LightFM for purchase data: how should I prepare interaction weights? Current approach uses per-user frequency normalized by max item frequency

I am training a recommendation model with LightFM on implicit purchase data.

In my dataset, each interaction represents a transaction: a user buys an item.

If a user buys the same item multiple times, I first aggregate the transactions by (user, item) and compute the purchase frequency.

My current approach for preparing interaction weights is:

For each (user, item) pair, compute the purchase frequency.

For each user, find the maximum purchase frequency across all items that user has bought.

Define the interaction weight as:

weight(user, item) = frequency(user, item) / max_frequency_of_any_item_for_that_user

So the weight is normalized per user and always falls in (0, 1].

For example, if a user has:

item A: 10 purchases

item B: 4 purchases

item C: 1 purchase

then the weights become:

item A: 10 / 10 = 1.0

item B: 4 / 10 = 0.4

item C: 1 / 10 = 0.1

My goal is top-K recommendation from purchase history.

My current LightFM settings are:

loss = warp

max_sampled = 20

no_components = 32

item_alpha = 0.00001

user_alpha = 0.0000001

What I would like advice on is:

Is this a reasonable way to prepare interaction weights for LightFM on purchase data?

Does normalizing by each user’s maximum item frequency remove too much useful signal?

Is there a better principle for preparing interaction weights for repeated purchases in implicit-feedback recommendation?

In practice, when using LightFM on purchase data, should interaction weights mainly reflect:

relative preference within a user,

or absolute purchase intensity?

I would appreciate advice from anyone who has worked with LightFM or implicit-feedback recommendation on transaction data.