A Video Recommendation Engine

The customer runs a large digital cinema platform that contains a vast collection of various movies, TV series, and short clips. After careful analysis, it was identified that most of their users were insufficiently engaged in comparison with competitor platforms and didn't spend much time consuming content. The customer decided to address this situation with multiple actions, including the improvement of the content library and enhancement of its recommendation engine. Due to a lack of expertise in the AI/ML area, the company contacted our team to optimize its video recommendation system and hence provide more relevant "watch next" proposals to end-users, encouraging them to spend more time in the system.

The digital cinema company was struggling with keeping end-users engaged so that they spent more time consuming content via its platform. This was a key business challenge that needed to be addressed to improve client satisfaction and revenue generation rates.

The customer found that their content library didn't meet the needs and preferences of their target audience, which was contributing to low user engagement and retention. This was a specific area of focus that could be tackled with better content acquisition and production strategies.

Based on initial analysis, our specialists identified that their existing recommendation engine was too simple and couldn't effectively handle complex segmentation and customer behavior patterns. Andersen's target was to develop a truly cutting-edge content-based recommendation system example. The following things were planned:

Over-the-top (OTT) platforms are becoming increasingly popular, with more consumers cutting the cord on traditional cable TV and opting for streaming services. This trend is expected to continue, and OTT revenues are projected to reach $159 billion by 2025.

The number of streaming platforms continues to grow, and consumers are becoming more selective about the content they choose to watch. This has led to a growing demand for personalized content recommendations tailored to individual viewing habits and preferences.

While the major streaming platforms like Netflix, Amazon Prime Video, and Disney+ continue to dominate the market, there is also a growing trend towards niche content platforms that cater to specific interests and demographics. Examples include Crunchyroll for anime fans, Shudder for horror enthusiasts, and BroadwayHD for theater lovers.

Many streaming platforms are now integrating with social media platforms like Facebook, Twitter, and Instagram to make it easier for users to share and discover new content. This can include personalized content recommendations based on the user's social media activity or social media sharing features that enable users to share their favorite shows or movies with their friends and followers.

In order to stay competitive, many streaming platforms are investing heavily in creating original content production – movies, TV shows, and documentaries – as well as acquiring exclusive distribution rights covering popular titles. The goal is to build a loyal fan base and differentiate themselves from competitors by offering unique and engaging content.

It was decided that the target solution of this movie recommendation system project would include:

Main components of the platform:

Because of that, there's no guarantee that this assignment was randomized – therefore, the historical data can have arbitrary biases.

Andersen's team addressed the above challenge with four steps:

ML algorithms predict the probability of a customer’s action by learning patterns based on historical data. In the context of video streaming platforms, this could involve analyzing a customer's viewing history, preferences, ratings, and other factors to predict the probability of them taking a particular action, such as watching a specific movie or TV show.

An individual prediction model is trained for each of the products/services to be recommended to clients.

The resulting model scores are then calibrated across content to generate comparable probability scores in order to rank the videos for each client.

For the first version, each model applies the same classification algorithm (e.g., gradient boost) trained on a standardized dataset (client demographics, interaction with products and customer service, product purchases, etc.).

Later, each individual model is improved (the parameters are tuned).

This approach can be particularly useful in situations where the environment is constantly changing, as it allows the system to adapt and optimize its recommendations in real time based on client feedback and other signals. However, reinforcement learning can also be more complex and computationally intensive than supervised learning and may require more data and resources to be effective.

By using A/B testing in the following way, we were able to make data-driven decisions about which features to add to the platform and how to optimize them for the best user experience:

First, we needed to develop a hypothesis about the changes we wanted to test. For example, we might have hypothesized that changing the algorithm used to generate recommendations would result in higher engagement among users.

Next, we needed to create two variations of the platform to test against each other. One variation would use the new recommendation algorithm, while the other would use the legacy algorithm.

Then, we needed to choose a representative sample of users to test the variations on. This sample should've been large enough to provide statistically significant results but not too large to negatively impact the user experience.

The sample of users should've been randomly divided into two groups, with one group seeing recommendations by the new algorithm and the other group seeing recommendations by the legacy algorithm.

It was necessary to choose a metric to measure the success of each variation. This could be anything from click-through rates to user engagement time.

The variations were then launched and tested by our specialists on the sample of users. Data on the chosen metric was collected and analyzed.

Finally, we analyzed the data to determine if there was a statistically significant difference between the two variations. If the new algorithm performed better, we would implement it on the platform. If not, we would continue testing until we found an improvement.

For each of these steps, there are many methods and techniques that provide different trade-offs in terms of complexity, the restrictiveness of assumptions, and the quality of results. Andersen evaluated all these options and came up with the solution.

This implementation has significantly improved client segmentation practices and made it possible to show more relevant recommended offers. The following results have been achieved: