Similar dataset but recommendations always the same for all users

Having successfully got my data into the correct format I've been able to run the quickstart recommender on my own dataset. However for some reason on a given run every user gets identical recommendations for the same item over and over again.

Looking at the movielens data in the same I see that it has the following properties:

• 944 users
• 1665 movies
• 100,000 mappings
  (each user rating ~100 movies)

My own data set has the following properties

• 4867 users
• 76 interests
• 24540 mappings
  (each user having ~5 interests)

I think I've gotten the data into the correct format. For example, the movie data mappings are like this:

b"One Flew Over the Cuckoo's Nest (1975)" b'138'
b'Strictly Ballroom (1992)' b'92'
b'Very Brady Sequel, A (1996)' b'301'
b'Pulp Fiction (1994)' b'60'
b'Scream 2 (1997)' b'197'

while I've now got the interest data similarly structured like so

b'Books' b'1242047'
b'Dance' b'91242048'
b'Sustainability' b'2870269'
b'Books' b'3970361'
b'Photography' b'3970362'

however the recommended interests with my dataset are locked into the same thing each time.

---

Running with the movielens data there's a nice spread of recommendations, and the training output is like this:

Epoch 1/3
25/25 [==============================] - 8s 260ms/step - factorized_top_k/top_1_categorical_accuracy: 7.0000e-05 - factorized_top_k/top_5_categorical_accuracy: 0.0015 - factorized_top_k/top_10_categorical_accuracy: 0.0047 - factorized_top_k/top_50_categorical_accuracy: 0.0445 - factorized_top_k/top_100_categorical_accuracy: 0.1001 - loss: 33082.5255 - regularization_loss: 0.0000e+00 - total_loss: 33082.5255
Epoch 2/3
25/25 [==============================] - 6s 246ms/step - factorized_top_k/top_1_categorical_accuracy: 1.9000e-04 - factorized_top_k/top_5_categorical_accuracy: 0.0047 - factorized_top_k/top_10_categorical_accuracy: 0.0136 - factorized_top_k/top_50_categorical_accuracy: 0.1065 - factorized_top_k/top_100_categorical_accuracy: 0.2112 - loss: 31007.2517 - regularization_loss: 0.0000e+00 - total_loss: 31007.2517
Epoch 3/3
25/25 [==============================] - 6s 242ms/step - factorized_top_k/top_1_categorical_accuracy: 2.3000e-04 - factorized_top_k/top_5_categorical_accuracy: 0.0078 - factorized_top_k/top_10_categorical_accuracy: 0.0212 - factorized_top_k/top_50_categorical_accuracy: 0.1432 - factorized_top_k/top_100_categorical_accuracy: 0.2669 - loss: 30418.3815 - regularization_loss: 0.0000e+00 - total_loss: 30418.3815
Top 3 recommendations for user 42: [b'Rent-a-Kid (1995)' b'Only You (1994)' b'Just Cause (1995)']

and we can see a nice pattern of recommendations for users:

thing, titles = index(np.array(["1"]))
print(f"Top 10 recommendations for user {user_ids_vocabulary.get_vocabulary()[1]}: {titles}")
print(f"thing: {thing}")
thing, titles = index(np.array(["2"]))
print(f"Top 10 recommendations for user {user_ids_vocabulary.get_vocabulary()[2]}: {titles}")
print(f"thing: {thing}")

