Summary

• Recommendations on smart television sets shared by multiple users whose tastes may vary.
• Temporal and sequential context.
• Empirical evaluation of several recommendation methods.
• WatchItNext recommendation system. Verifies methods on top of two learning models.

limited DATa

<deviceID, itemID, timeStamp>

• itemID doesn't detect a specific episode
• no user identification
• length of TV program unknown
• Inventory unknown

Recommenders

Collaborative filtering methods

• Output two matrices as their resulting models:
  • |D| × n matrix M_D
  • n × |I| matrix M_I
  • where D and I are the sets of all devices and items respectively.
• Each row of M_D and each column of M_I are vectors corresponding to a device and an item respectively.
• n is the selected latent dimension that represents n latent factors or topics.
• Recommendation Score:

LFM with a stochastic gradient descent

• The cost function used for optimization is a log-sigmoid function that penalizes watched items with a low score and non-watched items with a high score:

• Early-stop used to avoid overfitting and determine number of training iterations.

LDA as a Collaborative Filtering Recommender

• Every device considered to be an input document for LDA and every item watched by that device is a word in the document.
• Models each device as a mix of topics that relate to the combinations of entities that share the device.
• Used an existing LDA implementation to produce M_I matrix from which they inferred the M_D matrix.
• Probability that i will be watched on d:

Contextual Personalization

Sequential Context

• Add the context of a currently watched item c

• LFM model values often negative so M_I was normalized by adding the absolute value of M_I minimal entry to all entries.

TempoRal Context

TEMPORAL and sequential CONTEXT

• Combine temporal context on top of a sequential context recommender

Experimental Setting

Emulate Inventories

• Go over pairs of items watched on a device d.
• Emulate inventory of items available after watching c at time t as the set of all shows j that were watched by some device while d watched i.

two scenarios

(1) exploratory setting

      items which have yet to be consumed by a user 

(2) habitual setting

      include items previously watched on the device

Experimental Setting

Evaluation metric

• Average Rank Percentile (ARP) : how high was the show that was actually watched next ranked by the recommender.
• Rank percentile: 

where I_c,t is the generated inventory and r(i) is i's rank on the output of the model

Results

Assume a correlation between the taste diversity and the number of users sharing a device.

Results

Results

Conclusion

• Context significantly improves recommendation accuracy.
• Collaborative filtering schemes outperform the corresponding memory-based counterparts.
• Temporal context alone performs well but adding sequential context performs even better.
• Contribution? Useful? What about just identifying the user and tracking more data.