fani-lab / OpeNTF

Neural machine learning methods for Team Formation problem.
Other
20 stars 14 forks source link

2021 - WWW '21 - STAN: Spatio-Temporal Attention Network for Next Location Recommendation #140

Closed karan96 closed 2 years ago

karan96 commented 2 years ago

Title: STAN: Spatio-Temporal Attention Network for Next Location Recommendation Year: 2021 Venue: WWW '21: The Web Conference 2021

Introduction: Personallized recommendations are being used more and more due to its capability of prediciting users' behavior and recommending useful stuff that mostly assist users with favorable options. In this papers, Authors take into account of the effect that spatio-temporal charactersistcs have on a recommendation and to the extent that it affects next Point-of-Interest (POI) recommendation. The author also points out the shortcommings that exisiting method approaches have while considering spatial-temporal charactersistcs.

Main problem: To address the spatio-temporal correlation for a user's trajectory, take into account the spatial distance which earliery techniques overlooked and use personalized item frequency for next Point-of-Interest recommendation.

Input: Set of user, location and time.

Output: The desired recommendation.

Previous works and their gaps:

  1. Sequential Recommendation: -
    1. Markov based models mainly focuses on transition probabilty between two consecutive visits and does not capture the transition of intermittent visits.
    2. Deep Learning techniques such as RNN based sequential models are unable to effectively capture personalized item frequency(PIF).
  2. Next POI Recommendation: -
    1. Different methods as mentioned by the author does not effectively consider non-trivial correlations between non-adjacent locations and non-contiguous visits.
    2. Not very sutiable for modeling PIF information.

Proposed Method: The authors propose a four module package STAN - Spatio-Temporal Attention Network which tries to address the following: -

  1. To fully consider the spatio-temporaleffect for aggregating relevant locations.
  2. Using og linear interpolation technique for spatial discretization instead of GPS to recover spatial distances and reflect user spatial preference.
  3. A bi-attention architecture for PIF.

The four modules are explained as follows: -

  1. A multimodal embedding module: - This module encodes user, location and time into latent representations together and is able to capture spatial distances.
  2. Self attention aggregation layer: - Follows self attention mechanisms. It aggregates relevant visited locations and updates the representations of each visit by using user embedded trajectory and spatio-temporal relation matrix.
  3. Attention Matching Layer: - To calculate the probability of each location candidate for next location from weighted check-in representations.
  4. Balanced Sampler: - Computes cross-entropy loss.

Experimentation: - Authors conducted on four real-world datasets: -

  1. TKY
  2. Gowalla: - http://snap.stanford.edu/data/loc-gowalla.html
  3. SIN: - https://www.ntu.edu.sg/home/gaocong/data/poidata.zip
  4. NYC: - http://www-public.imtbs-tsp.eu/~zhang_da/pub/dataset_tsmc2014.zip Comparison with other baselines models were done and recall scores were measured for topk probability samples. -Could have been better if they used other metrics as well.

Code: - Unavailable

Gaps of the Work: - Might have overlooked other factors while considering location such as personal perefrence, type of visit and assumes that the users follows a set pattern in a definite period of time which might not be true all the time.