SLIDE 1
Outline
- Task
- Background
- Motivation
- Our model
structure of HRM connections to previous methods
- Experiments
- Summary
2
Learning Hierarchical Representation Model for Next Basket - - PowerPoint PPT Presentation
Learning Hierarchical Representation Model for Next Basket - - PowerPoint PPT Presentation
Learning Hierarchical Representation Model for Next Basket Recommendation Pengfei Wang, Jiafeng Guo, Yanyan Lan, Jun Xu, Shengxian Wan, Xueqi Cheng CAS Key Lab of Network Data Science and Technology Institute of Computing Technology,Chinese
SLIDE 2
SLIDE 3
Problem: Next basket recommendation
3 Next basket recommendation : given a sequence of purchases, what items are purchased sequentially?
potato beers bread phone maternity butter bread earphone milk baby-car nursing bottle battery
transaction i-1 transaction i transaction i+1
?
SLIDE 4
4
Background
- Sequence models
- Collaborative Filtering
- Hybrid methods
MC (Markov Chain) [Zimdars et al. UAI01, Chen et al. KDD12] NMF(Non negative Matrix Factorization) [Daniel D. Lee et al. NIPS01] FPMC (Factorized Personalized Markov Chain) [S Rendle et al.
WWW10]
SLIDE 5
5 a a d b transactions of u1 b c a transactions of u2
Weakness of Previous methods
weakness of MC: Lack users` general interests
Sequence models based on Markov Assumption：MC
a,b,c, d represent items
SLIDE 6
6 a a d b transactions of u1 b c a transactions of u2
Weakness of Previous methods
Global one:
weakness of NMF: Lack sequence behaviors
Collaborative filtering matrix factorization：NMF
Global set
SLIDE 7
7 a a d b transactions of u1 b c a transactions of u2
Weakness of Previous methods
Hybrid method：Factorized Personalized Markov Chain
[S Rendle et al. WWW10 best paper]
weakness of FPMC: linear combination of different factors
vi vj u v vj vi
U represent users, V represent items
SLIDE 8
8 a a d b transactions of u1 b c a transactions of u2
Weakness of Previous methods
Hybrid method：Factorized Personalized Markov Chain
[S Rendle et al. WWW10 best paper]
weakness of FPMC: linear combination of different factors General taste Sequential behavior
user next item last item next item
+
<vi
I,L,v1 L,I>
<vi
I,L,v2 L,I>
<vi
I,L,v3 L,I>
linear combination
Independent Influence!
SLIDE 9
9
Weakness of Previous methods
+
Is that linear combination enough for a good recommendation?
pumpkin potato pumpkin cucumber chips chocolate candy candy
next transaction next transaction Last transaction Last transaction
Halloween!
SLIDE 10
We need a model that is capable of incorporating more complicated interactions among multiple factors. This becomes the major motivation of our work.
Motivation
10
SLIDE 11
11
Hierarchical Representation Model
The structure of HRM
Level 2 Level 1
general interest
sequential behavior
SLIDE 12
Aggregation Method
Linear method
average pooling
Nonlinear method
max pooling
- ther types of operators(top-k average pooling, k-max
pooling, hadamard pooling) (V is a set of input vectors to be aggregated)
SLIDE 13
13
Learning and Prediction
Objective function Negative sampling
negative count sample distribution all users all trans all items
The probability of purchasing one item next transaction:
sum of all items: too large! exponent of item and users` hybrid interest
SLIDE 14
14
Connection to Previous methods
Degradation To MC
=
Select copy: copy item when constructing the transaction representation from item vectors, the operation randomly selects one item vector and copies it
softmax select copy select copy
SLIDE 15
Connection to Previous methods
15
Degradation To MF
=
Select copy : always select and copy user vector in the second layer, ignoring the sequential information
select copy softmax
SLIDE 16
16
Connection to Previous methods
Degradation To FPMC
Avg Pooling is used , each instance corresponds to 1 negative sample
softmax avg pooling avg pooling
SLIDE 17
17
Experiments
Ta-Feng BeiRen T-Mall #transactions
67964 91294 1805
#items
7982 5845 191
#users
9238 9321 292
#avg.transaction size
5.9 5.8 1.2
#avg.transaction per user
7.4 9.7 5.6
Data sets
retails ecommerce
SLIDE 18
F1-score: Hit-ratio: NDCG: 18
Experiments
Evaluation Metric
a ranking measure harmonic mean of precision and recall coverage
SLIDE 19
19
Experiments
Comparison among Different HRMs
Avg pooling perform worst When apply max pooling on any layer, the performance improved a little When apply max pooling on all layers, HRM performed best
observation
SLIDE 20
20
Experiments
Comparison with baselines
top popular sequential behavior hybrid method general interest
Top method performed worst NMF and MC performed better than top method FPMC performed better than NMF and MC HRM performed best
observation
SLIDE 21
21
Experiments
Comparison over groups
NMF perform better than MC on active group, while MC performs better than NMF on inactive group HRM performed best
observation
SLIDE 22
22
Experiments
The Impact of Negative Sampling
More negative count we choose ,the more F1-score we obtain The sampling number k increases, the performance gain between two consecutive trials decreases
observation
SLIDE 23
23
Summary
A next basket recommendation task A Hierarchical Representation Model
- model both sequential behavior and users` general taste
- Aggregation operators to connect two level factors.
- HRM can produce multiple recommendation models by
introducing different aggregation operations
Furture works
- More aggregations operations will be analyzed
- Integrate other types of information, e.g. timestamp of
transaction
SLIDE 24
24
Thank You！
EMAIL:wangpengfei@software.ict.ac.cn
SLIDE 25
SLIDE 26
- Poin twise mutual information (PMI) is
a widely used word similarity measure
SLIDE 27
27
Weakness of Previous methods
the sum of score in general recommend and score of sequential recommend