TIFUKNN

Two patterns in PIF (personalized item frequency): \(user_i'\ s \ record \ \ \{ v_1, v_2, ..., v_i, ..., v_t\} \\\ where \ v_i = [0,1,1,1,1,0....]^T \\ aiming \ to \ predict \ v_{t+1}\)

• item frequency - repeated purchase pattern

  • next basket contains user’s historical records
  • has been used, get considerable performance gain

• similar user - colllaborative purchase pattern

  • use vector to represent each user, top 300 nearest neighbors

  • \(PIF_{vector} = \sum_{i=1}^{t}{v_i}\)

  • caculate 4 ratios

    • captured by repeated: 0.1 - 0.6
    • Captured by collaborative - most, 0.5 - 0.98
    • captured by both: 0.1 - 0.6 around but smaller than repeated since intersection
    • Not captured by both: 0.01-0.4, most 0.1: 1- union of repeated and collaborative

  • RNNs has difficulty to learn vector addition

code return

$ python TIFUKNN.py ./data/TaFang_history_NB.csv ./data/TaFang_future_NB.csv 300 0.9 0.7 0.7 7 10           
start dictionary generation...
{'MATERIAL_NUMBER': 11997}
# dimensions of final vector: 11997 | 0
finish dictionary generation*****
Number of training instances: 10000
Num. of top:  10
recall:  0.12933969065750625
NDCG:  0.09859257577070348

temporal-item-frequency-baseduser-KNN (TIFUKNN)

Existing methods:

• Markov chain: cannot capture high-order information from long time ago
• KNN can capture unseen pattern but it is only a small part in the target baske
  • item embedding
  • model temperal relation across all the baskets - user representation

Integrating Temporal Dynamics

item with same frequency cannot be discriminated

gap distibution - decay weight: early item appears, smaller weight to the final frequency

Nearest Neighbors based method

• similarity calculation
• prediction based on similarity

User similarity calculation

• Sum up; hierarchical time decayed weights

• Steps \(\{ v_1, v_2, ..., v_i, ..., v_t\}, \ where \ v_i = [0,1,1,1,1,0....]^T\)

  • partition historical 20 baskets into 5 groups \(Group_j = [v_1, v_2, ..., v_{x = \frac{t}{m}} ]\)

    • if cannot equally partitioned, each full basket take ceiling t/m and the first group can be less

  • Within group: each vector j in each group multiply a time-decayed weight r_b^{x-j} \(v_{group} = v_1 \times r_b^{x-1} + v_2 \times r_b^{x-2} + ... + v_{x} \times r_b^0\)

  • Cross group: each group vector multiply a weight \(v_{user} = v_{g_1} \times r_g^{m-1} + v_{g_2} \times r_g^{x-2} + ... + v_{g_x} \times r_g^0\)

  • Now get the user representation and use for similarity calculation - Euclidean distance or other

  • Final predict, hyperparameter \alpha \(P = \alpha \times v_{user} + (1-\alpha) \times u_{nearest\ neighber}\)

Drawbacks

• If customer A only buys bananas again and agian and no other items. It can be the nearest neighbors to anyone else.
• A’s bananas will have super high score and be in the final top_k for anyone else.