From Past to Present: Personalized Attention Session- Aware RNN Recommender System Presenter: Mei Wang Mentors: Weizhi Li, Yan Yan 08/22/2018
.�����6 �� �����5�4���� � �� �6�3�65 2��� � �� ��1�� ��56� � �� ��3��3���� � �� ���4������ � �� ���4������ � 1
������������� ����������� ������� ↑↓Expanding ←→Wide-Ranging e-commerce music social media app store video 2 1�. ��2� ,.2�0 �����2���.�2� �811�� .2���2 ���� ������1
������������� ������������ �������������� Influen.al Factors Past Present Internal External Periodic Purchases Current Needs Personality Environments Makeups Umbrella Culture Weather Car wiper blade Birthday Fashion Style Festivals Favorite brands Travel suits Aesthe3c Taste Location Preferred Color Mood Age Ads Clothing styles Promotions Educa3on Friends Fast Moving Hang around Figure Family Consumer Goods ... Marriage Income ... ... ... 3 1�. ��2� ,.2�0 �����2���.�2� �811�� .2���2 ���� ������1
������������� ��� ��� ����������� Context-Aware •temporal information •spatial location •user profiles •cross-domain features CF- NN- based based Sequence- Time-Aware Aware • easy-to-obtain •the ordering of • indicative interactions information of •outperform popular user's needs alternatives 4 1�. ��2� ,.2�0 �����2���.�2� �811�� .2���2 ���� ������1
������������� ���������� ��� ���������� ( ) i 1 i 4 2 i 2 3 i 3 i 2 1 i 3 i 1 i 1 i 2 ? . . . . . . . . . . . . 1 1 1 2 3 3 User 1 Session 1 Session 2 Session 3 The sequential interactions between users and items are crucial data sources for recommender systems. ���������� 2��L �PC�LCGA ����C�G������ ��� F�L���� ��E�E� @��M� �G ���JL�L�JF M��J CGL�J��LC�G� OCL�CG � �CGAE� ����C�G �G� ��FIE�L�E� �C���J� �EE L�� �L��J E�GA�L�JF M��J CGL�J��LC�G ��L� �J��� �C@@�J�GL ����C�G�� ���������� 8�� A��E �@ L�C� O�JD C� L� F�D� �@@��LCN� M�� �@ ��L� CGLJ������C�G �G� CGL�J�����C�G IJ�@CE�� �G� ��G�LJM�L � ��LL�J I�J��G�ECR�� ����C�G��O�J� J���FF�G��J ���L�F� 5 2,� ���. -���1 ��,�,�8�8���� ��22,� ��8,�� ���� �0����2
��31����3���� ���������2 � ,�N� ALIG NC@ J�MN M@MMDI�M MCIOF� JF�S �M �DAA@L@�N LIF@M �M � JL@M@�N M@MMDI�� �ON �C@ D�N@L��NDI� ��N� DM NC@L@ DM �I MJ@�DAD� LOF@ P@LS �IDMS� MIG@ �FD�EM AIL D�N@�L�ND�� NC@ �L@ G@��D��AOF� MIG@ M@MMDI����M@� MCILN� � �L@ �FD�E@� �S MG�FF N@LG JLIADF@M ��� �L��DNDI��F ��� ����IN D�N@L@MN� �CDF@ MIG@ M@MMDI�����L@ FI��� NL�D� �DNC NII FI�� G�S @P@� �@ �FD�E@� �S N@LG JLIADF@M� M@�O@��@ F@��NC� �CD�C GDMN�E@� �DFF L@MOFN D� @RNL@G@ NL�D�D�� F�N@��S ��� F�L�@ G@GILS �IMN� 6 �-0 ���� .0��2 ��-8-����0��� 8:��-� 0��-�� ���� �1,����
������������� ��������� �� ��������� ���� �������� �DA GIMEO�MEI� I� INK GI@AF @ALEC� EL E�LJEKA@ �� KA�F @�M� R I�LAKO�MEI�L ��@ ���F�LEL� 0M �IGAL �KIG I�FE�A �E���DE A��IGGAK�A ��OEC�MEI� FIC R @�M� D�OE�C �KIN�@ ���4 E�MAK��MEI�L� �4 NLAKL ��@ �4 EMAGL �KIG ��2 ��MACIKEAL� �A �IGA NJ PEMD MDA �IFFIPE�C MDKAA I�LAKO�MEI�L� R 7 4-0 ���� .0��� �8-�-����0��� ��44-8 0��-8� ���� �1,���4
