Over the last decade, popularity and fascination for social networks have exponentially increased. This rapid growth has triggered cybercriminals to utilize social networks for their malicious activities like social network spam. This risk and danger in social networking sites urge the need for an accurate and efficient spam detection model. Traditionally, supervised and unsupervised classification algorithms are used to identify social network spam. But spammers often change their behavior to evade spam filtering techniques. This results in huge data volatility. Traditional techniques are, therefore, sometimes ineffective in filtering spam in the social network. Hence, this work proposes to use a feed-forward neural network that can use the hidden relationship in this complex data for spam detection model. To improve the model’s accuracy and to speed up the training process, important hyperparameters such as learning rate, momentum term, architecture of neural network, activation function, training algorithm, initial weights ranges, and initial tuning of weights are necessary. As there is no general and predefined method available for this process, a reinforcement learning and k-Norm factor-based shuffled frog leaping algorithm to find the optimum set of neural network parameters is proposed in this paper. In the first stage, learning rate and momentum parameters in the continuous variable range are tuned using reinforcement learning. In the second stage, the best possible combinations of remaining parameter values are chosen using the proposed modified shuffled frog leaping algorithm that uses k-Norm to improve the exploitation. Experiments were carried out for the Tip spam dataset and Twitter dataset on a feed-forward neural network with tuned parameters. The results prove that the proposed algorithm achieves higher accuracy and lower false positive rate when compared to other existing techniques.