Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Document Classification | Copyright © 2020 Tata Consultancy Services Limited
Product age based demand forecast model for fashion retailRajesh Kumar Vashishtha, Vibhati Burman, Rajan Kumar, Srividhya Sethuraman, Abhinaya R Sekar, Sharadha Ramanan
TCS
KDD 2020 Workshop, August 2020, San Diego, California-USA
(r.vashishtha, vibhati.b, rajan.7, srividhya.sethuraman1, abhinayar.sekar, sharadha.ramanan)@tcs.com
TCS Research and InnovationChennai, India
Document Classification | Copyright © 2020 Tata Consultancy Services LimitedTCS2
Introduction
Comparative Methods
Contents
1
2
3
4
5
6
Problem Statement
Problem Formulation
Methodology
Data Characterstics
7
8
Results
Conclusion
Document Classification
Introduction
3
1. Introduction 2. Contributions● Accurate forecasts are important and
challenging
● Important to ensure retailer’s profitability and to reduce environmental damage caused by disposal of unsold inventory
● Challenging because most products are new in a season with no historical data and have short life cycles, huge sales variations and long lead-times
● Leveraging age of a product in its demand forecasting
● Unique and significant feature engineering to achieve accurate demand forecasts
● Extending our approach to recommend product launch time for the next season
● Assortment optimization
Document Classification
Problem Statement and Problem formulation
4
Accurate Demand Forecast for a fashion retailer for the upcoming season, 6 months in advance.
I/p
o/p
● Product Attribute data from previous year and current year● Historical Sales data of previous year’s products● Product pricing data for current year● Store capacity data
● Demand forecast for each store-article for the next season
Size-profiling
I/p
o/p
● Demand forecasts at store-article level
● Demand forecast results at store-article-size level
Document Classification
Problem Statement and Problem formulation
5
Ideal Time For Launch
I/p
o/p
● Demand Forecast models● I/p data with age shifted by 5 weeks
● Ideal time(phase) for launch of a product
Assortment Optimization
I/p
o/p
● Demand forecast model output● Business Constraints
● List of articles that is to be kept in the store for each phase
Document Classification
Methodology – Demand Formulation
6
Di , s=f (Ai , t , Age i , s , Pi , s , t , nholidayst , S . I .)
In the age based model for demand forecasting of fashion items, the demand D of an article i in store s at time t, is formulated as:
By using this formulation, we are collating the age, attribute values, temporal features and selling price of an article as demand estimators. This is done because the visual characteristics, selling price and the temporal features alone are insufficient to accurately predict the future demand.
Document Classification
Feature Engineering – Age
7
Effect of translation from launch time to age
Document Classification
Feature Engineering – Color and Usage
8
“We've combined the comfort of a t-shirt with the elegance of a midi dress to bring you one dreamy summer staple. Regular cut, with plenty of wiggle room through the body. Striped design, with round neckline, short sleeves and splits at the hem for easier movement ”
1. Usage 2. Color
Usage category: Summer
Color name to equivalent RGB value
Document Classification
Methodology – Train and test flow diagram
9
TrainFlowDiagram
TestFlowDiagram
Document Classification
Methodology – Size Profiling and Launch Date
10
Size Profiling
● Estimate demand at size level, i.e. quantity of XS, S, M, L, XL, XXL
● Group previous year’s sales data at store-size level. Cluster similar stores based on the sales contribution by each size within a store
● Assumption : Customer size profile in a region is consistent year on year
Product Launch date
● Four different launch times are considered in a season, i.e. the beginning of each phase
● We iteratively forecast demand for an article considering different launch times. We then select the best launch time for an article based on the criteria of maximizing revenue
Document Classification
Methodology – Fashion Assortment Optimization
11
