[email protected]
[email protected]
!" # $ % !& ' ( ) * " + , % - ' ( ). / 0 1 2 3 4 5 678 9: ; < = > ? ( @ A-B & C D8 9E F G C H I J K L M * NO P QR S 4 5 67T U ! & V W X H Y Z[ \ 67 ] ^ _-` ,Aa b c d e f V g h i j k ]l m n o pqr )e f s t [ u v i w 67 ]x y / i z qr )e f ^ _{| }C ~ l r \ [ \ e f ] -Aa b s t }H 4 (profile)[ \ l m p e f { }H }H | }8 9 (Item-based)n o p { D8 9 c j k ]8 9 n o ¡ / 8 9l m ¡ ¢ £ * ¤" [ , j e f -` Aa b ¥ ¦ §(Precision)¨© ª §(Recall) / F1 « ¬ ¥ ® ¯ a b * Nr \ | } * N[ \ ° ±-²³ k ´ µ ¶ Aa b 3 i j s t µ Q· C k ]l m ® n o p ¨q ¸ p / r \ ¹ \ | } -` º e f V ¨n o p¨l m p¨' ( 67 ]Abstract
With the rapid development of information technology and popularization of the internet, traditional job-finding services have declined and replaced by the on-line job service. Job seekers can query job vacancy information they like easily and economically by using on-line job finding service. However, due to huge job vacancy information, job seekers still have to examine those jobs manually or through matching system which provides job matching service based on resumes.
In this paper, we introduce recommender system concept, propose a hybrid model-based recommender that achieves better performance of job matching and provides personalized service. We use data from a famous domestic job hunting company to evaluate our model. Our hybrid model first processes content
filtering using user profile to produce candidate recommendations. We also partition users based on submitting resume records. Then, we use item-based collaborative filtering to compute item similarities and content item similarities. Finally, we find a regression formula to combine the similarities to produce top N recommendations.
The experimental results show that our system can improve the recommendation performance significantly. Besides, our system also achieves better performance when comparing with other combining approaches.
Keywords: recommender system, collaborative
filtering, content-based filtering, online job hunting
1.
» " ¼ ½ ¾ e)¿ F ^ _À Á  ( à / Ä )Å Æ y Ç È É : ; R S ( = > ? ( @Ao ' ( ) * " y ® Ê Ë Ã / Ì Í i w Ê Ë Î Ï §-,G C r Ð Ñ I Ò Ó ²ÔÕ Ñ ^ _( Ö< × Ø À Ù - r Ð Ñ Ú U »Û Ü Ý Þ ß à á 3 â r ã ä @ å æ C »ç è G C á I L M é ê ëF > ? C »Ø Û Ü ß à D,G C H I Bì í ê > ? î ï ð ñ Do ëí ê = ¯ R S 3 4 D8 9-Å Æ r Ð Ñ ò À Á  ( Bì G á I y : ; ß à é ê o ¼ i z ëF Dó ô · õ r D, H Æ ò y ö Ú * ÷ øù @Aú û ü D/ : ; ý þ »Û Ü = @A- y Ç È ÷ 1 2 3 â 67B ¼ Å Æ ä @Ç È @ -Ö@ ( D }H u u p j r y y Å Æ ä @ - -Ö C Æ Aa b c d e f V g h k ]l m pn o p j u e f s t {/ Æ s t i z qr )6 7 ] e f ^_-2.
