0. Purpose of the Project¶

This analysis uses real-world data provided by a digital advertising agency, who shared an anonymized version of the dataset they are actively analyzing, with the goal of proposing more effective marketing strategies based on actual customer behavior. The dataset (available at this link) contains detailed event logs from an online shopping mall, including visit timestamps, advertising media, device type, and purchase-related actions. Using this data, we construct user-level features from raw logs, compare conversion performance across media and devices, segment customers via K-means clustering, and model cart-stage conversion with logistic regression, ultimately deriving actionable insights for retargeting, onboarding, and loyalty-focused campaigns.

1. Intro¶

In this project, we work with an online shopping mall visit dataset that records each customer’s interactions with the site. The data eMFORCE_GA_yonsei.csv include information such as purchase counts, purchase amounts, visit timestamps, advertising media, and behavioral event types (event_name). Using these variables, we aim to analyze which factors influence customers’ purchasing activities and to classify customers into groups based on their purchase patterns.

2. Data Preprocessing¶

(1) row : id & col : event names, media, device¶

We transforms the raw log-level access records into a user-level dataset. First, we create a crosstab so that each row corresponds to a user and each column represents an event_name, summarizing how often each event occurs for that user. Then, for each user, we extract their most frequently used media channel and device. These variables, together with a later-defined purchase conversion indicator (converted), will form a structured table where each user is described by their main media, device, and event-based behavior.

This code merges the event-based features with each user’s main media and device to create a single user-level table. It then defines a binary conversion variable, converted, which is set to 1 if the user has at least one purchase event and 0 otherwise, and displays the first five rows of the resulting dataset.

(2) time¶

Now, we calculates how long each user stayed on the site in total by defining and aggregating sessions. First, it converts event_timestamp to a proper datetime format and sorts all records by user and time. Then, for each user, it computes the time difference between consecutive events and starts a new session whenever the gap exceeds 24 hours (or when there is no previous event). Using these session IDs, it measures each session’s duration as the difference between the first and last timestamp. Finally, it sums the session durations for each user to obtain total_stay_time_sec, and displays the first 10 users’ results.

This code sets user_pseudo_id as the index of the total session time table and then left-joins it to user_df so that each user’s total stay time (total_stay_time_sec) is added as a new column to the user-level dataset.

3. Conversion rate per media¶

In this section, we examine how conversion performance differs across advertising media. By grouping users by media and calculating the mean of the binary converted variable, we obtain the conversion rate for each channel and sort them in descending order to see which media sources are most effective at driving purchases. We broadly divided them into four groups and analyzed the characteristics of each group.

(1) High conversion rate (> 30%)¶

• advertiser_apppush: 57.7%, advertiser_braze: 35.9%
• App Push: notification messages sent to app users
• Braze: a CRM marketing platform that automates and personalizes app push, email, and SMS campaigns
• CRM (Customer Relationship Management)–based retargeting channels show outstanding conversion performance.

(2) Moderate conversion rate¶

• direct: 23.1%, advertiser_naver: 20.0%, advertiser_homeshopping: 19.8%
• Because these users already know the brand and visit directly, they have a higher likelihood of purchasing.
• Other advertisers (e.g., advertiser_sms, advertiser_tv, etc.) also show conversion rates of around 14%.

(3) Low conversion rate¶

• Finance/payment-related channels (payco: 8.6%, TOSS: 4.4%)
• Search-related channels (naver_search: 8.0%, google_SA: 3.9%)
• Kakao channels (kakao_dm: 6.3%, kakao_share: 4.0%, kakao_SA: 3.3%)
• Advertising strategies in these areas need to be re-evaluated.

(4) Extremely low conversion rate (≪ 0.1%)¶

• google_DA (Google Display Ads), ADN, appier (AI-based targeting ads), Facebook, Instagram
• These channels are mainly visual and click-oriented, so users tend to browse, show interest, or click lightly for fun rather than proceed to purchase.

4. Device¶

In this section, we investigate whether conversion behavior differs by device type. We first construct a contingency table that cross-tabulates device and the binary converted variable, which summarizes how many users on each device did or did not convert. The conversion rates are approximately 11.28% for mobile users (460,497 / (3,621,810 + 460,497)) and 2.28% for desktop users (9,091 / (390,444 + 9,091)). We next examine whether this substantial difference in conversion rates between mobile and desktop is statistically significant. This table will then be used to statistically test the association between device and conversion.

(1) chi squared test¶

To formally test whether conversion is independent of device type, we perform a chi-square test of independence with the null hypothesis that conversion rate does not differ by device (H₀: device and conversion are independent) and the alternative that it does (H₁: not independent). The resulting chi-square statistic (≈ 31,459.94) and p-value (≈ 0.0) lead us to reject the null hypothesis, indicating that conversion behavior differs significantly between mobile and desktop users.

(2) t - test¶

To further compare conversion performance by device, we conduct Welch’s t-test on the binary converted variable for mobile and desktop users. Under the null hypothesis that the two groups have the same mean conversion rate and the alternative that they differ, the test yields a very large t-statistic (≈ 318.03) and a p-value close to 0, so we reject the null and conclude that the average conversion rate for mobile users is significantly higher than that for desktop users.

(3) conclusion¶

Both the chi-square test and Welch’s t-test consistently show that conversion behavior differs significantly by device, with mobile users exhibiting a much higher conversion rate than desktop users. This gap is likely related to the strong performance of mobile-focused channels such as App Push, which drive especially high conversion among app users.

5. K-means¶

In this section, we analyze how customers move through the shopping journey—from visiting and browsing the site to making a purchase—by applying K-means clustering to group users with similar behavioral patterns. To focus the clustering purely on event-based activity features, we remove media and device (categorical attributes used later for interpretation) as well as converted, which is an outcome variable rather than an input feature for unsupervised clustering.

