Towards Network Aware Recommendations Savvas Kastanakis - PowerPoint PPT Presentation

Towards Network Aware Recommendations Savvas Kastanakis Postgraduate Student @ CSD UOC Supervisor: Xenofontas Dimitropoulos Advisor: Pavlos Sermpezis

Agenda ❖ Motivation ❖ The concept of Joint Caching and Recommendations ❖ Simulation Based Evaluation ❖ Real Users Evaluation ❖ Future Work

Agenda ❖ Motivation ❖ The concept of Joint Caching and Recommendations ❖ Simulation Based Evaluation ❖ Real Users Evaluation ❖ Future Work

The problem Internet mobile traffic, especially ❖ for online video services (e.g., YouTube, Netflix), increases exponentially Mobile networks struggle to ❖ atuain high QoE while serving all content requests This brings current network ❖ systems and architectures to the test

Caching: A traditional solution Win-Win (user & network) ❖ reduces access latency ➢ offloads network load (distributes it to the edges) ➢ Low Cache Hit Ratio (CHR ~= 15%) ❖ small caches ( cache size ~GB vs. catalog size ~PB) ➢ volatility in users’ preferences ➢ caching algorithms limitations (variable traffic, frequent ➢ changes of users) 5

Recommendation Systems: A modern solution Why recommendation systems (RS)? ● Help users explore the enormous ○ content space Drive content consumption ○ (~80% in Netflix, >50% in YouTube) Integrated in popular services ○ (YouTube, Netflix, Spotify, etc.) How to leverage RS? ● Recommend contents that are cached ○ e.g., [ ToMM’15 , WoWMoM’18 ] Cache contents that can be recommended ○ e.g., [Globecom’17, JSAC’18] Jointly decide caching and recommendations ○ e.g., [INFOCOM’16] 6

Agenda ❖ Motivation ❖ The concept of Joint Caching and Recommendations ❖ Simulation Based Evaluation ❖ Real Users Evaluation ❖ Future Work

Caching and Recommendations: An example low high CHR CHR 8

Caching and Recommendations: How about content similarity? Biased Initial Recommendations Recommendations 9

CABaRet: C ache- A ware and B FS related R ecommendations initial content cached content directly related content indirectly related content 10

Questions Raised ★ Is this approach going to improve CHR? ★ Are users going to follow your recommendations?

Agenda ❖ Motivation ❖ The concept of Joint Caching and Recommendations ❖ Simulation Based Evaluation ❖ Real Users Evaluation ❖ Future Work

Does this thing really work? Conducted a simulation-based evaluation ❖ CABaRet outperformed YouTube in terms ➢ of Cache Hit Ratio (CHR) CABaRet shows promising gains, ➢ providing an overall 10x increase in CHR 13

Yes, but... How about real users?

Agenda ❖ Motivation ❖ The concept of Joint Caching and Recommendations ❖ Simulation Based Evaluation ❖ Real Users Evaluation ❖ Future Work

CABaRet Experimental Testbed: Real world data Video Player Cache Friendly Recommendations Original Ratings Recommendations 16

Do users follow the “biased” recommendations? We opt to measure whether users are ● willing to select the “nudged” recommendations To quantify this, we define two metrics: ● the hit ratio, HR : this is the ratio of ○ high-QoS (cached) videos that users selected, over the total number of viewed videos the recommendation ratio, RR : this is ○ the ratio of recommended high-QoS videos over the total number of recommended videos

Win Win Situation ★ Is this approach going to improve CHR? ★ Are users going to follow your recommendations?

Agenda ❖ Motivation ❖ The concept of Joint Caching and Recommendations ❖ Simulation Based Evaluation ❖ Real Users Evaluation ❖ Future Work

Future Work QoS Interest QoE QoE = f(QoS, Interest)

21

Back-Up Slides 22

Results from real users’ ratings Win for the Average CHR 58% Network/Content Provider Ratings (in %) Cached Contents Non Cached (CABaRet Contents Recommendations) QoS 88% 33% Interest 69% 72% Win for the End User QoR 68% 71% QoE 70% 42%