Telephony technologies (mobile, VoIP, and fixed) have potentially improved the way we communicate in our daily life and have been widely adopted for business and personal communications. At the same time, scammers, criminals, and fraudsters have also find the telephony network an attractive and affordable medium to target end-users with the advertisement, marketing of legal and illegal products, and bombard them with the huge volume of unwanted calls. These calls would not only trick call recipients into disclosing their private information such as credit card numbers, PIN code which can be used for financial fraud but also causes a lot of displeasure because of continuous ringing. The fraudsters, political campaigners can also use telephony systems to spread malicious information (hate political or religious messages) in real-time through audio or text messages, which have serious political and social consequences if malicious callers are not mitigated in a quick time. In this context, the identification of malicious callers would not only minimize telephony fraud but would also bring peace to the lives of individuals. One way to classifies users as a spammer or legitimate is to get feedback from the call recipients about their recent interactions with the caller, but these systems not only bring inconvenience to callees but also require changes in the system design. The call detail records extensively log the activities of users and can be used to categorize them as the spammer and non-spammer. In this paper, we utilize the information from the call detailed records and proposed a spam detection framework for the telephone network that identifies malicious callers by utilizing the social behavioral features of users within the network. To this extent, we first model the behavior of the users as the directed social graph and then analyze different features of the social graph i.e. the Relationship Network and Call patterns of users towards their peers. We then used these features along with the decision tree to classify callers into three classes i.e. human, spammer and call center. We analyzed the call record data-set consisting of more than 2 million users. We have conducted a detailed evaluation of our framework which demonstrates its effectiveness by achieving acceptable detection accuracy and extremely low false-positive rate. The performance results show that the spammers and call center numbers not only have a large number of non-repetitive calls but also have a large number of short duration calls. Similarly, on the other hand, the legitimate callers have a good number of repetitive calls and most of them interacted for a relatively long duration.