Intelligent and effective incentivization strategies towards proactive recommendations

Recommendation systems have been increasingly deployed and studied in many domains. These systems can generate appropriate recommendations to users through passively learning their behavior patterns and preferences. However, such passive recommendations can only satisfy users' requirements and demands, but cannot realize system-level objectives. Therefore, it is crucial to propose proactive recommendations that can satisfy users' requirements and implement system goals in the meanwhile.

One straightforward solution is to provide incentives to users, since providing users with incentives has been verified as an effective method to promote users' specific behaviors, and has been widely adopted in many real-world applications. To effectively incentivize users, many research works have designed approaches that generate suitable incentives for each user in different scenarios. However, most existing approaches only focus on incentivizing users directly, which may require more budget to implement. Therefore, it is necessary to propose new approaches that can incentivize users effectively when the budget is insufficient.

In this thesis, I focus on how to realize proactive recommendations in different scenarios by incentivizing users' behaviors by spending a limited budget. First, I conduct a preliminary investigation on the feasibility of utilizing social influence to facilitate the effect of incentives in the social network, as intuitively we can directly incentivize influential users in the social network first, and then indirectly incentivize their neighbors via social influence. Second, I further study how to incentivize users in social networks under a limited budget, where the information about the network is completely unknown. Next, motivated by the fact that most existing incentive approaches only consider incentivizing users' one-off behaviors, a novel incentive mechanism aiming to incentivize users' long-term engagement is proiv posed. Last but not least, a deep reinforcement learning framework is proposed for incentivizing users' behaviors in unknown social networks, where only the topological structure of the network is given, and users‚' attributes are unknown.