reD-B61.dvi

Size: px

Start display at page:

Download "reD-B61.dvi"

えみかに
7 years ago
Views:

1 D 211 Original Paper Customer Behavior Prediction System by Large Scale Data Fusion in a Retail Service Ishigai Tsuasa Taeshi Taenaa Yoichi Motomura Tohou University. isg@econ.tohou.ac.jp, isg/ National Institute of Advanced Industrial Science and Technology taenaa-t@aist.go.jp, y.motomura@aist.go.jp, eywords: service engineering, large scale data modeling, latent class model, ID-POS data, Bayesian networ Summary This paper describes a computational customer behavior modeling by Bayesian networ with an appropriate category. Categories are generated by a heterogeneous data fusion using an ID-POS data and customer s questionnaire responses with respect to their lifestyle. We propose a latent class model that is an extension of PLSI model. In the proposed model, customers and items are classified probabilistically into some latent lifestyle categories and latent item category. We show that the performance of the proposed model is superior to that of the -means and PLSI in terms of category mining. We produce a Bayesian networ model including the customer and item categories, situations and conditions of purchases. Based on that networ structure, we can systematically identify useful nowledge for use in sustainable services. In the retail service, nowledge management with point of sales data mining is integral to maintaining and improving productivity. This method provides useful nowledge based on the ID-POS data for efficient customer relationship management and can be applicable for other service industries. This method is applicable for mareting support, service modeling, and decision maing in various business fields, including retail services. 1. [ 7, 9, 9] [Ueda 9, Taenaa 1] [ 8]

motomura/ eywords: service engineering, large scale data modeling, latent class model, ID-POS data, Bayesian networ Summary This paper describes a computational customer behavior modeling by Bayesian

2 671 [ 9] [Nielsen 5, 4] 1 ID 1 ID-POS 4 (i) (ii) (iii) [ 6, Pearl 9] x 3 アンケートデータ : 商品にこだわりがある 1 : 安売りが大好き 2 : 健康に興味がある 3 : 料理が好き N x 1 x 2 季節 N 安売り潜在クラスモデルによる大規模データ融合潜在顧客カテゴリ家庭生活充実派節約消費派こだわり消費派カテゴリマイニング潜在商品カテゴリ健康的料理用高品質低価格 ID-POSデータ ID-POS データベース大規模な顧客購買履歴カテゴリベースの顧客行動予測モデル時間帯ベイジアンネットワーク 1 カテゴリ A カテゴリ B x 4 x 5 1 x 5= x 5= 2 3 商品 1 商品 2 商品 3 N 商品 M CPT of p(x 5 x 1 ) x 1= 春 X 1= 夏 x 1= 冬条件付きテーブル (CPT) [ 9] [Ishigai 1, 1b, 1c, 1d] 2. ID-POS ICT 199 ID-POS ID

理用高品質低価格 ID-POSデータ ID-POS データベース大規模な顧客購買履歴カテゴリベースの顧客行動予測モデル時間帯ベイジアンネットワーク 1 カテゴリ A カテゴリ B x 4 x 5 1 x 5= x 5= 2

3 D 211 ID-POS RFM [ 1] [Saunders 8, Xu 9] [Goodman 78, 3] [Ramaswamy 96] [Weng 3, 2] [Greenacre 6, 5] ID-POS [ 8] [ 9] [Yada 7] [ 5] [Iwata 9] PLSI [Hofmann 99, Hofmann 1] [ 11a] ID-POS PLSI [ 1a] ID-POS PLSI PLSI ID-POS ID-POS 2 2 (1) (2) (3) 3 X Y i j x i (i =1,,X) y j (j =1,,Y) U V l u ( =1,,U) v l (l =1,,V) x,y,u,v p p(x i,y j,u,v l )=p(u )p(x i u )p(v l u )p(y j v l ) (1) i j N ij l = = X i i Y N ij log p(x i,y j ) j { X Y U } V N ij log p(u )p(x i u )p(v l u )p(y j v l ) j l EM X U p(x u) U V p(v u) Y V p(y v) U p(u) (2) E = U p(u,v l x i,y j )= p(x i,y j,u,v l ) p(x i,y j ) p(u )p(x i u )p(v l u )p(y j v l ) V l p(u )p(x i u )p(v l u )p(y j v l ) (3) (3) (2)

3 1 ID-POS 2 2 (1) (2) (3) 3 X Y i j x i (i =1,,X) y j (j =1,,Y) U V l u ( =1,,U) v l (l =1,,V) x,y,u,v p p(x i,y j,u,v l )=p(u )p(x i u )p(v l u )p(y j v l ) (1) i j N ij