Top 10 recommendations for user 405: [[b'Doom Generation, The (1995)'
  b'Brother Minister: The Assassination of Malcolm X (1994)'
  b'Theodore Rex (1995)' b'Nadja (1994)'
  b'Turbo: A Power Rangers Movie (1997)'
  b'All Dogs Go to Heaven 2 (1996)' b'Kansas City (1996)'
  b'Maya Lin: A Strong Clear Vision (1994)' b'White Balloon, The (1995)'
  b'Flipper (1996)']]
thing: [[4.8432746 4.619417  4.452347  4.371699  3.4825168 3.1067772 3.0736032
  3.0411866 3.0028658 2.8818445]]
Top 10 recommendations for user 655: [[b'3 Ninjas: High Noon At Mega Mountain (1998)' b'Promesse, La (1996)'
  b'For the Moment (1994)' b'City of Angels (1998)'
  b"Antonia's Line (1995)" b"Marvin's Room (1996)"
  b'Once Upon a Time... When We Were Colored (1995)'
  b'Unhook the Stars (1996)' b'Kolya (1996)' b'Secrets & Lies (1996)']]
thing: [[9.494508  6.5379915 5.68907   5.3878336 5.3482184 5.290077  5.2577763
  5.2102613 5.1238346 5.1227913]]

but for the interests data we get this:

Epoch 1/3
6/6 [==============================] - 10s 2s/step - factorized_top_k/top_1_categorical_accuracy: 0.1213 - factorized_top_k/top_5_categorical_accuracy: 0.1213 - factorized_top_k/top_10_categorical_accuracy: 0.1213 - factorized_top_k/top_50_categorical_accuracy: 0.1238 - factorized_top_k/top_100_categorical_accuracy: 0.1293 - loss: 31238.8574 - regularization_loss: 0.0000e+00 - total_loss: 31238.8574
Epoch 2/3
6/6 [==============================] - 10s 2s/step - factorized_top_k/top_1_categorical_accuracy: 0.2418 - factorized_top_k/top_5_categorical_accuracy: 0.2418 - factorized_top_k/top_10_categorical_accuracy: 0.2418 - factorized_top_k/top_50_categorical_accuracy: 0.2457 - factorized_top_k/top_100_categorical_accuracy: 0.2508 - loss: 37982.0569 - regularization_loss: 0.0000e+00 - total_loss: 37982.0569
Epoch 3/3
6/6 [==============================] - 9s 2s/step - factorized_top_k/top_1_categorical_accuracy: 0.1624 - factorized_top_k/top_5_categorical_accuracy: 0.1624 - factorized_top_k/top_10_categorical_accuracy: 0.1624 - factorized_top_k/top_50_categorical_accuracy: 0.1667 - factorized_top_k/top_100_categorical_accuracy: 0.1732 - loss: 31066.0243 - regularization_loss: 0.0000e+00 - total_loss: 31066.0243
Top 3 recommendations for user 42: [b'Spa' b'Spa' b'Spa']

and here's the details for the first two users:

Top 10 recommendations for user 990000155054: [[b'Spa' b'Spa' b'Spa' b'Spa' b'Spa' b'Spa' b'Spa' b'Spa' b'Spa' b'Spa']]
thing: [[0.194 0.194 0.194 0.194 0.194 0.194 0.194 0.194 0.194 0.194]]
Top 10 recommendations for user 990000154983: [[b'Spa' b'Spa' b'Spa' b'Spa' b'Spa' b'Spa' b'Spa' b'Spa' b'Spa' b'Spa']]
thing: [[0.194 0.194 0.194 0.194 0.194 0.194 0.194 0.194 0.194 0.194]]

On different runs the recommendation might recommend a different interest, but it always gets stuck with the same one for all users. I've tried longer training runs, but I'm starting to wonder if there's a need to have some minimum number of "ratings" per user? I think that's the main difference between the data sets, i.e. 100 ratings for each user in the movielens dataset, but only five in my interests dataset. Or could it be the number of possible interests is just too small?

Or I'm making some stupid mistake in the code (in the interest vocabulary lookup table perhaps?). What are the requirements on the dataset size to allow this model to work?

Is there anything in the output I show above to indicate what's going wrong? The net getting stuck in a local minima perhaps?

I've tried longer training runs but that doesn't seem to make any difference. Perhaps the settings for the retrieval model need to be different given the fewer "ratings" or some other difference in the dataset proportions?