥≥q«⌦¢"˚q 'Ñ / ® ¶˚q˝k©(⇧|ò˙[œ(«ô· º(⇧Ü™!’~0FÊõ⌦ £Ñi¡. $v(Û⇥2A. Ÿ-ÙoÕÅ 20. ‡∫LÚÑxœ>⇢. ,(. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(28) ⇧}NÑLÚ £ÑLÚ Ñ®˙. Nx~ñ'⇢∫⇤±ÄÑLÚ⇥®¶˙(⇧í πbÔÂk©(⇧|ò˙∑ÄÑLÚ £ÑLÚ /∆-(±ÄLÚ⌦⇥F. ∞LÚ'⇢/í. ´(⇧}NÑ. ÂÂ≥qÑU0π✏Ü↵Ñq. s. _ ©º®¶˚q˝ º(⇧Ü™ ∑Ä. (⇧Ü™´⌃^∫!¯‹. –.↵¶⌦/⇤x<↵MÑ. ⌘⌘. ⇣ k = 10. œM(⇧. 20. ‡d. F˚qÔ˝‡d. ˝®˙Ù„€ÑLÚ º ŸÜ™/Ù}Ñ⇥∞N¶1–õÜ (ÜU0⇡↵®¶˚q/& T0„€®¶ÑHú⇥ (,f«ô∆-. }NF{. .Uœπ✏. 30 Ñ˙ºÂbÑ®. ¶⇧Æ ç)(⇡õ⌥⇡U0!ãÑ}fi⇥. ÊWPú⌃ê⌥. 政治大. 立. 學. io. P@10 0.0402 0.0552. P@20 0.0337 0.0430. P@30 0.0302 0.0372. P@10 Improvement 37.3%. 0.0528 0.0522. 0.0438 0.0425. 0.0387 0.0356. -1.1%. 0.0407 0.0775. 0.0333 0.0614. 0.0298 0.0525. n. al. Ch. engchi. y. ‧. Nat. Method Deepwalk Superhighway+ Deepwalk LINE Superhighway+ LINE HPE Superhighway+ HPE. sit. ñ∫áh˛. ‧ 國. 4.4.1. ÷. 90.4%. er. 4.4. i Un. v. h 4.3: ñ∫áh˛. 1h 4.3-⌘⌘ÔÂóÂ 20 —2. ( Deepwalk ⌥ HPE -. ⌘⌘Ñ˙. π’( k = 10. 30 Ñ h ˛ ˝ o 0 › N ü , Ñ π ’ $ v ( HPE - ⇣ w E ¶ • _h:⌘⌘Ñπ’(–.↵¶⌦ º®¶ÑHú öÑ–G F. /( LINE Ñπ’⌦. ⌘⌘Ñ˙. VeÑh˛÷6í. –G. FÓp&í. à. o⇥. 4.4.2. Ïfiáh˛. ⌥ñ∫á⇣>^<. (h 4.4-⌘⌘ÂS. ( Deepwalk ⌥ HPE -. ⌘⌘Ñ˙. π. ’( k = 10 20 30 Ñh˛_˝ o0›Nü,Ñπ’ ( HPE - ⇣wE ¶_ÖN 7 ⇣ N #0 ( LINE Ñπ’⌦ ⌘⌘Ñ˙ VeÑh˛/↵M Ñ. NÓp&í à o⇥ 21. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(29) Method Deepwalk Superhighway+ Deepwalk LINE Superhighway+ LINE HPE Superhighway+ HPE. R@10 0.0528 0.0670. R@20 0.0540 0.0646. R@30 0.0576 0.0681. R@10 Improvement 26.9%. 0.0660 0.0628. 0.0693 0.0675. 0.0782 0.0702. -5.1%. 0.0518 0.0889. 0.0543 0.0915. 0.0595 0.0978. 71.6%. h 4.4: Ïfiáh˛ ⌘⌘|˛( Deepwalk HPE Ñπ’K↵. 1⌦bi.U0π✏. )(⌘⌘. Ñ˙ Ve Ü˝ óñ∫á⌥ÏfiáÑh˛⌦G ⇡_/⌘⌘@ ÖÑP ú h:)(⌘⌘π’®¶˙ÜÑLÆ˝ Ù}0⌥(⇧ÑO}f˜÷˙Ü T0↵∫ h˛&í. 政治大 ®¶ÑÓÑ⇥ N<ó¢ Ñ/ LINE (†e⌘⌘Ñ˙ Veå 立 ä} Õ /ÆE↵M Ô˝Ñü‡/⌘⌘ç∫flœ↵π’x“Ñ. 學. (6ìÑäÕ. ⇣∫‹⌦. Ñ Deepwalk fl HPE. Ê. πb. ì0ieÂåÑ¿fi. y. ø˝✏NÖ#PÑ¿fiä. sit. sG ñ∫á G<ah˛. iv l C n h e n g cMAP@30 h i U MAP@10 Improvement MAP@10 MAP@20 n. Method Deepwalk Superhighway+ Deepwalk LINE Superhighway+ LINE HPE Superhighway+ HPE. ✏N®_Jpπ✏. er. io. 4.4.3. 1º˝. s. Nat. x“Ù⇢⌦o⇥. óx“Hú. ‧. ‧ 國. π✏ ‹⇢⌘⌘Ñ˙ Ve@˙˙ÜÑ Úì ó/ÆÑå⌃ ()(å é¯<¶Zx“Ñ LINE ⌦⇤‡∫Í˝ì0ieÊÛÑ¿fix“ ó(6 —. 