������������� ��������� ���� �����1�1 Short-term profile predominates in the selection of the recommendations � The blue bar represents the mean percentage of user interactions 50% hanging in the top 10 categories during the same session, and the orange one represents for items. 20% Overall both of them are subject to exponential decrease, which proves that user's short-term shopping goal plays a predominant role for the intra-session interaction choices. 8 1�. ��2� ,.2�0 �����2���.�2� �811�� .2���2 ���� ������1
��2�����2���� ��������� ��2� �������� Longer-term behavioral patterns and user preferences can also be important � The figure shows mean percentage value of a user clicking some 65% repeated categories and items that he/she had clicked before in the previous 10 sessions. 20% Inter-session informa?on provides 30 to 60 percent of informa?on for next-basket category predic?on and 5 to 20 percent of knowledge about repeated items. 9 1�. ��2� ,.2�0 �����2���.�2� �811�� .2���2 ���� ������1
������3������ ��������� ���� �������� The click gap time (item dwell time) helps in connecting short-term and long-term � The figure shows the normalized histogram of the click gap (view 10s dwell <me) of user interac<ons, which follows gamma distribu<on with maximum around 10 seconds. Generally speaking, the longer time a user spends on the item, the more interest he has in it. 300s This perfectly bridges the gap of discrete interaction sequence data with potential weights. 10 1�. ��2� ,.2�0 �����2���.�2� �811�� .2���2 ���� ������1
�2��3�����32� �3������� �: �1������� ���� �2��:��� In this paper, we want to quantify, exploit and integrate the effectiveness � of user's intra-session and inter-session profiles with temporal dynamics. Short-term profile predominates � The very last actions in the present session should Session-based RNN recommender GRU4REC represent an important piece of context information system as the basis of our model design Longer-term profile counts � long-term profiles are important for recommender We choose to use an efficient embedding layer to system, while current state-of-art session-based automatically train and activate short and long term approaches fail to model them effectively. profiles from session representations. The dwell time helps � Finally, with the help of temporal dynamics scheme, we incorporate temporal context in the RNN and perform efficient combination for short-term session sequence information and long-term user and item profiles. 11 1�. ��2� ,.2�0 �����2���.�2� �811�� .2���2 ���� ������1
���:��������� ����:�������� -����:�� �������� �� ��������� ��� ��:������A�� �����������:� �� �� �IG�GJ� ������ � FGM�D �� G@@�I �F �O��F���D� ����F�CGF J���E� �G D�M�I�A� ��E�GI�D �PF�EC�J ��IJGF�DC��� ����F�CGF ��JJCGF��N�I� ���GEE�F��I JPJ��E EG��D� �G J���E� �O�DGC�CFA EGI� CF�I��J�JJCGF J��ED�JJDP CF��AI��� CF�I��J�JJCGF �F� CF@GIE��CGF JG �J �G �F��F�� J�JJCGF� CF��I�J�JJCGF �IG@CD�J� ��J�� �� CF �CE� �CE�FJCGF� ����:���� ������� �������� �������:� �� ������� �� �� �GF�L�� �O��FJCM� �O��ICE�F�J GF �� CF�DL�� DGFA���IE LJ�I �IG@CD�J @GI @GLI I��D ����J��J �F� ��EGFJ�I��� ��� J�JJCGF���J�� �� �G D��IF ��� �IGJJ� �@@���CM�F�JJ G@ ����� @GI J�JJCGF �����IF �F� LJ�I @�MGIC�� ��IJGF�DC��� I��GEE�F���CGF� �MGDL�CGF CF � J��ED�JJ N�P� 12 2,� ���. -���1 ��,�,�8�8���� ��22,� ��8,�� ���� �0����2
Recommend
More recommend