Master's Thesis Aamna Najmi

Imputation of missing Product Information using Deep Learning: A Use Case on the Amazon Product Catalogue

The last couple of years have seen massive advancement in deep learning across many tasks
such as computer vision (CV), natural language processing (NLP) and speech recognition.
This advancement can be observed by end users across various online platforms, one such
platform being the e-commerce domain where giants like Amazon are providing users
voice assistants, personalized recommendations and efficient product search options. The
advancement in the field of deep learning has been catalyzed by the availability of enormous
annotated datasets like the Wikipedia corpora in various languages [1] and the Imagenet
dataset [2]. However, there have not been appropriate amount of pre-processed datasets in
the e-commerce domain that are available for research. With the customer being the most
significant part on any e-commerce platform, there is a rising need of natural language and
computer vision enabled applications to improve user experience and increase organizational
benefits.
In order to overcome the shortage of publicly available datasets in the e-commerce domain
in both textual and visual form, we propose the use of domain adaptation to leverage
existing advancements in approaches like transfer learning and multi-task learning by using
state of the art techniques. Domain adaptation exploits task-independent commonalities and
overcomes the problem of dataset shortage [3], especially in the e-commerce domain. Through
this work, we have tried to improve product catalog quality by predicting missing product
information such as category, color, brand and target gender on the e-commerce platform
thus enabling efficient product search and improving user experience on the platform.
As part of the dataset generation phase, we have created three different e-commerce dataset
in languages including English, German and French for text based problems and English,
German and Italian for image based problems. The dataset has been used to predict missing
product information using deep learning approaches like transfer learning and multi-task
learning. We have also compared single task approaches for image classification tasks with
transfer learning and discussed benefits. In the natural language processing front, we have
compared single task learning with both transfer learning and multi-task learning. We
observed that for image classification tasks, single task is on equal footing with transfer
learning however the latter is trained and implemented in less than half the time invested
in training a deep learning model from scratch. For text classification the text corpora was
trained on a state-of-the-art deep learning model, the Transformer. In addition, we compared
two types of domain adaptation techniques, transfer learning and multi-task learning and
found that both approaches are on an equal footing in terms of accuracy. We show that
multi-task and transfer learning is advisable in situations where training data is sparse
through experiments in which a jointly trained transformer is able to outperform a single-task
trained transformer.

After the predictions, we conducted a survey to see if including the predicted features in
the product detail pages helps online customers in making buying decisions. Majority of the
respondents prefer the predicted features to be included on the product detail page. Hence,
suggesting that the predictions made through transfer learning and multi-task learning are
useful and applicable in the e-commerce domain to enhance user experience.
Through this thesis, we show how domain adaptation techniques outperform single task
learning for text based datasets in terms of accuracy and f1-score and converges way faster
for image classification tasks using the e-commerce datasets. These techniques are better
options when dealing with dataset shortage, imbalanced classes and in cases where we do
not want to train a model from scratch for a prolonged period of time.