0.0295 0.0346. 0.0254 0.0276. 0.0246 0.0253. 17.2%. 0.0358 0.0367. 0.0304 0.0292. 0.0291 0.0266. 2.5%. 0.0292 0.0509. 0.0262 0.0410. 0.0252 0.0379. 74.3%. h 4.5: sGñ∫áG<h˛ 1h 4.5-⌘⌘ÔÂ↵0 ≤Ôh:’⌦. k = 10 ÑB⇡. ⌘⌘Ñ˙. h˛˝‘ü,Ñπ’Å}⇥ HPE. π’((⌦. Í(ñ∫á. Ïfiá⌦. WÑ–G (sGñ∫áG<h˛⌦_à} h: HPE (⌘⌘˙ ↵ Í˝ Ù& (⇧Ñ B~0 £ÑLÚ ⇡õ ¯‹ 22. . o. π’Ñk© ÑLÚ_'. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(30) ( Deepwalk Ñh˛⌦÷í. ⇢˝⇤*H˙˛(LÆMµ. 'E¶⇣w. F_. ö↵¶Ñ⌦G y%ÅË✏Ñù6/ LINE ⇥(MµÊWPú- ⌘⌘ÚìÂ S LINE ((⌘⌘Ñ˙ Ve↵Õ /⇤↵MÑ F(sGñ∫áG<h˛ ⌦G. sÑh˛. ñ∫á⇣>. ⇢Û( k = 10 B. @⌦G. ✏EÑ⌦G. F(íè⌦Ñ/⇤. h:⌘⌘Ñπ’÷6!’. ÆEÑ9o. (®¶˚q-. 6Ü™ ÕÅÑø/ ÷@ £ÑLÚ⇤(LÆMb1˙˛ Ë⌃ ⌘⌘Ñπ’Ü˝®¶LÆ ⇤}Ñ¡Í⇥. (⇡. ∞N¶h˛ Novelty@10 0.666 0.710. Novelty@30 0.670 0.726. 0.690. 0.700. 0.631 0.710. 0.746 0.613. 0.750 0.622. 政治大 0.587 0.617. 立. y. Nat. h 4.6: Top-N ∞N¶h˛. VeåÑ!ã. (∞N¶⌦Ñ. er. io. (h 4.6-⌘⌘⌫˙ü,Ñ!ã⌥†e⌘⌘˙. 0.750 0.632. ‧. ‧ 國. Novelty@20 0.671 0.721. 學. Method Deepwalk Superhighway+ Deepwalk LINE Superhighway+ LINE HPE Superhighway+ HPE. sit. 4.4.4. _‡d. (. h˛⇥Ç⌦ 4.3‡¿-@œÑ. n. ∞N¶;Å/(,fd!ã(®¶⇤∫∑˚ a iv l C Ÿ-K@ÂÕÅÑü‡∫⇢÷6⌘⌘(}Û ¢"∞LÚÑ˝õ ⇡(Û⇥2A n h e nF‘wF}Úì}NÑLÚ ⇥ÑB⇡æ⌘ºF}⌘⌘úaÑÛ⇥ ⌘⌘⇤ÙÛ gchi U Å}0í ¯. }N. F//&. Ñ®¶LÆÜ™. ↵˝. WÙ⇢#. ⌘⌘ú}ÑLÚ⇥‡d. ºi.sGñ∫áG<. ∞N↵¶⇤ÿÑLÆ_„hW(&. ÑLÚ. _˝. (⇧ú}ÑP6K. ö↵¶⌦k©(⇧¢"v÷*ÂÑLÚ⇥Ç. ⌦h@U:Ñ ( Deepwalk fl LINE ⌦ ⌘⌘Ñπ’(∞N¶⌦˝›Nü,Ñπ ’ F( HPE ⌦ ⌘⌘Ñπ’(∞N¶⌦Õ h˛ s⇥⌘⌘ç∫ü‡/‡∫∞ N¶(–.↵¶⌦/⌥ñ∫á aÑ‹o‹¬. Ïfiá sGñ∫áG<. ãÇ⌘⌘ÛìsGñ∫áG<˝–G. .⇡ñ. ↵¯ís. F–GÑq. 1h:LÆ. g(⇧í }NÑL !¯‹ ‘á1⇤↵M _1 ⌘(⇧¢" ”UÑ _⇤⇥ HPE Ñh˛1h:W⇡ fi LINE ÷6(ñ∫áIU0⌦H˛↵M® ‚. F(∞N¶⌦{›Nü,Ñπ’. U0⌦h˛›Nü,Ñπ’ ú. Â⌦U0⇡ñ. ‰⌘⌘Z. Ñ/ Deepwalk. ≈(. ⇧. (∞N¶⌦h˛_⇤s⇥. ⌘⌘ç∫⌘⌘Ñπ’˝. ≤Ôh:’x“π’⌥≤Ôx“ÑÙ}. & 23. ✏NÖ#PÑ˙äÜk©⌅^. H˜÷(⇧O}Âó0Ù. ¡Í. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(31) Ñ≤Ôh:’ÜZ®¶⇥. 4.5. Êã⌃ê. 4.5.1. Êã ⌃ê-Â Â® ¶ ˚ q ∫ ã. (d✏¿&˝. ⌘⌘✏NÊõÑãPÜU:⌘⌘Ñ®¶˚q(ÊõÑ®¶OL⌦/. & ⌘⌘Ñ⇣. Â ⌦π’å†e⌘⌘Ñπ’(x“⌦ÑPú⇥. ∫ÜÅ⌃ê⌘⌘®¶ÑPú/&&. ⌘⌘Ñ. ⌘⌘ºÓ⇡«ô∆-⌘x. (⇧ &)(÷ÑwÚF} ⌅Ü˙À®¶Ñ⇧Æ⇥h 4.7⌫˙⌘⌘ π’Ñ!ã@®¶˙ÜÑPú⇥9⁄⇡ (⇧÷⌘wÚ ⌅Ü®∑. Ü ( ⇡. 