Purely speculative and all conjecture. I know nothing of the algorithms involved and make the following assumptions about the meta-data and the algorithms.
1: Geo-location of the event/person
2: Time of the event/person
3: Compare location +1 correlation.
4: Compare time +1 correlation.
5: Location compares street.
6: Time within 1 day.
Given these very simplistic assumptions. We have two people. We'll call them Good Steve and Evil Steve. They have never met, never seen each other. One lives at Street A #15 and is homebound (GS), the other, ES works at Street A #12 and is plotting embezzelment. Abreviated to GS and ES for the purpose of the demonstration.
Day 1, City A, Street A #12: ES makes 4 calls, which get logged.
Day 1, City A, Street A #15: GS makes 2 calls, which gets logged.
Correlation between ES and GS: 6.
Already, the correlation between ES and GS is 6 after one day. Because they're on the same street, just a few street #s away from each other.
Suppose this goes on in the same way for a few months. Say 3. The correlation is 540 after three months. Now, say that the person that ES was calling has half that, assuming calling the same person. In the ensuing metadata analysis after the embezzlement is discovered, there is a link formed between ES and GS that is GREATER in this admittedly VERY simple model than that of ES and the person ES was conspiring against. Another example, say this sort of thing happened but ES called a bank, and GS called the same bank after or before ES. There becomes a tenuous link between the bank, ES and GS based on both location and time and even number called, a stat not directly recorded by this algorithm.
The actual reality should be far more complex, but I would imagine a meta-data analysis would rely on more rules with finer resolutions among other things... At least I hope so, so that the probability correlations of a connection between two people or a person and a group of people is more solid and worth investigation than the example I demonstrated as a worst case scenario.
In some cases metadata can be useful, but I do not think it is for any reasonable, serious leg for investigations to stand on. Certainly it is useful in an investigatory sense to draw lines between connected people and groups, but an investigation is necessarily an activity that takes place AFTER something has gone down that requires investigation. It is NOT for government to do an ongoing investigation into its citizens without due cause, oversight and a full accounting after the fact.
To do otherwise would be to invite the temptation to use the knowledge and insight such an ongoing investigation would make available to tamp down on things the government in power would really not prefer to allow. It's not hard to imagine a far religious right government doing so, but we must also be wary of the far left as well. To allow the left also to tamp down on private and personal freedoms would be as bad as the far right doing the same.
To make it more amenable to the lovers of LOTR out there...
It is analogous to Frodo offering the ring to Gandalf. Here's the quote:
I dare not take it. Not even to keep it safe. Understand, Frodo. I would use this ring (knowledge) from a desire to do good... But through me, it would wield a power too great and terrible to imagine.