A Utility-Mining-Driven Active Learning Approach for Analyzing Clickstream Sequences

Danny Y. C. Wang; Lars Arne Jordanger; Jerry Chun-Wei Lin

cs.LGOct 9, 2024

A Utility-Mining-Driven Active Learning Approach for Analyzing Clickstream Sequences

Authors: Danny Y. C. Wang, Lars Arne Jordanger, Jerry Chun-Wei Lin

Organizations: Department of C. E. M., Høgskulen på Vestlandet, Bergen, Norway

Abstract

In rapidly evolving e-commerce industry, the capability of selecting high-quality data for model training is essential. This study introduces the High-Utility Sequential Pattern Mining using SHAP values (HUSPM-SHAP) model, a utility mining-based active learning strategy to tackle this challenge. We found that the parameter settings for positive and negative SHAP values impact the model's mining outcomes, introducing a key consideration into the active learning framework. Through extensive experiments aimed at predicting behaviors that do lead to purchases or not, the designed HUSPM-SHAP model demonstrates its superiority across diverse scenarios. The model's ability to mitigate labeling needs while maintaining high predictive performance is highlighted. Our findings demonstrate the model's capability to refine e-commerce data processing, steering towards more streamlined, cost-effective prediction modeling.

Figures & tables

Fig. 1: Motivational overview illustrating the problem and limitations of existing active learning methods.

Fig. 2: A high-level overview of the clickstream analysis workflow employed in this study

Note: The ”Purchase” action is used as the label $\mathcal{Y}$ in this study.
Action	Symbol	Common Sequence Patterns
Page view	1	$\langle 1,3\rangle$ , $\langle 1,2,3\rangle$
Browsing product detail	2	$\langle 2,3\rangle$
Add product to shopping cart	3	$\langle 1,2,3\rangle$ , $\langle 1,3\rangle$ , $\langle 2,3\rangle$
Remove product from cart	4	$\langle 4,1\rangle$ , $\langle 4,2\rangle$

TABLE I: Rules for assigning symbols to clickstream actions

Table 4

Fig. 4: Graphical illustration of the calculation logic of HUSPM-SHAP in pool-based active learning

Layer (Type)
Embedding (Input dim=200, Output dim=40)
LSTM (100 units, return_sequences=True)
Dropout (rate=0.2)
Batch Normalization
LSTM (50 units, return_sequences=False)
Dense (2 units, Softmax activation)

TABLE IV: Baseline Model Architecture

S2L Last Model
	10% P, 90% NP	30% P, 70% NP	50% P, 50% NP
Coveo dataset: 10,000 points, Early window length: 5
Precision	97.70%	97.31%	89.61%
Accuracy	93.57%	85.33%	80.80%
Recall	95.36%	84.34%	70.72%
F1	96.51%	90.36%	79.05%

TABLE V: Performance using S2L Last method with various sample sizes and window lengths on Coveo dataset

Fig. 5: A stacked bar chart of mean of sampling SHAP values

Fig. 6: A SHAP force plot example showing the individual clickstream sequence contribution

Fig. 7: Sum of sampling SHAP values for each element

Rank	Sequence	Utility by SHAP value
Top 1	$\langle 3\rangle$	9.68
Top 2	$\langle 1,3\rangle$	5.81
Top 3	$\langle 1\rangle$	4.36
Top 4	$\langle 2,3\rangle$	3.87
Top 5	$\langle 2,1,3\rangle$	3.80

TABLE VI: Top 5 high utility sequences calculated by HUSPM-SHAP

Metrics	RD	UNC-1	Core-set	LLAL	SHAP	Proposed
Purchase Ratio: 90%
P	98.51%	98.57%	98.39%	97.70%	98.39%	98.57%
ACC	93.30%	93.93%	94.00%	93.90%	93.93%	94.00%
RC	94.27%	94.96%	95.19%	95.70%	95.12%	95.03%
F1	96.34%	96.73%	96.76%	96.69%	96.73%	96.77%
MCC	58.65%	56.18%	57.30%	58.55%	34.85%	56.80%

TABLE VII: Early Window Length 5: Comparison of model performance across various active learning models (best in each row: bold and underlined). P: Precision, RC: Recall, ACC: Accuracy

Metrics	RD	UNC-1	Core-set	LLAL	SHAP	Proposed
Purchase Ratio: 90%
P	97.04%	96.86%	96.27%	96.97%	96.39%	96.53%
ACC	93.77%	93.77%	93.80%	93.63%	93.67%	93.80%
RC	96.16%	96.33%	96.91%	96.09%	96.67%	96.67%
F1	96.60%	96.59%	96.59%	96.52%	96.53%	96.60%
MCC	59.66%	60.31%	62.69%	58.80%	60.14%	61.24%

TABLE VIII: Early Window Length 14: Comparison of model performance across various active learning models (best in each row: bold and underlined).

