Recommender System Project

Welcome on the AI316 Recommender System project page.

This project have to be done by team of 2.

Each week a new environment is released. Each environment simulate different aspect of user behavior. Your goal as an agent is to implement a recommender system that get the best performances on each environments.

Environments

Each environment will be accessible through a different IP address. Next examples used 1.2.3.4 but make sure to use the good ones instead.

Each environment is specific to a user session. Logging have been send to you by email.

To be identified each request should contain your user_id

user_id=aaaa

First Environment

This environment encode the classic explicit feedback setting with user, item and rating.

The first thing to do is to call the reset method. It will initialize the environment and return a new sample of historical data

Reset

http://1.2.3.4/reset?user_id=aaaa

• Restart the environment with new random values.
• Return historical data already temporally aligned.

Predict

Call

http://1.2.3.4/predict?user_id=aaaa&predicted_score=0.761

Return

{"next_item":158,"next_user":25,"rating":1}

Require:

• user_id
• predicted_score (for previous (user, item))

Return:

• rating
• next_user
• next_item

Notice the lag here. Rating returned by the environment correspond to the previously send (user, item).

Second environment

The second environment is a variation of the first one that include covariates in the inputs (a.k.a. content based).

Methods are the same but you receive also a vector of features.

• next_variables

Third environment

The third environment generate implicit feedback.

Reset

output:

• nb_users
• nb_items
• state_history
• reward_history
• action_history
• next_state

Predict

input:

• recommended_item (int. item position in previous state, not item_id)

output:

• reward
• state

Notes

State is a vector of features for k available items. Let denote j one of them.

• state[j][0] = user
• state[j][1] = item
• state[j][2] = price
• state[j][3:] = variables

Performance evaluation

Explicit feedback environments

The performance metric will be mse over 1000 steps on 3 independent run.

Implicit feedback environments

Those environment explicitly gives you the reward at each steps.

The performance metric will be the cumulative reward over 1000 steps on 3 independent run.

Deploy your own API

On the first environment you're asked to develop an API that can be requested to get prediction.

We should be able to deploy this API by running only

docker compose up --build

Evaluation method is the same than before but we will send queries every 100ms. Non responding request will count as +1000.

train

input:

• nb_users
• nb_items
• user_history
• item_history
• rating_history

do the training in less than 1 minute.

predict

input:

• user
• item

output:

• rating

can be called every 100ms.