A link stream is a set of links {(t, u, v)} in which a triplet (t, u, v) models the interaction between two nodes u and v at time t. In many situations, data result from the measurement of interactions between several million of nodes over time and can thus be studied through the link stream's formalism. This is the case, for instance, of phone calls, email exchanges, money transfers, contacts between individuals, IP traffic, online shopping, and many more. The goal of this thesis is the detection of sets of abnormal links in a link stream. In a first part, we design a method that constructs different contexts, a context being a set of characteristics describing the circumstances of an anomaly. These contexts allow us to find unexpected behaviors that are relevant, according to several dimensions and perspectives. In a second part, we design a method to detect anomalies in heterogeneous distributions whose behavior is constant over time, by comparing a sequence of similar heterogeneous distributions. We apply our methodological tools to temporal interactions coming from retweets of Twitter and IP traffic of MAWI group.