How to design a system with low density?

I have read through (most of) the issues and have the recommendation up and running but the result is not as good as I hoped. I have a few questions looking for your answer. Below is info about my current dataset:

user_features.shape = (79770, 26035) # avg features/user = 2
item_features.shape = (21093, 20157) # avg features/item = 4
interactions.shape = (79770, 21093)

I train with this code, which is similar to your example:

model = tensorrec.TensorRec()
model.fit(interactions, user_features, item_features, epochs=1000, verbose=True)

However, my hardware is not powerful enough so I had to picked samples from the full dataset instead of running it in full:

sample_user_features.shape = (4, 26035)
sample_interactions.shape = (4, 21093)

predicted_ranks = model.predict_rank(user_features=sample_user_features, item_features=item_features)
r_at_k = tensorrec.eval.recall_at_k(predicted_ranks, sample_interactions, k=10)
print(np.mean(r_at_k))

The recall@10 value is 0.00006. I manually inspected the predicted item for these users and it’s not good. So my questions are:

1. I think my avg features/user of 2 and avg features/item of 4 are too small. Do you think so? Should the recommendation works with such low density?
2. I choose 4 user ids from the dataset and build sample matrix to run the recall. I’m not confident about this step. Is this mathematically and statistically correct?
3. I understand that you have implemented fit_partial. In my situation, the item features can increase over time. Will fit_partial works with resized matrix?

Thank you.

[daohoangson]

@EricC91 I have had redone my input matrices so it’s denser and switched to use BalancedWMRBLossGraph. The result is good enough for our use case.

[ericcanadas]

Hi @daohoangson,

I am confronted with almost the same results as you… Did you finally find parameters that work well?

Thank you