H&M Personalized Fashion Product

Recommendation Using LightgbmRanker

Dan Xian1 , Shaozan Cui2 , Bo Wang3 , and Lishuai Cui4(B)

1 Northeastern University, San Jose, USA
[email protected]
2 Nanjing University of Finance & Economics, Shanghai, China
3 College of Information and Electrical Engineering, China Agricultural University, Beijing,
China
4 King’s College London, London, UK

[email protected]

Abstract. The product recommendation system is a system that analyzes user

preferences, searches for a large number of product information in the e-mall, and
recommends products that may be of interest to users, providing an intelligent
shopping experience for online shopping users. It can help users more accurately
and quickly discover interesting and high-quality product information, enhance
the value of information, and improve the user’s online shopping experience.
In this article, we use data provided by H&M to construct a recommendation
system using LightgbmRanker. To evaluate our experiment’s performance, we do
compared competitions. We use map as the evaluation metric. The result shows
that our LightgbmRanker owns the highest 0.0282 among these models, which is
0.063, 0.056 higher than SVD, TF recommender respectively.

Keywords: recommendation system · LightGBM Ranker · MAP · feature

engineering

1 Introduction
With the advent of the Internet era, the characteristics of mobile Internet sharing at
will have catered to the convenient and diversified needs of modern consumers, online
shopping has become a very common but important shopping method in life, and the
e-commerce market has been stimulated with great vitality. The core competition point
of major e-commerce companies has also shifted from basic services to providing per-
sonalized product recommendation services to users, and the product recommendation
system has become one of the core functions of e-commerce platforms. As the category
of goods grows, users will spend more money to proactively obtain the items they want
through simple retrieval, and the product recommendation system is the most promis-
ing way to solve this problem. The product recommendation system is a system that
analyzes user preferences, searches for a large number of product information in the e-
mall, and recommends products that may be of interest to users, providing an intelligent

© The Author(s) 2023

J. Yen et al. (Eds.): ICBIS 2023, AHCS 14, pp. 201–208, 2023.
https://doi.org/10.2991/978-94-6463-198-2_23
202 D. Xian et al.

shopping experience for online shopping users. It can help users more accurately and
quickly discover interesting and high-quality product information, enhance the value of
information, and improve the user’s online shopping experience.
In this article, we use data provided by H&M to construct a recommendation system
using LightgbmRanker. Related work is described in sect. 2, and we introduce our
methodology and experiment in sects. 3 and 4.

2 Related Work

In the context of the rapid development of e-commerce, personalized recommendation

algorithms that tap user interest in massive items have become the core of the recom-
mendation system. Today’s mainstream recommendation technologies are CB (Content-
Based), CF (Collaborative Filtering), deep learning-based recommendation, and hybrid
recommendation.
1) Content-Based recommendation
CB is the earliest technology used in the recommendation system. The basic idea
is: according to the user’s historical behavior, to obtain the user’s interest preferences,
to recommend users to similar to their interest preferences. For feature extraction for
targets, [1] introduced a vector space model technique, [2] comprehensively considered
the distribution of word frequency and document frequency in the target and the target,
and used the word frequency-inverse document frequency technology TF-IDF (Term
Frequency- Inverse Document Frequency, TF-IDF) for text preprocessing, and targeted
recommendation [3, 4] for users. [5] used the idea of relevance feedback to explore the
implicit relationship between user behavior and preferences to improve the effect of
target recommendation5.
The advantages are that the principle is easy to understand, the implementation is
simple, and the recommendation results can be clearly interpreted: but the disadvantage
is: it is impossible to actively discover the potential interest that the user has not shown:
it can only process text, and it is impossible to process video, audio, and other targets
that are difficult to directly summarize the features.
2) Collaborative filtering recommendation
CF is a very classic technique in recommender systems. The basic idea is that users
with similar interests may be interested in the same things. The basic operation of the
collaborative filtering algorithm is to analyze user preferences, find similar user groups,
and then determine the preference degree of a single user to the target according to the
preference program of the user group to the target.
Lu Z and others have proposed a recommendation algorithm based on Naive Bayes
to classify target types [7] proposed a collaborative filtering algorithm with uncertain
neighbors, which effectively alleviated the problem of data sparsity by calculating the
similarity between users and products and using similar objects of products as the
recommendation results.
Recommendation based on collaborative filtering can make more accurate target
recommendation in the case of learning a large amount of data. But the disadvantages
are also obvious: a large amount of data is needed to generate a more accurate user
portrait; there is a cold start problem for new users or new targets: the algorithm has
 H&M Personalized Fashion Product 203

