12. Station-based model (Yahoo)
L S;Θ( )= logP i s;t,Θ( )
(i,t)∈PS
∑
PS ∈S
∑
Правдоподобие плейлиста станции:
P i s;t,Θ( )=
exp(rsi;t )
exp(rsj;t )
j
∑
rsi;t = bi + qi
T
vs + vs
slot(t)
+
1
Ps
(t,w)
qj
j∈Ps
(t,w)
∑
$
%
&
&&
'
(
)
))
Запись для самой станции:
13. Station-less model (Yahoo)
L S;Θ( )= logP i s;t,Θ( )
(i,t)∈PS
∑
PS ∈S
∑
Правдоподобие плейлиста станции:
P i s;t,Θ( )=
exp(rsi;t )
exp(rsj;t )
j
∑
rsi;t = bi + qi
T 1
Ps
qj
j∈Ps
∑ +
1
Ps
(t,w)
qj
j∈Ps
(t,w)
∑
$
%
&
&&
'
(
)
))
Запись для самой станции:
15. Обучение моделей — трюк
Δθ =η
∂rsi;t
∂θ
− w j s( )
∂rsj;t
∂θj∈J
∑
&
'
((
)
*
++
Шаг SGD:
Sampling:
P i S( ) J w i S( )=
exp(rsi;t )
P i S( )
exp(rsj;t )
P j S( )j∈J
∑
21. Регуляризация
X = argmax
X∈ℜ|S|×d
L D | X( )− λ X F
2
(V,U) = argmax
V,U∈ℜ|S|×d
L D |V,U( )− λ V F
2
+ U F
2
( )
Single-point model:
Dual-point model:
22. Регуляризация
X = argmax
X∈ℜ|S|×d
L D | X( )− λ X F
2
(V,U) = argmax
V,U∈ℜ|S|×d
L D |V,U( )− λ V F
2
+ U F
2
( )−
−v Δ2 s,s( )2
s∈S
∑
Single-point model:
Dual-point model:
25. Ссылки
• N. Aizenberg, Y. Koren, and O. Somekh. 2012. Build your own music
recommender by modeling internet radio streams. In Proceedings of the
21st international conference on World Wide Web (WWW '12). ACM, New
York, NY, USA, 1-10. http://doi.acm.org/10.1145/2187836.2187838
• B. McFee, and G. Lankriet. 2011. The natural language of playlists. In
Proceedings of 12th International Society for Music Information Retrieval
Conference (ISMIR 2011). http://cosmal.ucsd.edu/~gert/papers/nlp.pdf
• S. Chen, J. Moore, D. Turnbull, and T. Joachims. 2012. Playlist prediction via
metric embedding. In Proceedings of the 18th ACM SIGKDD international
conference on Knowledge discovery and data mining (KDD '12). ACM, New
York, NY, USA, 714-722. http://doi.acm.org/10.1145/2339530.2339643