Datasets: TaFeng (E-commerce); and Delicious (Bookmark Tag)
Recommendation: !"#$% ← {(|r+(-) ≤ 0}, where 2+
(-) is the ranking of item (; and 2(-): 4(-) → 1, 2, … , :Methodology: For a given testing basket sequence ;, hide last basket ! and generate the next-basket recommendation given < ;\B >Metric: Half-life Utility (HLU) measures the overall ranking performance. Higher is better.
Conclusion: Experiments on the two datasets show that the modeling of correlation information contributes statistically significant improvements as compared to traditional basket-sequence models in terms of top-K recommendations.
Basket Sequence Correlation Networks (Beacon)
Correlation-Sensitive Next-Basket RecommendationDuc-Trong Le, Hady W. Lauw, Yuan Fang
Problem
Experiments
Item-Item Correlation Matrix
v Input: a set of items V; A ∈ ℝ|D|× D ; each basket BF → GF ∈ {0, 1}|D|,
Salmon, Wasabi,Japanese Rice
Crab, Pepper,Melted Butter, Garlic
Fresh Oyster,Fresh Milk, Wasabi
Fresh Oyster,Lemon, Mint
Leaf
Independent
Recommendation
Correlation-Sensitive RecommendationT=1 T=2 T=3
?
Task: Modeling concurrentlyv Correlative associations among items of a basketv Sequential associations across baskets of a sequence
to predict the next basket of correlated items.
Motivating example:
Food RecommendationCount
Co-occurrence 0 0 2 10 0 0 12 0 0 11 1 1 0
Co-occurrence Matrix !
Correlation Matrix "
{Milk, Eggs, Bread}
{Jam, Bread}{Milk, Eggs}
0 0 .67 .33
0 0 0 .58.67 0 0 .33.33 .58 .33 0
Normalize
Objective: Leverage correlations between item-item pairs
Properties:
v A pair with frequent co-occurrence has a higher score than lessfrequent ones.
v A pair with exclusive connection has a higher score than non-exclusive ones.
1 0 1 1
!0
Correlation-Sensitive Basket Encoder
… 0 1 0 1
B
…
LSTM LSTM…
…!ℓ(%)
Sequence Encoder
'0 'ℓ(%)
(0 (ℓ(%)
)ℓ(%)
Basket Sequence S
Correlation-Sensitive Score Predictor
Item Scores *(%)
Correlation Matrix +
0 0 .67 .33
0 0 0 .58
.67 0 0 .33
.33 .58 .33 0
,- .
/.
# Module Operations Parameters
1
Corr
elat
ion-
Sens
itive
Ba
sket
Enc
oder
• The immediate representation IF ∈ ℝ D of BF:
JF = G% ∘ M + ReLU(G%A − TU)
• The L-dimensional latent representation V% ∈ ℝW of BF:
XF = ReLU IFY + Z
• Item importance M ∈ ℝ D
• Noise-cancelling T ∈ ℝ[
• Y ∈ ℝ|D|×\, Z ∈ ℝ\
2
Sequ
ence
En
code
r • The H-dimensional recurrent hidden output ]% ∈ ℝ^:
]F = tanh VFc + ]Fdecf + g
• c ∈ ℝW×^,g ∈ ℝ^
• c′ ∈ ℝ^×^
3
Corr
elat
ion-
Sens
itive
Sc
ore
Pred
icto
r • The sequential signal for next-item adoptions i(-) ∈ ℝ|D|:
i(-) = j ]ℓ(l)m• The correlation-sensitive score 4(-) ∈ ℝ D :
4(-) = n i - ∘ M + i - A + (1 − n)i -
• m ∈ ℝ^×|D|
n ∈ [0,1]
L=8 L=32L=8
L=32
L=32, H=16 L=64, H=32
L=64L=32
L=8,H=64 L=H=64
4.55.05.56.06.57.07.58.0
TaFeng Delicious
HLU
POPMCMCNDREAMBSEQtriple2vecBeacon
Target BookmarkDelicious Tag Basket Prediction (K=5)
Beacon MC POP
Manual de jQuery (1)
web, design, programming, javascript, tools
digital, sociales, web, internet, periodismo
art, design, education, video, tools
The $300 Million Button (2)
twitter, ux, propinquity, critical, writing
design, peace, education, blog, tips
art, design, education, video, tools
(1) https://desarrolloweb.com/manuales/manual-jquery.html (2) https://articles.uie.com/three_hund_million_button