G C 67 ]. / u p¤-Å Æ ÔA : ü ! / p -` `
2.1
! (Information Retrieval) " U }H # d ü $ % ¨& ' T U ( à / * " p }H ) Ò * -² # d ü $ % T & ' * " + , }C »d - . } * N/ºYahoo¨Openfind¨Google0 R ô 1 -,0 R ô 1 o2 3 ª Ø 4 G C } H I J K U 5 Ø 6 7 -, * " o2 Q{8 NÖx pj 4 J K 3 Ø 4 - ü a b [2]¼ 9 j 0 R {8 N S j } H : Ô â / 4 -; Æ % < ! * " {= i z > q r )¤§Å Æ ¿ F ? H i j p ü @ [ qr )¤§-, p ¼ + , }C e f V -`2.2
p G C ! ëA B qr )² " C Ø ºl m pn o p-l m p " U }H 4 (profile)® 9 D 8 9 ° G [ , e f E ] } H 4 8 9 [7-8]-Å Æ ¿ F ? H F G H u qI ¦ U ü $ " J V @K ® L ® U L M N QO P j }H Ò * ÖQ R ü S [1]- ,l m pT Öu ñ U V /8 N W X Y F ¨BT Ö Z [ \] y / 8 N^ _ ` Ba 4 -` & C l m p3 u ñ Å ,¿ F ? H b c i j d u e p n o p (Collaborative Filtering)[11][13]-3 f n o pg U h Ú i j k = Ì e f l- m x | n (Word-of-Mouth)o Í ) ¤[15]-n o pp / } H (User-based)q / 8 9 (Item-based)C e e f * " ² }H * " g U S j ® Ar 4 - s t u v/ Æ t u 4 à / e f -B & C }H * " L M B w x y }H z Æ = ¡ 3 / 3 { | Ø x y } Ý § -` ,8 9 * " U x y 8 9= ¡ , 8 9= ¡ G ~ C Å Æ 8 9 n o p. / ( x y 8 9= ü S s t {/ Æ s t P @0 ) e f K y -; Æ % < ëÖ¿ F a b [5][13]9 j 8 9 ° }H n o pe f ¥ ¦ -n o p° l m pT Ö Z [ \ · o2 U u ñ U V /º [ 8 9 e f (cold start)¨« c x y B / x y } Ý § -, G C l m n o p ¯ ¯ ¿ F b c ? H i j k ]l m n o * N z Æ [1][3-4][9][17]-` k ]l m ® n o p ü a b [10][16] " U 8 9 * " ¸ x y 8 9l m ¡ / n o ¡ v/ u ° à / ]-Æ * Nµ ¶ y i z ± e f [ \B = y i z ± qr )¤§/ = X k ]C * N Å J ,ì U ° ]Å Æ . y 3 ä @ ñ 8 9 § ¡ Ø / B ¡ -` Å Æ Aa b i j u k ]l m ® n o p s t - }H 4 / l m p * " e f Æ * " . S j }H 4 ¡ à ¢ e f qr )¤§-% }H à / {x y £ u 8 9n o ¡ -o2 Q * " . ¢ ) = 8 9 ü S Ô e f . i w qr )¤§-`; Æ % < ¼ x y 8 9 l m ¡ { u k ]Å J c j ¢ £ * ¤" 8 9 n o ¡ / 8 9 l m ¡ à / ] - ¤ e f + ¥ Nq® ¯ } H } 8 9[ \ e f -
3.
Aa b i j k ]l m n o pe f s t | }C 67 ]/ Æ i w ¦ ]¥ ¦ §o @ ¢ qr )e f ^ _-AV " § ¨ © Ø s ª º } H s ª ¨ }H 4 s ª ¨8 9 ¡ x y s ª / e f s ª - ` 1 `
3.1
8 9 n o pÔe f ¥ ¦ § ( ° }H n o p [5][13] U = y i z > qr )¤§-Å Æ ( G } H [ \ 4 }H « Ôu Ó ë y O ¹ j » « 8 9= ¬ ¡ ¼ Å Æ y à ¢ e f qr ) ¤§-, ü a b [6]¼ 9 j ( G }H [ \ ¦ ³ y i w e f [ \-,Aa b }H ® ¯ } k-means ° y NG }H [ \ i w qr