4 673 潜在顧客カテゴリu u x 顧客 x i p(x u) p(u) 潜在商品カテゴリv l p(v u) 潜在変数観測変数 2 v y 商品 y j p(y v) M Y V j l N ij p(u,v l x i,y j ) p(x i u )= X Y V i j l N ij p(u,v l x i,y j ) X Y i j p(v l u )= N ijp(u,v l x i,y j ) X Y V i j l N ij p(u,v l x i,y j ) X U i p(y j v l )= N ijp(u,v l x i,y j ) X Y U i j N ijp(u,v l x i,y j ) X Y V i j l N ij p(u,v l x i,y j ) p(u )= V l N ij p(u,v l x i,y j ) X i Y j U 1 1, X = 3965 Y = 1 (AIC) [ 4] Para = U(X 1) + V (U 1) + V (Y 1) + (U 1) AIC (4) (5) (6) (7) AIC = 2l +2 Para (8) ID-POS ID-POS ,511,467 ID-POS N ij ID-POS ID ID 1 1 MySQL ID-POS , 3,965 [ 97,, 1, 1, 1] 5 [Goldberg 9] A 35 3 Kaiser B i q Q (i) q {1,2,3,4}

) p(u )= V l N ij p(u,v l x i,y j ) X i Y j U 1 1, X = 3965 Y = 1 (AIC) [ 4] Para = U(X 1) + V (U 1) + V (Y 1) + (U 1) AIC (4) (5) (6) (7) AIC = 2l +2 Para (8) 3 2 1 ID-POS 15 28 1

5 D L (i) L(i) 1 = {Q(i) 5,Q(i) 13,Q(i) 14,Q(i) 15 }, L(i) 2 = {Q (i) 1,Q(i) 2,Q(i) L (i) 4 = {Q(i) 3,Q(i) 18 }, L(i) 3 = {Q(i) 12,Q(i) 16,Q(i) 19,Q(i) 7,Q(i) 8 }, L(i) 5 = {Q(i) 6,Q(i) 1,Q(i) 11,Q(i) 17 2 }, }, L(i) 6 = {Q (i) 4,Q(i) 9 } Q 1 Q 11 Q 18 L i L (i) = Q (i) q L (i) L Q (i) q. (9) 1 L (i) ,965 ID-POS 4,175, p(u x i )= p(u )= p(x i )= L(i) L(i) i L(i) i L(i) i j N ij j N ij, (1), (11). (12) 3 p(u x i ) i ライフスタイルカテゴリ 6 因子 p(u) p(x u) 潜在顧客カテゴリu u x 制約条件顧客 x i p(v u) 潜在商品カテゴリv l p(u)とp(x u)をアンケート結果に基づき固定 3 v 潜在変数 y 商品 y j p(y v) 観測変数 (1) (11) p(u ) (12) p(x i ) p(x i u ) (4) PLSI

(9) 1 L (i) 4 4 2 3,965 ID-POS 4,175,441 3 3 p(u x i )= p(u )= p(x i )= L(i) L(i) i L(i) i L(i) i j N ij j N ij, (1), (11).

6 (a) p(v u 1 ) の推定結果 (こだわり消費派 ) (b) p(v u 2 )の推定結果 ( 家庭生活充実派 ) (c) p(v u 3 )の推定結果 (アクティブ消費派 ) (d) p(v u 4 )の推定結果 ( 節約消費派 ) [ 11b] [ 11] [ 9] p(v l x i ) i i (e) p(v u 5 )の推定結果 ( 堅実消費派 ) (f) p(v u 6 )の推定結果 (パパっと消費派 ) 4 p(v u ) 3 4 (3) p(u )p(x i u ) (1) (12) p(u )p(x i u )=γ i (3) p(u,v l x i,y j )= U γ i p(v l u )p(y j v l ) V l γ i p(v l u )p(y j v l ) (13) EM M (5),(6) p(u,v l x i,y j ) M (5),(6) (13) x i u y j v l p(u ) x i 3 5 Mac OS X GHz Quad-Core Intel Xeon 32GB PC U =6 V =[2,5,1,15,2,3,5] 3 AIC V =1 V =[6,7,8,9,1,11,12,13,14] 5 AIC V =12 U =6,V =12 1 U =6,V = p(v u ) u 1 u 2 u 4 u 6 u 3 u 5 i j p(u x i ) p(v l y j ) l 2 PB

