Video Analytics @Scale - Junction TV Analytics Whitepaper.pdf · The processing engine leverages...

7
Video Analytics @Scale Junction TV Media Insight Framework New generation of video analytics Existing video analytics solutions are fragmented and limited to reporting logged data through various segmentation approaches. Fragmented analysis of data limits the level of holistic insights and its strategic actionable value. Such solutions are generic and therefore do not specically address need for specic service category or vertical. Junction TV media insight is built to address a specic media vertical: multi-device over-the-top TV service. Beyond traditional video analysis, we have built intelligent deep correlation and mining features that provide both long-term and real-time actionable insights. Such holistic insights provides extensive analysis into improving users experience, forecasting users wish, reducing user churn, as well as understanding the business impact. To support deeper level of insights amid rapid penetration of OTT devices, there is critical need for hyper-scale analytics data pipelining and processing infrastructure. Mature open source big data technologies such as Kafka and Spark allow real-time processing of large data streams. JunctionTV Media Insight platform is built leveraging these technologies and following the lambda architecture as recently proven by web-content companies such as Twittter, Facebook, Netix. In order to have rapid elasticity in the processing infrastructure, our analytics system leverages the Amazon Web Services (AWS). In fact, Netix reportedly handles over 550 billion events per day over AWS infrastructure using lambda architecture. ©Juncon TV Inc www.juncontv.com Juncon TV Media Insight

Transcript of Video Analytics @Scale - Junction TV Analytics Whitepaper.pdf · The processing engine leverages...

Page 1: Video Analytics @Scale - Junction TV Analytics Whitepaper.pdf · The processing engine leverages mature big-data technologies like Kafka, Spark and Cassandra. Scaling is achieved

Video Analytics @Scale Junction TV Media Insight Framework New generation of video analytics

Existing video analytics solutions are fragmented and limited to reporting logged data through various segmentation approaches. Fragmented analysis of data limits the level of holistic insights and its strategic actionable value. Such solutions are generic and therefore do not specifically address need for specific service category or vertical.

Junction TV media insight is built to address a specific media vertical: multi-device over-the-top TV service. Beyond traditional video analysis, we have built intelligent deep correlation and mining features that provide both long-term and real-time actionable insights. Such holistic insights provides extensive analysis into improving users experience, forecasting users wish, reducing user churn, as well as understanding the business impact.

To support deeper level of insights amid rapid penetration of OTT devices, there is critical need for hyper-scale analytics data pipelining and processing infrastructure. Mature open source big data technologies such as Kafka and Spark allow real-time processing of large data streams. JunctionTV Media Insight platform is built leveraging these technologies and following the lambda architecture as recently proven by web-content companies such as Twittter, Facebook, Netflix. In order to have rapid elasticity in the processing infrastructure, our analytics system leverages the Amazon Web Services (AWS). In fact, Netflix reportedly handles over 550 billion events per day over AWS infrastructure using lambda architecture.

©Junction TV Inc www.junctiontv.com Junction TV Media Insight

Page 2: Video Analytics @Scale - Junction TV Analytics Whitepaper.pdf · The processing engine leverages mature big-data technologies like Kafka, Spark and Cassandra. Scaling is achieved

A major challenge faced by OTT businesses is that they are no longer confined to a walled garden. Quality of experience is a new core driver in addition to quality of content. Howsoever well a subscriber’s interests align with the content, frequent buffering during playback would drive her away. In current scenario, if a subscriber complains about “bad quality”, the provider has no idea about what could be wrong. Knowing that multiple subscribers from a particular ISP are facing quality issues, the provider could take up the issue with the ISP and/or update peering agreements. The business models have evolved and so have the factors that affect the business. But traditional video analytics haven’t evolved to understand this reality. Junction TV Media Insight is a comprehensive video analytics framework that provides a 360 degree view of the whole OTT service. Sample these typical issues in OTT services:

- The subscriber churn has increased and you don’t know why? Media Insight can help.

- You want to know the growth trends as a result of a new campaign? Media Insight has it.

- You need to troubleshoot an individual subscriber’s bad QoE problem? Media Insight will allow you to drill down to each playback session of the subscriber.

- Is your live event not being viewed as expected? Media Insight provides live viewership information including numbers / locations of viewers and any QoE problems they face

- How well is the AVOD service working? Media Insight can tell you whether your ad-provider is doing a good job or not.

- You simply need the raw numbers of devices and subscribers like the traditional systems? We give that too.

Know Thy Subscriber

Media Insight’s fundamental tenet is – Know Thy Subscriber. So rather than just working with counts, Media Insight has user-level tracking at its core. It tracks all events of a subscriber from various sources including view reports, device reports, subscription information and QoE reports.

With Media Insight, it is possible to answer questions like

• Which subscribers have reduced their video consumption by 50% in the last month? • Which subscribers are facing bad QoE that is affecting their viewing behavior? • How many subscribers, who faced QoE problem in more than 30% of their sessions in last month, left the

service?