)¤§- k-means 2
3.2
Ô k ]l m ® n o pe f V l m p3 ± ° ç ² ³ Ú U ® n o p´ H µ ¶ 7 v/ · ]° à / ]-Å Æ {8 NÖx ¸ l m p3 T ¹ K y o ¼ = y Ö x i z ± qr )¤§-Ö C Æ Aa b }H Y Z 4 {/ Æ [ \ l m p e f Æ * " . / Öx º » }H 4 ¡ / Ç È · | } H 4 Ù Í {i w e f qr ) ¤§-Aa b ¤ }H Y Z ¼ + ? Z¤§¨67! ¨67Q q _ ¸ 7 }H 4 % ¯ }l m p½ & ' ® D8 9l m [ \ ° G p }H e f -² l m p½ /¾ º` ` ¿À }H ? Z¤§L M Á C T C D8 93 â ? Z¤§`ø`  À }H 67! L M Ø C T C D8 9 3 â 67! `ø` à À }H Ä 67Q L M Ä C D8 967Q ¡ l `ø` Å À }H Ä _ ¸ « ]8 9L M " ® D8 93 â _ ¸ « ]8 9Â Æ E ]u 8 ` 3
3.3
}H 4 s ª . / ( + Ç } H 4 ¡ @ ¢ qr )¤§ g /ou È l m p J K . y 3 ª Ø 4 / 8 NÖx Z 8 9= [ \ ¬ ¡ -Å Æ Aa b 8 9 (Item-based) n o p x y 8 9 = ¡ [ u v Z e f 8 9= [ \¬ và ( x y 8 9l m ¡ à ¢ 8 9= ü S -Å Æ ÔÆ s ª }H s ª » É D Æ {¯ }cosine similarity° y N[14] x y j » 8 9 ¡ {Ê ( u Ë Ì Å Í Î ù Å ±Æ , ±Á ¡ ²Ï Ì Ð " /¾ º 2 2 1 ij ik C jk i ij ik i i r r n m n r r × =
∑
× +∑
∑
(1) ² mCjkO ¶ Ô « C 8 9jÑk ¡ , 8 9rijqrik }H uiU L Ò Y ZÓ0 = Ò Y Z¿ Ò Y Zøui∈C, nO ¶ 8 9 j qk Ô « Å Æ x y j § ± nÓ3 j ± §$ % G ±Á §-; x y 8 9 ¡ < L M x y j » 8 9 Õ / 7 e f s ª x y }H ® » = ¤§-²Ð " /¾ º` ` j k k u C ji C i k r r C ∈ =∑
(2)` ² Ck i r 8 9iÔ « k Õ Ó,Ck O ¶ « }H q rjiO ¶ }H jU L ÖÒ 8 9i Y ZÓ0 = Ò Y Z¿ Ò Y Z- ; < n o* " x y j 8 9 ¡ % < o 2 Q¼ x y j 8 9l m ¡ -Aa b D8 9 67 _ ¸ / 67Q 7 x y ¡ ¥ -²Ð " /¾ º` `( )
1 jk jk jk CM =DM × +λ LM × −λ(3) ² CMjk 8 9jq8 9k l m ¡ DMjk 6 7 _ ¸ 3 x y j ¡ jk LM 67Q 3 x y j ¡ λ ]C l m ~ ¡ ° Ó-² DMjk / ¾ Ð " x y ¡ º 1 1 n i i i jk n i i D W DM W = = × =
∑
∑
D =
{ }
0,1(4) nO¶89j q 89k3 Ö ~ 6 7 _ ¸ × « q ,D =i 1O¶C 89T Ö o _ ¸ $ % D =i 0 B o -Wi O¶Ø _ ¸ Ù ×
Ó,A a b Ú % Û Ü (text mining)Ú }
TFIDF[12]g h Ý jWi Ù × Óº 2 log ( ) i i N W idf O = =
(5) N O3 Ö D 89q ,O O¶Ø _ ¸ Þi q Å Æ Ð " (5). / i w 89= o ¸
ø ± Æ j¹ C D 89 ¡ $ % . / Î ù o ¸ ø ± Ú j¹ C D 89 ¡ - C LMjkx y ¡ * " & C ¤ a b Ç } H 3 Ä ` 6 7 Q ® ß I 8 96 7 Q B u Ï à á ] Ø F è Ô o u Q â lÅ Æ A a b ¾ã Ð " x y j ¡ º 1, 0.5, 0, jk if LM if if = (6) C D 89 6 7 Q Ï à o (/ºòä å Ø æ â) ç è 1 § ,Ö ~ C o u âé (/:ò ä å ê â® òä å Ø æ â) ç è 0.5 § , Ï à B E ë (/ºòä å ® Á ì å ) ç è 0 - 4
