same recommendations to all the users

[yestoprince]

@takuti
I tried to build a recommendation engine using matrix factorization but after training my model i found that it is giving same recommendations to all the users is there any issue with it i followed all the instructions as per your tutorial.
can you please check my code to see is there any issue with the code ? or with your prediction method.
Thanks :)

Here is my code

`using CSV

train = CSV.read("training_data_book.csv", delim = ',', escapechar = '\' )
test = CSV.read("testing_data_book.csv", delim=',', escapechar='\')

Created dictionaries in below code for mapping the original ids with sequential ids

user_mappings = Dict{Int, Int}()
book_mappings = Dict{String, Int}()
user_counter, book_counter = 0,0
events = Event[]

for row in eachrow(train)
global user_counter, book_counter
user_id, book_id, rating = row[:UserID], row[:ISBN], row[:Rating]
haskey(user_mappings, user_id) || (user_mappings[user_id] = (user_counter += 1))
haskey(book_mappings, book_id) || (book_mappings[book_id] = (book_counter += 1))
push!(events, Event(user_mappings[user_id], book_mappings[book_id], rating))
end

Calling to data Accessor

da = DataAccessor(events, user_counter, book_counter)
recommender = MF(da, Parameters(:k => 2))

Here i build my model

build(recommender, learning_rate=15e-4, max_iter=100)

Predicting 5 recommendation of user whose sequential id is 10

for i in recommendations_to_books(recommend(recommender, 35, 10, [1:book_counter...]))
main_data[main_data.ISBN .== i, :BookTitle] |> println
end

for i in recommendations_to_books(recommend(recommender, 15, 10, [1:book_counter...]))
main_data[main_data.ISBN .== i, :BookTitle] |> println
end
`

[yestoprince]

Here the recommender for user 35 as well as user 15 above giving same recommendation and i tested for other users ids as well the result is still the same :( Please help
@takuti

[yestoprince]

Check this screenshot @takuti

[takuti]

@yestoprince Sorry for the late response.

Can you first check if recommend(recommender, user, 10, [1:book_counter...]) gives the same predicted ratings to all user-item pairs? For example, our tiny example in README actually gives same recommendation to user 1, 2, and 3, but predicted values are different:

julia> recommend(recommender, 1, k, candidates)
2-element Array{Pair{Int64,Float64},1}:
 6 => 0.0890438
 2 => 0.0465087

julia> recommend(recommender, 2, k, candidates)
2-element Array{Pair{Int64,Float64},1}:
 6 => 0.207168
 2 => 0.108206

julia> recommend(recommender, 3, k, candidates)
2-element Array{Pair{Int64,Float64},1}:
 6 => 0.0890386
 2 => 0.0465061

In this case, recommender works fine, and problem is just a lack of training samples or inappropriate configuration of Parameters().

Also, please make sure if all users and items have at least 1 rating event. To give an example, the README example has no events for user 4 and 5:

const n_user = 5
const n_item = 10

events = [Event(1, 2, 1), Event(3, 2, 1), Event(2, 6, 4)]

da = DataAccessor(events, n_user, n_item)

Consequently, both prediction and recommendation results for those two users become same, because there is no chance to update their model parameters:

julia> recommend(recommender, 4, k, candidates)
2-element Array{Pair{Int64,Float64},1}:
 6 => 0.0643689
 2 => 0.0336207

julia> recommend(recommender, 5, k, candidates)
2-element Array{Pair{Int64,Float64},1}:
 6 => 0.0643689
 2 => 0.0336207

Note that missing events are equally filled by 0.1 by default:

Recommendation.jl/src/model/mf.jl

Lines 25 to 28 in 1828df9

	# initialize with small values
	# (random is also possible)
	P = ones(n_user, rec.hyperparams[:k]) * 0.1
	Q = ones(n_item, rec.hyperparams[:k]) * 0.1

For further investigation, we need to take a look at your data and implementation detail of recommendations_to_books function.

[yestoprince]

Ok thanks but now i moved to Python for building recommendation engine. Do you have experience in that ? i also have set of errors in that too ?

[takuti]

Yes I do. As long as dataset is same, I believe you'll face the same situation. Otherwise, it's a bug of Recommendation.jl. If so, it'd be appreciated if you could report the issue.

[takuti]

Feel free to reopen this issue if you have additional comments on this package.