From an OTT business perspective, these insights are critical. However, these are not available in conventional analytics solutions.

Media Insight goes one step further and relies on deep correlation to provide actionable insights.

©Junction TV Inc www.junctiontv.com Junction TV Media Insight

Page 3: Video Analytics @Scale - Junction TV Analytics Whitepaper.pdf · The processing engine leverages mature big-data technologies like Kafka, Spark and Cassandra. Scaling is achieved

What is Deep Correlation? Media Insight takes data from diverse sources. These include :

- Subscriber Management System: Information about subscriber account status. - Device Logs: Content view logs, Ad views and QoE information for each play session. -

-

CDN Access Logs: Data transfer logs for each URL request from a playback session.

Content Metadata: Various metadata associated with an asset like its genre, cast, directors along with content-access information like its position in the playlist

-

Subscriber’s behavior is modeled using data from all these inputs. Similarly, Content access profile is created. Together these serve as basis for advanced correlation reports like recommendation, churn analysis and cohort analysis.

Causality AnalysisCausality Analysis

Causality AnalysisPredictive Analysis

Causality AnalysisRecommendation

Causality AnalysisTrends and Statistics

Subscriber Management

Device Logs

ContentMetadata

CDN Logs

Subscriber & CohortBehavior Models

Content & DeviceUsage Models

In OTT world, 1 billion events a day is a small number. Just 10M subscribers with an average of 5 devices (in a household), can generate that many events (assuming a mere 20 events per day per device). Media Insight is designed to scale to billions of events to serve the needs of emerging OTT world

Does it scale?

The processing engine leverages mature big-data technologies like Kafka, Spark and Cassandra. Scaling is achieved using Amazon Web Services. We implement Lambda architecture to achieve hyper-scale while providing both real-time information as well as deep insights with batch processing. Our deployment architecture is shown below.

©Junction TV Inc www.junctiontv.com Junction TV Media Insight

Page 4: Video Analytics @Scale - Junction TV Analytics Whitepaper.pdf · The processing engine leverages mature big-data technologies like Kafka, Spark and Cassandra. Scaling is achieved

AWS S3

Amaz

on K

ines

is /

Kafk

a

Repo

rt S

ervi

ng L

ayer

REST

API

Distributed Logger DistributedHigh Performance

Pub-Sub Messaging

Real-timeData Validation

Data Sink

Cassandra

Hadoop/EMR

AWS Redshift

ApacheSpark

AnalyticsProcessing

So what can Media Insight do? Some of the capabilities are given below.

Media Insight: Deep dive

Media Insight provides a 360 degree holistic view of the prime actors in the OTT world– devices, subscribers and content

Understand what is going on?

- Install penetration: How many devices installed the video delivery application is the most fundamental information However, Media Insight provides more pertinent information: How many devices are “truly” active? How many devices haven’t been used in a long time to watch any content?

- Consumption patterns: How many videos are watched on a typical device in a month? What time of day are the videos watched?

©Junction TV Inc www.junctiontv.com Junction TV Media Insight

Page 5: Video Analytics @Scale - Junction TV Analytics Whitepaper.pdf · The processing engine leverages mature big-data technologies like Kafka, Spark and Cassandra. Scaling is achieved

- Subscriber Behavior: Correlating across devices, how many devices does a subscriber use? How much video does a typical subscriber consume in a given duration?

- Content Value: What type of content is popular? Are titles getting the traction they should considering the licensing cost?

Bad QoE is precursor to a subscriber leaving even if the content is great. The goal of QoE analytics is to enable elimination of “QoE as a concern” altogether. To achieve this, if a subscriber complains about bad quality, the provider has to have actionable input to tackle the problem. Media Insight provides this input:

Improve Quality of experience

- All information down to the frequency and duration of buffering in each playback session is known. For each subscriber, Media Insight provides ability to highlight good and bad QoE for each playback.

- Isolation of ISPs at specific locations whose subscribers are consistently having bad QoE. Businesses can upgrade/change their CDNs or peering.

- Real-time QoE assessment helps in identifying and possibly fixing QoE problems in live setting

- If specific content playback results in bad quality sessions, the files of video may be corrupt (in case of VoD) and have to be properly re-encoded. Media Insight can identify the set of possible re-encoding candidates for you.

-

Understand the business impact The bottom-line of any business is to sustain and grow revenue. In context of OTT, this involves keeping tab on revenue drivers like Ads and Subscribers. Media Insight provides understanding of these drivers:

- The efficacy of Ad revenue depends on: 1) Ability of the ad-provider to fill the ad-slots: Media Insight highlights the fraction of ad-slots that is left

unfilled. This enables content owners to negotiate with or change ad providing partners.

2) Ad viewing behavior of subscribers: Subscriber’s ad watching behavior allows tailoring of marketing strategies appropriately. For example, subscribers who tend to avoid ads can be targeted to become SVOD consumers.

