Personalized Video Recommendation Based on Cross-Platform User Modeling

Zhengyu Deng, Jitao Sang and Changsheng Xu*

Summary

---

(1) Most personalized video recommendation methods are based on single-platform user modeling, which suffers from data sparsity and cold-start issues.
(2) We introduce cross-platform user modeling as a solution by smartly aggregating user information from different platforms.
(3) Unlike traditional recommendation methods where sufficient user information is assumed available in the target platform, this proposed method works well when there is little knowledge about users' interests in the target platform.

---

Framework

---

We use YouTube as the target platform where to perform the recommendation task, and Google+ as the auxiliary platform where user information is transferred. Two strategies are designed to strengthen the understanding of user interest in the target platform: one is profile enrichment and the other is collaborative relationship transfer.

---

Approach

---

--Assumption

We assume that users who have similar profiles in Google+ are very likely to have similar profiles in YouTube, so we transfer the collaborative relationship in Google+ to YouTube. Furthermore, we model the user similarity in Google+ from different perspectives and assign different weights to them.

--Multiple kernel learning

We adopt Multiple Kernel Learning (MKL) to obtain the weights of the different perspectives.

(1) Innovative: transfer enrichable user relations from social network to multimedia application platform;

---

Experiments

---

Examined strategies include:

1) Recommend only by YouTube Profile (S1);

2) Recommend by Profile Enrichment (S2);

3) Recommend by YouTube Profile with Collaborative Transfer (S3);

4) Recommend by Profile Enrichment with Collaborative Transfer (S4).

Experimental Results

The comparison of average F-score at different depths by different strategies is illustrated in the figure below . It shows that the strategy that enrich user profile with part of Google+ information has the best performance.

We assume that it's not feasible to directly aggregate all profiles of a user across platforms. We have designed a experiment to validate this hypothesis. We combine all the information in Google+ and YouTube for each user and obtain the compounded "Global Profile" for each of them. And then the video recommendation list is generated based on this profile. The performance is shown in the figure below. We can see that the performance by global profile is much worse than that by YouTube profile enriched with only certain information in Google+ (S2).

---

Publication

---

Personalized Celebrity Video Search Based on Cross-space Mining [pdf] [slides] [poster] [code] [data]

Zhengyu Deng, Jitao Sang and Changsheng Xu
In IEEE International Conference on Multimedia and Expo (ICME), California, Jul. 2013, pp.1-6.