With the development of Internet of Things (IoT), Mobile CrowdSensing (MCS) platform will release projects consisting of heterogeneous tasks, requiring participants with different skills to collaborate to develop such systems. In this paper, a heterogeneous multi-project multi-task allocation model is proposed based on the group collaboration mode to cater for this problem state. Our method would distinguish the roles of members within the group, and incorporate the inherent attributes of participants like skill level and social competence. With the constraints of skill matching and completion time, one needs to simultaneously maximize the sensing quality and to minimize the platform cost by finding an optimal task-participant allocation schedule. To solve the established model, a multi-objective fireworks algorithm with dual-feedback ensemble learning framework is proposed. The weight of the weak optimizer would be adjusted automatically by the evolutionary significance, for which the individual generation method more suitable for the current state would be chosen. The individual evaluation mechanism is updated by the objective exploration degree, so that the evolutionary direction can be adaptively adjusted. To experimentally evaluate the proposed approach, it would be compared with five representative algorithms on 12 real-world instances. Experimental results show that our algorithm can assist platform managers in making better decisions.