a performance bottleneck, and when there is too much data, the time complexity of
calculating the user group feature portrait will be greatly increase.
3) Deep learning recommendation
In addition to the above two traditional recommendation techniques, the application
of deep learning technology to the recommendation system has also become a research
hotspot. Deep learning is a branch of machine learning and an extension of traditional
artificial neural networks, which achieve complex function approximations by building
nonlinear mapping functions in multiple hidden layers, and build recommended mod-
els that are more in line with user interests. [6] proposed an algorithm that combines
collaborative filtering with restricted Boltzmann machine RBM (Restricted Boltzmann
Rachine, RBM), which is the earliest recommended technique based on deep learning,
using a two-layer RBM model to extract data features, training the user’s preference
data as input, calculating the hidden factor vector of the hidden layer through condi-
tional probability, and using the hidden factor vector of the hidden layer to reverse the
user’s liking degree when predicting.
The advantage of the recommended technology of deep learning is that it is possible
to learn more abstract essential characteristics from the complex multidimensional data
of the user and the target: layer-by-layer training in multiple hidden layers effectively
reduces the training difficulty; Features can be learned and extracted automatically with-
out the need for manual operation. However, the recommendation technology based on
deep learning is extremely dependent on machine computing power, and there is no
effective interpretation of the recommendation results.
4) hybrid recommendation
Hybrid recommendation is to combine the advantages of various recommendation
techniques to solve the problem that the accuracy of single-model recommendation is
not high. As a semantic web information tool, the knowledge graph contains semantic
information and mesh path structure information to provide a new way of thinking for
solving the target recommendation problem. This paper mainly examines recommended
techniques based on knowledge graphs.
• Our Contribution

– We utilize LightgbmRanker to create product recommendations.

– We introduce our dataset and do feature engineering.
– In the experiments process, we do the comparing experiments and the result shows
that our model performed better than the other models.

3 Methodology

The traditional GBDT and XGBoost algorithms have quite good efficiency, but in the
case of large amount of data or too high feature dimension, the efficiency and scalability
of these algorithms cannot meet the actual needs. Timeliness is important for real-time
online consumer behavior analysis systems. The LightGBM algorithm can solve these
204 D. Xian et al.

problems very well, so LightGBM is applied to the online consumer behavior prediction
system.
• Principles of LightGBM
LightGBM (Lightweight Gradient Decision Tree) is a gradient boosting framework
based on the decision tree algorithm proposed by Microsoft to support efficient well line
training. The algorithm has not made many changes in principle, and the principle of
the algorithm is the same as that of XGBoost, which uses the Taylor expansion of the
loss function as an approximation of the current error to obtain a new decision tree. The
innovation of this algorithm lies in the application of GOSS algorithm and EFB algorithm
to speed up operation. When achieving the same accuracy as GBDT, LightGBM is about
10 times faster. The following mainly introduces the GOSS algorithm and the EFB
algorithm.
1) GOSS algorithm
The GOSS (Gradient based One-Side Samplin, gradient-based, one-sided sampling)
algorithm optimizes the algorithm speed from a sample perspective. Sample importance
is measured differently in different algorithms. In the SVM, the support vector closest to
the separated hyperplane is the most important: in the AdaBoost algorithm, the sample
weight is a measure of importance, and increasing the weight of the mispartitioned
sample makes the next iteration focus on training the mispartitioned sample. In GBDT,
the gradient of the sample can be used as a basis for evaluating the importance, if the
gradient of the sample is small, it means that the sample has been trained and the training
residual is small. Similarly, LightGBM can use the gradient of a sample as a measure of
its importance.
However, if all the small gradient samples are deleted, it is easy to lack the data set
information and cause a loss of accuracy. The GOSS algorithm can solve this problem by
retaining large gradient samples, randomly selecting a part of the small gradient samples
and multiplying the weights, so that without changing the original data set information,
more attention can be paid to those undertrained samples.
The specific step of the GOSS algorithm is to first sort the samples according to the
absolute value of the gradient, retain the large gradient sample of the first a × 100%,
then randomly select b × 100% of the small gradient sample in the remaining dataset,
and finally multiply the small gradient sample by the constant (1 − a)/b. Therefore,
the LightGBM algorithm uses a gradient based on each sample to optimize the training
samples.
2) EFB algorithm
The EFB (Exclusive Feature Bunding) algorithm optimizes the algorithm speed from
a feature.
perspective. In high-dimensional sparse spaces, most features take non-zero values
almost at different times, meaning they are mutually exclusive. The EFB algorithm takes
advantage of the mutexity of high-dimensional sparse features to bundle mutex features
together to form a feature and build it into a histogram, which can reduce the training
complexity and further speed up the operation speed.
There are two difficulties in the implementation process: how to determine which
features are mutually exclusive, and how to bundle these mutually exclusive features.
For the first problem, it can be solved by a greedy algorithm: for the second problem,
 H&M Personalized Fashion Product 205

Fig. 1. LightGBMRANK mode

the goal of placing the mutex features into buckets of different histograms is achieved
by increasing the offset of the mutex features. The EFB algorithm bundles and merges
sparse mutex features, reduces the feature set, and can further improve the efficiency of
the algorithm.
• Pros and cons of LightGBM
LightGBM adopts optimization strategies such as histogram-based sorting algorithm
and tree generation based on leaf nodes. The advantages of the algorithm are that using
histogram has lower training execution time and higher operating efficiency, and storing
discrete values instead of continuous values makes memory Reduced consumption,
supports category features and parallel processing, and is often used in industrial big
data.
• LightGBMRANK model
Our LightGBMRANK. Model is shown in Fig. 1. There are three kinds of data:
product features, customer feature and transaction features. And these float data are
converted into bin data. We construct histogram and get the decision tree.

