Basic Information

Link

When purchasing clothes, it is intuitive that we often have preferences for detailed semantic attributes (such as neckline, heel height, skirt length) in addition to global impressions. -> Semantic Attribute Explainable Recommender System (SAERS)
problem:
- 1) it is difficult to obtain clothing semantic attributes features without the manual attribute annotation in the large-scale E commerce data.
- 2) the user preferences are sophisticated, while traditional methods usually transform the item image into a latent vector directly.
solution:
- 1) develop a Semantic Extraction Network (SEN), which is used to extract the region-specific attribute representations in a weakly-supervised manner (multi-task classification loss: with 12 attributes like high neck (semi-high, turtle, ...), skirt length (short, knee, midi, ankle, ...), ...). Through SEN, each item is projected to the semantic attribute space.
- 2) to capture the diversity of the semantic attribute preference among items, we design a Fine-grained Preferences Attention (FPA) module to automatically match the user preference for each attribute in semantic attribute space, and aggregate all attributes with different weights.
optimization: Bayesian Personalized Rank (BPR) framework.
interpretation: favorite attribute by attention from FPA, region by Grad-AAM(?)

dataset

http://vision.cs.utexas.edu/projects/finegrained/utzap50k/
https://tianchi.aliyun.com/competition/entrance/231671/ information
evaluation: recommendation accuracy & quantitative evaluation that compares the area of interest (attention) with the images purchased in the past