Body measure-aware fashion product recommendations: evaluating the predictive power of body scan data

Abstract

Fashion product consumer are faced with large and fast changing product o‚erings. Œe fashion purchase decision process is complex, as the consumer has to consider various inƒuencing factors like current fashion trends, what fashion products €t to their personality, and what products €t to their physical appearance like hair colors or body measures. Based on novel technologies, 3D body avatars can be reconstructed from 3D or 2D data. From these avatars, body measures can be determined. Œe objective of this research is to investigate the predictive performance of body measures extracted from a 3D body scanner for predicting fashion item preferences. Œerefore, item preferences and body scans from 200 users were collected. From the body scans, 11 body measures are extracted and integrated into a prediction model using Factorization Machines. Œe results from a cross-validation show, that including body measurements signi€cantly improves the prediction performance of the recommendation model, especially in new user scenarios, when no information about the fashion product preferences of the active user is known.

INTRODUCTION

Fashion products have the highest turnover of all product categories sold via e-commerce worldwide. Online fashion retailer o‚er their consumers large product ranges. For instance, the German online retailer Zalando indicates to have 150,000 products from 1,500 brands in their assortment 1.

Consumer behavior research indicates, that having a too broad product o‚er can cause a choice overload for the consumer, what leads to a delayed or no €nal purchase decision [ 7 ], and can decrease the consumers’ satisfaction regarding their purchase decision [ 17 ]. Œis e‚ect has also been identi€ed within fashion online purchase decisions [ 11 ]. To avoid choice overload, fashion online stores should limit the options for consumers in a way, so they are displayed with a sucient product variety but in the same time, they 1hŠps://www.zalando.de/presse zahlen-und-fakten/ ComplexRec 2017, Como, Italy. 2017. Copyright for the individual papers remains with the authors. Copying permiŠed for private and academic purposes. Œis volume is published and copyrighted by its editors. Published on CEUR-WS, Volume 1892.. can make informed decisions [ 11 ]. Especially fashion purchase decisions are complex, as multiple aspects of the user are inƒuecing factors like the personality, emotion, or general fashion trends as well as their physical appearance like the height [ 9 ], skin color, or body type [ 12 ]. Œe user has to decide not only based on what fashion products she or he in general likes, but also e.g., what colors €t his or her hair and skin color, and what cut of the clothes emphasizing or covers certain body parts.

Usually, recommender systems are applied to €lter relevant products for the visitors of online stores, leveraging techniques like collaborative €ltering or content-based €ltering. Œe central assumption of collaborative €ltering is, that users having demonstrated similar preferences in the past will have similar preferences in the future [ 3 ]. Œerefore, conventional collaborative €ltering based recommender systems only consider the product preferences per user in the form of ratings, views, or purchases and consequently derive similarities between users on these data. However, due to the wide adoption of new technologies like social networks, or mobile phones by the users as well as novel technologies to store and processing of large datasets, rich side information about users, items, and their interaction are available. Œerefore, further research is needed to investigate mechanisms for integrating such rich side information into recommendation systems, as well as to assess the impact of such side information on the prediction quality [ 18 ].

Based on novel technologies, information about the users’ physical appearance can be acquired. For instance, 3D models of bodies or faces can be constructed based on low-cost 3D scanners [ 21 ] [ 22 ], or reconstructed from 2D photos [ 8 ] [ 5 ]. User studies indicate that they perceive 3D scanner as useful for online shopping in general [ 20 ], as well as in the fashion product context [ 12 ], as long as data protection is ensured. However, according to a recent literature review about apparel product recommendation research, there is a research gap regarding the potential of body scan measurements to enrich consumer pro€les in apparel recommendation [ 4 ].

Œe objective of this research is to investigate the predictive power of body measures for predicting fashion product preferences of individual users. Œerefore, a dataset containing fashion product preferences as well as body measurements extracted from body scans are collected, and o„ine experiments are conducted to compare the impact of these measurements on the prediction accuracy. In this paper, €rst the data collection and the resulting data set is illustrated in Section 2, and preliminary results are demonstrated in Section 3. Finally, in Section 4 the results are discussed as well as planned next steps for further research are illustrated. 2 2.1