7 D 211 相関係数 PB : 高価格帯商品 : 低価格帯商品こだわり消費派アクティブ消費派堅実消費派家庭生活充実派節約消費派パパっと消費派 5 ライフスタイルカテゴリ p(v u) U =6,V =12 2 p(v u ) 3 -means PLSI u p(high)= {p(high 1 ),...,p(high l ),...,p(high V )} p(low)={p(low 1 ),...,p(low l ),...,p(low V )} p(v u ) p(high) p(v u ) p(low) PLSI Perplexity F-measure [Hofmann 1] Perplexity F-measure ID-POS fold l Ñl Ñl Ñ l Train Ñl Ñ l Test Ñ Train l Ñ l Train Ñ l Test

5 ライフスタイルカテゴリ p(v u) 1 2 9 7 1 2 3 5 U =6,V =12 2 p(v u ) 3 -means PLSI 8.63 2.15 1.5 1.3.89 71 5 21 3 u p(high)= {p(high 1 ),...,p(high l ),.

8 677 Ñ l Test Ñl Train 11 Ñl Test 1 Ñ l Train Ñ l Test E l E l E l = Ñl train l Ñ train l Ñl test l Ñ test l (14) -means [?] PLSI Ñ l 12-fold U =6 V =12 -means PLSI 3.2 [ 11a] PLSI UV l E l 12-fold LoopyBP [ 6] 4 2 AIC AIC N v z z = {z 1,z 2,,z Nv } P (z) z i pa(z i ) S S N v P (z S)= p(z i pa(z i ),S) (15) i=1 z i pa(z i ) N pa(i) pa(z i ) z i N K(i) Θ S = {θ ij } S pa(z i ) j z i = D pa(z i ) j z i N ij D l BN (Θ S D) N v l BN (Θ S D) i=1 N pa(i) j=1 N K(i) (N ij )logθ ij (16) =1 AIC N v AIC BN = 2l(Θ S D)+2 N pa(zi ) (17) i=1

4 1 LoopyBP [ 6] 4 2 AIC AIC N v z z = {z 1,z 2,,z Nv } P (z) z i pa(z i ) S S N v P (z S)= p(z i pa(z i ),S) (15) i=1 z i pa(z i ) N

9 D 211 ライフスタイルカテゴリ状況変数こだわり消費派家庭生活充実派アクティブ消費派節約消費派堅実消費派パパっと消費派月上旬午前潜在商品カテゴリ月中旬月下旬昼夕方カテゴリ1 カテゴリ5 カテゴリ2 カテゴリ6 カテゴリ3 カテゴリ7 カテゴリ4 カテゴリ8 春夏夜平日カテゴリ9 カテゴリ1 カテゴリ11 カテゴリ12 秋休日冬商品特徴 PB1 保存食品果物野菜菓子ダイエット低価格帯肉主食系惣菜その他高価格帯お手軽品魚日配品調味料 PB2 PB3 国内産 ABC 6 12 Bayonet[ 3] Greedy search AIC 条件付き確率値夏夜 7 調味料 (a) 構築されたベイジアンネットワークの部分グラフ夏潜在商品カテゴリ12 夜 (b) 潜在商品カテゴリ12に関する確率値お手軽品

価格帯肉主食系惣菜その他高価格帯お手軽品魚日配品調味料 PB2 PB3 国内産 6 4 3 42 ABC 6 12 Bayonet[ 3] Greedy search AIC 6 45 58 条件付き確率値.

10 679 8 条件付き確率値休日午前 (a) 構築されたベイジアンネットワークの部分グラフ午前パパっと消費派休日 (b) パパっと消費派に関する確率値 2 8 [ 91, Michman 91] 5. ID-POS GUI ID-POS 46 PLSI [Das 7] ID-POS