3.4
} H 4 s ª / 89 ¡ x y s ª 3 ( ü s t e f s ª . / 0 ) (on-line) * " [ \ e f - e f s ª } H ® x y j } H ® » = ¡ 7 í Ù × J î ² Ð " /¾º` ` 2 2 ( , ) ( ) k k C ai i a k i C ai i i r r sim u C r r × =∑
∑
(7) ² sim u C( , )a k } H a® k ¡ rai } H aU L Ö Ò Y Z Ó0= Ò Y Z 1 Ò Y Z - Ck i r Ð " (2)3 x y j k ´ -+ Ç } H ® » ¡ % ¤ } H e f + je f 89 ® } H ® « [ \ n o Þ q x y ² Ð " /¾º ( , ) ( , ) ( , ) k C ip a k k a p i a k k m sim u C CF u r sim u C × =∑ ∑
∑
(8) ² CF u r( , )a p O¶ } H a G C e f 89 p n o ¡ Þ q ,i } H q k q Ck ip m Ð " (1)3 x y j ¡ ( , )a k sim u C Ð " (7)x y j ¡ -; x y n o ¡ Þ q < ¼ x y lm ¡ Þ q à ¢ 89= ü S ² Ð " /¾º ( , ) n ip i a p CM CB u r n =∑
(9) ( , )a p CB u r O¶ } H a G C e f 89 p l m ¡ Þ q ,n } H q CMip Ð " (3)3 x y j lm ¡ - A a b D 89 c jk ] 89n o ¡ / 89lm ¡ ¢ £ * ¤ " { / Æ * ¤ " j ï ¡ ² Ð " /¾º , , ( , )a p ( ( a p), ( a p), p) CBCF u r = f CB u r CF u r N (10) ² N 89 - x y j3 Ö e fp 89 ¡ { / Î ð * " ñ P + Ç ¥ N q 7 Top-N e f - !"# $%&'("# )*&'("# + ,-./ + 5
4.
4.1
A a b 3 } ã ~ l r \ ò -² § ¨ } H Y Z ¨ D 89 lm / } H -A a b G £ Ä } H Ø C 10 / D 89 Ø C ó [ \ a b « ¬ ² Þ r 796 r, D 1431 Þ 15656 -A a b { ô + 5-fold cross-validation * " õ @ ö { £ u ÷ + u N 7 ø ù « (test set) ² ú © N 7 H e f s t û ü « (training set)[ \ ö ÷ « ¬ , G £ u ÷ A a b ¤ ø ù « J ý £ Ä } H = + Ç ¥ 3 (Given 3)7 e f s t # d ² þ ú } 7 C « ¬ s t -`4.2
A a b ¥ ¦ § (precision)¨© ª § (recall) / F1-metric G s t [ \ « ¬ -² ¥ ¦ § 3 Ö e f 89 } H ß I 3 4 ° © ª § I ¦ e f 89 3 Ö } H 4 5 89q ° -,& C ¥ ¦ § ® © ª § Ô A \ ( T Ö Ô(trade-off) ¡ \ @ à e f q Ç ¥ ¦ § ¾ Î ,© ª § i w -Å Æ A a b } F1-metric ç è ¥ ¦ § q © ª § o Ù × [ \ « ¬ ² Ð " / ¾º` _ _ _ _
correctly recommended items precision
total recommended items
=
(11)_ _
_ _ _ _
correctly recommended items recall
total items liked by users
= (12) 2 1 recall precision F recall precision × × = + (13)
4.3
