How to filter data with the highest number

Using awk, as implied by your tags:

$ awk -F_ '
    {m[$1] = $2 > m[$1] ? $2 : m[$1]} 
    END {OFS = FS; for(x in m) print x, m[x]}
  ' file | sort
abc_4
def_3
ghi_2

If you have GNU awk > 4.0 you can do the sorting internally

$ awk -F_ '
    {m[$1] = $2 > m[$1] ? $2 : m[$1]} 
    END {
      OFS = FS; PROCINFO["sorted_in"] = "@ind_str_asc"; 
      for(x in m) print x, m[x]
    }
  ' file
abc_4
def_3
ghi_2

Alternatively, if you have Miller you can do a groupwise max using the stat1 verb:

$ mlr --nidx --fs _ stats1 -a max -f 2 -g 1 file
abc_4
def_3
ghi_2

or with GNU datamash

$ datamash -t_ groupby 1 max 2 < file
abc_4
def_3
ghi_2

or for your specific input, assuming that the prefix strings are fixed length, using standard utilities sort and uniq:

$ sort -t_ -k1,1 -k2,2nr file | uniq -w4
abc_4
def_3
ghi_2