3) Frequency of occurrence of ad-slots (or cue-points) in a content timeline: Aggregated subscriber behavior can be used to reschedule cue-points to maximize the ad revenue.

- Subscriber growth is a critical indicator in an SVOD (it is important in any model) model:

1) Trend of subscriber growth: The bottom-line is knowing the number of subscribers and the growth trend. This information is of core strategic importance in planning growth.

©Junction TV Inc www.junctiontv.com Junction TV Media Insight

Page 6: Video Analytics @Scale - Junction TV Analytics Whitepaper.pdf · The processing engine leverages mature big-data technologies like Kafka, Spark and Cassandra. Scaling is achieved

2) Subscriber’s viewing behavior: How much content a typical subscriber consumes drives the package pricing. (don’t refer to package pricing alone) something else

Understand content performance Content licensing and production (OTT service company does not produce content) is typically the biggest cost in the business. Hence understanding whether the content is worth its cost is critical to the business.

- Getting the raw numbers of views helps in understanding the content popularity.

- Deeper analysis can provide breakup of AVOD and SVOD consumption of the content. This can help in making decisions such as whether moving an asset to SVOD-only tier would increase revenue.

- Semantic analysis of popular tags helps in content procurement/producing decision. For example, if Sci-fi with “space” theme is gaining popularity, adding “Star-trek” to the offerings may be beneficial.

- Navigation Category analysis helps in determining content placement for ease of discovery. For example, if users do not navigate through a category “Space” but search for content “Star trek”, moving related content to “Sci-fi” or creating a new category “Star trek library” may make sense.

Understand the root causes behind churn A core feature of our analytics is the multi-dimensional churn analysis. Retaining subscribers is a primary goal of any content owner. Subscribers leaving service is a cause of concern for the business. Rather than just showing that the number of subscribers is dwindling, Media Insight analyses view data, QoE data and subscription information and correlates them to identify potential reasons for the churn:

- Intent analysis: Identifying those subscribers who were in it purely to try out the service. These pseudo-conversions typically serve to inflate subscription numbers whereas they were never potential long-term subscribers.

- QoE analysis: Subscribers leaving due to bad QoE can be identified by Media Insight using a subscriber’s viewing and QoE history. With proper QoE management, these subscribers could have been retained.

- Cohort analysis: Insight reports group subscribers along various dimensions to provide a comprehensive cohort-analysis:

o ISP-level cohorts are useful in identifying groups of subscribers from same ISPs that are part of the churn. This helps in identifying bad QoE from corresponding ISPs leading to churn.

o Geographic cohorts tend to identify if users from common location are part of churn. This helps in understanding if the users from a particular geography are more or less likely to watch your content.

o Temporal cohort analysis is used to identify if the content itself is engaging users or not. For example, treating users who subscribed during the same month as a cohort, comparison across cohorts from different months can provide the consumption trends.

For each of these cohorts, QoE and Content consumption dimensions are evaluated to narrow down the causes.

©Junction TV Inc www.junctiontv.com Junction TV Media Insight

Page 7: Video Analytics @Scale - Junction TV Analytics Whitepaper.pdf · The processing engine leverages mature big-data technologies like Kafka, Spark and Cassandra. Scaling is achieved

Actionable Intelligence Media Insight takes a quantum leap in providing deeper insights into your operations. Some of the reports with deep business impact are as follows:

- High-potential conversion targets: In businesses which have both AVOD and SVOD offerings, converting AVOD users to SVOD is desirable. Media Insight can analyze viewing behavior of AVOD subscribers and recommend the subset of users who the marketing team could pursue with SVOD offers for a good conversion rate.

- Need for fresh content: Based on subscriber’s viewing model, Media Insight can provide information of when fresh content is needed to keep a subset of subscribers engaged.

- Troubleshooting sudden QoE issues: The ability to correlate events down to device make as well as firmware level allows Media Insight to isolate root causes of problems. For example, It can identify that iPAd users are facing problems as indicated by reduced activity or that a new android release in some phones is creating a lot of QoE issues

- ISP profiling: Information like list of ISPs that have over 30% of users facing QoE problems is useful in taking up the issues with corresponding ISPs.

- Content Recommendation: Based on subscriber viewing behavior, Media Insight can assist in recommending content for individual users as well as placement of content in navigation structures.

- Encoding recommendation: Analyzing the stream rates of content viewed and corresponding QoE, recommendations on re-encoding content and required set of encoding bitrates can be generated.

Ad Insertion Logic: Typically ad-slots in a video are determined at arbitrary locations based on thumb-rules.Media Insight can assist you in determining good ad-slots that can increase revenue. For example, arecommendation can be to have extra ads at the start of the content rather than at equal intervals.

With its deep analytics capabilities, Media Insight is equipped to be an able assistant for any content business. As the content business evolve, data-driven intelligence is the only way forward and Media Insight leads the way.

-

©Junction TV Inc www.junctiontv.com Junction TV Media Insight