The objective is to obtain a list of articles to be kept in a store such that overall revenue is maximized and business constraints are satisfied. The assotment optimization is formulated as integer linear programming as follows:
Document Classification
Data Characterstics
12
● Multinational fashion retailer
● 300 B&M stores, 40 categories and around 35k items
● Product attributes, historical performance data, store size
● Two categories (owing to their high contribution to sales): - Dresses in Women's wear - Kids wear
● Analysis of 3 years data - 1 year = 2 seasons - A season = 6 months = 4 phases of 6-7 weeks
● Train: 1 January 2018 to 15 June 2018
● Test: 1 January 2019 to 15 June 2019
Document Classification
Comparative Methods
13
Average model:
Average value of similar items from the previous season
Baseline XGBoost model:
All features except age, RGB encoded color and Usage
Baseline DL model:
All features except age, RGB encoded color and Usage
Document Classification
Experimental Setting – Age based Model
14
Architechture of the age based DL model
Document Classification
Results - Kidswear
15
Model Loss Function
RMSEArticle-
Store-Week
RMSEArticle-
Store-Phase
RMSEArticle-Store-Season
MAEArticle-
Store- Week
MAEArticle-
Store- Phase
MAEArticle-Store-Season
Averagemodel
MSE 3.56 10.55 39.27 3.07 7.25 28.8
Baseline XGboost
MSE 2.97 7.61 31.89 2.53 6.13 22.76
Baseline DL MSE 2.84 7.63 30.23 2.61 6.11 21.31Age- based
XGBoostMSE 1.65 4.92 15.38 1.21 2.83 7.12
Age- based XGBoost
Huber 1.64 4.96 15.43 1.19 2.68 6.93
Age- based XGBoost
logcosh 1.61 4.89 15.41 1.17 2.54 6.85
Age-based DL MSE 1.54 4.32 13.04 1.15 2.41 6.53Age-based DL Huber 1.56 4.35 13.15 1.21 2.34 6.68Age-based DL logcosh 1.51 4.29 12.95 1.13 2.35 7.34
Document Classification
Results – Dresses in Women’s wear
16
Model Loss Function
RMSEArticle-
Store-Week
RMSEArticle-
Store-Phase
RMSEArticle-Store-Season
MAEArticle-
Store-Week
MAE Article-
Store-Phase
MAEArticle-Store-Season
Averagemodel
MSE 6.83 13.62 39.20 3.54 8.09 26.80
Baseline XGboost
MSE 5.14 12.36 34.70 3.81 7.82 22.93
Baseline DL MSE 5.21 11.93 32.12 2.99 7.53 21.38Age- based
XGBoostMSE 1.98 9.89 22.34 1.61 5.71 12.13
Age- based XGBoost
Huber 4.23 10.36 23.71 2.48 7.95 12.13
Age- based XGBoost
logcosh 3.95 8.29 21.17 2.37 7.63 11.46
Age-based DL MSE 2.86 9.67 22.79 2.71 6.84 15.67Age-based DL Huber 4.09 10.81 25.12 2.64 7.93 15.43Age-based DL logcosh 4.21 9.55 24.93 2.64 6.05 15.78
Document Classification
Results- Kidswear and Dresses in women’s wear
17
Actuals vs predicted for kidswear
Actuals vs predicted for dresses in women’s wear
Document Classification
Size profiling
18
The different plot legends labelled 0, 1 and 2 represents three different size profiles across the stores. We observe that although the three size profiles have almost similar distribution, there is some variation in percentage share for each size.
Document Classification
Ideal time for launch of a product in a store
19
Document Classification
Robust performance of the age based forecast model - real world use case of a multinational fashion retailer
Recommending the ideal time of launch for each new item in the next season
List of items and number of units to be listed in a store for next season
Average uplift of 41% in revenue when compared to the retailer’s plan.
Future Work : Usage of trend, competitor data, Images, footfalls, and number of changing rooms
Conclusion and Future Work
Document Classification21
References
[1] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, San- jay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dandelion Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. 2015. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems. https://www.tensorflow.org/ Software available from tensorflow.org.
[2] Kin-Fan Au, Tsan-Ming Choi, and Yong Yu. 2008. Fashion retail forecasting by evolutionary neural networks. International Journal of Production Economics 114, 2 (2008), 615–630.
[3] Leonie Barrie. [n. d.]. Apparel overproduction a risk firms seem to accept. Ac- cessed: 2020-02-13.
[4] James Bergstra, Daniel Yamins, and David Daniel Cox. 2013. Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures. (2013).
[5] Dirk M Beyer, Fereydoon Safai, and Farid AitSalia. 2005. Profile-based product demandforecasting. USPatent6,978,249.
