Research map / Reinforcement learning / Visual and representation RL

The POWER of Ikigai: Optimizing Life Fulfillment with an Integrated User Simulator and Adaptive Hobby Recommender

AAAI 2025 · Reinforcement learning › Visual and representation RL · approach: RL with tree search for discovery · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes POWER, which pairs an MLP-based ikigai simulator (with a probabilistic user-acceptance module) with an A2C-trained hobby recommender, and validates it offline and in a 30-day user study with 40 participants.

Problem

Ikigai (sense of purpose/life worth) is dynamic but existing scales (ikigai-9, four-factor) require tedious questionnaires, making frequent tracking impractical; the paper seeks automatic ikigai estimation and hobby recommendations that raise it.

Approach

Train an MLP ikigai simulator to predict questionnaire responses from profiles, wrap it with a probabilistic acceptance/frequency module as an environment, and optimize a hobby recommender with advantage actor-critic using the predicted ikigai change as reward.

Key results

• Ikigai simulator achieves lower distance to ground truth than Random and Probabilistic baselines on ikigai-9 responses.
• In the simulator, POWER gives the highest average ikigai increase per hobby recommendation, though its acceptance rate is slightly lower than the supervised model and random.
• In the 30-day user study, Group III (POWER with daily encouraging messages) showed a significant ikigai-9 improvement (p=0.017) with 9.114% relative improvement in Group II and 4.117% in Group III, while the supervised group (Group I) showed no significant change.

Measured effects reported in the paper

Surprising findings

• POWER's acceptance rate was slightly lower than the supervised model and even random recommendation, because the reward of 0 for rejection leads it to favor ikigai gain over acceptance.

Limitations

• User study is small (40 participants, all ethnically Chinese) and simulator-based evaluation relies on predicted rather than measured ikigai.
• Self-reported questionnaire changes with no control group beyond the supervised comparison; relative improvements vary across scales and groups.
• Group I (supervised) and the four-factor scale results were inconsistent, limiting strength of conclusions.

Experimental setup

Datasets: In-house questionnaire of 514 respondents (profiles, hobbies, ikigai-9 and four-factor scores)

Benchmarks: In-house questionnaire dataset (514 responses) 30-day user study (40 participants)

Metrics: distance to ground truth (ikigai-9 item Likert) average ikigai change acceptance rate paired t-test p-values relative improvement

Baselines: Random Probabilistic Supervised hobby recommender

Builds on: A2C ikigai-9 scale four-factor ikigai scale

Closest papers by meaning

• LLM-Powered User Simulator for Recommender System (AAAI 2025 · similarity 0.79)
• Recommender Ecosystems: A Mechanism Design Perspective on Holistic Modeling and Optimization (AAAI 2024 · similarity 0.78)
• Enhancing Personalized Multi-Turn Dialogue with Curiosity Reward (NeurIPS 2025 · similarity 0.78)
• RoME: A Robust Mixed-Effects Bandit Algorithm for Optimizing Mobile Health Interventions (NeurIPS 2024 · similarity 0.78)
• Bites of Tomorrow: Personalized Recommendations for a Healthier and Greener Plate (AAAI 2025 · similarity 0.78)
• Effect Size Estimation for Duration Recommendation in Online Experiments: Leveraging Hierarchical Models and Objective Utility Approaches (AAAI 2024 · similarity 0.78)