Majority of the top N recommended items for users are mostly similar

This is my code:

import numpy as np
import pandas as pd
from lightfm import LightFM
from lightfm.data import Dataset


def make_model_for_recommendation_with_dataset():
    _dataset = Dataset()
    _data_pd = pd.read_csv("../ml-latest-small/ratings.csv")
    movies_pd = pd.read_csv('../ml-latest-small/movies.csv')
    _dataset.fit(users=_data_pd["userId"], items=_data_pd["movieId"])
    ratings_gp = _data_pd["rating"].apply(lambda x: x / 5.0)  #-> 0.0 >= weight >= 1.0
    interaction_list = (zip(_data_pd["userId"], _data_pd["movieId"], ratings_gp))
    interactions, weights = _dataset.build_interactions(interaction_list)
    model = LightFM(loss='warp')
    model.fit(interactions, num_threads=2)
    items_id_mapping = _dataset.mapping()[2]
    max_inner_id = max(items_id_mapping.values()) + 1
    item_labels = np.empty(max_inner_id, dtype=np.object)
    for movie_id, title in list(zip(movies_pd['movieId'], movies_pd['title'])):
        iid = int(movie_id)
        if iid in items_id_mapping:
            item_labels[items_id_mapping[iid]] = title
    return (model, {
        "train": interactions,
        "item_labels": item_labels,
        "mapping": _dataset.mapping()
    })


def sample_recommendation(_model, _data, user_ids):
    n_users, n_items = _data['train'].shape
    for user_id in user_ids:
        if 'mapping' in _data.keys():
            user_id_mapping, _, _, _ = _data['mapping']
            user_id = user_id_mapping[user_id]
        known_positives = _data['item_labels'][_data['train'].tocsr()[user_id].indices]
        scores = _model.predict(user_id, np.arange(n_items))
        top_items = _data['item_labels'][np.argsort(-scores)]
        print("User %s" % user_id)
        print("     Known positives:")
        for x in known_positives[:6]:
            print("        %s" % x)
        print("     Recommended:")
        for x in top_items[:6]:
            print("        %s" % x)


def recommendations():
    model, data = make_model_for_recommendation_with_dataset()
    while True:
        _user_id = int(input("Enter user id (0 for exit):\n").strip(" \n"))
        if _user_id == 0:
            break
        sample_recommendation(model, data, [_user_id])


recommendations()

But the majority of the top N recommended items for users are mostly similar. like this:

userId = 112
     Known positives:
        Pulp Fiction (1994)
        Twister (1996)
        Birdcage, The (1996)
        Willy Wonka & the Chocolate Factory (1971)
        Star Trek: First Contact (1996)
        Grumpier Old Men (1995)
     Recommended:
        Forrest Gump (1994)
        Star Wars: Episode IV - A New Hope (1977)
        Silence of the Lambs, The (1991)
        Jurassic Park (1993)
        Toy Story (1995)
        Schindler's List (1993)

userId = 113
     Known positives:
        GoldenEye (1995)
        Aristocats, The (1970)
        Happy Gilmore (1996)
        Get Shorty (1995)
        Dead Man Walking (1995)
        Jumanji (1995)
     Recommended:
        Forrest Gump (1994)
        Star Wars: Episode IV - A New Hope (1977)
        Jurassic Park (1993)
        Silence of the Lambs, The (1991)
        Toy Story (1995)
        Schindler's List (1993)

userId = 25
     Known positives:
        Beavis and Butt-Head Do America (1996)
        Trainspotting (1996)
        Star Wars: Episode IV - A New Hope (1977)
        Willy Wonka & the Chocolate Factory (1971)
        Star Trek: First Contact (1996)
        Grumpier Old Men (1995)
     Recommended:
        Forrest Gump (1994)
        Star Wars: Episode IV - A New Hope (1977)
        Silence of the Lambs, The (1991)
        Schindler's List (1993)
        Pulp Fiction (1994)
        Jurassic Park (1993)

[SimonCW]

@merrcury hard to say from here. I don’t have much experience with the MovieLens dataset. Have you compared your results with the example in the documentation to verify that your code is fine?

In general, popularity bias (over-emphasizing very popular items in the recommendations) is a common thing in recommender systems. One easy way that helped in my case was to “ignore” the item biases (you can think of them as “encoding popularity”) when scoring the model like so:

import numpy as np

recipe_biases = model.get_item_representations(features=None)[0]
zero_recipe_biases = np.zeros_like(recipe_biases)
model.item_biases = zero_recipe_biases

Afterwards if you call predict on the model, the scores won’t be influenced by the biases.

[shakeel-appyhigh]

I have the same problem. My matrix size is Num users: 287823, num_items 56393 Gives the same recommendations for all users.