In the midst of vast amounts of available fashion items, consumers today require more e cient recommendation services. A system that sorts out items that form a stylish ensemble with already selected or possessed items would provide them with greater convenience. In this paper, we propose a fashion item recommendation method that learns the way the fashion items are matched from a large ensemble database. We empirically show that the proposed method can explain factors that a ect item matching and recommend the most suitable items to the given set of items.
Today, as massive amounts of fashion items are available
in both online and o ine market, needs for e cient
recommendation services has grown signi cantly. One of the most
important factors in recommending a fashion item is how
well the item combines with a set of other items to form
stylish ensemble. A number of works have been proposed in
matching fashion items using web-scraped out t
combination dataset from sites such as Pinterest. However, they are
mostly based on color matching and are not exible enough
to exploit other relevant features[
In this paper, we propose a fashion item recommendation method that learns from a large ensemble database. The items and their attributes, and the ensembles are modeled as a heterogeneous information network that allows for exible semantic analysis. We de ne meta-paths on the network as patterns of relationships between items with respect to attributes and ensembles. Relative importance of each meta-path in matching items is learned from the ensemble database. We show through experiments that our proposed method outperforms baseline algorithms. 2.
We have collected 18,449 fashion items and 7,458 ensembles from an online shopping mall. Each ensemble contains about 2.5 items. The ensembles, which are presented by professional fashion coordinators of the shopping mall, consists of clothes, shoes, and fashion accessories as shown in Fig 1. We extracted and re ned 4 attributes - category, material, pattern, and color - from item descriptions and item images. Table 1 shows value sets of each attribute. Weighted multi-color vectors are extracted from data images using a color extraction tool. The color vectors are then grouped into 3000 clusters using k-means clustering. pattern material category item color ensemble 3.
Fig 2 shows the network schema for fashion item ensemble
dataset. There are 6 types of nodes, namely, item,
category, pattern, material, color, and ensemble.
Unlabeled edges represent direct associations between the nodes.
We use the concept of meta-path[
To learn the coe cients of each meta-path, we sample
2,000 ensembles among 6,500 training ensembles (the rest
is used for evaluation). Then for each sampled ensemble,
we randomly choose one item as the target item and use
the rest as query items. We choose to use normalized path
count(NPC)[
N P Cpi (Q; c) =
X N P Cpi (q; c)=jQj q2Q where NPCpi (q; c) is normalized path count between q and c along meta-path pi , Q is the set of query items, and c is the candidate item. The candidate items are sampled from the items that are released in the same month as the target item. And the according label becomes: lQ;c = (1; if c = target item
0; otherwise The coe cient of each meta-path is learned using logistic regression on the feature vector and label pairs, (fQ;c; lQ;c).
Table 2 shows the important meta-paths and corresponding coe cients. Negative coe cient of (a) means the items that belong to the same category are rarely matched, which is trivial. In case of (d), the positive coe cient indicates that categories matched frequently on the network are actually important in item matching. Meta-paths for color attribute ((c) & (h)) show similar result with the meta-paths for category attribute ((a) & (d)), while those for the pattern attribute ((b) & (e)) turn out to be in the opposite. Also, we can infer from (f) and (g) that pattern and color are tightly related in styling. 4.
The e ectiveness of recommendation have been evaluated using the remaining 958 ensembles. As in the training stage, one item per ensemble is chosen as the target item and the remaining used as query items. Items nearest to the query items are recommended using the trained regression model. Random selection (Random) and personalized pagerank (PPR) based recommendations are used as baseline methods. Table 3 shows the results where performance is measured in terms of precision at k (k=1,3,5; P@1, ..., P@5) and mean reciprocal rank (MRR). The performance of PPR is lower than Random since PPR assigns higher scores to items near the query items. Consequently, the items of the same category or color with the query items tend to be recommended. Meanwhile, the meta-path based recommendation exploits the learned weights of the meta-paths, resulting in more effective recommendation.
This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea Government(MSIP) (No. 20110030812).