We standardize all features using StandardScaler so that each variable has mean 0 and unit variance, ensuring that no single feature dominates the K-means clustering due to its scale.

We determine the number of clusters k using the elbow method:

• We examine how inertia (within-cluster sum of squares) decreases as k increases.
• The optimal k is where the rate of decrease starts to level off.
• After k = 6, the reduction in inertia becomes relatively small, suggesting diminishing returns.
• With k = 3, clusters mainly separate users into high-conversion, low-conversion, and outlier groups, which provides limited behavioral insight.

We fit a K-means model with k = 6 clusters on the standardized features and assign each user to one of the six clusters. Then, we compute the mean of all numeric variables by cluster to summarize and compare the characteristic behavior of each customer group.

Cluster 0 : long-term dormant users¶

• Behavioral indicators (excluding outliers) are generally the lowest among all clusters.
• Both login rate (0.08) and conversion rate (0.069) are low, and total stay time is also short.
• Interpretation: customers who showed only minimal interest without purchasing → candidates for retargeting campaigns.

Cluster 1 : VIP customers (big spenders)¶

• Very high activity levels: add_to_cart ≈ 49.5, purchase ≈ 22.2, and extremely long stay_time (over 10 million seconds).
• Number of sessions (≈ 406) and logins (≈ 8.9) are very high, indicating highly engaged users.
• Conversion rate converted ≈ 0.99, meaning almost all users in this cluster make a purchase.
• Interpretation: core revenue-driving customers → strong candidates for loyalty programs and premium benefits.

Cluster 2 : Outliers or abnormal behavior¶

• Extremely high activity counts, e.g., login ≈ 16,072 and begin_checkout ≈ 15,612.
• Despite the large number of actions, converted = 0 and purchase = 0, showing a highly imbalanced pattern.
• Interpretation: likely bot traffic or logging errors → should be considered as outliers and potentially removed from analysis.

Cluster 3 : Loyal regular buyers¶

• Solid activity levels: add_to_cart ≈ 8.96, purchase ≈ 7.35, with around 191 sessions and relatively long total stay_time.
• High conversion rate converted ≈ 0.84, meaning most users in this cluster end up purchasing.
• Interpretation: valuable customers who visit frequently and make repeat purchases → key targets for retention strategies.

Cluster 4 : Low-loyalty general customers¶

• Relatively low activity levels: add_to_cart ≈ 0.42, purchase ≈ 1.77, about 6.75 sessions, and short total stay_time.
• Conversion rate converted ≈ 0.25, indicating that purchases are relatively infrequent.
• Interpretation: at-risk or churning users → candidates for remarketing campaigns and discount coupons.

Cluster 5 : First-time visitors¶

• Highest values for “회원가입완료” (signup completed) and “first_visit”, with a moderate conversion rate (converted ≈ 0.5).
• Short total stay_time → users who are likely visiting the site for the first time.
• Interpretation: customers who need marketing tailored to first-time visitors (onboarding guides, welcome offers, etc.).

6. Cart¶

In this section, we focus on users who have reached the cart stage and examine which factors drive them to complete a purchase. We restrict the data to customers with at least one add_to_cart event and build a logistic regression model with converted as the target and various behavioral features (and device information) as predictors. This allows us to quantify how different actions in the cart stage affect the probability of final conversion and to identify leverage points for improving checkout completion.

For the cart-stage analysis, we construct the feature matrix X by removing purchase, media, converted, and cluster. The variable converted is our target and must not be included as an input, while purchase is effectively an outcome-like variable that would leak information about conversion. We also exclude media and the derived cluster label so that the logistic regression focuses on the direct effects of users’ behavioral features in the cart stage, without confounding from channel attributes or previously computed segments.

From the logistic regression coefficients, we can draw the following conclusions about which factors most strongly influence conversion at the cart stage:

• begin_checkout (β ≈ 0.46)

  • This is the strongest positive predictor: users who proceed to the checkout step are much more likely to complete a purchase.

• 회원가입완료 – signup completed (β ≈ 0.20)

  • Completing membership registration is also strongly associated with higher conversion, suggesting that registered members are more likely to finish the purchase than non-members.

• login and session_start (small positive coefficients)

  • More logins and session starts slightly increase the probability of conversion, indicating that repeated engagement with the site helps, but their effect is weaker than checkout or signup.

• first_visit (β < 0)

  • First-time visitors are less likely to convert than returning users, which aligns with the idea that users often need multiple visits before purchasing.

• device (β ≈ -0.21, mobile = 0, desktop = 1)

  • The negative coefficient means that desktop users have a lower conversion probability than mobile users, even among those who added items to the cart, consistent with the earlier device-level analysis.

• total_stay_time_sec and add_to_cart (coefficients close to 0)

  • Their effects are very small in magnitude in this model, suggesting that spending more time or adding slightly more items to the cart does not, by itself, strongly change the conversion probability, once other behaviors (like checkout and signup) are controlled for.

Overall, the model highlights that moving forward in the funnel (begin_checkout), completing signup, and using mobile are the key drivers of conversion at the cart stage, whereas first-time visits and desktop usage are associated with lower purchase completion.

7. Conclusion¶

In this analysis, we used detailed event-level logs from an online shopping mall to understand how customers move from visit to purchase and how different channels and behaviors affect conversion.

Overall, the analysis shows that personalized CRM channels, mobile usage, and deeper funnel engagement (checkout and signup) are critical drivers of conversion, while user behavior patterns allow us to identify VIPs, loyal buyers, at-risk customers, and newcomers. These insights can be used to design targeted marketing and retention strategies that allocate resources more efficiently and improve overall sales performance.