Page 98 - My FlipBook
P. 98
Collaborative Projects所
內
合
作
計
畫
promoted items, which can affect users’ preferences through the economic concept of cross-elasticity of
demand. Secondly, perceptions of the relationships between items are usually personal. Thirdly, personal
perceptions of item relationships are dynamic, with newly-adopted items usually opening new experiences
to users and changing their perceptions regarding related items. Fourthly, changes in personal perceptions of
item relationships may a ect users’ preferences and their social in uence over friends.
Therefore, to address the important need to quantify social influence, we will leverage KG to capture
relationships between items, where KG represents facts, a meta-graph for a certain relationship describes
the semantics of connectivity between items, and a weighting re ects the importance of that meta-graph to
that relationship. By learning and tailoring the meta-graph weighting according to previously adopted items,
dynamic personal perceptions of item relationships can be captured through personal item networks, with the
vertices as items and the edges as item relationships (e.g., complementary and substitutable) along with their
relevance scores. Accordingly, we will formulate a new task, In uence Maximization with Dynamic Personal
Perception (DPP-IM), to establish a sequence of promotions for relevant items wherein the relationships
between items and the time-varied personal preferences and social influences due to previously adopted
items are both considered. DPP-IM will identify correlated items to be promoted at di erent times to suitable
users within a budget, thereby maximizing the bene ts of chosen items under dynamic user perceptions and
social in uence.
Another limitation of the XR recommendation system to be developed in the first year is that it assumes
a fully centralized system, i.e., the displayed item configuration can be dictated by the retailing platform.
However, in reality, retailers sell a plethora of goods from various vendors, who naturally compete for
limited retailing resources such as item display slots in XR malls and advertisement slots in e-commerce.
Therefore, it is important to carefully examine the strategic and competitive behaviors of suppliers in the XR
recommendation system. Accordingly, in the second year of the project, we will integrate Real-Time Bidding
(RTB) mechanisms to help determine the rights of item display in our XR recommendation system. It is
essential to devise a subsystem to precisely predict the winning price (generally the highest bidding price
among all competitors) for the supply-side platform (SSP, which is our XR recommendation platform) to help
infer auction results and con gure the item display, as well as for the demand-side platforms (DSPs, the agents
for advertisers) to guide their bidding behavior with limited budgets. Under common sealed-bid auction
mechanisms such as the second-price auction, a major challenge for DSPs is the lack of complete information
about the winning price, especially for lost bids in the past. That is because the winning price is visible
only to the winner. This problem is exacerbated when the rst-price auction is adopted, which has become
increasingly popular over the second price auction. In the rst-price auction, the winning price is invisible even
when a DSP wins the impression. The only information available is if the DSP wins the impression with its bid
price. Therefore, we aim to help DSPs to design an e ective prediction method for the winning price of various
distributions and under di erent auction mechanisms. To devise a new winning price model, we will evaluate
the performance and in uence of di erent loss functions based on a deep model structure. Moreover, we will
design a model layer or component and a loss function for learning a double-censored winning price from the
winning rate. We will also analyze the regularization term that makes the distribution smoother and in uences
the performance of the winning price model. Furthermore, we will assess how the winning price prediction
in uences the winning rate and revenue, and we will study how to use the winning price model to construct a
bidding strategy. Our model and bidding strategy will be evaluated by means of metrics such as revenue and
cost curve.
Third Year
We will shift our focus to intensive data collection and real-time event/location-based searching through
the Internet of Things (IoT) to facilitate real-time detection of user feedback and provide immediate user
support in the XR shopping environment. Detecting real-time user feedback is essential for customized
item con guration in XR stores. Moreover, XR customers unfamiliar with the environment or interfaces may
have technical problems and require help. It is crucial to exploit SIoT to provide immediate user support
and guidance across XR stores to avoid massive sta deployment. Therefore, it is vital to devise scalable and
e cient deployment of SIoT. Accordingly, we will study SIoT group construction and individual SIoT selection
tasks for XR stores in terms of SIoT communication, computation, and MEC network coverage to provide
heterogeneous feedback detection and location-based searching for user support in dynamic environments.
We will formulate a new optimization problem to optimize search results under dynamic environments for
location-based detection and recommendations.