治政 ˛⇥Ú 大⌘⌘⇣. (⇧àúaF}80t„ÑAL . ⇤®¶÷ 80 t„ÑAL. / ˛LÚ⇥ 立. ®¶˚q_…r. (ü,Ñπ’®¶↵. LÆ·b_˜. ⇡_/&. ⌘⌘. Ñ®¶⇥1dÊãÔÂÂS⌘⌘Ñπ’ÔÂ˙º. N˛Â w˚ÂXÜóLÚÑyµç. y. Nat. io. sit. Êã ⌃ê-Â Â≤ Ô h : ’ x “ ∫ ã. er. 4.5.2. (≤Ôh:’·b. ‧. T. ‧ 國. ˛⇥⇠. 學. ‹Üœ/ For Your Entertainment Oh My My ⇡^˛„LÚ ⇢Û/ùªWÑL Ú Princess and the Pea II ⌘⌘Ñπ’G~N˝⇤®fi 80 t„ÑALLK . al. v. n. (⌘⌘Ñ®¶˚qK- ≤Ôh:’x“Ñ≈¡⇤àÙ•ÑqˇLÚÑ®¶H ú º(6Ñ≤Ôh:’_˝9⁄ Å2L¯‹Ñ⌃ê≈É⇥✏Nh:’x “ ⌘⌘. Ch. ⇢. • ˝ ✏NÜê«⌦ÑÂX •. Â↵. ºLÚÜ™ h:✏. engchi. i Un. eÜ„zÓ⇡«ôÑ. è'OL. ®<^<ÑLÚ…rÅä¯—ä}⇥_1/. (6ìT. LÚÑ≤Ô. N˛Ñyµ(·b. ⌘⌘( Selena Gomez Ñ Fetish ⇡ñLZ∫ÂbÑLÚ. ›. π✏↵. Ñ≤Ôh:’2L⌃ê⇥ h 4.8 ⌥h 4.9∫⌘⌘)( Fetish ⇡ñLÑ≤Ôh:’( Ñπ’ï↵2L ⇠&¯<¶Ñ⌧↵@~˙ÜÑ 10 ñL ÔÂ↵0 (ü,Ñπ’K- dÜ~ 0Üœ Demi Lovato. Ariana Grande ⇡^⌥ Selena Gomez ¯<ÑsLK®<ÚK. <NÑ˜‹Ü-áí≈. ”áAL⇢Û/R 24. LÚI®<ÂpÑLÚ. h:. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(32) Method Deepwalk. LINE. sit. n. al. er. io. Superhighway+ LINE. Superhighway+ HPE. ‧. Nat. Superhighway+ Deepwalk. 學. HPE. ‧ 國. 立. 政治大. LK George Michael Guns N’ Roses George Michael LK Essential - 80’s Bon Jovi Scorpions Drowning Pool Adam Lambert Backstreet Boys JTR Backstreet Boys Elvis Presley Sugar Ray Drowning Pool Bon Jovi Bon Jovi The Offspring Sesame Street Europe Peter Cetera Bonnie Tyler Phil Collins Modern Talking George Michael Europe Eagles Europe George Michael Europe Air Supply Phil Collins Bon Jovi Air Supply. y. (⇧F}⇠⌅. LÚ 1 Careless Whisper Sweet Child O’ Mine Wake Me Up Before You Go-Go LÚ 1 Can’t Fight This Feeling Livin’ On A Prayer Rock You Like A Hurricane Bodies For Your Entertainment Roll With It Oh My My Nobody But You I Believe Every Morning Bodies I’m With You Never Say Goodbye Pretty Fly for a White Guy Princess and the Pea Carrie Glory Of Love Total Eclipse Of The Heart Against All Odds Cheri Cheri Lady Last Christmas Carrie Desperado The Final Countdown Fastlove Carrie Without You Against All Odds Always Making Love Out Of Nothing At All. Ch. engchi. i Un. v. h 4.7: ®¶˚qKÊã. 25. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(33) LÚ 1 Fetish LÚ 1 My Everything Cry Baby Dark Side q ✏◆}⌧Ë Eye of the Needle Reaper Unstoppable NOMAD Rainbow 2! 