4 Experiments

• Experiments data

In our paper, we use the data from previous transactions and customer and
product meta data provided by H&M. The available meta data spans from simple
data, such as garment type and customer age, to text data from product descrip-
tions, to image data from garment images. We need to predict what articles each
customer will purchase in the 7-day period immediately after the training data
ends. The metadata for each article_id available for purchase are 25 kinds: arti-
cle_id, product_code, prod_name, product_type_name, product_group_name, graphi-
cal_appearance_no, graphical_appearance_name, colour_group_code, etc., which are
detailed describe in articles.csv. “Images” is a folder of images corresponding to each
article_id. The metadata for each customer_id in dataset are 7 kinds: customer_id,
206 D. Xian et al.

FN, Active, club_member_status, fashion_news_frequency, age, postal_code which are

presented in customers.csv.
We do feature engineering, and the number of product types per each product group
is shown in Fig. 2, which shows accessories owns the most product types. And the top
5 product types are accessories, shoes, garment upper body, underwear and swimwear
respectively. Furthermore, we draw product name wordcloud in Fig. 3, which shows the
most frequent product name in the data.

• Training parameters

The LightGBM’s parameters according to empirical methods and grid search are
shown in Table 1.

Fig. 2. Number of product types per each product group.

Fig. 3. Product name Wordcloud

 H&M Personalized Fashion Product 207

Table 1. Training parameters

n_estimators 3000
learning_rate 0.001
num_leaves 20

Table 2. Experiment result

Models MAP@12
SVD 0.0219
TF recommender 0.0226
LightgbmRanker 0.0282

• Evaluation metrics

The evaluation metric is mean Average Precision @ 12 (MAP@12):

1  
U min(n,12)
1
MAP@12 = P(k) × rel(k)
U min(m, 12)
u=1 k=1

where U is the number of customers, P(k) is the precision at cutoff k, n is the number
predictions per customer, m is the number of ground truth values per customer, and rel(k)
is an indicator function equaling 1 if the item at rank k is a relevant (correct) label, zero
otherwise.
• Experiment result
To evaluate our experiment’s performance, we do compared competitions. The higher
metric is, the better our model will be. The experiment result is shown in Table 2. Our
LightgbmRanker owns the highest 0.0282 among these models, which is 0.063, 0.056
higher than SVD, TF recommender respectively.

5 Conclusion

In our paper, we proposed a product Recommendation system using the LightgbmRanker

model. The feature engineering is conducted for three kind of features: Product features,
Customer features and Transaction features. These features will be as an input for the
ranker algorithm using Goss and EFB algorithm to calculate the histogram and build the
decision trees. Compared to other models or methods like SVD, TF recommender, the
proposed model owns the highest recall score which is 0.0282. In the future, we will dig
into more features especially the text features to build a better recommendation system.
208 D. Xian et al.

References
1. Salton G, Wong A, Yang C S. A vector space model for automatic indexing. Communications
of the ACM, 1975, 18(11): 613-620
2. Agnihotri D, Verma K, Tripathi P. Variable global feature selection scheme for automatic
classification of text documents. Expert Systems with Applications, 2017, 81: 268-281
3. Tang L, Long B, Chen B C, et al. An empirical study on recommendation with multiple types of
feedback. in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining. 2016: 283–292
4. Xia Z, XuS, Liu N, et al. Hot News Recommendation System from Heterogeneous Websites
Based on Bayesian Model. The Scientific World Journal, 2014, (2014–6–26), 2014, 2014:7
5. Wei J, He J, Chen K, et al. Collaborative filtering and deep learning based recommendation
system for cold start items. Expert Systems with Applications, 2017, 69: 29-39
6. Liang P, Lan Y, Guo J, et al. Text Matching as Image Recognition. in: Proceedings of the AAAI
Conference on Artificial Intelligence. 20 l 6,30 (1):2793–2799

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-
NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/),
which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any
medium or format, as long as you give appropriate credit to the original author(s) and the source,
provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter’s Creative
Commons license, unless indicated otherwise in a credit line to the material. If material is not
included in the chapter’s Creative Commons license and your intended use is not permitted by
statutory regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder.