Early Bird Catches the Worm: Predicting Returns Even Before Purchase in Fashion E-commerce

Manchit Madan; Sajan Kedia; Sumit Borar

arxiv: 1906.12128 · v1 · pith:3UG4OCG5new · submitted 2019-06-28 · 💻 cs.LG · stat.ML

Early Bird Catches the Worm: Predicting Returns Even Before Purchase in Fashion E-commerce

Sajan Kedia , Manchit Madan , Sumit Borar This is my paper

Pith reviewed 2026-05-25 13:36 UTC · model grok-4.3

classification 💻 cs.LG stat.ML

keywords return predictionfashion e-commercedeep neural networkproduct embeddingsBPRskip-grampre-purchase intervention

0 comments

The pith

A deep neural network predicts the probability that a customer will return a fashion item before the order is placed.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that return likelihood can be estimated in real time at the cart stage by feeding a deep neural network with two kinds of embeddings plus hand-crafted features. One embedding set comes from Bayesian Personalized Ranking and captures latent user tastes and product attributes; the second comes from a skip-gram model and encodes body shape and size. If the predictions are reliable, platforms can intervene before checkout to lower the volume of returns and the associated reverse-logistics costs.

Core claim

The central claim is that a deep neural network that combines Bayesian Personalized Ranking product embeddings, skip-gram body-size embeddings, and engineered customer and product features produces usable estimates of return probability prior to purchase, and that these estimates can be deployed in live experiments that measurably reduce overall returns on a large fashion e-commerce site.

What carries the argument

Deep neural network that fuses Bayesian Personalized Ranking product embeddings, skip-gram body embeddings, and engineered features to output a return probability score.

If this is right

Platforms can surface targeted messages or alternative recommendations at the cart page to discourage likely returns.
Return-probability scores can be computed for millions of customers in real time without waiting for post-purchase data.
Early detection enables preemptive actions that lower reverse-logistics and liquidation costs.
The same pipeline can be used to test different intervention strategies whose effectiveness is measured by change in realized return rates.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the embeddings capture stable user and product signals, the same architecture could be retrained on other high-return categories such as electronics or furniture.
The body-size embedding might also improve size-recommendation models that reduce returns caused by fit problems.
Because predictions occur before the order, the method could be combined with dynamic pricing or shipping offers that further steer customers away from low-fit items.

Load-bearing premise

The chosen embeddings and engineered features already contain enough information to forecast whether a given customer will return a specific product before any purchase occurs.

What would settle it

A controlled live experiment in which the model-driven pre-purchase interventions produce no reduction in return rate relative to a matched control group.

Figures

Figures reproduced from arXiv: 1906.12128 by Manchit Madan, Sajan Kedia, Sumit Borar.

**Figure 1.** Figure 1: First two images from the left demonstrate personalized delivery charges for the same user having different carts. The third image shows a non-returnable option by giving a small additional coupon for the product having high return probability 2. RELATED WORK In this section, we have briefly reviewed the existing literature on return prediction. In [11] authors used Mahalanobis Feature Extraction to pr… view at source ↗

**Figure 2.** Figure 2: BPR-MF Architecture 3.2 Personalized Sizing latent features: Skipgram approach Size and fit is one of the biggest reason for returns in fashion. It’s difficult to address this problem since all brands have different fit even for the same size number. From a customer’s point of view, it’s a major challenge to decide the perfect size number for different brands. To address this problem, sizing vectors pers… view at source ↗

**Figure 3.** Figure 3: A fully connected deep neural network model Probability. Subject to this Return Probability, appropriate action is taken by the Decision Engine which is then passed back to the Cart API. This real-time architecture empowers us to predict returns at the cart page before an order is placed and take a preemptive action to minimize loss [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Real-time Architecture 4. RESULTS AND ANALYSIS This section demonstrates the results and analysis of the experiments. First, various data sources used for the modeling purpose are discussed. Thereafter some interesting insights derived from the analysis of the data are presented. Further, the evaluation metrics are discussed. At last, the comparison of different models on key evaluation metrics is done. … view at source ↗

**Figure 6.** Figure 6: Precision Recall value at different thresh [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗

**Figure 7.** Figure 7: depicts the performance of all the models in terms of ROC curve [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗

read the original abstract

With the rapid growth in fashion e-commerce and customer-friendly product return policies, the cost to handle returned products has become a significant challenge. E-tailers incur huge losses in terms of reverse logistics costs, liquidation cost due to damaged returns or fraudulent behavior. Accurate prediction of product returns prior to order placement can be critical for companies. It can facilitate e-tailers to take preemptive measures even before the order is placed, hence reducing overall returns. Furthermore, finding return probability for millions of customers at the cart page in real-time can be difficult. To address this problem we propose a novel approach based on Deep Neural Network. Users' taste & products' latent hidden features were captured using product embeddings based on Bayesian Personalized Ranking (BPR). Another set of embeddings was used which captured users' body shape and size by using skip-gram based model. The deep neural network incorporates these embeddings along with the engineered features to predict return probability. Using this return probability, several live experiments were conducted on one of the major fashion e-commerce platform in order to reduce overall returns.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Standard BPR and skip-gram embeddings fed into a DNN for pre-purchase return prediction, with the main evidence coming from live A/B tests on a production fashion platform.

read the letter

The paper takes known embedding methods and applies them to forecast returns before an order is placed. BPR captures product preferences, skip-gram captures body shape from user data, and both go into a DNN alongside engineered features to output a return probability. They then use that score in live experiments on a major e-commerce site to drive interventions at the cart stage. That deployment story and the reported reduction in returns are the concrete parts worth noting. The work shows the pipeline can run at scale for millions of users in real time, which is a practical constraint many academic models ignore. The live A/B results give some external check on whether the predictions translate to fewer returns, rather than just higher offline accuracy. On the modeling side, nothing here is new methodologically. BPR and skip-gram are established, the DNN is standard, and the claim of novelty rests on the specific use case rather than any algorithmic advance. The abstract does not include equations, loss details, or ablation controls, so it is difficult to isolate how much the embeddings add over simpler features. If the full paper has those controls and shows the embeddings matter, that strengthens the case; otherwise the contribution stays at the level of a solid engineering application. No obvious data leakage or circularity appears in the described setup. This is useful reading for teams already working on e-commerce return models who want an example of production integration. It is less relevant for readers looking for new theory or generalizable techniques. The live results make it worth sending to referees for a closer check on the experimental design and effect sizes.

Referee Report

0 major / 2 minor

Summary. The paper claims that a deep neural network incorporating Bayesian Personalized Ranking (BPR) product embeddings (to capture user taste and latent product features), skip-gram embeddings (to capture user body shape and size), and engineered features can predict product return probabilities prior to purchase in fashion e-commerce. Several live experiments on a major platform are reported to demonstrate that using these probabilities enables preemptive measures that reduce overall returns.

Significance. If the live A/B experiments isolate the contribution of the return-probability model and show a meaningful reduction in returns, the work has clear practical significance for e-commerce platforms facing high reverse-logistics and liquidation costs. The combination of BPR and skip-gram embeddings with engineered features is a reasonable way to extract pre-purchase signals, and the use of production-platform experiments strengthens the applied claim.

minor comments (2)

[Abstract] The abstract states that 'several live experiments were conducted' but provides no quantitative results (e.g., return-rate deltas, statistical significance, or A/B test sizes). Adding these numbers would strengthen the central claim without altering the modeling pipeline.
The description of the DNN (how the two embedding sets and engineered features are concatenated, the network depth/width, loss function, and training procedure) is referenced only at a high level. A diagram or explicit equations would improve clarity and reproducibility.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our manuscript, the accurate summary of our contributions, and the recommendation for minor revision. No specific major comments were raised in the report.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The manuscript presents a standard supervised DNN pipeline that ingests pre-trained BPR product embeddings, skip-gram body-shape embeddings, and hand-engineered features to output a return-probability score; this score is then used in live A/B tests. No equations, fitted parameters, or self-citations are shown that would make the model output definitionally identical to its inputs. The derivation chain is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no explicit free parameters, axioms, or invented entities are stated. The central claim implicitly rests on the unstated assumption that historical purchase data contains predictive signal for future returns.

pith-pipeline@v0.9.0 · 5721 in / 1057 out tokens · 17672 ms · 2026-05-25T13:36:51.649835+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 2 internal anchors

[1]

Early Bird Catches the Worm: Predicting Returns Even Before Purchase in Fashion E-commerce

INTRODUCTION E-commerce sector is the fastest growing sector, with an ex- pectation to reach $4 trillion by 2020 1. In today’s compet- itive market, to increase customer experience, most compa- nies are coming up with hassle-free return policies [1]. Cus- tomers are usually allowed a return within a month or a sim- ilar time period. This policy [2] has im...

work page internal anchor Pith review Pith/arXiv arXiv 2020
[2]

In [11] authors used Maha- lanobis Feature Extraction to predict product returns in e- commerce

RELA TED WORK In this section, we have brieﬂy reviewed the existing lit- erature on return prediction. In [11] authors used Maha- lanobis Feature Extraction to predict product returns in e- commerce. Here a prediction can only be made once the user has placed an order, due to which e-tailer can’t take preventive measures to reduce the returns. Also, this ...

work page
[3]

Nike-Men- Shoes-Sports-10

METHODOLOGY Given a useru and current cartC which can contain multiple products, our objective is to come up with return probabil- ity score denoted by f(u,C ) for a user & cart combination. This probability score is used to implement the action items mentioned in Section 1 to reduce the overall returns. Anal- ysis showed that number of similar items in a...

work page
[4]

• n style group as the number of same products with diﬀerent colors in a cart

Product-level features • brand, category, product age • product level return score for diﬀerent periods of time- last month, quarter, half-year, year. • n style group as the number of same products with diﬀerent colors in a cart. • n similar products as the number of similar prod- ucts in a cart

work page
[5]

Cart level features • cart size, order day (weekday/weekend), order time (morning/evening) • brand, category, delivery city, platform type

work page
[6]

Only important values of the features were transformed by using 65th percentile as the threshold of their occurrence in the data

User level features • Based on the lifetime data of all users, features like return count, revenue, order count, quantity, payment mode, purchase frequency, were created • Applied clustering techniques at diﬀerent levels to design new user features like user persona, en- gagement level, discount aﬃnity score, customer type, purchase intent One hot encoder...

work page
[7]

First, various data sources used for the modeling purpose are discussed

RESULTS AND ANALYSIS This section demonstrates the results and analysis of the ex- periments. First, various data sources used for the modeling purpose are discussed. Thereafter some interesting insights derived from the analysis of the data are presented. Further, the evaluation metrics are discussed. At last, the compari- son of diﬀerent models on key e...

work page
[8]

Experiments were conducted on three action items from the suggested ones in Section 1

EXPERIMENTS In this section, the live experiments performed on one of the leading fashion platform are discussed. Experiments were conducted on three action items from the suggested ones in Section 1. First experiment comprised of applying person- alized delivery charge which was diﬀerent for all customers depending on their respective live cart. The dual...

work page
[9]

To solve this, a novel ap- proach is proposed in this paper to detect returns and imple- ment certain actions items

CONCLUSION E-tailers experience return-rate problem leading to increased costs and lower proﬁt margins. To solve this, a novel ap- proach is proposed in this paper to detect returns and imple- ment certain actions items. Depending on the current cart conﬁguration of a user, a hybrid dual-model is built using a deep neural network to detect returnable cart...

work page
[10]

Product re- turns on the internet: A case of mixed signals?

C. Boniﬁeld, C. Cole, and R. L. Schultz, “Product re- turns on the internet: A case of mixed signals?” Jour- nal of Business Research , vol. 63, no. 9-10, pp. 1058– 1065, 2010

work page 2010
[11]

Equilibrium returns policies in the presence of supplier competition,

S. Bandyopadhyay and A. Paul, “Equilibrium returns policies in the presence of supplier competition,” Mar- keting Science, vol. 29, pp. 846–857, 09 2010

work page 2010
[12]

Fraudulent consumer returns: exploiting retailers’ return policies,

L. C. Harris, “Fraudulent consumer returns: exploiting retailers’ return policies,” European Journal of Market- ing, vol. 44, no. 6, pp. 730–747, 2010. [Online]. Avail- able: https://doi.org/10.1108/03090561011032694

work page doi:10.1108/03090561011032694 2010
[13]

The impact of customer returns on pricing and order decisions,

J. Chen and P. C. Bell, “The impact of customer returns on pricing and order decisions,” European Journal of Operational Research, vol. 195, pp. 280–295, 2009

work page 2009
[14]

Comparative analysis of the carbon footprints of conventional and online retailing: A last mile perspective,

J. B. Edwards, A. C. McKinnon, and S. L. Cullinane, “Comparative analysis of the carbon footprints of conventional and online retailing: A last mile perspective,” International Journal of Physical Distribution & Logistics Management , vol. 40, no. 1/2, pp. 103–123, 2010. [Online]. Available: https://doi.org/10.1108/09600031011018055

work page doi:10.1108/09600031011018055 2010
[15]

Predictive model selection for forecasting product returns,

J. Ma and H. M. Kim, “Predictive model selection for forecasting product returns,” Journal of Mechanical Design, vol. 138, no. 5, p. 054501, 2016

work page 2016
[16]

Toktay, Forecasting product returns

B. Toktay, Forecasting product returns. INSEAD, 2001

work page 2001
[17]

Modeling and fore- casting product returns: An industry case study,

A. Canda, X. Yuan, and F. Wang, “Modeling and fore- casting product returns: An industry case study,” in 2015 IEEE International Conference on Industrial En- gineering and Engineering Management (IEEM) , Dec 2015, pp. 871–875

work page 2015
[18]

Bpr: Bayesian personalized rank- ing from implicit feedback,

S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme, “Bpr: Bayesian personalized rank- ing from implicit feedback,” in Proceedings of the twenty-ﬁfth conference on uncertainty in artiﬁcial in- telligence. AUAI Press, 2009, pp. 452–461

work page 2009
[19]

Efficient Estimation of Word Representations in Vector Space

T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Eﬃcient estimation of word representations in vector space,” CoRR, vol. abs/1301.3781, 2013. [Online]. Available: http://arxiv.org/abs/1301.3781

work page internal anchor Pith review Pith/arXiv arXiv 2013
[20]

Predicting prod- uct returns in e-commerce: The contribution of maha- lanobis feature extraction,

P. Urbanke, J. Kranz, and L. Kolbe, “Predicting prod- uct returns in e-commerce: The contribution of maha- lanobis feature extraction,” in ICIS, 2015

work page 2015
[21]

Probabilistic ma- trix factorization,

A. Mnih and R. R. Salakhutdinov, “Probabilistic ma- trix factorization,” in Advances in neural information processing systems, 2008, pp. 1257–1264

work page 2008
[22]

Decoding fashion contexts using word embeddings

S. Arora and D. Warrier, “Decoding fashion contexts using word embeddings.”

work page
[23]

Non-negative matrix factorization with sparseness constraints,

P. O. Hoyer, “Non-negative matrix factorization with sparseness constraints,” Journal of machine learning research, vol. 5, no. Nov, pp. 1457–1469, 2004

work page 2004
[24]

Distributed representations of words and phrases and their compositionality,

T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean, “Distributed representations of words and phrases and their compositionality,” in NIPS, 2013

work page 2013

[1] [1]

Early Bird Catches the Worm: Predicting Returns Even Before Purchase in Fashion E-commerce

INTRODUCTION E-commerce sector is the fastest growing sector, with an ex- pectation to reach $4 trillion by 2020 1. In today’s compet- itive market, to increase customer experience, most compa- nies are coming up with hassle-free return policies [1]. Cus- tomers are usually allowed a return within a month or a sim- ilar time period. This policy [2] has im...

work page internal anchor Pith review Pith/arXiv arXiv 2020

[2] [2]

In [11] authors used Maha- lanobis Feature Extraction to predict product returns in e- commerce

RELA TED WORK In this section, we have brieﬂy reviewed the existing lit- erature on return prediction. In [11] authors used Maha- lanobis Feature Extraction to predict product returns in e- commerce. Here a prediction can only be made once the user has placed an order, due to which e-tailer can’t take preventive measures to reduce the returns. Also, this ...

work page

[3] [3]

Nike-Men- Shoes-Sports-10

METHODOLOGY Given a useru and current cartC which can contain multiple products, our objective is to come up with return probabil- ity score denoted by f(u,C ) for a user & cart combination. This probability score is used to implement the action items mentioned in Section 1 to reduce the overall returns. Anal- ysis showed that number of similar items in a...

work page

[4] [4]

• n style group as the number of same products with diﬀerent colors in a cart

Product-level features • brand, category, product age • product level return score for diﬀerent periods of time- last month, quarter, half-year, year. • n style group as the number of same products with diﬀerent colors in a cart. • n similar products as the number of similar prod- ucts in a cart

work page

[5] [5]

Cart level features • cart size, order day (weekday/weekend), order time (morning/evening) • brand, category, delivery city, platform type

work page

[6] [6]

Only important values of the features were transformed by using 65th percentile as the threshold of their occurrence in the data

User level features • Based on the lifetime data of all users, features like return count, revenue, order count, quantity, payment mode, purchase frequency, were created • Applied clustering techniques at diﬀerent levels to design new user features like user persona, en- gagement level, discount aﬃnity score, customer type, purchase intent One hot encoder...

work page

[7] [7]

First, various data sources used for the modeling purpose are discussed

RESULTS AND ANALYSIS This section demonstrates the results and analysis of the ex- periments. First, various data sources used for the modeling purpose are discussed. Thereafter some interesting insights derived from the analysis of the data are presented. Further, the evaluation metrics are discussed. At last, the compari- son of diﬀerent models on key e...

work page

[8] [8]

Experiments were conducted on three action items from the suggested ones in Section 1

EXPERIMENTS In this section, the live experiments performed on one of the leading fashion platform are discussed. Experiments were conducted on three action items from the suggested ones in Section 1. First experiment comprised of applying person- alized delivery charge which was diﬀerent for all customers depending on their respective live cart. The dual...

work page

[9] [9]

To solve this, a novel ap- proach is proposed in this paper to detect returns and imple- ment certain actions items

CONCLUSION E-tailers experience return-rate problem leading to increased costs and lower proﬁt margins. To solve this, a novel ap- proach is proposed in this paper to detect returns and imple- ment certain actions items. Depending on the current cart conﬁguration of a user, a hybrid dual-model is built using a deep neural network to detect returnable cart...

work page

[10] [10]

Product re- turns on the internet: A case of mixed signals?

C. Boniﬁeld, C. Cole, and R. L. Schultz, “Product re- turns on the internet: A case of mixed signals?” Jour- nal of Business Research , vol. 63, no. 9-10, pp. 1058– 1065, 2010

work page 2010

[11] [11]

Equilibrium returns policies in the presence of supplier competition,

S. Bandyopadhyay and A. Paul, “Equilibrium returns policies in the presence of supplier competition,” Mar- keting Science, vol. 29, pp. 846–857, 09 2010

work page 2010

[12] [12]

Fraudulent consumer returns: exploiting retailers’ return policies,

L. C. Harris, “Fraudulent consumer returns: exploiting retailers’ return policies,” European Journal of Market- ing, vol. 44, no. 6, pp. 730–747, 2010. [Online]. Avail- able: https://doi.org/10.1108/03090561011032694

work page doi:10.1108/03090561011032694 2010

[13] [13]

The impact of customer returns on pricing and order decisions,

J. Chen and P. C. Bell, “The impact of customer returns on pricing and order decisions,” European Journal of Operational Research, vol. 195, pp. 280–295, 2009

work page 2009

[14] [14]

Comparative analysis of the carbon footprints of conventional and online retailing: A last mile perspective,

J. B. Edwards, A. C. McKinnon, and S. L. Cullinane, “Comparative analysis of the carbon footprints of conventional and online retailing: A last mile perspective,” International Journal of Physical Distribution & Logistics Management , vol. 40, no. 1/2, pp. 103–123, 2010. [Online]. Available: https://doi.org/10.1108/09600031011018055

work page doi:10.1108/09600031011018055 2010

[15] [15]

Predictive model selection for forecasting product returns,

J. Ma and H. M. Kim, “Predictive model selection for forecasting product returns,” Journal of Mechanical Design, vol. 138, no. 5, p. 054501, 2016

work page 2016

[16] [16]

Toktay, Forecasting product returns

B. Toktay, Forecasting product returns. INSEAD, 2001

work page 2001

[17] [17]

Modeling and fore- casting product returns: An industry case study,

A. Canda, X. Yuan, and F. Wang, “Modeling and fore- casting product returns: An industry case study,” in 2015 IEEE International Conference on Industrial En- gineering and Engineering Management (IEEM) , Dec 2015, pp. 871–875

work page 2015

[18] [18]

Bpr: Bayesian personalized rank- ing from implicit feedback,

S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme, “Bpr: Bayesian personalized rank- ing from implicit feedback,” in Proceedings of the twenty-ﬁfth conference on uncertainty in artiﬁcial in- telligence. AUAI Press, 2009, pp. 452–461

work page 2009

[19] [19]

Efficient Estimation of Word Representations in Vector Space

T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Eﬃcient estimation of word representations in vector space,” CoRR, vol. abs/1301.3781, 2013. [Online]. Available: http://arxiv.org/abs/1301.3781

work page internal anchor Pith review Pith/arXiv arXiv 2013

[20] [20]

Predicting prod- uct returns in e-commerce: The contribution of maha- lanobis feature extraction,

P. Urbanke, J. Kranz, and L. Kolbe, “Predicting prod- uct returns in e-commerce: The contribution of maha- lanobis feature extraction,” in ICIS, 2015

work page 2015

[21] [21]

Probabilistic ma- trix factorization,

A. Mnih and R. R. Salakhutdinov, “Probabilistic ma- trix factorization,” in Advances in neural information processing systems, 2008, pp. 1257–1264

work page 2008

[22] [22]

Decoding fashion contexts using word embeddings

S. Arora and D. Warrier, “Decoding fashion contexts using word embeddings.”

work page

[23] [23]

Non-negative matrix factorization with sparseness constraints,

P. O. Hoyer, “Non-negative matrix factorization with sparseness constraints,” Journal of machine learning research, vol. 5, no. Nov, pp. 1457–1469, 2004

work page 2004

[24] [24]

Distributed representations of words and phrases and their compositionality,

T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean, “Distributed representations of words and phrases and their compositionality,” in NIPS, 2013

work page 2013