Here, any recommendation depends on what you’re trying to do.

If the goal is categorization and related downstream analyses at the levels of repeat families and classes, then fractional assignment (or other similar simple, conservative strategies; for some examples, see the htseq-count documentation) can be used.

An EM-based or other probabilistic/model-based approach is to try to infer the origin of ambiguous (multi-mapping) alignments through iterative estimation of expression levels (signal), then use that information to reallocate them. These methods depend on the model and its initialization, as well as the quality and quantity of information in the sequenced read data (e.g., read length, sequencing depth, the presence/absence of library-prep issues, etc.). As a result, they can concentrate multi-mapping alignments in ways that reflect model assumptions or data quality (or lack thereof), rather than “real” signal.

So, for family- or class-level quantification of repeats, fractional assignment (or other similar simple, conservative strategies) is acceptable and commonly used in the literature. For locus-level inference—e.g., in the context of repetitive signal amid otherwise non-repetitive sequence (e.g., TE insertion signal among unique sequence)—EM-style (and related probabilistic/model-based) methods are reasonable as long as the model assumptions are met and the data contain sufficient information (and not “too much” noise) to support them. In that more local/context-specific case, EM-style methods are arguably better suited than simple, conservative quantitation methods, which tend to “blur” differences across loci.

BTW, when I was first learning about all this, this was a really important review for me: Lanciano and Cristofari, Nat Rev Genet 2020. In the process of trying to find a link to the preceding review, I also stumbled on this one, which looks helpful: Deschamps-Francoeur, Simoneau, and Scott, Comput Struct Biotechnol J 2020.