DATA COLLECTION

Data collection process

For data collection, volunteers were recruited from the School of Business at the University of Erlangen-Nuremberg. To obtain one rather large sample instead of two smaller ones, data collection was concentrated on only one gender. Œe decision was made to focus on female participants, as a higher variation in body shapes are expected as well as female participants are assumed to have stronger opinions according to their fashion preferences. Œe data collection process consists out of two steps. In the €rst step, participants were scanned using a low-cost body scanner which creates threedimensional polygonal mesh data. Based on the mesh data, 11 key body measurements per participant were extracted. In the second step, the participants €lled out an online questionnaire. In the online questionnaire, the participants gave information about their height and weight as well as indicated the color of their hair, eyes, and skin. Also, they classi€ed their body shape into one of eight shapes. Finally, every participant indicated their preference towards 36 fashion products by answering the question, whether they would buy the displayed clothing or not on a seven-point Likert scale (1=do not want to buy it at all; 7= would buy it de€nitely). For the set of fashion products, upper an lower apparel that rather accentuate or mask certain body features were selected. 2.2

Resulting data-set

In total, 200 persons participated in the survey and body scanning. Œe distribution of the eleven body measures are illustrated in Figure 3. From the 200 valid scans, the average age is 22.35 and the standard deviation 2.94 years. 3 3.1

EVALUATION

Evaluation protocol

For determining the prediction quality, the rating is converted from the seven-point Likert scale to a binary rating, indicating whether a consumer would buy the product or not, by transforming ratings 5 as not interested (=0), and ratings > 5 as interested (=1). In the following, the term rating prediction is therefore interpreted as a classi€cation problem. During the evaluation, the data is split up in a test and training set using a 10-fold cross-validation, as it is illustrated in Figure 4. Each user is randomly assigned to one of the ten test and train user-sets. Furthermore, to simulate the new user scenarios, the 36 items are randomly divided into 24 test and 12 train item-sets. Œe split is based on a random selection of items, which is illustrated in Figure 3, where the darker gray bars indicate the test items belonging to the test-set. For each iteration through the ten folds, the algorithm is provided with all 36 ratings of all of the users in the train user-set. Furthermore, to investigate the new user scenarios, the algorithm was provided with no, two, or €ve item ratings randomly selected from the train item-set. For all of the three scenarios, the ratings for the same 20 products of the test item-set are predicted.

For determining the prediction quality, the metrics precision (Equation 1), recall (Equation 2), and the F-measure (Equation 3) are used. In machine learning research, various aggregation approaches n ito100 u b i r t s i d t n e m e r u s a e m50 t s a e r b t s i a w p i h y ll e b t u o g e l n i g e l k c a b Body measures t s a e r b k c e n m r a h t d i w k c a b s e v l a c are used to aggregate the F-measures from the individual crossvalidation results, which lead to diverging results. In this research, we use the formulation of the F-measure suggested by Forman and Scholz (2010) which resulted in the least bias [ 2 ].

Body measure-aware fashion product recommendations by the additional information. Another approach is to build multidimensional models, also referred as contextual models, where the additional information is directly integrated into the prediction model [ 1 ]. Recently, especially tensor decomposition approaches gained aŠraction, enabling the direct modeling of multi-modal information. In previous research, the focus was especially on the Tucker decomposition [ 19 ], the Parallel Factor Analysis (PARAFAC) [ 6 ], and Pairwise Interaction Tensor Factorization (PITIF) [ 16 ], which demonstrated high prediction qualities. However, the main drawback of these methods is that their adaption to non-categorical factors is dicult and error-prone [ 14 ].

As an alternative, Factorization Machines (FM) were introduced which combine the advantages of the tensor decomposition approaches to make predictions in highly sparse and multi-modal conditions, with the ability of support-vector-machines (SVM) to be a general predictor [ 13 ]. With FM, the user, rating, and additional information can be modeled as feature vectors, having categorical or continuous values. Œe target variable y can be a real-valued rating or a binary value [ 13 ]. FM have been demonstrated to be an e‚ective approach to integrate contextual information, like mood, into movie recommender systems [ 15 ]. Œe key aspect of FM is, that the interactions between the input variables are not calculated directly, but a low-rank approximation is used. Within this paper, the FM model shown in Equation 4 is considered, which models binary interactions between the low-rank approximations V = hvi ; vj i 2 Rn k . Œe variable k 2 N0+ represents the number of latent variables. Appropriate values of k have to be determined empirically. On the one side, the value should be large enough so relevant interactions in the data can be captured. On the other side, restricting the value of k, and therefore its expressiveness might lead to beŠer generalization of the model. [ 13 ] n Õ i=1 Õn Õn i=1 j=i+1 yˆ¹x º := w0 + wi xi + hvi ; vj ixi xj (4)

In this paper, the reference implementation of FM within the LibFM library is used [ 14 ]. For learning models based on FM, optimization models based on Stochastic Gradient Descent (SGD), Alternative Least Squares (ALS), and Markov-Chain Monte Carlo (MCMC) are proposed, and implemented in LibFM. We decided to use the MCMC optimization, as this approach has the less hyperparameter which have to be tuned.