11 D ID-POS [ 5] - 25 [ 4] 24 [Das 7] A. Das, M. Datar and A. Garg Google news personalization: scalable online collaborative filtering, Proc. 16th international conference on World Wide Web, pp , 27 [Goldberg 9] L.R. Goldberg, An alternative description of personality: The Big-Five factor structure, Journal of Personality and Social Psychology. Vol.59,.6, pp , 199. [Goodman 78] L. Goodman, Analyzing Qualitative/Categorical Data: Log-Linear Models and Latent-Structure Analysis, Abt Boos, 1978 [Greenacre 6] M. Greenacre and J. Blasius (ed.), Multiple Correspondence Analysis and Related Methods, Chapman & Hall, 26 [Hofmann 99] T. Hofmann and J. Puzicha, Latent class models for collaborative filtering, Proc. 16th international joint conference on Artificial intelligence, 1999 [Hofmann 1] T. Hofmann, Unsupervised Learning by Probabilistic Latent Semantic Analysis,Machine Learning, Vol.42,.1-2, pp , 21 [ 91] 1991 [ 1a] ID-POS, Vol.19,.461, NC29-16, pp , 21 [ 1b], 3J1-NFC1a-2, 21 [ 1c] 2, vol.11,.76, IBISML21-24, pp , 21 [ 1d] ID-POS 8 SIG-FPAI, pp.15-18, 21 [Ishigai 1] T. Ishigai, T. Taenaa and Y. Motomura, Category Mining by Heterogeneous Data Fusion Using PdLSI Model in a Retail Service, Proc. IEEE International Conference on Data Mining, pp , 21 [ 11a] ID POS Vol56,.2, pp.77-83, 211 [ 11b] 211 1B [Iwata 9] T. Iwata, S. Watanabe, T. Yamada, and N. Ueda, Topic tracing model for analyzing consumer purchase behavior. Proc. 21st International Joint Conference on Artifcial Intelligence, pp , 29. [ 7] (27) [ 5] 25 [ 4] 24 [Michman 91] R.D. Michman, Lifestyle Maret Segmentation, Praeger Pub, 1991 [ 3] BayoNet Vol42,.8, pp , 23 [ 6] (26) [ 9] Vol.53,.9, pp , 29 [ 11] 211 1B [ 9] Vol.24,.5, pp , 29 [ 9] 29 [ 8] 28 [ 9] DNA.477, pp [Nielsen 5] A.C. Nielsen and A. Heller, Consumer-Centric Category Management: How to Increase Profits by Managing Categories Based on Consumer Needs, Wiley, 25. [ 1] 21 [Pearl 9] J. Pearl, Causality: Models, Reasoning and Inference (2nd ed.), Cambridge University Press, 29. [Ramaswamy 96] V. Ramaswamy, R. Chatterjee and S.H. Cohen, Joint segmentation on distinct interdependent bases with categorical data, Jorunal of Mareting Research, vol. 33, no. 3, pp , 1996 [ 3] Vol. 3,.1 pp [ 8] Vol.38,.1, pp.1-19, 28 [Saunders 8] J.A. Saunders, Cluster Analysis for Maret Segmentation, European Journal of Mareting, Vol.14, no.7, , 198 [ 97] 1997 [ ] 2

Goldberg, An alternative description of personality: The Big-Five factor structure, Journal of Personality and Social Psychology. Vol.59,.6, pp.1216-1229, 199. [Goodman 78] L.

12 681 [Taenaa 1] T. Taenaa, K. Fujita, N. Nishino, T. Ishigai and Y. Motomura, Transdisciplinary approach to service design based on consumer s value and decision maing, International Journal of Organizational and Collective Intelligence, vol. 1, no. 1, pp , 21 [Ueda 9] K. Ueda, T. Taenaa, J. Vancza and L. Monostori Value creation and decision-maing in sustainable society CIRP Annals- Manufacturing Technology, vol. 58, no. 2, pp , 29 [ 1] 21 [ 9] 29 [ 2] 22 [Weng 3] S.S. Weng and M.J. Liu, Feature-based recommendations for one-to-one mareting, Expert Systems with Applications, Vol.26,.4, 23 [Xu 9] R. Xu and D. C. Wunsch, Clustering, Wiley, 29 [Yada 7] K. Yada, E. Ip, N. Katoh, Is this brand ephemeral? A multivariate tree-based decision analysis of new product sustainability, Decision Support Systems, Vol.44, pp , 27 [ 1] 1 21 [ 1] A. Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q1 Q11 Q12 Q13 Q14 Q15 Q16 Q17 Q18 Q19 Q2 B.1 F1 F2 F3 F4 F5 F6 Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q Q B. B.1 F1 F6 27 CREST IEEE PD

58-75, 21 [Ueda 9] K. Ueda, T. Taenaa, J. Vancza and L. Monostori Value creation and decision-maing in sustainable society CIRP Annals- Manufacturing Technology, vol. 58, no. 2, pp.

IT,, i

IT,, i 22 Retrieval support system using bookmarks that are shared in an organization 1110250 2011 3 17 IT,, i Abstract Retrieval support system using bookmarks that are shared in an organization Yoshihiko Komaki