Employee Satisfaction Analysis

Brief

A 142-row HR dataset — demographics, tenure, income, a 0–100 satisfaction score, and an open-ended text review per employee. Three angles on the same question — what's actually driving satisfaction: unsupervised clustering to find natural employee segments, a decision tree to rank which features predict dissatisfaction, and sentiment analysis on the free-text reviews themselves.

Employees

142

9 fields each

Age

37 yrs

range 19–58

Monthly income

$8.35K

range $1.56K–$20.24K

Satisfaction

59.5

range 20–100

Cleaning

Two missing values in the free-text review column were filled with "No feedback provided" rather than dropped, to keep every row for the numeric analysis. No duplicate rows. Columns were renamed to consistent snake_case, and the row_id identifier was dropped before modelling since it carries no signal.

Exploring the data

Business travel splits roughly 52% non-travel, 37% frequent travellers, and the rest rarely-travel. Education skews toward A-Level and Higher Certificate holders, with Bachelors, Diploma, Masters and Doctorate each in the minority. The workforce is 62% male, 38% female.

Histograms of age, monthly income, total working years and satisfaction score — Fig 1. — Age and tenure skew younger/junior; income is bimodal (a distinct lower- and higher-earning group); satisfaction is fairly evenly spread rather than clustered near one end.

Bar charts of business travel, education and gender counts — Fig 2. — Category counts for business travel, education level and gender.

What correlates with satisfaction

The standout relationship in the data isn't with satisfaction at all: age and total working years are strongly correlated (0.93) — unsurprising, since tenure accumulates with age. Income correlates moderately with both age (0.57) and tenure (0.55). Satisfaction score, though, barely moves with any of them — age (0.03), tenure (0.11), income (0.22) — which suggests seniority and pay aren't what's driving how satisfied people say they are.

Correlation heatmap of age, monthly income, total working years and satisfaction score — Fig 3. — Correlation heatmap: strong age↔tenure relationship, weak links to satisfaction.

Scatter plot of monthly income against satisfaction score — Fig 4. — Income vs. satisfaction isn't a smooth trend — it reads more like distinct groups, which is what motivated formal clustering rather than eyeballing the scatter.

Preparing for clustering

Numeric features (age, income, tenure, satisfaction) were Min-Max scaled to a 0–1 range, and the three categorical fields (business travel, education, gender) were label-encoded, since K-Means needs purely numeric input.

K-Means: how many segments?

Running the elbow method across k = 1 to 10, inertia drops sharply through k = 4 and flattens out after that. A silhouette analysis over the same range agrees — k = 4 scores highest at 0.74, well clear of every other option — so both methods point to the same number of segments.

Elbow method plot showing inertia against number of clusters — Fig 5. — Elbow method: the bend at k=4 marks where adding more clusters stops meaningfully reducing inertia.

Silhouette score plot peaking at 4 clusters — Fig 6. — Silhouette analysis agrees: 4 clusters scores highest (0.74).

Four employee segments

Fitting K-Means with k = 4 on scaled age, income, tenure and satisfaction splits the workforce into four fairly distinct groups when plotted on income against satisfaction — and the chart makes one thing obvious straight away: income alone doesn't predict satisfaction.

Scatter plot of the four K-Means clusters by monthly income and satisfaction score — Fig 7. — Four segments by income and satisfaction. Two of the higher-income groups aren't the most satisfied ones.

01High income, highly satisfied — reads as the longer-tenured, senior group. Worth retaining through mentorship roles and long-term incentives rather than pay alone.
02Lower income, moderately satisfied — an earlier-career profile still building tenure. Clear progression paths and competitive entry pay would matter more here than they currently seem to.
03Lower-to-moderate income, clearly dissatisfied — the group most worth intervening on, with performance-based recognition and skill development rather than assuming pay is the only lever.
04Top earners, only moderately satisfied — the clearest evidence that compensation and satisfaction aren't the same thing; better suited to an engagement survey than another pay rise.

Predicting satisfaction with a decision tree

Beyond describing segments, a decision tree classifier was trained to predict whether an employee sits above or below the satisfaction midpoint, using demographic and role features. On a held-out 20% test split it reached 86.2% accuracy — 90% precision and recall on "Satisfied", 78% on "Dissatisfied" (the softer number likely reflects the smaller class in a 142-row dataset).

Bar chart of decision tree feature importances, led by frequent business travel — Fig 8. — What the tree actually splits on: frequent travel dominates, then tenure, then income.

Frequent business travel is the single strongest predictor of dissatisfaction — by a wide margin over everything else the model had access to, including pay. Total working years and monthly income follow well behind it.

What employees are actually saying

The free-text reviews were cleaned, tokenised and scored for sentiment with TextBlob. Read across all 142 reviews, feedback skews positive but not overwhelmingly so:

Pie chart of sentiment distribution across employee reviews: 55% positive, 27.5% neutral, 17.6% negative — Fig 9. — 55% positive, 27.5% neutral, 17.6% negative across all reviews.

Word cloud of the most frequent words in employee feedback, dominated by work, pay, travel, management and good — Fig 10. — Most frequent words after stopword removal — "pay," "travel" and "management" sit right alongside "good" and "happy."

Cutting sentiment by business travel frequency lines up with the decision tree's top predictor almost exactly: frequent travellers post the highest share of negative reviews by a clear margin, while non-travellers and rare travellers skew far more positive.

Stacked bar chart of sentiment share by business travel frequency, with Travel Frequently showing the most negative sentiment — Fig 11. — Sentiment by travel frequency — the "Travel Frequently" group carries the most negative and least positive feedback of the three.

The two clearest extremes: "ServiceFirst is a great start for my career, learning a lot" (most positive) against "Office politics are the worst" (most negative) — career growth and workplace culture are pulling in opposite directions.

Where this points

Three independent methods — clustering, a decision tree, and sentiment analysis of free text — converge on the same story:

01 Travel is the biggest lever, not pay — reworking travel policy (rotation, wellness support, flexible scheduling) would likely move satisfaction more than a compensation review would.
02 Two of the four segments earn well above average and are still only moderately satisfied, which points at engagement rather than income as the gap to close for higher earners.
03 Career growth and workplace culture dominate the positive reviews; office politics and stagnant pay dominate the negative ones — both are addressable without a blanket pay rise.