0
$\begingroup$

I am trying to build a hybrid recommender system using lightFM that only recommends one of $3$ items. In my case, they are marketing campaigns that a company would like to recommend for users at a risk of churn. I have a bunch of useful user features and I can build a user item interaction matrix based on the users' feedback since I have access to historical data where in the past marketing campaigns were given. The feedback is computed based on whether the client stayed with the company or not after having received the campaign (binary response).

My question is, with very few items such as in my case, can item features be useful? For instance, say I collect a bunch of binary, categorical and numerical features for the items. I would have to one hot encode and normalize / binarize my features. However, if I have $10$ categories of one feature and only $3$ observations in my item dataset, I would end up with a quite short and quite sparse dataframe that contains little information (if not useless since I only have $3$ rows). I was wondering if there is any merit to feeding them as item features to my lightFM model in this case?

$\endgroup$

1 Answer 1

1
$\begingroup$

I would say that although you are building a recommender of sorts, really you are trying to predict which marketing pathway would be most beneficial to your customers.

In both cases you would need to have some past uses to provide similarity, so you must have some list of customers who have used each and these variables for them. I would suggest building a straight up classifier with 3 levels of the target variable. Choose the specific type of model based on the peculiarities of your data (width, length, how much categorical, how will it be deployed, how fast must it respond etc)

You will very likely get better performance this way so long as you have a sufficient number of known consumers of the three campaigns already with supporting data to train upon.

$\endgroup$
1
  • $\begingroup$ Thank you for your insight. I do indeed have access to historical data where in the past customers have recieved some campaigns randomly and an implicit feedback is computed based on whether they stayed with the company afterwards or not... I thought about a typical classifier but I still needed to explore this idea of hybrid recommenders since I can gather data for both users and items... Can you point me to any documentation or references or a comparison study of both approaches in case there are little items to recommend? Thanks again and +1 ! $\endgroup$
    – bmasri
    Commented Oct 26, 2022 at 13:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.