Haskell, 42 bytes

f x=and[elem(a,d)x|(a,b)<-x,(c,d)<-x,b==c]

Usage example: f [(1,2), (2,4), (6,5), (1,4)]-> True.

(Outer)loop over all pairs (a,b) and (inner)loop over the same pairs, now called (c,d) and every time when b==c check if (a,d)is also an existent pair. Combine the results with logical and.

JavaScript (ES6), 69 67 bytes

a=>(g=f=>a.every(f))(([b,c])=>g(([d,e])=>c-d|!g(([d,c])=>b-d|c-e)))

Saved 2 bytes thanks to an idea by @Cyoce. There were four previous 69-byte formulations:

a=>a.every(([b,c])=>a.every(([d,e])=>c-d|a.some(([d,c])=>b==d&c==e)))
a=>!a.some(([b,c])=>!a.some(([d,e])=>c==d&a.every(([d,c])=>b-d|c-e)))
a=>a.every(([b,c])=>a.every(([d,e])=>c-d|!a.every(([d,c])=>b-d|c-e)))
(a,g=f=>a.every(f))=>g(([b,c])=>g(([d,e])=>c-d|!g(([d,c])=>b-d|c-e)))

MATL, 27 25 bytes

7#u2e!l6MX>thl4$XQttY*g<~

Input format is a matrix (using ; as row separator) where each pair of the relation is a column. For example, test cases

[(1, 2), (2, 4), (6, 5), (1, 4)]
[(7, 8), (9, 10), (15, -5)]
[(5, 9), (9, 54), (0, 0)]

are respectively input as

[1 2 6 1; 2 4 5 4]
[7 9 15; 8 10 -5]
[5 9 0; 9 54 0]

Truthy output is a matrix formed by ones. Falsy is a matrix that contains at least one zero.

Try it online!

Explanation

The code first reduces the input integers to unique, 1-based integer values. From those values it generates the adjacency matrix; matrix-multiplies it by itself; and converts nonzero values in the result matrix to ones. Finally, it checks that no entry in the latter matrix exceeds that in the adjacency matrix.

CJam (22 bytes)

{__Wf%m*{z~~=*}%\-e_!}

Online test suite. This is an anonymous block (function) which takes the elements as a two-level array, but the test suite does string manipulation to put the input into a suitable format first.

Dissection

{         e# Begin a block
  _       e#   Duplicate the argument
  _Wf%    e#   Duplicate again and reverse each pair in this copy
  m*      e#   Cartesian product
  {       e#   Map over arrays of the form [[a b][d c]] where [a b] and [c d]
          e#   are in the relation
    z~~=* e#     b==c ? [a d] : []
  }%
  \-      e#   Remove those transitive pairs which were in the original relation
  e_!     e#   Test that we're only left with empty arrays
}

Pyth, 14 bytes

!-eMfqFhTCM*_M

Test suite

Input format is expected to be [[0, 0], [0, 1], ... ]

!-eMfqFhTCM*_M
!-eMfqFhTCM*_MQQQ    Variable introduction
            _MQ      Reverse all of the pairs
           *   Q     Cartesian product with all of the pairs
         CM          Transpose. We now have [[A2, B1], [A1, B2]] for each pair
                     [A1, A2], [B1, B2] in the input.
    f                Filter on
       hT            The first element (the middle two values)
     qF              Being equal
  eM                 Take the end of all remaining elements (other two values)
 -              Q    Remove the pairs that are in the input
!                    Negate. True if no transitive pairs were not in the input

 Prolog, 66 bytes

t(L):-not((member((A,B),L),member((B,C),L),not(member((A,C),L)))).

The relation is not transitive if we can find (A,B) and (B,C) such that (A,C) doesn't hold.

Axiom 103 bytes

c(x)==(for i in x repeat for j in x repeat if i.2=j.1 and ~member?([i.1, j.2],x)then return false;true)

ungolfed:

c(x)==
  for i in x repeat
    for j in x repeat
       if i.2=j.1 and ~member?([i.1, j.2],x) then return false
  true

                                                                   Type: Void

the exercises

(2) -> c([[1,2],[2,4],[6,5],[1,4]])
   Compiling function c with type List List PositiveInteger -> Boolean
   (2)  true
                                                                Type: Boolean
(3) -> c([[7,8],[9,10],[15,-5]])
   Compiling function c with type List List Integer -> Boolean
   (3)  true
                                                            Type: Boolean
(4) -> c([[5,9],[9,54],[0,0]])
   Compiling function c with type List List NonNegativeInteger ->
      Boolean
   (4)  false

Pyth - 22 21 bytes

.Am},hdedQfqFtPTsM^Q2

Test Suite.