推荐系统论文 Fine-grained Interest Matching for Neural News Recommendation

Fine-grained Interest Matching for Neural News Recommendation

ACL 2020

Main Idea

Instead of aggregating user’s all historical browsed news into a unified vector, we hierarchically construct multi-level representations for each news via stacked
dilated convolutions

Model

Three major components:

1. news representation module

2. cross interaction module
   to exploit and aggregate matching information from
   each pair of news at each level of granularity

3. prediction module

news representation module

HDC(hierarchical dilated convolution) encoder to learn representations of news from multiple semantic views

空洞卷积
应用于CV中的语义分割，可以扩大视野同时保持分辨率

1. input word embedding sequence

$x_i$ is a d-dimensioned word emb

2. use HDC

to capture multi-grained semantic features

3. output

The output of each stacked layer l is preserved as feature maps of the news text at a specific level of granularity

Suppose there are L layers stacked, the multi-grained news representations can be defined as [d0, d1, . . . , dL].

Cross Interaction Module

Given representations of the k-th browsed news:

and the candidate news:

to build a segment-segment matching matrix for each granularity:

• fuses all interaction matrices into a 3D mathching image Q

n: total number of browsed news
each pixel $Q_{k,i,j}$ is defined as:

note: each pixel is a concatenated vector with L + 1 channels

• 3D CNN

• output integrated matching vector $s_{u,c}$

predictor

Dataset

MSN click logs

• $FI{M}_{first}$: use CNN instead of HDC
• $FI{N}_{last}$: only use the last layer of HDC