User Behavior Modeling and Applications in Dynamic Marketing Systems
用戶行為建模及其在動態營銷系統中的應用
Student thesis: Doctoral Thesis
Author(s)
Related Research Unit(s)
Detail(s)
Awarding Institution | |
---|---|
Supervisors/Advisors |
|
Award date | 26 Aug 2024 |
Link(s)
Permanent Link | https://scholars.cityu.edu.hk/en/theses/theses(967ea54c-87bb-4874-974a-1ea1fc1af71f).html |
---|---|
Other link(s) | Links |
Abstract
Modeling and understanding user behaviors play a key role in successfully promoting marketing results of dynamic systems, including social media influencer marketing, heating, ventilation, and air conditioning (HVAC) systems, and personalized recommender systems. For these three domains, (1) the participation of human decisions has a high impact on long-term marketing results, (2) understanding latent patterns during the decision process helps facilitate the whole system to run more rewardingly, and (3) machine learning techniques are crucial in determining the intrinsic interaction mechanism between users and other components. Therefore, in this dissertation, we dive into three concrete user-involved dynamic systems and aim to provide instructions on how to effectively improve marketing performance on the basis of recognizing sequential user behaviors.
First, we develop some content demonstration strategies in the social media influencer marketing area. To achieve this, an empirical study is investigated on the characteristic and semantic factors that are associated with a high level of user engagement. Specifically, we first collect a cross-platform opinion leaders' blogs dataset, which includes 344,643 blogs published by 93 opinion leaders who have accounts on both Weibo and WeChat, the two most popular social media platforms in China. Then, to find out the associations between user engagement and practically accessible attributes of opinion leaders as well as their blogs, we adopt sentiment analysis models to process the blog data, devise a normalization technique to alleviate the issue caused by the heterogeneity of fall-out intervals of blogs, and design a gradient-based saliency method to highlight the important ingredients in blog sentences. Utilizing these computational tools, we reveal that user engagement on the two platforms agrees on some common factors, such as the number of blog tokens. Meanwhile, the two platforms differ in some aspects, such as the semantic patterns that can improve the level of user engagement.
The second focus of this dissertation is HVAC marketing systems, where we propose to design automated control strategies in order to enhance users' thermal comfort and save energy at the same time. In particular, we noticed that the training process of existing algorithmic HVAC controllers could be improved by leveraging the huge amount of historical data. To this end, we comprehensively evaluate the strengths and limitations of offline algorithmic HVAC controllers from two perspectives: algorithms and dataset characteristics. The experiments on two building environments reveal some critical observations. On the one hand, the effectiveness of sequential user behavior modeling is confirmed in reducing violations of indoor temperatures and taking rewarding control actions. On the other hand, the results indicate that datasets of a certain suboptimality level and relatively small scale can be utilized to effectively train a well-performed dynamic HVAC controller. In particular, such controllers can reduce at most 28.5% violation ratios of indoor temperatures and achieve at most 12.1% power savings compared to the baseline controllers.
Third, the dissertation explores a more efficient manner of realizing real-time personalization in the sequential recommendation scenario, where user preference is dynamically inferred from sequential historical behaviors. To optimize the marketing goal in terms of long-term user involvement, offline RL-based recommendation strategies are commonly adopted as they provide an additional advantage in avoiding online explorations that may harm users' experiences. However, previous studies mainly focus on discrete action and policy spaces, which might have difficulties in handling dramatically growing items efficiently. To mitigate this issue, we design a new algorithmic framework applicable to continuous control manners, where low-dimensional unified actions are composed of normalized user/item representations. Accordingly, the policy evaluation and policy improvement procedures are carefully designed with strategic exploration and directional control, respectively. Finally, extensive experiments are conducted to demonstrate the effectiveness of our framework in gaining more rewarding marketing results, compared to the discrete baselines.
The studies in this dissertation will contribute to the understanding of sequential user behavior modeling, from the perspective of strategic content demonstration, data-driven automated controlling, and efficient real-time personalization. In the future, other than the potential improvement in each single marketing system, we also hope to find a way to build close connections between different systems. A promising direction is to combine the design of more intelligent decision-making systems with the pattern knowledge extracted from content understanding.
First, we develop some content demonstration strategies in the social media influencer marketing area. To achieve this, an empirical study is investigated on the characteristic and semantic factors that are associated with a high level of user engagement. Specifically, we first collect a cross-platform opinion leaders' blogs dataset, which includes 344,643 blogs published by 93 opinion leaders who have accounts on both Weibo and WeChat, the two most popular social media platforms in China. Then, to find out the associations between user engagement and practically accessible attributes of opinion leaders as well as their blogs, we adopt sentiment analysis models to process the blog data, devise a normalization technique to alleviate the issue caused by the heterogeneity of fall-out intervals of blogs, and design a gradient-based saliency method to highlight the important ingredients in blog sentences. Utilizing these computational tools, we reveal that user engagement on the two platforms agrees on some common factors, such as the number of blog tokens. Meanwhile, the two platforms differ in some aspects, such as the semantic patterns that can improve the level of user engagement.
The second focus of this dissertation is HVAC marketing systems, where we propose to design automated control strategies in order to enhance users' thermal comfort and save energy at the same time. In particular, we noticed that the training process of existing algorithmic HVAC controllers could be improved by leveraging the huge amount of historical data. To this end, we comprehensively evaluate the strengths and limitations of offline algorithmic HVAC controllers from two perspectives: algorithms and dataset characteristics. The experiments on two building environments reveal some critical observations. On the one hand, the effectiveness of sequential user behavior modeling is confirmed in reducing violations of indoor temperatures and taking rewarding control actions. On the other hand, the results indicate that datasets of a certain suboptimality level and relatively small scale can be utilized to effectively train a well-performed dynamic HVAC controller. In particular, such controllers can reduce at most 28.5% violation ratios of indoor temperatures and achieve at most 12.1% power savings compared to the baseline controllers.
Third, the dissertation explores a more efficient manner of realizing real-time personalization in the sequential recommendation scenario, where user preference is dynamically inferred from sequential historical behaviors. To optimize the marketing goal in terms of long-term user involvement, offline RL-based recommendation strategies are commonly adopted as they provide an additional advantage in avoiding online explorations that may harm users' experiences. However, previous studies mainly focus on discrete action and policy spaces, which might have difficulties in handling dramatically growing items efficiently. To mitigate this issue, we design a new algorithmic framework applicable to continuous control manners, where low-dimensional unified actions are composed of normalized user/item representations. Accordingly, the policy evaluation and policy improvement procedures are carefully designed with strategic exploration and directional control, respectively. Finally, extensive experiments are conducted to demonstrate the effectiveness of our framework in gaining more rewarding marketing results, compared to the discrete baselines.
The studies in this dissertation will contribute to the understanding of sequential user behavior modeling, from the perspective of strategic content demonstration, data-driven automated controlling, and efficient real-time personalization. In the future, other than the potential improvement in each single marketing system, we also hope to find a way to build close connections between different systems. A promising direction is to combine the design of more intelligent decision-making systems with the pattern knowledge extracted from content understanding.