Page 99 - My FlipBook
P. 99
Brochure 2020
Finally, we aim to enhance our recommendation system such that it is robust to biased data. It is inevitable
that relevant real-world data, collected from previous e-commerce purchase logs and advertisement
bidding logs, will contain noise and bias. Previous studies optimized the recommendation model based
on assumptions that data is not biased and used o ine assessment results to represent online results.
However, the collected data may deviate from a specific tendency. Training an XR recommendation
system based on such biased data is likely to cause two main problems: 1) the results of o ine evaluation
cannot truly reflect the online results, diminishing user satisfaction and experience in XR stores; and
2) the allocated display slots are biased toward existing advertisers or popular items, inhibiting the
comprehensiveness and soundness of the recommendation results. Therefore, it is important for our
XR recommendation system to balance training data bias. Accordingly, we will propose a method to
overcome bias in the data. More speci cally, we will design a weighting system to re ne the model so that
we can obtain more comprehensive information from biased data. Moreover, we will generate biased data
to conduct hypothesis veri cation. Our experimental environment will simulate the e ects of biased data
and non-biased data on model training. Then, we can develop a negative sampling method to simulate
the contents recommended by the existing models that users do not select, allowing us to improve the
items recommended by the model.
XR-enabled e-commerce vendors such as eBay, Myers, IKEA, Lowe, and Amazon will benefit from our
XR recommendation system because it will facilitate flexible and balanced item configurations to boost
sales via social in uences and interactions while not sacri cing individual user preferences. Moreover, for
local vendors (e.g., PChome, MOMO, Buy123, and Rakuten) that have not adopted XR technologies, their
promotions can be enhanced by utilizing correlations among items and understanding customers’dynamic
perceptions of items. Our SIoT deployment subsystem can bene t companies interested in SIoT platforms
and services, e.g., Amazon AWS IoT platform and MinSphere (jointly built by Alibaba and Siemens), as they
connect numerous SIoTs to provide ubiquitous and location-based services. Finally, our techniques for
unbiased and robust learning will prove valuable to all general applications of learning-based prediction
models, including but not limited to recommendation systems (such as Google, Yahoo, YouTube, and
Alibaba).
97
Finally, we aim to enhance our recommendation system such that it is robust to biased data. It is inevitable
that relevant real-world data, collected from previous e-commerce purchase logs and advertisement
bidding logs, will contain noise and bias. Previous studies optimized the recommendation model based
on assumptions that data is not biased and used o ine assessment results to represent online results.
However, the collected data may deviate from a specific tendency. Training an XR recommendation
system based on such biased data is likely to cause two main problems: 1) the results of o ine evaluation
cannot truly reflect the online results, diminishing user satisfaction and experience in XR stores; and
2) the allocated display slots are biased toward existing advertisers or popular items, inhibiting the
comprehensiveness and soundness of the recommendation results. Therefore, it is important for our
XR recommendation system to balance training data bias. Accordingly, we will propose a method to
overcome bias in the data. More speci cally, we will design a weighting system to re ne the model so that
we can obtain more comprehensive information from biased data. Moreover, we will generate biased data
to conduct hypothesis veri cation. Our experimental environment will simulate the e ects of biased data
and non-biased data on model training. Then, we can develop a negative sampling method to simulate
the contents recommended by the existing models that users do not select, allowing us to improve the
items recommended by the model.
XR-enabled e-commerce vendors such as eBay, Myers, IKEA, Lowe, and Amazon will benefit from our
XR recommendation system because it will facilitate flexible and balanced item configurations to boost
sales via social in uences and interactions while not sacri cing individual user preferences. Moreover, for
local vendors (e.g., PChome, MOMO, Buy123, and Rakuten) that have not adopted XR technologies, their
promotions can be enhanced by utilizing correlations among items and understanding customers’dynamic
perceptions of items. Our SIoT deployment subsystem can bene t companies interested in SIoT platforms
and services, e.g., Amazon AWS IoT platform and MinSphere (jointly built by Alibaba and Siemens), as they
connect numerous SIoTs to provide ubiquitous and location-based services. Finally, our techniques for
unbiased and robust learning will prove valuable to all general applications of learning-based prediction
models, including but not limited to recommendation systems (such as Google, Yahoo, YouTube, and
Alibaba).
97
