Edited in response to very constructive criticism from @DavidHammen and @CallMeTom. I agree with them, but I didn't say those things in my initial answer, and I should have.
If the only source of data you have is a TLE, then you are starting from a low-quality initial state, which you should expect to be wrong by several kilometers.
All a high-quality propagator can do from there is tell you where something that actually was where the TLE claimed your satellite of interest was will go. You don't know where your satellite actually was, so nothing can tell you where it will actually go. The other propagator will do a better job than SGP4 of estimating where an imaginary object at the TLE's initial state will end up, but that doesn't mean the imaginary object will evolve into a state closer to the state of the real satellite. The error built into the very approximate nature of a TLE is not recoverable without a better source of data. If you have something else, then use it instead, because TLEs are terrible.
However, with all that in mind, if all you have is a TLE, and you are interested in what happens to a notional satellite that really was where the TLE claimed something was, then yes, that is the best you're going to be able to do. TLEs exist for the purpose of being easily distributed. SGP4 exists for the purpose of turning TLEs into something more useful, like Cartesian position and velocity. Once you have those as the initial state at your desired times, handing them to a different propagator with better models for gravity, drag, solar pressure, and everything else is the best way to proceed, as long as you remember that trusting the TLE to begin with may well be your biggest source of error.
I do this routinely at work, but only in design studies to model sensor performance on a moderately realistic simulated satellite environment. In that case, propagating years into the future is not my goal. I just use a bunch of TLEs to give me a lifelike distribution of initial states, because being off by tens or even hundreds of kilometers in-track at the starting point doesn't matter to the simulation results; all that matters is how the states evolve from their imaginary starting conditions, for which I would never use SGP4. If I am doing anything with a currently operational satellite, I always have something much better than a TLE to start with.
If you have not just another propagator, but also an orbit determination tool, then you can play with using the SGP4 output to simulate observations, and determine your own orbit from that. I stress "play", because the only question this answers is "I wonder what would happen if..." You're not going to make a TLE-derived orbit better without real data; but if simulation is all you're after, then it can be interesting to explore this option. Real data is available from several commercial vendors, but it's not cheap -- except perhaps in comparison to the painfully expensive commercial orbit determination tools.
The process of converting out of TLE & TEME seems long and convoluted if you plan to type it all in yourself, but you don't have to. You can download SGP4 from https://www.space-track.org/documentation#/sgp4 and use it to process a bunch of TLEs into long lists of position and velocity; osculating Keplerian elements; latitude, longitude, and altitude; or a variety of other formats. Then you can do whatever you want with them.
@uhoh: Never take a TLE at face value! Its components are mean elements, and so is part of the definition of its coordinate system. At face value, they describe the motion of a fictional satellite with respect to a fictional equinox. However, everything is carefully arranged to combine and cancel in just the right way to get something reasonable out, but only if you use SGP4 to do it. In the words of Spacetrack Report #3,
The NORAD element sets are “mean” values obtained by removing periodic variations in a particular way. In order to obtain good predictions, these periodic variations must be reconstructed (by the prediction model) in exactly the same way they were removed by NORAD.
The point cloud approach might produce some interesting results, but I think the main flaw is we're missing some important data that space-track does not provide, namely the covariance. If we had that, we could replace each point in time with not a single state vector, but rather a large ensemble normally distributed around that point, and see how a particular confidence volume grows over time.