Training personalized based machine learning models

Question

i work on a php project along with python which uses flask as api which predict user like on a post based on the previous engagement on other posts and its purely user based.

my requirement is suppose there are 1000`s of users in our system. and they have done likes for old posts before.when new posts arrive i need to somehow identify whether user likes it or not.and this is done via a cron job

approach 1

i am using Logistic regression as model so probably need dynamic pkl file for each user.because different users engagement on same post is different so i need to save some thing like model_{user_id}.pkl file where user_id is the user id of the user

approach 2

use content based recommended system.but as far as i know it can't store like a pkl file in production. so for each users from the 1000`s of users i need to run the recommender function.

approach 1 drawback

creating dynamic pkl file for each user which means more files.i never seen this approach on internet

approach 2 drawback

calling the recommender function for each user is probably a bad idea i believe.that will heavily affect cpu usage etc.

can somebody please help me how to properly solve this problem.i am new in machine learning. please consider my question. thanks in advance.

Answer 1

I would suggest something like this:

• Create the user models as an array (or data frame) of models
• Save this array as a pkl
• When loading the app (not on each API call), load the array of models into the memory
• When an API is called, the model is already in the memory - use it to predict the result

Something like this (not tested - just a notion):

#for saving the model
model_data = pd.DataFrame(columns=['user','model'])
temp_model = RandomForestClassifier().fit(X,y)
new = pd.DataFrame({'user':[user_id],'model':[temp_model]})
model_data = model_data.append(new)
packed_model = jsonpickle.pickler.Pickler.flatten(model_data)

#for loading the model
unpacked_model = jsonpickle.unpickler.Unpickler.restore(packed_model) #this should be in the begining of your flask file - loaded into the memory
user_model=unpacked_model.at(user_id,'model') #this should be inside every api call