96
內
合
作
計
畫
promoted items, which can affect users’ preferences through the economic concept of cross-elasticity of
demand. Secondly, perceptions of the relationships between items are usually personal. Thirdly, personal
perceptions of item relationships are dynamic, with newly-adopted items usually opening new experiences
to users and changing their perceptions regarding related items. Fourthly, changes in personal perceptions of
item relationships may a ect users’ preferences and their social in uence over friends.
Therefore, to address the important need to quantify social influence, we will leverage KG to capture
relationships between items, where KG represents facts, a meta-graph for a certain relationship describes
the semantics of connectivity between items, and a weighting re ects the importance of that meta-graph to
that relationship. By learning and tailoring the meta-graph weighting according to previously adopted items,
dynamic personal perceptions of item relationships can be captured through personal item networks, with the
vertices as items and the edges as item relationships (e.g., complementary and substitutable) along with their
relevance scores. Accordingly, we will formulate a new task, In uence Maximization with Dynamic Personal
Perception (DPP-IM), to establish a sequence of promotions for relevant items wherein the relationships
between items and the time-varied personal preferences and social influences due to previously adopted
items are both considered. DPP-IM will identify correlated items to be promoted at di erent times to suitable
users within a budget, thereby maximizing the bene ts of chosen items under dynamic user perceptions and
social in uence.
Another limitation of the XR recommendation system to be developed in the first year is that it assumes
a fully centralized system, i.e., the displayed item configuration can be dictated by the retailing platform.
However, in reality, retailers sell a plethora of goods from various vendors, who naturally compete for
limited retailing resources such as item display slots in XR malls and advertisement slots in e-commerce.
Therefore, it is important to carefully examine the strategic and competitive behaviors of suppliers in the XR
recommendation system. Accordingly, in the second year of the project, we will integrate Real-Time Bidding
(RTB) mechanisms to help determine the rights of item display in our XR recommendation system. It is
essential to devise a subsystem to precisely predict the winning price (generally the highest bidding price
among all competitors) for the supply-side platform (SSP, which is our XR recommendation platform) to help
infer auction results and con gure the item display, as well as for the demand-side platforms (DSPs, the agents
for advertisers) to guide their bidding behavior with limited budgets. Under common sealed-bid auction
mechanisms such as the second-price auction, a major challenge for DSPs is the lack of complete information
about the winning price, especially for lost bids in the past. That is because the winning price is visible
only to the winner. This problem is exacerbated when the rst-price auction is adopted, which has become
increasingly popular over the second price auction. In the rst-price auction, the winning price is invisible even
when a DSP wins the impression. The only information available is if the DSP wins the impression with its bid
price. Therefore, we aim to help DSPs to design an e ective prediction method for the winning price of various
distributions and under di erent auction mechanisms. To devise a new winning price model, we will evaluate
the performance and in uence of di erent loss functions based on a deep model structure. Moreover, we will
design a model layer or component and a loss function for learning a double-censored winning price from the
winning rate. We will also analyze the regularization term that makes the distribution smoother and in uences
the performance of the winning price model. Furthermore, we will assess how the winning price prediction
in uences the winning rate and revenue, and we will study how to use the winning price model to construct a
bidding strategy. Our model and bidding strategy will be evaluated by means of metrics such as revenue and
cost curve.
Third Year
We will shift our focus to intensive data collection and real-time event/location-based searching through
the Internet of Things (IoT) to facilitate real-time detection of user feedback and provide immediate user
support in the XR shopping environment. Detecting real-time user feedback is essential for customized
item con guration in XR stores. Moreover, XR customers unfamiliar with the environment or interfaces may
have technical problems and require help. It is crucial to exploit SIoT to provide immediate user support
and guidance across XR stores to avoid massive sta deployment. Therefore, it is vital to devise scalable and
e cient deployment of SIoT. Accordingly, we will study SIoT group construction and individual SIoT selection
tasks for XR stores in terms of SIoT communication, computation, and MEC network coverage to provide
heterogeneous feedback detection and location-based searching for user support in dynamic environments.
We will formulate a new optimization problem to optimize search results under dynamic environments for
location-based detection and recommendations.
96