&Cnop{ =i z > q r ) e f¤ § Å Æ A a b ( G } H Ã / -, < + Ç A a b 3¨5¨10¨ 15¨20 / 5q e f [ \³ « ¬ -O 1µ ¶ 10 (k=10)² e « ¬ D ¥ P a Ø Ó -Å Æ A a b / 10 ® ² þ * N [ \° ± -, 6O¶ A a b ³ j 8 9 ® no] ° ü S { / Æ c j k ] l m ® no ¡ ¢ £ * ¤ " -; Æ % < Ð " (3) λ Ó ³ k ´ µ ¶ Ó λ =0.11y k ´ Å Æ A a b / Æ Ó [ \³ « ¬ -` 1 k Precision Recall F1 3 0.2113 0.0639 0.1060 5 0.2017 0.0612 0.1022 10 0.2118 0.0640 0.1061 15 0.2085 0.0630 0.1051 20 0.2070 0.0626 0.1046 0.83 0.88 0.93 0.98 0 10 20 30 40 50 6 , ¤ O2 . / 5 µ j A a b 3 i j s t µ · C² þ * N ² k ] n o® l m * N 0 U no ¡ ® l m ¡ / u ° Ã / ] (A a b î ü a b [10][16][ \³ ² k ´ no° 0.8 e f[ \ )/ Æ 7 e fJ î -, 8 9 n o p ì U x y n o ¡ [ \ e f{ =k ] ë l m p-Â Cl m p U 8 9 l m ¡ x y + Ç e f-; Æ % < A a b { ® ~ l r \¹ \e fõ V / Ú } e f[ \° ± -&Cr \e fõ V U } H Ä Y Z & ' [ \p{ î D = / Î * " e f ( e f î ¤ § ñ P )Å Æ e f q } H Ä Y Z & ' } Ý § , -Ô Æ A a b e fq NÃ / Ø ® ¹\* N ° ± -o2 Q ¤ O3 . / µ j A a b s t . / Ø i e f[ \ { ø y Ô ¥ q e f8 9 + Ç ± ¥ ¦ § Å Æ G C } H Ô + Ç e f y ë Ç Ç a 3 â , 8 â Ø 4 ª - 2 !!!! """" #### A a b Top N precision recall F1 1 0.2651 0.0161 0.0351 3 0.2311 0.0420 0.0811 5 0.2118 0.0640 0.1061 7 0.1976 0.0839 0.1248 10 0.1755 0.1063 0.1364 k ] no® l m * N ` Top N precision recall F1 1 0.1645 0.0099 0.0242 3 0.1658 0.0302 0.0606 5 0.1550 0.0469 0.0833 7 0.1481 0.0626 0.0985 10 0.1395 0.0843 0.1134 Item-based nop` Top N precision recall F1 1 0.1481 0.0090 0.0220 3 0.1466 0.0267 0.0549 5 0.1407 0.0424 0.0752 7 0.1342 0.0567 0.0902 10 0.1239 0.0748 0.1017 l m p` Top N precision recall F1 1 0.1143 0.0069 0.0162 3 0.1139 0.0206 0.0411 5 0.1055 0.0318 0.0544 7 0.0967 0.0408 0.0629 10 0.0878 0.0527 0.0697
3 $$$$ %%%% &&&& '''' (((( """" #### A a b Top N precision recall F1 5 0.2118 0.0640 0.1061 10 0.1755 0.1063 0.1364 20 0.1438 0.1743 0.1564 40 0.1104 0.2665 0.1504 50 0.0999 0.3014 0.1438
r \¹\e fõ V ` Top N precision recall F1 5 0.0819 0.0248 0.0416 10 0.0721 0.0435 0.0556 20 0.0604 0.0726 0.0636 40 0.0459 0.1101 0.0612 50 0.0417 0.1249 0.0588 e f` Top N precision recall F1 5 0.0952 0.0288 0.0406 10 0.0795 0.0481 0.0544 20 0.0699 0.0844 0.0689 40 0.0594 0.1435 0.0767 50 0.0561 0.1694 0.0775
5.
