Application in recommender system

[Tych0n]

Hi, aksnzhy! Thank you for this library. Can you please guide me a bit? I have a dataset with four columns: transaction_count, user, item, item_colour. I want to recommend some items to users, based on transaction_count. I can use ALS with transaction_count, user and item columns, for example with "implicit" library. But if i want to take in account item_colour i need to use for example ffm. So, i create ffm-formatted file:

transaction_count user_id:value_id:1 item_id:value_id:1 item_colour_id:value_id:1

5 0:0:1 1:3:1 2:5:1
3 0:1:1 1:4:1 2:6:1
8 0:2:1 1:3:1 2:7:1

and train my model. But, if i want to recommend top-5 items with some colours to a user, i need to create all combinations of user:item:colour rows, score them and then sort among each user all variants of item:colour by modeled probabilities and select 5 best among them. The problem is that such a list of all possible combinations explodes with my dimensions (users=80000, items=14000, colours=5), and impossible to operate. Is there any hack for implementation?

[aksnzhy]

@Tych0n I'm not sure if I get your pointer correctly. You want to recommend top-5 items to users, but now xLearn can only do the binary classification task. One possible solution is that you can use one-vers-all to do a multi-class classification task, and then to compare the probability calculated by xLearn, and select the highest probability. It will only spend 5x cost compare to one binary classification task.

[Tych0n]

Yes, you got it right. I'm using 'task':'reg' and transaction_count as continious target, but I sure can scale it to binary one-vers-all form per user. The problem is, that to compare probabilities calculated by xLearn, i need to score 14000*5 variants per user. And do it 80000 times - it ends in creating dataset with 5.6 billion rows, and then scoring it with xLearn. I think may be i'm missing some point and doing it wrong.

[aksnzhy]

@Tych0n If you do this job like what you said, you actually facing a classification problem with 14000*5 labels. I think it is not a very good. If you can convert it to a binary task, e.g., given an item, predict the probability how you like it. Then you can predict all the probability of items for each user.

[Tych0n]

@aksnzhy got your point. Correct me please, if i'm wrong: i convert task to binary, score all user-item pairs to get probability of each pair, then sort top-5 items for each user. In this approach i still face the problem of creating 5.6 billion rows dataset to score and select from?

[BrianMiner]

I have the same exact question - do you have to create all possible permutations of items for each user and score all of them and order by probability?

[BrianMiner]

@Tych0n did you ever figure out a method?

[Tych0n]

@BrianMiner, no. I wasn't able to use this library in production for my problem, too many items to select recommendations from. I ended up using implicit.ALS. By the way, i tried LightFM - it showed comparable MAP@k, but still wasn't able to outperform ALS. At least in my setup.

[BrianMiner]

This must be a common issue for these types of models though. There must be some way to pre-filter the candidates. You see FM and FFM used for CTR problems. That could be as large as issue when you need to score so many ads * placements * versions (copy, color etc) for each user.

[sumitsidana]

Hi, you could use negative sampling to convert the problem from one class classification to binary classification. For every positive (transaction), sample a non-transaction. This way, you can still use this library or FM or FFM. Do sampling only during the training time. While prediction time, consider all the the items for prediction.