4 RESULTS AND DISCUSSION

Œe resulting F-measure values are illustrated in Figure 5. Œe models were built having latent variables values of k 2 f16; 32; 64; 128g. As mentioned in subsection 3.1, three new user scenarios are investigated, in which no, two, or €ve ratings of the user are known. In the €rst scenario, the models having the measurement information of the user have a considerable beŠer performance than the models without this information. In this scenario, the best model without information has a F-value of 0.40, whereas the best model with F-value a value of 0.49. Œis clearly beŠer performance shrinks in case of the second scenario, where two items are known and vanishes in the last scenario, where €ve items are known. Œe precision of the models having body measurement information is in all cases higher, but at the same time, the recall is lower compared to items

Fold 1 Fold 2 Fold 3 Fold 4 Fold 5 Fold 6 Fold 7 Fold 8 Fold 9 t e s r e s u n i a r t test item-set train item-set Fold 10 test user-set

3.2 Factorization Machines

In recommender systems research, various approaches were suggested to consider additional information besides the ratings of users per item. Œe additional information can be integrated via pre-€ltering or post-€ltering, where conventional recommendation algorithms are applied and the input data or the results are €ltered (1) (2) (3) 0.7 0.6 0.3 0.2 no measure with measure the model having no measure information. In total, the empirical results indicate, that body measures possess signi€cant predictive power in the context of apparel recommendation, especially in new users scenarios, where no previous user product preference information are available. In practice, one situation could be when consumers are geŠing scanned in a store the €rst time.

Further research is needed to identify which speci€c body measures have predictive power and which rather introduce noise to the model. From the eleven measures used, it can be expected, that some measures like the hip or belly measure give more information to the model than for example the calves measure. Another possibility to integrate body scan information is to assign each scan to a distinctive body shape class [ 10 ], or use the principal components from the morphable model approach [ 22 ]. In addition, the impact of the hair and skin color nuances on the predictive performance will be investigated in further research.

References

[1] Gediminas Adomavicius and Alexander Tuzhilin . 2008 . Context-aware Recommender Systems . In Proceedings of the 2008 ACM Conference on Recommender Systems (RecSys '08) . ACM, New York, NY, USA, 335 - 336 . DOI:hŠp: //dx.doi.org/10.1145/1454008.1454068

[2] George Forman and Martin Scholz . 2010 . Apples-to-apples in Cross-validation Studies: Pitfalls in Classi€er Performance Measurement . SIGKDD Explor. Newsl . 12 , 1 ( 2010 ), 49 - 57 . DOI:hŠp://dx.doi.org/10.1145/1882471.1882479

[3] David Goldberg , David Nichols, Brian M. Oki , and Douglas Terry . 1992 . Using Collaborative Filtering to Weave an Information Tapestry . Commun. ACM 35 , 12 ( 1992 ), 61 - 70 . DOI:hŠp://dx.doi.org/10.1145/138859.138867

[4] Congying Guan , Shengfeng Qin, Wessie Ling, and Guofu Ding . 2016 . Apparel recommendation system evolution: an empirical review . International Journal of Clothing Science and Technology 28 , 6 ( 2016 ), 854 - 879 . DOI:hŠp://dx.doi.org/10. 1108/IJCST-09-2015-0100

[5] Peng Guan , Alexander Weiss, Alexandru O. Balan , and Michael J. Black . 2009 . Estimating human shape and pose from a single image . In Computer Vision , 2009 IEEE 12th International Conference on. 1381 - 1388 .

[6] Richard Harshman . 1970 . Foundations of the PARAFAC procedure: Models and conditions for an “explanatory” multi-modal factor analysis . UCLA Working Papers in Phonetics 16 ( 1970 ).

[7] Sheena S. Iyengar and Mark R. Lepper . 2000 . When choice is demotivating: Can one desire too much of a good thing? Journal of personality and social psychology 79 , 6 ( 2000 ), 995 .