Fig. 8: Comparative performance of active learning strategies across data utilization levels

				HUSPM-SHAP
Metrics	RD	SHAP	SPM	LSTM	TR
Precision	89.00%	88.61%	88.68%	89.26%	88.35%
Accuracy	88.20%	87.87%	88.10%	88.23%	87.03%
Recall	88.09%	87.81%	88.16%	87.95%	86.61%
F1	88.54%	88.21%	88.42%	88.60%	87.47%

TABLE IX: Ablation study: Performance comparison of different backbones for HUSPM-SHAP (Sequence Length = 15, Purchase Ratio = 10%).

				HUSPM-SHAP
Metrics	RD	SHAP	SPM	LSTM	TR
P	97.04%	96.39%	96.23%	96.53%	97.40%
Acc	93.77%	93.67%	93.67%	93.80%	93.37%
RC	96.16%	96.67%	96.80%	96.67%	95.42%
F1	96.60%	96.53%	96.52%	96.60%	96.40%

TABLE X: Ablation study: Performance comparison of HUSPM-SHAP using different backbone models (Sequence Length = 15, Purchase Ratio = 50%).

Fig. 9: Comparison of purchase and non-purchase percentage between random and HUSPM-SHAP methods using 10% purchase data

Active Learning

Active learning selects the examples, labels, features, or experiments most useful for improving a model under limited budgets.

See all 296 papers

QueryMarket: Cost-Aware Online Active Learning in Data Markets
Jun 16, 2026Xiwen Huang, Pierre PinsonActive Learning
Dyson School of Design Engineering, Imperial College London, London, SW7 2DB, U.K. · Dyson School of Design Engineering, Imperial College London, London, U.K. · Halfspace (part of Accenture) +2
Are Candidate Models Really Needed for Active Learning?
May 14, 2026Harshini Mridula Mohan, Maanya Manjunath, Vipul Arya +2Active Learning
a SoCSE, RV University, Bengaluru, India. · b School of Engineering and Technology, Vidyashilp University, Bengaluru, India. · c Dataplex Inc., USA.
Batch Bayesian Active Learning with Partial Batch Label Sampling
Oct 10, 2025Kangping Hu, Stephen MussmannActive Learning
Department of Computer Science, Georgia Institute of Technology, Atlanta, GA 30332

Jun 16, 2026J/K move · Enter open · S save

QueryMarket: Cost-Aware Online Active Learning in Data Markets

Xiwen Huang, Pierre Pinson

Dyson School of Design Engineering, Imperial College London, London, SW7 2DB, U.K. · Dyson School of Design Engineering, Imperial College London, London, U.K. · Halfspace (part of Accenture)+2

Active Learning

Open paper AskPDF arXiv

Data acquisition is a major bottleneck for learning in real-time streams: analysts must decide on the fly which labels to purchase while respecting a rolling budget. However, existing online active learning rarely unifies pricing, information gain, and rolling budget constraints under concept drift. We introduce QueryMarket, a market-inspired framework that queries each incoming data point based on its estimated utility to the model and its price. Within this framework, we propose OVBAL (online variance-based active learning), which integrates data pricing with information-driven selection by estimating each sample's marginal utility via a D-optimality criterion with exponential forgetting and executing cost-aware purchases under rolling budget constraints. OVBAL yields a simple, fully online decision rule that adapts to nonstationary streams and heterogeneous label costs. Experiments on synthetic data and a real-world solar power generation forecasting task show that OVBAL is particularly effective under seller-centric pricing and yields a more favorable long-run error-cost trade-off in the real-world task under both pricing schemes.

Abstract

Active Learning

Explore similar work

QueryMarket: Cost-Aware Online Active Learning in Data Markets

Are Candidate Models Really Needed for Active Learning?

Batch Bayesian Active Learning with Partial Batch Label Sampling