Hashtags Prediction for Image Post in Social Networks

Abstract

Currently social media plays an important role in day to day life of people. Young generation as well as the middle-aged generation tend to share their experiences via social media. This could be a place the visit, could be a food they try out, could be a new dress or could be anything they wish to share with their network. Considering the engagement with social networks, posting images is a good way to share their experiences to a wider audience. Because of that the image posting is very popular in major social networks like Facebook, Instagram, Twitter, Tumbler and etc. Users tend to post some hashtags along with the images they are posting. This can be a small description to an image or a sentiment related with it. This study is supposed to identify the images post in social networks and predict hashtags for the identified images while considering the users age, gender and geographic location. A CNNLSTM model is proposed to accomplish the above purpose. CNN model is trained with the public Instagram posts collected and CNN model is able to classify images. For this study we have consider only four image classes which are belonged to two geographic locations. Relevant preprocessed hashtags were fed to the LSTM model along with the output of CNN model. Hashtags were tagged with the users age category, gender and geo location. Output of the LSTM model is the predicted hashtags. The model was trained 5,10,15,20, 25 and 30 epochs to identify the accurate prediction. The model shows the best accuracy after 20th epoch. This study considered the single object images. As a future work we can consider the images contains more than one object. Also, in this study we consider the age ranges of users due to limited availability of data. So as a next phase we can predict more user specific hashtags. Also, in this study based on the availability of data we consider the base location as New York and London, as a future work we can improve the granularity of the location. Other than that, in this study we only consider the hashtags with English language, as a future work we can consider the prediction of hashtags with native languages.