[8] Luo Jiang , Juyong Zhang, Bailin Deng, Hao Li , and Ligang Liu . 2017 . 3D Face Reconstruction with Geometry Details from a Single Image . ( 2017 ). hŠp://arxiv. org/pdf/1702.05619

[9] Michelle R. Jones and Valerie L. Giddings . 2010 . Tall women's satisfaction with the €t and style of tall women's clothing . Jnl of Fashion Mrkting and Mgt 14 , 1 ( 2010 ), 58 - 71 . DOI:hŠp://dx.doi.org/10.1108/13612021011025438

[10] Jeong Yim Lee , Cynthia L. Istook, Yun Ja Nam, and Sun Mi Park. 2007 . Comparison of body shape between USA and Korean women . Int Jnl of Clothing Sci & Tech 19 , 5 ( 2007 ), 374 - 391 . DOI:hŠp://dx.doi.org/10.1108/09556220710819555

[11] Komal Nagar and Payal Gandotra . 2016 . Exploring Choice Overload, Internet Shopping Anxiety, Variety Seeking and Online Shopping Adoption Relationship: Evidence from Online Fashion Stores . Global Business Review 17 , 4 ( 2016 ), 851 - 869 . DOI:hŠp://dx.doi.org/10.1177/0972150916645682

[12] Jaekyung Park , Yunja Nam, Kueng-mi Choi , Yuri Lee, and Kyu-Hye Lee . 2009 . Apparel consumers' body type and their shopping characteristics . Jnl of Fashion Mrkting and Mgt 13 , 3 ( 2009 ), 372 - 393 . DOI:hŠp://dx.doi.org/10.1108/ 13612020910974500

[13] Ste‚en Rendle . 2010 . Factorization Machines . In Proceedings of the 2010 IEEE International Conference on Data Mining (ICDM '10) . IEEE Computer Society, Washington, DC, USA, 995 - 1000 . DOI:hŠp://dx.doi.org/10.1109/ICDM. 2010 .127

[14] Ste‚en Rendle . 2012 . Factorization Machines with libFM . ACM Trans. Intell. Syst. Technol. 3 , 3 ( 2012 ), 57 : 1 - 57 : 22 .

[15] Stefen Rendle , Zeno Gantner, Christoph Freudenthaler, and Schmidt-Πieme Lars . 2011 . Fast Context-aware Recommendations with Factorization Machines: 34th International ACM SIGIR Conference on Research and Development in Information Retrieval ; July 24 - 28 , 2011 , Beijing, China. ACM, New York, NY.

[16] Ste ‚en Rendle and Lars Schmidt- Œ ieme . 2010 . Pairwise Interaction Tensor Factorization for Personalized Tag Recommendation . In Proceedings of the Šird ACM International Conference on Web Search and Data Mining (WSDM '10) . ACM, New York, NY, USA, 81 - 90 . DOI:hŠp://dx.doi.org/10.1145/1718487.1718498

[17] Barry Schwartz . 2016 . Še paradox of choice: Why more is less (revised edition ed .).

[18] Yue Shi , Martha Larson , and Alan Hanjalic . 2014 . Collaborative Filtering Beyond the User-Item Matrix: A Survey of the State of the Art and Future Challenges . ACM Comput. Surv . 47 , 1 ( 2014 ), 3 : 1 - 3 : 45 . DOI:hŠp://dx.doi.org/10.1145/2556270

[19] Ledyard R. Tucker . 1966 . Some mathematical notes on three-mode factor analysis . Psychometrika 31 , 3 ( 1966 ), 279 - 311 . DOI:hŠp://dx.doi.org/10.1007/BF02289464

[20] Christian Zagel and Jochen Su¨ßmuth. 2013 . Nutzenpotenziale maßgetreuer 3D Avatare aus Low-cost Bodyscannern . HMD : Praxis der Wirtscha‡sinformatik 50 ( 2013 ), 48 - 57 . DOI:hŠp://dx.doi.org/10.1007/BF03342068

[21] Christian Zagel , Jochen Su¨ßmuth, and Freimut Bodendorf . 2013 . Automatische Rekonstruktion eines 3D Ko¨rpermodells aus Kinect Sensordaten . In Wirtscha‡sinformatik Proceedings 2013. Paper 35 . hŠp://aisel.aisnet. org/wi2013/35

[22] Michael Zollho ¨fer, Michael Martinek , Gu¨nther Greiner, Marc Stamminger, and Jochen Su¨ßmuth. 2011 . Automatic Reconstruction of Personalized Avatars from 3D Face Scans . Computer Animation and Virtual Worlds (Proceedings of CASA 2011 ) 22 , 2 - 3 ( 2011 ), 195 - 202 .