Adaptive pointwise-pairwise learning to rank

Patent metadata

Field	Value
Publication number	US-2021383254-A1
Application number	US-202117215323-A
Country	US
Kind code	A1
Filing date	Mar 29, 2021
Priority date	Jun 3, 2020
Publication date	Dec 9, 2021
Grant date	—

Abstract

A method of ranking items for a given entity uses sets of triplets <u, i, j>, each set of triplets including an entity u and a pair of items i and j with a known relative relevance for entity u, to train a learnable scoring function ƒ and to learn optimized values of a first set θ of learnable parameters. The training includes optimizing a loss function depending on θ, on a second set of learnable parameters θg, and on a probability of having the item i preferred to the item j by the entity u. The probability defines a continuum between pointwise and pairwise ranking of items through a learnable mixing function depending on θg. After training, the trained learnable scoring function ƒ is applied to all input pairs <u′, i′> to rank all items i′ for an entity u′.

First claim

1 . A computer-implemented method of constructing a ranking model for an information retrieval system by ranking items for a given entity with a learnable scoring function, wherein an entity is one of a user, a query, or a context, the method comprising:
(a) training a learnable scoring function ƒ, which depends on a first set of learnable parameters θ and on a set of triplets <u i, j>, the set of triplets comprising an entity u from a set of entities and a pair of items i and j from a set of items, wherein a relative relevance of each pair of items i and j from the set of items for each entity u from the set of entities is known, to learn optimized values of the first set of learnable parameters θ; said training the learnable scoring function ƒ including (a1) optimizing a loss function depending on the first set of learnable parameters θ and on a second set of learnable parameters θ g , wherein the loss function defines a continuum between pointwise ranking and pairwise ranking of items through its dependency on a learnable mixing function g depending on θ g ; (b) applying, using the learned optimized values of the first set of learnable parameters θ, the trained learnable scoring function ƒ to all input pairs <u′, i′> formed by a given entity u′ from the set of entities and by every item i′ from the set of items to obtain a relevance score for each pair; (c) ranking the items i′ for the entity u′ based on the respective relevance scores; and (d) constructing, using the ranked the items i′ for the entity u′, a ranking model for an information retrieval system.

2 . The method as claimed in claim 1 , wherein said training the learnable scoring function ƒ, further comprises:
(a2) computing, for each triplet <u, i, j> of the set of triplets, a weighting coefficient g(u, i, j; θ g )∈[0,1] by applying the learnable mixing function g to the entity u, the two items i and j, and the second set of learnable parameters θ g ;
(a3) computing, for each triplet <u, i, j> of the set of triplets, a first relevance score ƒ(u, j, θ) defining a relevance of the item j for the entity u, by applying the learnable scoring function ƒ to the item j and the entity u;
(a4) computing, for each triplet, <u, i, j> of the set of triplets, a second relevance score ƒ(u, i; θ) defining a relevance of the item i for the entity u, by applying the learnable scoring function ƒ to the item i and the entity u;
(a5) applying, for each triplet <u, i, j> of the set of triplets, the weighting coefficient g(u, i, j; θ g ) to the first relevance score ƒ(u, j; θ);
(a6) computing, for each triplet <u, i, j> of the set of triplets, a probability of having the item i preferred to the item j by the entity u as a function of a difference between the second relevance score ƒ(u, i; θ) and the weighted first relevance score, wherein the computed probability corresponds to a pairwise relevance probability of having the item i preferred to the item j by the entity u if g(u, i, j, θg)=1, to a pointwise relevance probability defining how relevant the item i is to the entity u if g(u, i, j; θ g )=0, and defines said continuum between pointwise ranking and pairwise ranking of items if 0<q(u, i, j; θg)<1; and
(a7) learning optimized values of the first and second sets of learnable parameters θ and θ g by optimizing the loss function, depending on θ and θg, through gradient descent optimization, the loss function being defined as a sum over all triplets <u, i, j> of a function derived from the probability of having the item i preferred to the item j by the entity u.

3 . The method as claimed in claim 2 , wherein the function of the difference between the second relevance score and the weighted first relevance score is a sigmoid function.

4 . The method as claimed in claim 1 , wherein the loss function is a negative log-likelihood loss function and optimizing the loss function minimizes the loss function over the set of entities and the set of items.

5 . The method as claimed in claim 2 , wherein the loss function is a negative log-likelihood loss function and optimizing the loss function minimizes the loss function over the set of entities and the set of items.

6 . The method as claimed in claim 4 , wherein the negative log-likelihood loss function is defined by:

L

a
⁢
d
⁢
a
⁢
p
⁢
t
⁢
i
⁢
v
⁢
e

⁡

(

θ
,

θ
g

)

=

-

∑

(

u
,
i
,
j

)

∈
D

⁢

(

y

u
,

i
>
j

⁢
log
⁢

σ
⁡

(

f
⁡

(

u
,

i
;
θ

)

-

g
⁡

(

u
,
i
,

j
;

Assignees

• Naver Corp

Inventors

• Renders Jean-Michel
• Cinar Yagmur Gizem

Classifications

• G06N3/047

  Probabilistic or stochastic networks · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/0985

  Hyperparameter optimisation; Meta-learning; Learning-to-learn · CPC title

• G06N3/09

  Supervised learning · CPC title