A a b i j q r) e fs t Ö x Q } l m ® nopC H Ì ¡ Ã / k ] B ì Ö x Q | } C' ( 6 7 ] õ V o ¼ · Ck ] l m ® nop-² ü $ Ô CA a b } H 4 Ä } H 4 ¡ Ã ¢ l m p( q r) ¤ § / } } H Ã ¢ nop( q r) ¤ § u ¢ £ * ¤ " Æ C H Q k ] -` &CA a b G Ck ] no® l m ¡ ì ì U D 8 9 7 ] J î Ö X Ci w e f[ \ x ´ Ö U 3 / G C= a b * B . d ë F l m Ù / º D t 3 ° ¨ } H / Æ S j z Æ = ü S [ , c j ] * ¤ "
[1]M. Balabanovic and Y. Shoham, "Fab: Content-Based, Collaborative Recommendation," Communications of the ACM, vol. 40, pp. 66-72, 1997.
[2]K. Bradley, R. Rafter, B. Smyth , Case-Based User Profiling for Content Personalization, Lecture Notes In Computer Science, Vol. 1892, 2000, pp. 62-72.
[3]M. Claypool, A. Gokhale, T. Miranda, P. Murnikov, D. Netes, M. Sartin, “Combining content-based and collaborative filters in an online newspaper”, In Proc. ACM-SIGIR Workshop on Recommender Systems, 1999.
[4]J. Delgado, N. Ishii and T. Ura, “Content-based Collaborative Information Filtering : Actively Learning to Classify and Recommend Documents”, In Proc. Second Int. Workshop, CIA'98, pp.206-215, 1998.
[5]M. Deshpande and G. Karypis, “Item-based top-n recommendation algorithms”, ACM Transactions on Information Systems, 22(1):1–34, 2004.
[6]C.S. Hwang and P.J. Tsai, “A Collaborative Recommender System Based on User Association Clusters”, WISE, 463-469, 2005
[7]B. Krulwich and C. Burkey, “The InfoFinder agent: Learning user interests through heuristic phrase extraction”, IEEE Intelligent Systems Journal (Expert), vol.12, no. 5, pp. 22-27, 1997.
[8]K. Lang, “NewsWeeder: Learning to filter netnews”, In Proceedings of the Twelfth International Conference on Machine Learning, San Francisco, CA.Morgan Kaufman, pp. 331-339, 1995.
[9]P. Melville, R.J. Mooney and R. Nagarajan, “Content-boosted collaborative filtering”, In Proceedings of the SIGIR2001 Workshop on Recommender Systems, New Orleans, LA, September 2001.
[10] B. Mobasher, X. Jin and Y. Zhou, “Semantically Enhanced Collaborative Filtering on the Web”, Center for Web Intelligence School of Computer Science, Telecommunication, and Information Systems DePaul University, Chicago, Illinois, USA.
[11] P. Resnick, N.M. Iakovou, Sushak, P. Bergstrom and J. Riedl, “GroupLens:An open architecture for collaborative filtering of netnews”, In Proceedings of the 1994 Computer Supported Copperative Work Conference, 1994.
[12] G. Salton and C. Buckley, “Text Weighting Approaches in Automatic Text Retrieval”, Cornell University Technical Report 87-881, 1987.
[13] B. Sarwar, G. Karypis, J. Konstan and J. Riedl, “Item-based collaborative filtering recommendation algorithms”, In Proceedings of the 10th International WWW Conference, Hong Kong, May 2001.
[14] B. Sarwar, G. Karypis, J. Konstan and J. Riedl, “Analysis of recommender algorithms for e-commerce”, In Proceedings of the 2nd ACM E-Commerce Conference (EC’00), Minneapolis, MN, October 2000.
[15] U. Shardanand and P. Maes, “Social Information Filtering: Algorithms for Automating ‘Word of Mouth’”, Proc. CHI 95, ACM Press, New York, pp.210-217, 1995.
[16] K. Tso and S.T. Lars, “Attribute-Aware Collaborative Filtering”, Computer-based New Media Group(CGNM), Iinstitute for Computer Science, University of Freiburg, 79110 Freiburg, Germany. 2004.
[17] A.M.A. Wasfi, “Collecting User Access Patterns for Building user Profiles and Collaborative Filtering”, In Int. Conf. on Intelligent User Interfaces, pp.57-64, 1999.
