hiltonrosenfeld/real-time-ai-movielens

Real-time AI Demo based on MovieLens movie recommendations

What is Real-time AI?

Applied Machine learning, driven by the most recent, most valuable user and operational data, at scale, to power more accurate decisions and actions in real-time. Real-time AI refers to the use of artificial intelligence algorithms and models to make decisions and predictions in real-time with real-time data.

Real-time AI is used in a wide range of applications, including fraud detection, predictive maintenance, autonomous vehicles, and real-time decision-making in finance, healthcare, and manufacturing.

Understanding Real-time in AI

Let’s take an example of a movie recommendation model to demonstrate and understand the value of Real-time in AI.

We will use Movielens dataset to build a recommendation model from scratch, see how it works and evaluate how we could leverage Real-time data to improve the recommendations and user experience.

Goal

Given a user, return top 5 movie recommendations

Pre-requsities

pip install numpy
pip install pandas
pip install tensorflow
pip install scikit-learn

Dataset recommended for educational purpose is already included here. If needed, explore / download Movielens dataset,

Sample Data

userId	movieId	rating	timestamp
1	1	4.0	964982703
1	3	4.0	964981247
1	6	4.0	964982224
1	47	5.0	964983815

Training

Let's train a recommendation model with this dataset

Feel free to review train.py on preprocessing required and the model design.

python train.py

..
Epoch 1/10
1261/1261 [==============================] - 2s 1ms/step - loss: 0.0427
Epoch 2/10
1261/1261 [==============================] - 2s 1ms/step - loss: 0.0379
Epoch 3/10
1261/1261 [==============================] - 2s 1ms/step - loss: 0.0363
Epoch 4/10
1261/1261 [==============================] - 2s 1ms/step - loss: 0.0350
Epoch 5/10
1261/1261 [==============================] - 2s 1ms/step - loss: 0.0338
Epoch 6/10
1261/1261 [==============================] - 2s 1ms/step - loss: 0.0325
Epoch 7/10
1261/1261 [==============================] - 2s 1ms/step - loss: 0.0313
Epoch 8/10
1261/1261 [==============================] - 2s 1ms/step - loss: 0.0303
Epoch 9/10
1261/1261 [==============================] - 2s 1ms/step - loss: 0.0294
Epoch 10/10
1261/1261 [==============================] - 2s 1ms/step - loss: 0.0285

After training completes, it will save the trained model to movie_recommendation.h5

Inference

> python serve.py
304/304 [==============================] - 0s 228us/step
[7235 1761  709  277  687]

Returns the Top5 movie recommendations for userId 1.

Success

Success!! While this is Great, let's talk about how this model could be improved.

This model uses only the ratings information captured when a user rates a movie for training and inference. What this means is,

1. The model will produce the same output until its trained on new data.

2. This also misses out on a lot of other actions a user could be doing, for eg,

• browsing
• checking out trailers
• abandoning a movie midway
• watching a movie but not leaving a rating etc these are useful signals that could help with improving the recommendations.

Many of them are time-bound or values that changes over time, for eg, a feature like last_watched_movie will change everytime a user have watched a movie. It will be important to capture the entire journey of such features for better recommendations.

Feature engineering

Now, lets do some feature engineering!

Feature engineering or feature extraction or feature discovery is the process of using domain knowledge to extract features (characteristics, properties, attributes) from raw data

Notice the timestamp field in our initial dataset. We know when a user had watched a movie. We could use this information to engineer 2 new features

last_watched_movie - movie watched by the user before watching movie_id

last_rating - movie rating provided to last_watched_movie

Also remember, these values change over time. These features are usually referred to as Real-time features.

We could populate the new features to our dataset,

userId	movieId	rating	timestamp	last_watched_movie	last_rating
1	1	4.0	2000-07-30 18:45:03	1023	5.0
1	3	4.0	2000-07-30 18:20:47	1777	4.0
1	6	4.0	2000-07-30 18:37:04	2000	4.0
1	47	5.0	2000-07-30 19:03:35	593	4.0

Training

Please review train_realtime.py on how we added new features to the model.

python train_realtime.py

..
Epoch 1/10
1253/1253 [==============================] - 3s 2ms/step - loss: 0.0419
Epoch 2/10
1253/1253 [==============================] - 3s 2ms/step - loss: 0.0369
Epoch 3/10
1253/1253 [==============================] - 3s 2ms/step - loss: 0.0351
Epoch 4/10
1253/1253 [==============================] - 3s 2ms/step - loss: 0.0331
Epoch 5/10
1253/1253 [==============================] - 3s 2ms/step - loss: 0.0313
Epoch 6/10
1253/1253 [==============================] - 3s 2ms/step - loss: 0.0297
Epoch 7/10
1253/1253 [==============================] - 3s 2ms/step - loss: 0.0283
Epoch 8/10
1253/1253 [==============================] - 3s 2ms/step - loss: 0.0271
Epoch 9/10
1253/1253 [==============================] - 3s 2ms/step - loss: 0.0261
Epoch 10/10
1253/1253 [==============================] - 3s 2ms/step - loss: 0.0250

After training completes, it will save the trained model to movie_recommendation_realtime.h5

Inference

>>> get_user_recommendations(1,1,3,5)
304/304 [==============================] - 0s 258us/step
array([2224,  602,  908,  904, 1700])
>>> get_user_recommendations(1,5,2,5)
304/304 [==============================] - 0s 281us/step
array([ 602, 9600, 2224,  277, 1700])

Notice, everytime Real-time features changes, in other words, based on user activity, this model is able to generate a more contextual, relevant recommendations.

This is a simple example. If we are able to capture the user activity in real-time, and use it for inference, we can deliver a contextual, relevant recommendations that will improve the overall user experience.

Congratulations!! We have built our first Real-time AI model!

Live Demo

UI: https://movie-recommendations-demo.web.app/

Batch AI button at the bottom right allows you to toggle between Batch and Realtime mode.

Citation

To acknowledge use of the dataset in publications, please cite the following paper:

F. Maxwell Harper and Joseph A. Konstan. 2015. The MovieLens Datasets: History and Context. ACM Transactions on Interactive Intelligent Systems (TiiS) 5, 4: 19:1–19:19. https://doi.org/10.1145/2827872