2019 FGCS SOS: A multimedia recommender System for Online Social networks

[ZahraTaherikhonakdar]

Main Problem: The increasing use of Social Networks results in producing a large number of data showing Big Data features. In this context, in recent years Recommender Systems have been introduced to help users to find their intentions within this ocean of information. This paper proposed a state-of-the-art recommending system for big data applications able to generate multi-media data like texts, images, videos, and audio in one or more social media networks.

Input: User's rating, comments, profile, log history, and items metadata like tags, title author and so on Output: Recomended Items (video, photo, text)

Previous Works and their Gaps: The authors recommend items to users based on a content-based approach. This approach is based on a user's rating for similar items in the past. A major drawback of this approach is that the system can only recommend items similar to those already rated by the user. Moreover, we have a difficulty with describing effective similarity criteria between two items. A collaborative filtering strategy is proposed by another researcher. In this approach rating made by other users is considered and the similarity between user's profile is computed. One of the challenges in this strategy is matching users' profile who have a similar profile. The important problem is that the system could not recommend proper items due to the initial lack of ratings.

Proposed Method They used a subset of the Yahoo Flickr Creative Commons 100 Million Data (YFCC100M) multimedia collection (containing about 500,000 images), provided by Yahoo in 2014. In particular, they exploited users’ social interactions (friendships, tags, publishing, comments, favorites) with the related multimedia data. Images and user actions necessary to build the user-content graph were retrieved using Flickr API7 and are related to several domains and topics (e.g., animal, landscape, nature, etc.).

1- First stage is pre-filtering stage which determines a set of proper candidate items. The candidate has to match user preferences and actual needs. 2-Ranking recommended items in the previous step is the second stage. 3-In the next step the list of most suitable items would be determined when a user selects one or more interesting items among candidate items. 4-Finally the presentation of the final recommended items as good candidates for the recommendation.

Results: The below figure shows the trend of RMSE for the proposed system as well as for the UPCC and IPCC algorithms (as baseline algorithm), as the sparsity of the rating matrix increases. The proposed approach outperforms UPPC and IPPC ones for each value of items’ sparsity – and especially for higher values – showing as social information can improve recommendations.

Code

Gaps of this work It just considered the available data without considering the changes that may occur in the semantic of words.

[ZahraTaherikhonakdar]

I will explain the results

[hosseinfani]

@ZahraTaherikhonakdar Also, choose an assignee, project, and a label

[hosseinfani]

@ZahraTaherikhonakdar Not sure I understood the gaps of the work. What do you mean by the "semantic of data"?

[ZahraTaherikhonakdar]

@ZahraTaherikhonakdar Not sure I understood the gaps of the work. What do you mean by the "semantic of data"?

My bad!! I mean semantic of words. I'll correct it

[hosseinfani]

Again, what do you mean by "words" in this work?