The assumption is that the wave function is a probability amplitude. In particular, it's a vector that is normalized. In Dirac's notation, this is the statement:
$$\langle \psi |\psi\rangle = 1.$$
This can be made more concrete with:
$$\begin{align}
\mathrm{ordinary\ vectors\ } &\sum_{i} \psi^\star_i \psi_i = 1, \\
\mathrm{wave\ functions\ } &\int \psi^\star(x) \psi(x) \operatorname{d}x = 1,\ \mathrm{or} \\
\mathrm{even\ wave\ functionals\ } & \int \left[\mathcal{D}\phi(x)\right] \Psi^\star[\phi(x)] \Psi[\phi(x)] = 1.
\end{align}$$
Dont' worry if that last one is cryptic - it's for when you're dealing with quantum field theory.
The important point is that the wave function is confined to exist in only a part of the vector space; like how unit vectors are confined to lie on the surface of a sphere. Transformations that respect this constraint are called unitary. Thus that constraint means that every allowed transformation of $|\psi\rangle$ is unitary. Rotations, spatial translations, reflections, etc, all must respect the requirement that the wave function remains normalized.
The rest follows from the requirement that the time translation operation is a continuous change in $|\psi\rangle$ and that quantum mechanics maps on to classical mechanics on average (see: the correspondence principle). That means that $\hat{H}$, the generator of time translations in quantum mechanics, has to correspond with the generator of time translations in classical mechanics, the Hamiltonian.
There is one exception I know of to the unitarity requirement. That is time reflections. Time refletion is anti-unitary. For details, see the Wikipedia article on $T$-symmetry.