3! Issues Chained To The Rhythm Cry Baby Rainbow Neon Lights Same Old Love Ikuyo Who Says Bed Impossible Cry Baby Rainbow Dark Side Eye of the Needle q ✏◆}⌧Ë NOMAD Mr Lonely (feat. Fat Lip) My Everything 2! 3! Stone Cold. n. engchi. y. i Un. h 4.8: ≤Ôh:’x“Êã. 26. sit. io HPE. Ch. ‧. Nat. al. 政治大. 學. LINE. ‧ 國. 立. LK Selena Gomez LK Ariana Grande Demi Lovato Bishop Briggs Xun Sia Sia Sia Walk Off The Earth Kesha BANGTAN BOYS Julia Michaels Katy Perry Demi Lovato Kesha Demi Lovato Selena Gomez Kyle Selena Gomez Nicki Minaj Shontelle Demi Lovato Kesha Bishop Briggs Sia Xun Walk Off The Earth Portugal. The Man Ariana Grande BANGTAN BOYS Demi Lovato. er. query Method Deepwalk. v. K. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(34) LÚ 1 Fetish LÚ 1 Bad Liar The Way I Are (Dance With Somebody) Sober Crying in the Club Good For You Woman Call It What You Want Kill Em With Kindness The Heart Wants What It Wants Remember I Told You Bad Liar The Way I Are (Dance With Somebody) Sober Call It What You Want Slow Down Down The Heart Wants What It Wants Crying in the Club Woman Gorgeous Crying in the Club Bad Liar The Way I Are (Dance With Somebody) Back to You Friends I Got You Swish Swish ...Ready For It? Down Gorgeous. 立. io. n. Superhighway+ HPE. y. er. Nat. al. ‧. ‧ 國. 學. Superhighway+ LINE. 政治大. LK Selena Gomez LK Selena Gomez Bebe Rexha Selena Gomez Camila Cabello Selena Gomez Kesha Taylor Swift Selena Gomez Selena Gomez Nick Jonas Selena Gomez Bebe Rexha Selena Gomez Taylor Swift Selena Gomez Fifth Harmony Selena Gomez Camila Cabello Kesha Taylor Swift Camila Cabello Selena Gomez Bebe Rexha Louis Tomlinson Justin Bieber, BloodPop R Bebe Rexha Katy Perry Taylor Swift Fifth Harmony Taylor Swift. sit. query Method Superhighway+ Deepwalk. Ch. engchi. i Un. h 4.9: ≤Ôh:’x“Êã. 27. v. Kå. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.

(35) (x“h:’ÑB⇡Ñ/⇤. ö↵¶Ñ◊0«ô. è'qˇ. ⇤. OÓ. ⌘⌘. Ñπ’@~fiÜÑG~N˝/®<¯—ÑLÚ ⇢Û˝/s'LKE⇢ ⇡‘⇤ & ⌘⌘⇣ ÑPú _IÊÜw˚Üê«ôÑÂX0Ó⇡«ô/ ©º(Ó⇡ «ô⌦Ñx“⇥. 立. 政治大. ‧. ‧ 國. 學. n. er. io. sit. y. Nat. al. Ch. engchi. 28. i Un. v. ΍ΘΒͧкйͨпскнͿΗΌΌΞлйктйклйк.