HUMAN ACTIVITY RECOGNITION FOR FRIEND RECOMMENDATIONS

Abstract

In this thesis, I have used human activity recognition to find out the performed activity of a user and thereby predicting the lifestyle of the user based on his activities which is hence used to construct a friend recommendation model. With the advent of technology and social network platforms, friend suggestions are mostly based on the social graphs. Greater the number of mutual friends you have with a person, higher the chances of getting the friend recommendation of that person. Such an approach does not reflect end user’s preference for friend selection. Thus, we have presented an approach for friend suggestions which is based on the similarity of the lifestyle of different users. With the variety of sensors which are embedded in mobile devices, it has become convenient to analyze human activities. To implement an intelligent friend recommendation system, we made use of Convolution Neural Network (popularly known as CNN) for activity recognition by extracting the local feature and superimposing the statistical features, though making sure that information related to global features is preserved. Subsequent to discovering activities of user, Latent Dirichlet Allocation (LDA) [1] algorithms used to identify the life style. We additionally propose a similar metric algorithm in order to compute the similarity of life styles among users by a friend-similarity graph. Whenever a user requests for friends, then my algorithm will return people’s list sorted by recommendation score, from which user can choose whom to send request. The recommendation friends can be user’s immediate friends or they can be friends of friends based on their scores. Open dataset has been taken from http://archive.ics.uci.edu/ml/datasets/SmartphoneBased+Recognition+of+Human+Activities+and+Postural+Transitions# which contains user data for 30 person having 10402 data entries of sensor data. 21 friends are taken as a test case and a pictorial graph for user similarity is constructed. In the graph, white color box corresponds to a high similarity score and the darker the color, lower is the similarity score. So, this pictorial representation gives a clear picture for the closeness in users. The one thing which has come out as a very strong argument and a contributing parameter in our model is the statistical parameters. Pictorial representation of the data features gave a pretty good idea about the different statistical features that can be used in order to classify among different classes more efficiently.