[6] Vivek F Farias, Srikanth Jagabathula, and Devavrat Shah. 2017. Building optimized and hyperlocal product assortments: A nonparametric choice approach. Available at SSRN 2905381 (2017).
https://www.tensorflow.org/
Document Classification22
References
[[7] AndrewGelman,JohnBCarlin,HalSStern,DavidBDunson,AkiVehtari,and Donald B Rubin. 2013. Bayesian data analysis. CRC press.
[8] Majd Kharfan and Vicky Wing Kei Chan. 2018. Forecasting Seasonal Footwear Demand Using Machine Learning. (2018).
[9] Na Liu, Shuyun Ren, Tsan-Ming Choi, Chi-Leung Hui, and Sau-Fun Ng. 2013. Sales forecasting for fashion retailing service industry: a review. Mathematical Problems in Engineering 2013 (2013).
[10] Ana LD Loureiro, Vera L Miguéis, and Lucas FM da Silva. 2018. Exploring the useof deep neural networks for sales forecasting in fashion retail.Decision SupportSystems114 (2018), 81–93.
[11] Paris A Mastorocostas, John B Theocharis, and Vassilios S Petridis. 2001. A constrained orthogonal least-squares method for generating TSK fuzzy models: application to short-term load forecasting. Fuzzy Sets and Systems 118, 2 (2001),215–233.
[12] Maria Elena Nenni, Luca Giustiniano, and Luca Pirolo. 2013. Demand forecasting in the fashion industry: a review. International Journal of Engineering Business Management 5 (2013), 37.1] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, San- jay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dandelion Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. 2015. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems. https://www.tensorflow.org/ Software available from tensorflow.org
https://www.tensorflow.org/
Document Classification23
References
[11]Paris A Mastorocostas, John B Theocharis, and Vassilios S Petridis. 2001. Aconstrained orthogonal least-squares method for generating TSK fuzzy models:application to short-term load forecasting.Fuzzy Sets and Systems118, 2 (2001),215–233.
[12]Maria Elena Nenni, Luca Giustiniano, and Luca Pirolo. 2013. Demand forecastingin the fashion industry: a review.International Journal of Engineering BusinessManagement5 (2013), 37.
[13]F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M.Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cour-napeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: MachineLearning in Python.Journal of Machine Learning Research12 (2011), 2825–2830.
[14]Kumar Rajaram. 2001. Assortment planning in fashion retailing: methodology,application and analysis.European Journal of Operational Research129, 1 (2001),186–208.
[15]Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. 2018.How does batch normalization help optimization?. InAdvances in Neural Infor-mation Processing Systems. 2483–2493.
[16]Pawan Kumar Singh, Yadunath Gupta, Nilpa Jha, and Aruna Rajan. 2019. FashionRetail: Forecasting Demand for New Items.arXiv preprint arXiv:1907.01960(2019).
[17]Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and RuslanSalakhutdinov. 2014. Dropout: a simple way to prevent neural networks fromoverfitting.The journal of machine learning research15, 1 (2014), 1929–1958.[18]Sébastien Thomassey and Antonio Fiordaliso. 2006. A hybrid sales forecastingsystem based on clustering and decision trees.Decision Support Systems42, 1(2006), 408–421.[19]WK Wong and ZX Guo. 2010. A hybrid intelligent model for medium-term salesforecasting in fashion retail supply chains using extreme learning machine andharmony search algorithm.International Journal of Production Economics128, 2(2010), 614–624
Document Classification24
References
[18]Sébastien Thomassey and Antonio Fiordaliso. 2006. A hybrid sales forecastingsystem based on clustering and decision trees.Decision Support Systems42, 1(2006), 408–421.
[19]WK Wong and ZX Guo. 2010. A hybrid intelligent model for medium-term salesforecasting in fashion retail supply chains using extreme learning machine andharmony search algorithm.International Journal of Production Economics128, 2(2010), 614–624
Thank You
Slide 1Slide 2Presentation GuidelinesSlide 4Slide 5Slide 6Slide 7Slide 8Slide 9Slide 10Slide 11Slide 12Slide 13Slide 14Slide 15Slide 16Slide 17Slide 18Slide 19Slide 20Slide 21Slide 22Slide 23Slide 24Slide 25