In this article, we construct two Halpern-type relaxed algorithms with alternated and multi-step inertial extrapolation steps for split feasibility problems in infinite-dimensional Hilbert spaces. The first is the most general inertial method that employs three inertial steps in a single algorithm, one of which is an alternated inertial step, while the others are multi-step inertial steps, representing the recent improvements over the classical inertial step. Besides the inertial steps, the second algorithm uses a three-term conjugate gradient-like direction, which accelerates the sequence of iterates toward a solution of the problem. In proving the convergence of the second algorithm, we dispense with some of the restrictive assumptions in some conjugate gradient-like methods. Both algorithms employ a self-adaptive and monotonic step-length criterion that does not require knowledge of the norm of the underlying operator or the use of any line search procedure. Moreover, we formulate and prove some strong convergence theorems for each of the algorithms based on the convergence theorem of an alternated inertial Halpern-type relaxed algorithm with perturbations in real Hilbert spaces. Further, we analyse their applications to classification problems for some real-world datasets based on the extreme learning machine (ELM) with the $\ell_{1}$-regularization approach (that is, the Lasso model) and the $\ell_{1}-\ell_{2}$ hybrid regularization approach. Furthermore, we investigate their performance in solving a constrained minimization problem in infinite-dimensional Hilbert spaces. Finally, the numerical results of all experiments show that our proposed methods are robust, computationally efficient and achieve better generalization performance and stability than some existing algorithms in the literature.